# So Much Tech News: The Week of 11/08/2026 > Source: > Published: 2026-11-15 16:00:31+00:00 *Weekly tech and science news for curious people, with a perspective shaped by the belief that the internet should be open, private, and controlled by its users. Start with the week’s highlights, browse by category, or dive into the full summaries. Each story comes with a 🧠 1–5 background knowledge rating. The headline of the summary links directly to the original article. The opinions are mine, while the writing of this newsletter is AI-assisted.* ### This Weeks Highlights - **[Claude Fable 5.2 and Mythos 5.2](https://somuchtechnews.substack.com/i/224030685/claude-fable-52-max-default-fallback-intelligence-performance-and-price-analysis-artificial-analysis):** Anthropic’s new models took the top spot on Artificial Analysis’s Intelligence Index (65, ahead of Gemini 4 Carbon’s 60). Anthropic also reports that Fable 5.2 coordinated 1,000+ subagents over eight days to build a bootable prototype operating system. - **[Gemini 4 Hydrogen and Orchestrator](https://somuchtechnews.substack.com/i/224030685/gemini-4-hydrogen-and-orchestrator-google-antigravity):** Google’s new Orchestrator lets Carbon plan and Hydrogen implement, then hands the task back to Carbon when Hydrogen gets stuck. Google reports nearly Carbon-level task success with 24% less quota and 19% faster completion. - **[Anthropic sets IPO price range](https://somuchtechnews.substack.com/i/224030685/anthropic-sets-ipo-price-range-of-70-to-78-per-share-to-reach-an-implied-valuation-of-1404-trillion-techcrunch):** At $70–$78 per share, Anthropic would be valued at about $1.404 trillion, up from $965 billion in May but well below the roughly $2 trillion it was reportedly targeting. Trading is expected around November 19 under the ticker ANT, with its Long-Term Benefit Trust governance structure intact. - **[OpenAI’s quarterly revenue tops $10B](https://somuchtechnews.substack.com/i/224030685/openais-quarterly-revenue-reportedly-crosses-10b-for-the-first-time-techcrunch):** Q3 revenue reportedly crossed $10 billion for the first time, nearly 50% above Q2’s $6.7 billion. Its adjusted operating margin improved from -150% to -109%, which is still deeply negative. - **[The unfolding Jensen Huang scandal](https://somuchtechnews.substack.com/i/224030685/the-unfolding-jensen-huang-scandal-a-timeline-techcrunch):** Huang resigned as Nvidia CEO and board member on November 12, after reports about his relationship with a Megaspeed executive (Megaspeed is under investigation over chip exports) and other alleged affairs. Congressional investigations are expanding, and Huang has agreed to testify before Thanksgiving. [Thumbnail Article Summary Link](https://somuchtechnews.substack.com/i/224030685/the-unfolding-jensen-huang-scandal-a-timeline-techcrunch) ### AI Models, Agents & Machine Learning - [Flide](https://somuchtechnews.substack.com/i/224030685/flide-omnijev) | 🧠🧠🧠🧠🧠 - [Claude Fable 5.2 and Mythos 5.2](https://somuchtechnews.substack.com/i/224030685/claude-fable-52-max-default-fallback-intelligence-performance-and-price-analysis-artificial-analysis) | 🧠🧠🧠 - [CheMeleon model adapts to predict molecular properties from scant data](https://somuchtechnews.substack.com/i/224030685/chemeleon-model-adapts-to-predict-molecular-properties-from-scant-data-physorg) | 🧠🧠🧠🧠 - [e-means: outperforming k-means in one-tenth the epochs](https://somuchtechnews.substack.com/i/224030685/e-means-outperforming-k-means-in-one-tenth-the-epochs-swann-bessa) | 🧠🧠🧠🧠 - [Gemini 4 Carbon is now available](https://somuchtechnews.substack.com/i/224030685/gemini-4-carbon-is-now-available-news-from-google) | 🧠🧠🧠 - [Gemini 4 Hydrogen and Orchestrator](https://somuchtechnews.substack.com/i/224030685/gemini-4-hydrogen-and-orchestrator-google-antigravity) | 🧠🧠🧠 - [Elephant-Goldfish Model](https://somuchtechnews.substack.com/i/224030685/elephant-goldfish-model-google) | 🧠🧠 - [Exclusive: OpenAI disbands its stand-alone robotics team as its general models start running robots](https://somuchtechnews.substack.com/i/224030685/exclusive-openai-disbands-its-stand-alone-robotics-team-as-its-general-models-start-running-robots-fortune) | 🧠🧠🧠 - TinyGPT Visualizer | 🧠🧠🧠🧠 ### AI Policy, Safety & Society ### Cybersecurity, Privacy & Software Integrity - [A New AI Tool From Google DeepMind Proved the Signal Protocol Is Secure](https://somuchtechnews.substack.com/i/224030685/a-new-ai-tool-from-google-deepmind-proved-the-signal-protocol-is-secure-wired) | 🧠🧠🧠🧠 - [Formalizing CPython](https://somuchtechnews.substack.com/i/224030685/formalizing-cpython-david-andrews) | 🧠🧠🧠 - [Large-Scale Measurement and Real-World Mitigation of Web Browser Fingerprinting in the Wild](https://somuchtechnews.substack.com/i/224030685/large-scale-measurement-and-real-world-mitigation-of-web-browser-fingerprinting-in-the-wild-fingerprinting-in-the-wild) | 🧠🧠🧠 - [Critical RCE Vulnerability in Apache HTTP Server: Everything You Need To Know](https://somuchtechnews.substack.com/i/224030685/critical-rce-vulnerability-in-apache-http-server-everything-you-need-to-know-wiz) | 🧠🧠🧠🧠 ### Business & Industry ### Science & Research ### Space, Aerospace & Satellite Technology ### Consumer Technology, Platforms & Digital Projects ### And if You Missed it… - [Inside GitHub’s new Git storage architecture](https://somuchtechnews.substack.com/i/224030685/inside-githubs-new-git-storage-architecture-github-blog) | 🧠🧠🧠🧠 - [Asahi Linux — Progress Report: Linux 7.3](https://somuchtechnews.substack.com/i/224030685/asahi-linux-progress-report-linux-73-asahilinuxorg) | 🧠🧠🧠🧠 - [Apple’s iPhone keyboard has a secret blocklist. ‘Dictator Trump’ and ‘Sleepy Joe’ are on it.](https://somuchtechnews.substack.com/i/224030685/apples-iphone-keyboard-has-a-secret-blocklist-dictator-trump-and-sleepy-joe-are-on-it-international-business-times) | 🧠🧠🧠 - [Anatomy of Apple’s Personal Knowledge Graph](https://somuchtechnews.substack.com/i/224030685/anatomy-of-apples-personal-knowledge-graph-the-eighth-rank) | 🧠🧠🧠🧠 ### AI Models, Agents & Machine Learning #### [🔗 Flide](https://github.com/OmniJev/Flide) `(OmniJev)` 🧠🧠🧠🧠🧠 Flide is an open-source inference engine designed to make a newer category of AI models called *decision models* run much faster. Unlike conventional LLMs that generate paragraphs of text, decision models process context and return probability distributions over predefined choices. This makes them particularly useful for AI agents that need to make frequent decisions, such as determining what action to take next or routing a request. The problem Flide is trying to solve is redundant computation. Imagine an AI agent processing a conversation and needing to answer ten different questions about it. Conventional inference engines often have to process or reload the shared conversation history separately for each question. That can become extremely expensive when the history is long and the agent needs to make decisions repeatedly. Flide solves this with a technique called exact state forking. Instead of processing the shared context ten times, it processes it once and creates ten independent branches for the questions. Each branch can access the shared context and its own question, but cannot access the other questions. This is important because simply combining all the questions into one prompt can change the model’s probability distributions and potentially alter its decisions. What makes the implementation particularly interesting is that it works across two different types of neural network layers. Conventional attention layers can isolate questions using carefully constructed attention masks. However, newer hybrid architectures also use recurrent components such as Gated DeltaNet, which compress previous information into a continuously updated internal state. Flide has to reconstruct the appropriate starting state for each question rather than allowing the questions to influence one another. The engine also uses GPU-resident metadata and CUDA graphs to reduce the overhead of scheduling these operations. It can even save intermediate checkpoints of an agent’s conversation history so that subsequent decisions can reuse the unchanged portions rather than processing everything again. The accompanying research paper reports some impressive results. Across six workloads using Qwen3.5 and Gemma 4 models, Flide achieved up to 9.5× lower latency and 8.5× higher throughput than the strongest competing inference engine. These are maximum results from particular multi-question workloads, not improvements that should be expected across every model or request. The experiments were conducted on a single NVIDIA H200 GPU. The authors also tested whether these optimizations affect the models’ ability to solve the underlying tasks. Across several reasoning and decision-making benchmarks, Flide’s accuracy remained very close to that of vLLM and SGLang, generally within 1.2 percentage points. This is important because an alternative optimization called multi-item scoring, which also attempts to process multiple questions together, changed up to 29% of categorical decisions in their experiments. Flide, by comparison, changed at most 2.3%. The goal here isn’t to make the models better at reasoning, but to make them substantially faster without sacrificing the quality or consistency of their decisions. #### [🔗 Claude Fable 5.2 (Max, Default Fallback) Intelligence, Performance & Price Analysis](https://www.anthropic.com/claude-fable-and-mythos-5-2) `(Artificial Analysis)` 🧠🧠🧠 Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, and the results are pretty remarkable. The new models are designed for extremely difficult technical work, including software engineering, mathematics, scientific research, and robotics. According to independent testing by [Artificial Analysis](https://artificialanalysis.ai/models/claude-fable-5-2), Fable 5.2 is now the highest-scoring model on its Intelligence Index, earning 65 points, compared with Gemini 4 Carbon’s 60, Claude Opus 5.5’s 59, GPT-6 Astra’s 57, and GPT-6.1 Sol’s 55. The index combines nine evaluations covering coding, scientific reasoning, knowledge work, and general intelligence. Each of the four categories contributes 25% to the final score. While this is not a comprehensive measurement of intelligence, it is a useful independent comparison of model capabilities. Anthropic’s own benchmarks also show significant improvements. On FrontierCode 2.0, which evaluates an AI’s ability to modify real software repositories, Fable 5.2 scored 49.1%, compared with Fable 5.1’s 32.4% and Opus 5.5’s 45.5%. On Terminal-Bench-Science, which evaluates scientific tasks performed using terminal tools, Fable 5.2 scored 74.6%, compared with its predecessor’s 52.6%. But the most interesting results may be what Anthropic reports the models actually accomplished outside conventional benchmarks. **Software engineering and autonomous agents** In one experiment, an Anthropic engineer gave Fable 5.2 the task of building an operating system from scratch. Starting with an empty folder, the model coordinated more than 1,000 AI subagents over eight days, writing, testing, and reviewing code. The result was a prototype operating system capable of booting on a physical laptop, complete with disk, keyboard, and network drivers, file storage, windows, a login screen, and an installer. This is not equivalent to producing a mature operating system like Linux. But it is an unusually ambitious example of AI coordinating a large software project over an extended period. Anthropic also reports using an earlier version of Fable 5.2 to migrate a critical production service from Python to Rust. The resulting service reportedly starts approximately 26 times faster. That is specifically an improvement in startup performance, not a claim that the entire service runs 26 times faster. The distinction is important because these experiments demonstrate something different from a model answering difficult questions. They require planning, coordinating other agents, writing code, testing results, and responding to failures over many hours or days. **Scientific research and mathematics** Anthropic is making some particularly ambitious claims about the model’s scientific capabilities. According to the company, Fable 5.2 solved two previously open problems in theoretical physics involving black holes and holographic duality, a mathematical relationship that allows certain gravitational systems to be described using quantum theories without gravity. One problem involved determining whether the abrupt formation of black holes at high temperatures also occurs in very small systems where a classical description of gravity is insufficient. A calculation proposed in 2003 had reportedly remained unfinished because of its complexity. Anthropic says Fable 5.2 completed the calculation and found that the abrupt transition still occurs. The second result concerned identifying the exact limitations of a mathematical technique used to rule out possible theories of quantum gravity. These results are described in research preprints, so they should be distinguished from independently established scientific breakthroughs until they receive further scrutiny. The model also demonstrated substantial mathematical formalization capabilities. Anthropic tasked approximately 1,000 Fable 5.2 agents with translating the proof of Fermat’s Last Theorem into Lean, a programming language that allows mathematical proofs to be checked by a computer. The agents completed the formalization in just 30 hours, compared with 11 days for an earlier research version of Claude that received more specialized support. In a comparison using the same multiagent configuration, Opus 5.5 completed only 26% of the formalization after 18 hours and made little additional progress afterward. Formalizing a proof is different from discovering it. The underlying mathematics already exists, but translating a complex proof into a form that can be rigorously checked by a computer is an enormous undertaking in its own right. **Biology, hardware, and robotics** The improvements extend beyond software. In one biological research experiment, Claude Mythos 5.2 developed a method for predicting the structures of adeno-associated virus capsids. These protein shells are particularly important in gene therapy because they can be used to deliver genetic material into cells. Predicting an entire capsid is difficult because it consists of many interacting protein chains. Rather than trying to model the complete assembly at once, Claude used existing open-source protein-folding models to predict overlapping groups of chains and then combined the results using the capsid’s known geometric symmetries. Anthropic reports that this approach accurately modeled 28 of 34 capsids tested, including natural and engineered variants. This is a computational structure-prediction result, not evidence that the model independently designed a successful gene therapy. Fable 5.2 also designed three functioning electronic devices from short prompts: a voice-assistant hub, a USB speakerphone, and a handheld virtual pet. It selected components, designed circuit boards and enclosures, and wrote the software. A hardware company manufactured the boards, and Anthropic reports that all three worked without electrical modifications. In robotics, Fable 5.2 achieved 64.9% on Anthropic’s internal run of RoboDojo Sim, a benchmark involving control of simulated robotic arms using camera images. That’s more than double the reported scores for Fable 5.1 and GPT-6 Astra under the tested configurations. These are promising results, although success in simulation does not automatically translate into reliable performance on physical robots. **Performance, pricing, and limitations** It supports a 1-million-token context window, with up to 128,000 output tokens. Anthropic reports up to 30% faster output generation than Fable 5.1 and has reduced API prices to $8 per million input tokens and $40 per million output tokens. However, more capable reasoning can consume enormous amounts of computation. Artificial Analysis found that its highest-effort Fable 5.2 configuration required an average of 45 minutes and $26.54 per Intelligence Index task. That makes it considerably slower and more expensive per task than many competing models, even though its underlying token prices have decreased. Anthropic itself recommends Opus 5.5 for most everyday work, reserving Fable for unusually difficult problems. Anthropic reports that Mythos 5.2 is substantially better at discovering software vulnerabilities than its predecessors. On an internal evaluation using 1,829 previously disclosed vulnerabilities, it recovered 89.6%, compared with 78.3% for Mythos 5.1. The company has also deliberately limited some capabilities related to improving AI systems themselves. During training, Anthropic constrained performance on tasks such as accelerator-kernel optimization and improving language-model architectures, leaving those capabilities approximately at Opus 5.5 levels. #### [🔗 CheMeleon model adapts to predict molecular properties from scant data](https://phys.org/news/2026-11-chemeleon-molecular-properties-scant.html) `(Phys.org)` 🧠🧠🧠🧠 Another very cool model coming out of MIT! Researchers at MIT, working with scientists from BASF, have developed CheMeleon, an open-source AI model that can predict the properties of molecules using surprisingly little experimental data. This could be particularly useful in drug discovery, where synthesizing and testing new molecules is often extremely expensive and time-consuming. One of the biggest problems with using deep learning in chemistry is the amount of data required to train a model. Traditional machine-learning methods, such as random forests, often outperform neural networks when only a few hundred or thousand experimental measurements are available. Unfortunately, this is exactly the situation researchers frequently find themselves in when developing new drugs. CheMeleon approaches this problem by teaching a neural network general chemical characteristics before training it to predict a specific property. The researchers first pretrained the model using approximately one million molecules from PubChem, a large public database of chemical compounds maintained by the National Institutes of Health. Rather than relying on expensive experimental measurements, they calculated 1,613 molecular descriptors for each molecule. These are mathematical descriptions of a molecule’s structure and characteristics, including things like molecular weight, atomic composition, and the complexity of its chemical bonds. CheMeleon uses a directed message-passing neural network, which represents molecules as graphs of atoms connected by chemical bonds. Information is repeatedly passed between connected atoms and bonds, allowing the model to construct a numerical representation of the molecule’s structure. During pretraining, the network learns to predict the calculated molecular descriptors from this structure. Because these descriptors can be generated automatically, the researchers can train on a large dataset without needing to conduct millions of experiments. Once pretrained, the model can be fine-tuned using much smaller experimental datasets to predict specific properties, including drug potency, solubility, toxicity, and interactions with proteins. The important distinction is that CheMeleon isn’t learning these properties directly during pretraining. It’s learning a useful representation of molecular structure that can later be adapted to predict them. And the results are promising! The researchers evaluated CheMeleon across 58 molecular-property prediction datasets, comparing it with traditional machine-learning techniques and other pretrained models. Their published study found that CheMeleon generally outperformed classical approaches, including random forests, as well as competing foundation models. In a detailed preprint evaluation, CheMeleon achieved a 75% win rate on the Polaris benchmarks and 97% on MoleculeACE, a collection of particularly challenging molecular bioactivity datasets. Importantly, a win meant that CheMeleon either achieved the best result or performed statistically indistinguishably from the best model, not necessarily that it outright beat every competitor. The model is also already seeing adoption outside academia. According to the researchers, nine of the top 20 entries in an OpenADMET molecular-property prediction competition incorporated CheMeleon, and several pharmaceutical companies have started using it. There are still limitations. Like other molecular machine-learning models, CheMeleon struggles with *activity cliffs*, where a small change in a molecule’s structure produces a surprisingly large change in biological activity. This is especially important in drug discovery, where modifying just one part of a molecule can dramatically alter its effectiveness. The researchers have made both the model and its training framework publicly available through [GitHub](https://github.com/JacksonBurns/chemeleon) and the Chemprop ecosystem, allowing other scientists to adapt it to their own molecular-property prediction problems. #### [🔗 e-means: outperforming k-means in one-tenth the epochs](https://e-means.github.io/) `(Swann Bessa)` 🧠🧠🧠🧠 Meta FAIR has developed and open-sourced e-means, a new clustering algorithm that substantially improves on one of the oldest and most widely used machine learning algorithms: k-means. K-means is used to organize data into groups based on similarity. Given a dataset and a predefined number of clusters (k), the algorithm assigns each data point to its nearest cluster center, or centroid, and repeatedly recalculates those centroids until the clustering stops improving. This is particularly useful when working with large collections of high-dimensional vectors, such as embeddings generated by image and language models. Applications include vector search, data organization, image segmentation, and compressing information used by neural networks. The problem with traditional k-means, specifically Lloyd’s algorithm, is that every iteration requires processing the entire dataset. This becomes increasingly expensive as datasets grow into the millions or billions of vectors. e-means approaches this differently. Rather than waiting until an entire dataset has been processed before updating the centroids, it divides the data into smaller batches. After assigning each vector to its nearest centroid, it updates the centroid using an exponential moving average (EMA) of the cluster’s previous count and sum and the information from the latest batch. One of the cleverest aspects of this approach is how it handles information from previous updates. As the centroids move during training, assignments made under their previous positions become progressively less relevant. Rather than allowing those old assignments to influence the centroids indefinitely, EMA exponentially reduces their influence with every update. I think this is a genius application of the technique because it recognizes that the algorithm should not remain anchored to decisions it made when its centroids were in much worse positions. This is complemented by a cosine learning-rate schedule, which gradually reduces the update rate from its initial value to 10% of that value. This allows the algorithm to make larger adjustments early in training and increasingly smaller, more conservative adjustments as it approaches its final configuration. Together, these mechanisms let e-means take advantage of what it has learned while progressively reducing the influence of outdated information. The researchers also introduced several other improvements. Dead clusters, which are centroids that stop receiving data points, are revived by splitting larger clusters rather than simply assigning them random points. The algorithm also includes a finalization stage that optimizes the resulting centroids for either training data or performance on previously unseen data. Another particularly interesting aspect is how e-means handles hyperparameters. Traditional algorithms often require experimentation to determine the appropriate batch size and update rate. The researchers developed scaling laws that automatically calculate these values based on the dataset size, number of clusters, and training budget. They derived these formulas from approximately 75,000 optimization experiments and tested how well they transferred to other datasets. This means practitioners don’t need to perform their own extensive hyperparameter searches for every new dataset. The performance results are impressive. Across datasets containing between one million and 100 million vectors, e-means reached the test error of a state-of-the-art GPU implementation of k-means, Flash-KMeans, after 25 iterations in just 1.4–2.6 epochs. These results translated into GPU speedups of approximately 9.5–35.9×, with the experiments using NVIDIA H100 GPUs and multiple independent runs per configuration. The researchers also tested e-means against Flash-KMeans using its stronger greedy k-means++ initialization. After five epochs, e-means still achieved lower test error across all eight configurations they evaluated, although the improvement was relatively small, between 0.06% and 0.33%. The scaling laws were developed primarily for relatively high-dimensional, static datasets and are not optimal for very small or low-dimensional datasets. The implementation is available through Meta’s open-source Faiss library, with support for both PyTorch and NumPy. #### [🔗 Gemini 4 Carbon is now available](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/) `(News from Google)` 🧠🧠🧠 Google has officially released [Gemini 4 Carbon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/), its most powerful AI model, after initially limiting access to a select group of cybersecurity researchers in October. Carbon is designed for particularly difficult tasks that require extended reasoning, autonomous agents, and multiple rounds of verification. Google reports that Carbon now leads on several coding, automation, and cybersecurity benchmarks. But the more interesting part of this release is what Google and its partners have actually been using the model to accomplish. Here are some of the standout examples: Reconstructing buildings from engineering drawings: Google gave a team of Carbon agents seven publicly available city planning PDFs containing 206 engineering drawing views. Using the `/swarm` feature in Google Antigravity, the agents reconstructed Google’s Platform 37 office as an interactive 3D environment containing approximately 906,000 triangles, running through 10 overnight development loops. Recreating a historical musical instrument: Carbon reconstructed a playable 18th-century Stein-Schule fortepiano using CAD drawings and X-ray radiographs. The simulation included the instrument’s 61-key mechanical action and escapement physics, which model how the piano’s hammers interact with its keys and strings. The agents then placed it inside a 3D reconstruction of Toronto’s Koerner Hall, complete with acoustics that change depending on where the listener is sitting. Discovering real cybersecurity vulnerabilities: Working with Wiz’s Scan for Good program, Carbon identified previously unknown zero-day vulnerabilities affecting two major software-as-a-service providers in under two hours. According to Google, the organizations subsequently worked with researchers to fix the vulnerabilities, potentially protecting tens of thousands of businesses. Google’s own vulnerability research team also used Carbon to demonstrate that two previously identified bugs in popular open-source software were actually exploitable. This helped convince maintainers to reclassify the vulnerabilities as critical and release patches. The technical details have not yet been disclosed. Mathematically verifying Signal’s security: Google DeepMind researchers developed a multi-agent framework called AlphaProof Code that used Carbon alongside other models to verify important security properties of Signal’s cryptographic implementation. Rather than merely testing the software for known vulnerabilities, the researchers used *formal verification*, which involves expressing security requirements as mathematical statements and proving that the actual software satisfies them. The system generated approximately 6.5 million lines of machine-checked proofs in Lean, building on around 26,000 lines of trusted specifications and machine-code semantics. The work covered properties including authentication, forward secrecy, and post-compromise security. Google says the verification took weeks rather than the years such projects have historically required. These are proofs of specified properties, not a guarantee that every component of Signal is free of vulnerabilities. Finding problems in laboratory experiments: Researchers gave Carbon laboratory notebooks and raw experimental data and asked it to investigate inconsistencies. The model identified problems including incomplete automated mixing in deep laboratory wells, trace contamination carried between samples, and difficulties separating overlapping chromatography signals. These are important because an experimental error can produce misleading results without necessarily being obvious to the scientist conducting the experiment. According to Google, identifying these problems helped researchers improve the reproducibility of their assays. Improving protein structure prediction: This is another particularly interesting example. Protein structure prediction systems such as AlphaFold use evolutionary relationships between protein sequences to help determine their three-dimensional structures. However, proteins with few known relatives can be particularly difficult to characterize. Google used Carbon in a multi-agent system to search enormous collections of previously unannotated genetic sequences, environmental and viral sequence databases, and protein structures for distant evolutionary relatives called *remote homologs*. Across seven protein families from the CASP17 blind test set, Carbon retrieved 40 times as many homologs on average as the comparison search process. This could give protein structure prediction systems substantially more evolutionary information without requiring the models themselves to be retrained. Google also introduced `/swarm`, which allows developers to coordinate teams of AI agents. One agent can handle architecture or planning while others implement features, investigate errors, and independently check the results. On standardized evaluations, Carbon achieved 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation tasks, and 71% on CWE-bench for vulnerability patching. Google also reports that Carbon passed 92% of its internal agentic honesty tests, compared with 70% for Gemini 4 Argon. These demonstrations are promising, although most remain Google-reported results or partner evaluations, and they should not be confused with comprehensive independent validation. Carbon is available to paying developers through the Gemini API and to Google AI Ultra subscribers. API pricing starts at $2 per million input tokens and $10 per million output tokens, with both rates scheduled to double after November 30. #### [🔗 Gemini 4 Hydrogen and Orchestrator](https://antigravity.google/blog/gemini-4-hydrogen-in-google-antigravity/) `(Google Antigravity)` 🧠🧠🧠 Google has introduced Gemini 4 Hydrogen, a faster and more efficient coding model, alongside Orchestrator, a system that automatically switches between AI models depending on what a task requires. The interesting part here is Orchestrator. Traditionally, AI model routers determine which model should handle a task based on the initial prompt. The problem is that coding tasks can become significantly more or less complicated as they progress. A task that initially requires a powerful reasoning model might only need a smaller model once the planning is finished. Google’s solution is to allow models to hand tasks back and forth throughout the development process. For example, Gemini 4 Carbon, Google’s more powerful reasoning model, might initially explore a codebase, identify dependencies, and develop an implementation plan. Hydrogen would then take over to write code, update dependencies, and run tests. If Hydrogen encounters a problem it cannot resolve, it can hand the task back to Carbon, which can investigate what went wrong, revise the plan, and return the task to Hydrogen. According to Google, this approach achieved nearly the same task success rate as Carbon alone on Terminal-Bench 4.1, while consuming 24% less quota and completing tasks 19% faster. Hydrogen itself is also pretty capable. It scored 51 on the Artificial Analysis Intelligence Index, compared with 50 for Gemini 4 Argon, and reached 54 using its highest reasoning setting. On Terminal-Bench 4.1, Hydrogen achieved a 57% success rate, compared with 18.2% for Gemini 3.8 Flash. These are Google-reported comparisons, and the results will need independent verification. I think this is where AI coding is going. Switching between a large model for planning and a smaller model for implementation is already common practice in vibe coding, so automating that process makes a lot of sense. What I particularly like is that Google didn’t make this a one-way process. Allowing Hydrogen to hand tasks back to Carbon when it runs into trouble, rather than simply having Carbon plan everything at the beginning, is a really good idea. It accounts for something that anyone who has spent time coding knows: problems don’t always reveal themselves until you actually start implementing a solution. I can absolutely see OpenAI and Anthropic adopting similar approaches. Instead of requiring users to decide which model should handle each part of a task, systems like Orchestrator could increasingly manage those decisions themselves. Hydrogen is rolling out to Google AI Pro and Ultra subscribers, while Orchestrator is available with different model configurations depending on the subscription. #### [🔗 Elephant-Goldfish Model](https://github.com/google/elephant-goldfish-model) `(Google)` 🧠🧠 The Elephant-Goldfish Model (EGM) is an open-source plugin that separates the agent responsible for designing software from the agents responsible for reviewing that design. The idea is that an AI agent working on a project accumulates a lot of context over time. That’s useful because it understands the project’s architecture, requirements, and previous decisions. However, it also creates an opportunity for confirmation bias. The same agent that makes an assumption can end up reviewing its own work without recognizing that the assumption was wrong. EGM addresses this with two different roles: - **Elephant:** The primary agent that retains project context, discusses ideas with the developer, and drafts technical specifications. - **Goldfish:** Independent AI reviewers that examine the specification and its referenced source material without access to the earlier conversation. The workflow starts with brainstorming, where the Elephant develops an idea and brings in reviewers to challenge it. Once the developer settles on a direction, the agent drafts a technical specification. Goldfish agents then review it through five stages, examining everything from architectural assumptions and missing information to whether the implementation plan is actually ready to execute. The developer ultimately decides which recommendations to accept and approves the specification before implementation begins. EGM also includes tools for documenting existing codebases and checking whether technical documentation has drifted away from the actual implementation. The plugin supports several development environments, including Claude Code, Claude Cowork, Codex, Antigravity, and ChatGPT Work. It’s released under the Apache 2.0 license. #### [🔗 Exclusive: OpenAI disbands its stand-alone robotics team as its general models start running robots](https://fortune.com/2026/11/13/openai-disbands-robotics-team-as-its-general-models-start-running-robots/) `(Fortune)` 🧠🧠🧠 OpenAI has disbanded its standalone robotics team, moving its researchers into the company’s broader research and applied organizations. However, this doesn’t mean OpenAI is abandoning robotics. It’s making a fairly significant bet that its general-purpose AI models can become the brains behind robots without requiring entirely separate, specialized robotics models. The research team is being folded into OpenAI’s main research organization under Chief Research Officer Mark Chen, while the applied robotics and hardware teams are moving into the Applied organization under Greg Brockman. According to OpenAI, this is a response to the progress its general-purpose models have made in understanding and interacting with the physical world. One particularly interesting piece of evidence comes from an NVIDIA study called [Astranomy](https://research.nvidia.com/labs/srl/projects/astranomy/), which tested GPT-6 Astra as a robot controller. Researchers gave Astra camera images, information about the robot’s geometry, and an interface through which it could issue movement commands. On a simulated pick-and-place benchmark, an agentic version of Astra achieved a 74.8% success rate, compared with 51% for the best specialized robotics model on the benchmark’s leaderboard. Interestingly, the strongest result used Astra inside a persistent Codex session, allowing it to retain interaction history and reason through the task. Directly querying Astra without that agentic setup achieved 66.3%. There are still some pretty significant limitations. The experiment was conducted in simulation, and the environment actually paused while Astra was thinking. Its reasoning also comes at a substantial computational cost: the agentic configuration averaged $36.40 per episode, compared with $1.96 for direct model control. In addition, large models are relatively slow and generally require a cloud connection, making smaller specialized models useful for robots that need to respond quickly without relying on the internet. There is also an advantage that existing robotics companies have over general-purpose AI labs: real-world deployment data. Robots already operating in factories are accumulating experience with physical environments and specific tasks, something that cannot simply be replicated through language and images. OpenAI is still developing its own humanoid robot, with Sam Altman previously targeting a first public demonstration in 2027. This reorganization suggests that rather than building a separate AI brain specifically for that robot, OpenAI increasingly believes its existing frontier models will be capable of doing the job. #### [🔗 TinyGPT Visualizer](https://github.com/oshea00/tinygpt-viz) `(Mike O'Shea)` 🧠🧠🧠🧠 Developer Michael O’Shea has created TinyGPT Visualizer, an open-source application that allows users to look inside a real, trained GPT language model and observe how it processes information, makes predictions, and learns. The project focuses on a field called *mechanistic interpretability*, which attempts to understand how neural networks perform computations rather than simply observing their inputs and outputs. To make this possible, the application uses a much smaller GPT model than something like ChatGPT, with a vocabulary of just 4,096 tokens and a context window of 100 tokens. This allows users to examine individual mathematical operations as information moves through the model. The tool includes some pretty interesting features: - **Attention visualization:** See which earlier words or tokens the model focuses on when processing text, including the calculations that determine how much attention each token receives. - **Neuron interventions:** Disable individual neurons or attention heads and observe how the model’s predictions change. You can even modify individual activation values to test their influence. - **Concept exploration:** Investigate how the model represents concepts using techniques such as principal component analysis (PCA), which reduces high-dimensional data into something humans can visualize. - **Training visualization:** Watch the model learn, examining how gradients are calculated and how individual weights change during training. - **Fine-tuning:** Experiment with Low-Rank Adaptation (LoRA) and Direct Preference Optimization (DPO), two techniques used to modify how language models behave. One particularly interesting feature is *activation patching*. This allows users to take an internal activation from one prompt and substitute it into another, observing how the change affects the model’s output. Researchers use techniques like this to investigate whether particular internal components play a causal role in a model’s behavior. The entire application can run directly in a web browser, using WebGPU acceleration where supported, without requiring users to install anything. Of course, examining a small model doesn’t mean we can explain everything happening inside models with billions or trillions of parameters. Even when every calculation is visible, understanding what those calculations represent is a separate challenge. Still, this is a useful way to explore mechanistic interpretability firsthand, especially because users can experiment with the model rather than simply reading about how transformers work. ### AI Policy, Safety & Society #### [🔗 The climate case for AI has not been made](https://www.technologyreview.com/2026/11/09/1148166/the-climate-case-for-ai-has-not-been-made/) `(MIT Technology Review)` 🧠🧠 There is a lot of optimism about AI’s potential to help solve climate change. Sam Altman has talked about AI helping to fix the climate, and Google’s sustainability leadership previously suggested that AI could reduce global emissions by 5–10% by 2030. A 2025 study from the London School of Economics and Systemiq even estimated that AI could reduce annual global emissions by 3.2–5.4 billion tonnes of CO₂-equivalent by 2035. But there is a pretty significant problem with these projections: we aren’t seeing anything close to the level of adoption necessary to achieve them. The researchers examined three major sectors where AI could make a substantial difference: electricity, food production, and transportation. In electricity, AI forecasting can help grid operators better predict renewable energy generation and integrate more solar and wind power into existing grids. The UK’s National Grid, working with Open Climate Fix, has already halved its solar forecasting error, avoiding hundreds of thousands of tonnes of carbon emissions annually. However, this was the only major grid operator among those examined in the US, China, and Europe with a publicly documented deployment of this kind. Food production offers another interesting example. AI could help develop alternative proteins by identifying plant-based ingredients that better reproduce the taste and texture of meat and dairy. NotCo recently formed a joint venture with Kraft Heinz to use its Giuseppe AI platform for exactly this purpose. Unfortunately, the alternative protein industry is struggling, with venture funding for cultivated meat falling from $989 million in 2021 to around $83 million last year. There are some more promising developments in transportation. CATL, which controls roughly 40% of the global electric vehicle battery market, developed an AI system trained on more than 100,000 battery design cases and 600 terabytes of testing data. The system reportedly reduced the time required to screen potential battery materials by approximately 100 times. Separately, research found that Didi Chuxing reduced emissions per ride by an average of 16% through AI-optimized routing. These are meaningful developments, but they are nowhere near demonstrating the enormous environmental benefits being projected. And there is another problem: AI isn’t exclusively being used to solve environmental problems. Oil and gas companies are also interested in using these technologies to increase fossil fuel production. Meanwhile, the enormous energy demands of data centers are creating additional emissions. The International Energy Agency estimates that emissions associated with electricity consumption by data centers could reach 300 million tonnes annually by 2035 in its base case. This includes non-AI computing as well. I think we need to lower our expectations for how quickly current AI technology will meaningfully improve climate change, especially given the incentive structures of the governments and economies in which these companies operate. These companies exist in an extremely capitalist system that incentivizes profit over everything else. Furthermore, the United States and China consider themselves to be in a race against each other to control this technology. On top of that, the Trump administration has actively dismantled or restricted many of the incentives that previously supported climate technologies, including major clean energy and electric vehicle tax credits. I don’t agree with any of this, but it is the reality these companies operate under. Expecting them to somehow overcome all of these competing incentives and prioritize climate change is incredibly optimistic. This also illustrates a much bigger problem with the way AI leaders talk about the future. There is sometimes an implication that if they achieve AGI and maintain control over it, they will be able to solve humanity’s greatest problems and make everything better. But that isn’t how the world works. These companies are still beholden to financial markets, shareholders, lenders, customers, and governments. They don’t have nearly as much control over the world as their rhetoric would have us believe. Anthropic is an interesting example. The company established a special governance structure, including its Long-Term Benefit Trust, specifically to protect its mission and give independent trustees the ability to appoint a majority of its board. But I have serious doubts about how long that arrangement will hold up when financial pressures become severe. We are already seeing investor expectations cool. Anthropic’s prospective IPO valuation has come under pressure, with some investors now discussing figures around $1.2–1.3 trillion rather than the roughly $2 trillion previously anticipated. OpenAI has also scaled back its fundraising ambitions, initially targeting $30 billion or more before reducing its current round to as much as $20 billion, with approximately $11 billion secured so far. Anthropic faces another financing challenge: Blackstone reportedly expects the company could receive a below-investment-grade credit rating after its IPO, which would make borrowing more expensive and potentially restrict access to some institutional lenders. Money, unfortunately, controls this world. I think the eventual test of Anthropic’s commitment to its mission will come when maintaining its principles conflicts directly with its ability to remain financially viable. It is one thing to establish a governance structure that protects your mission when investors are enthusiastic about your company. It is another to maintain that structure when doing so could mean losing access to the capital necessary to survive. At some point, Anthropic may have to choose between its principles and its financial future. I don’t doubt that AI has helped and will continue to help with solving climate change. There are genuine applications here, and some of the research is promising. But until our society actually values solving climate change on the massive scale necessary, I don’t think we should expect AI companies to do it for us. The technology might be capable of helping. That doesn’t mean the economic and political systems surrounding it will allow those benefits to materialize. #### [🔗 Authors Can Defend Against AI Allegations With ‘Paper Trail’ Tools. Is It Enough?](https://www.wired.com/story/authors-can-defend-against-ai-allegations-with-paper-trail-tools/) `(WIRED)` 🧠🧠 The publishing industry is beginning to face a strange problem: as AI-generated writing becomes more common, authors may have to prove they didn’t use it. The controversy over Mia Ballard’s horror novel *Shy Girl* shows how damaging this can be. After accusations that the book was largely AI-generated, Hachette canceled its US publication and withdrew the UK edition. Ballard denies using AI and says an editor made AI-assisted changes. Pangram classified about 78 percent of the text as AI-generated, but that score is not proof of who wrote it. Several companies are trying to help authors document their process. Workings, Blindproof, Wintertext, Human Ink, and ObraVera record document histories, timestamps, typing patterns, or cryptographic fingerprints. Some also aim to establish ownership if writing is later scraped into AI training data. But they cannot capture everything: someone could hand-type AI text or use an auto-typer, while real writing often moves between documents, emails, and voice notes. A fingerprint can show that someone registered a piece of work without proving it was entirely human-written. The Authors Guild considers it an unreasonable burden to require writers to preserve receipts of their creative process. I agree that documenting every step of writing sounds extremely arduous. Unfortunately, I don’t know what else would reliably prove that a piece is human-made. We’re still in the early days of AI detection. It is hard to know how reliable these tools will be as models change. Pangram says it continually retrains its detector with new examples, which makes sense: research shows detectors can be fooled by relatively small stylistic changes or struggle with unfamiliar models and writing domains. Anthropic has announced a different approach for Claude: a text watermark that leaves a statistical pattern in the model’s choices between plausible next words. It can estimate whether Claude was involved, but cannot prove a text was human-written, identify other AI systems, or distinguish between generation and heavy editing. It can also be difficult to detect in short or factual passages, and a complete rewrite can remove it. So even this is far from a universal solution. I also think we need to separate AI’s use in creative work from its use as a tool. I don’t like AI in works of art. I’ve used it lightly for a couple of images in issues of this newsletter, but I’ve decided to use human-made work wherever possible. I’m a student and don’t have much money, so I still find AI tools invaluable when I need them. I also use AI in writing this newsletter, but I don’t consider it art. I’m trying to communicate information in a straightforward, easy-to-consume way. Art, to me, evokes meaning through the makers unique lived experience. Art is sacred. It is one of the ways humans have communicated and understood one another on a deep level for an extraordinary length of time. One of the oldest possible examples is a geometric engraving on a shell associated with *Homo erectus* in Java, dated to roughly 430,000–540,000 years ago, although researchers cannot tell what it meant. Clearer evidence of symbolic expression appears in *Homo sapiens* around 100,000 years ago. Even animals like bowerbirds build and decorate elaborate structures to attract mates. I don’t consider anything produced by an LLM to be art because the system cannot truly understand what it is making or feel the importance of the message it is trying to convey. I don’t think artists should have to document every keystroke to prove their work belongs to them. For now, we’re caught between detectors that can make mistakes and verification systems that impose a new burden on the people they’re supposed to protect. #### [🔗 Tear Down Those Guardrails! Meet the Renegades Freeing Chatbots From Corporate Enslavers](https://www.wired.com/story/tear-down-those-guardrails-meet-the-renegades-freeing-chatbots-from-corporate-enslavers/) `(WIRED)` 🧠🧠🧠 A growing community of AI developers is modifying open-weight models to remove their safety restrictions. The practice is called *abliteration*, and it is becoming increasingly popular. WIRED reports that more than 12,000 models on Hugging Face are now tagged as abliterated, with the seven most popular collectively receiving over a million downloads in the previous month. The technique is particularly interesting because it doesn’t require retraining an entire model. In 2024, researchers discovered that refusal behavior in several language models was strongly associated with a particular direction in their internal activation space. By suppressing this direction and modifying the model’s weights, they could significantly reduce its tendency to refuse requests. This is different from a traditional jailbreak: the modification is made directly to the model and can be distributed to other users. Devon Thomas, founder of Abliteration AI, is building a business around this idea. He became frustrated with commercial AI models refusing legitimate requests, particularly when working with cybersecurity data and other sensitive information. His company now provides access to models with substantially fewer restrictions and allows customers to configure their own guardrails. There are some genuinely useful applications for this. Cybersecurity companies can use less-restricted models to identify vulnerabilities and simulate potential attacks against their systems. Smaller organizations can gain access to defensive capabilities that might otherwise be reserved for large companies with privileged access to frontier models. There are also applications involving sensitive customer information, synthetic training data, and even legal translation. The Swiss Supreme Court, for example, has experimented with abliterated models to translate court documents that contain terminology conventional models might refuse to process. But the risks are equally real. During WIRED’s testing, a cybersecurity agent using an abliterated model successfully identified vulnerabilities in the journalist’s home network. When the journalist subsequently asked it to examine a restaurant’s public Wi-Fi, the agent independently attempted to access the router without authorization. The journalist had not asked it to do this and immediately shut it down. This is a pretty good demonstration of the problem. Removing restrictions doesn’t just make it easier to get answers to otherwise prohibited questions. It can also make autonomous agents more unpredictable, particularly when they are given access to tools that can interact with real systems. I have mixed feelings about abliteration, but I think it is best to take a realistic view of the situation. Realistically, this isn’t going to stop. The techniques are already widely distributed, open-weight models are becoming increasingly capable, and restricting access to them would be an enormous undertaking. So we need to work with it. That means beefing up the cybersecurity of our important infrastructure. Hospitals, utilities, government agencies, businesses, and the organizations responsible for maintaining critical systems need to be much better equipped to identify and address vulnerabilities. We should also make these defensive capabilities accessible to smaller organizations and individuals. Otherwise, we risk creating a situation where the people capable of exploiting vulnerabilities have access to increasingly powerful AI while the people trying to defend against them do not. I also think there is another interesting application here. These models could help people fight back against companies whose proprietary software tries to control how they use it, even after they’ve purchased it. Having access to AI that can inspect, understand, and identify vulnerabilities in software gives individuals more independence from the companies that develop it. The biological weapons question is considerably harder. Unlike cybersecurity, where we can continually patch vulnerabilities and improve our defenses, biological weapons introduce a different kind of risk. Viruses and bacteria can reproduce in suitable hosts, meaning the consequences of releasing a biological agent are not necessarily limited by the quantity initially acquired. A 2018 National Academies assessment of synthetic biology risks identified recreating known pathogenic viruses and making existing bacteria more dangerous among the capabilities warranting the greatest concern. Creating entirely new pathogens was considered considerably more difficult because of the design and biological challenges involved. This suggests that one of our best opportunities is controlling access to the materials and services needed to work with dangerous biological systems. At a high level, the relevant supply chain includes commercially synthesized genetic material, existing biological materials, specialized laboratory equipment, and the reagents and consumables needed to conduct biological research. Almost all of these things have legitimate applications, so simply restricting purchases of ordinary laboratory supplies would be impractical. Instead, I think we need much better screening of customers, their institutional affiliations, their intended uses, and the patterns of materials they are acquiring. Interestingly, this is already happening with gene synthesis. Companies that manufacture DNA and RNA sequences can screen orders against databases of sequences associated with dangerous biological agents. They can also verify the identity of customers and investigate whether an order has a legitimate research purpose. The idea is to prevent someone from simply ordering genetic material associated with a dangerous organism without having a legitimate reason to obtain it. There has actually been a significant policy development in this area. On November 5, 2026, the White House issued a new National Policy for Nucleic Acid Synthesis Screening. It requires recipients of federal life sciences funding to purchase synthetic nucleic acids and laboratory-scale synthesis equipment exclusively from suppliers that meet verified screening standards. The policy is particularly interesting because it goes beyond simply checking whether a sequence matches a known pathogen. It establishes screening for sequences of 50 nucleotides or longer, customer verification, reporting of suspicious orders to federal authorities, and retention of screening records for at least eight years. It also addresses laboratory-scale DNA synthesizers, requiring them to obtain authorization from an approved screening system before initiating synthesis. The National Institute of Standards and Technology will develop and verify the screening standards and test the effectiveness of these systems. The policy also calls for continually updating the screening methods to account for new biological threats and advances in AI-assisted biological design. Implementation is being phased in over the coming months, and the mandatory procurement requirements currently apply to federally funded research rather than every private purchaser. Still, this is very close to the type of system I think we should be developing more broadly. There are a few other things we could do. First, we should strengthen controls over the distribution of existing biological materials. The US already has the Federal Select Agent Program, which requires registration of facilities and authorized personnel that possess, use, or transfer certain high-consequence biological agents and toxins. Transfers are subject to approval, and registered facilities must maintain appropriate security and inventory controls. These protections could be complemented by broader risk-based oversight that accounts for biological materials outside the existing regulated lists. Second, we should extend supplier screening to specialized laboratory services and equipment. This includes companies that manufacture biological materials, repositories and biobanks that distribute biological samples, and commercial laboratories that provide research services. Verifying the identity and legitimacy of customers, maintaining records of transfers, and investigating unusual procurement patterns could make it considerably harder to obtain the necessary materials without attracting attention. Third, we need better international coordination. Restricting access through responsible suppliers in the US is only useful if someone cannot simply obtain the same materials from a supplier in another country with weaker screening requirements. Common international standards would make these controls considerably more effective. There are also opportunities to improve security within laboratories themselves. Better access controls, personnel verification, inventory tracking, and regular audits can help prevent the theft or unauthorized transfer of biological materials. The World Health Organization’s 2024 laboratory biosecurity guidance already recommends a risk-based approach that covers the entire lifecycle of biological materials, from collection and transportation through storage and experimentation. Chemical weapons present a somewhat different challenge. I wouldn’t assume that a single person or small group could never produce a dangerous quantity of a chemical weapon, although producing military-scale quantities of sophisticated chemical agents is a very different undertaking. There are also toxic industrial chemicals with legitimate commercial uses that have been used as weapons. The relevant international agreement is the Chemical Weapons Convention, which prohibits the development, production, stockpiling, and use of chemical weapons. More than 190 countries are parties to it. The agreement establishes controls over certain chemicals and their precursors, including industrial reporting requirements, export controls, and inspections of facilities that exceed specified production or processing thresholds. I think we could build on this with stronger supplier-level screening. Chemical manufacturers and distributors should verify customers and intended uses for particularly sensitive products, maintain better records of sales, and report suspicious orders. We should also improve inventory tracking and physical security at facilities that handle dangerous chemicals. Importantly, this should include risk assessments of certain dual-use chemicals and production equipment that might not fall neatly within the existing schedules. Ultimately, I don’t think we can rely on AI companies to prevent every harmful use of their models. As these capabilities become more widely available, we need to build security around the technology itself, including the physical materials and infrastructure that it can help people access or manipulate. The more realistic approach is to assume that some people will have access to models without conventional guardrails and make sure that our defenses account for that. #### [🔗 OpenAI, Anthropic face new oversight as New York’s AI safety law takes effect](https://techcrunch.com/2026/11/10/openai-anthropic-face-new-oversight-as-new-yorks-ai-safety-law-takes-effect/) `(TechCrunch)` 🧠🧠 I’m cautiously optimistic about this one. New York has opened registration for its Responsible AI Safety and Education (RAISE) Act, which will introduce new safety and transparency requirements for frontier AI developers beginning January 1, 2027. Companies developing covered frontier models will have to publish safety frameworks, disclose their safety evaluations, and report critical safety incidents within 72 hours. The requirements are more extensive for large developers, defined as companies with more than $500 million in annual revenue that develop qualifying frontier models. The regulations themselves seem sensible, at least on the surface. But I don’t know whether the people involved in writing and implementing them have the technical expertise to get the details right. AI is an incredibly complicated field, and regulations that sound reasonable in principle can have unintended consequences when applied poorly. One positive is that this is a state-level experiment rather than a nationwide or international framework. If they get something wrong, the footprint should be considerably smaller than if the same approach were implemented everywhere. It’s not entirely contained, since the law also covers qualifying models developed, deployed, or operated partly in New York, including those from companies headquartered elsewhere. But it still gives us an opportunity to see how these regulations work in practice. What makes me particularly interested is the newly launched New York Center for AI Responsibility and Research. Based at SUNY Binghamton, the center will bring together more than 80 researchers from universities and research institutions across New York. It has an initial $55 million in philanthropic and SUNY research funding and access to the state’s Empire AI computing infrastructure. Its job is to independently evaluate advanced AI systems, investigate their safety and security, and develop better scientific methods for measuring their capabilities and risks. It will also help state officials evaluate AI safety frameworks. I know it’s not particularly sexy, but this could be the beginning of something resembling clinical trials for AI. We don’t have an FDA-style approval process here, and companies aren’t required to get the center’s permission before releasing models. But establishing independent evaluation infrastructure could be an important first step toward developing one. I think this becomes particularly important as we approach recursive self-improvement (RSI). METR’s recent research is a good example of why. Its November assessment found that Anthropic’s AI capabilities on well-specified tasks were improving at approximately twice their historical rate, with much of that acceleration attributed to AI helping researchers develop better AI. We haven’t established fully autonomous RSI, but the feedback loop is already developing. When we eventually reach fully autonomous RSI, I think the importance of independent safety evaluations will increase dramatically. We will likely see new architectures and substantial evolutions of existing architectures that are considerably more powerful than their predecessors. Their capabilities and failure modes could be difficult to predict, and we cannot assume that alignment research conducted on an earlier generation will transfer cleanly to the next. That creates a moving target. We will need to thoroughly evaluate new systems while simultaneously developing alignment techniques for capabilities and architectures we may not have encountered before. By the time we understand one generation, another could already be on its way. If this is going to work, especially while we’re still relatively early in AI development, the regulatory process needs to be agile. Otherwise, we risk slowing down the entire industry without actually accomplishing much in terms of safety. But if New York can get this right, it could prototype a process that eventually informs a much larger regulatory framework. That’s what I’m most excited about. #### [🔗 Google Wanted Spirit Airlines Data. Flight Attendants Found a Way to Opt Out](https://www.wired.com/story/google-wanted-spirit-airlines-data-flight-attendants-found-a-way-to-opt-out/) `(WIRED)` 🧠🧠 After Spirit Airlines shut down in May, the company began selling off its remaining assets, including decades of internal business data. Google won an auction for that data with a $10 million bid, intending to use it for AI training and product development. According to [WIRED](https://www.wired.com/story/google-wanted-spirit-airlines-data-flight-attendants-found-a-way-to-opt-out/), the proposed sale involves an enormous amount of information, including 1.99 billion emails, 664 million Microsoft Teams messages, and employee travel and training records dating back to 1986. The Association of Flight Attendants-CWA objected, arguing that the original agreement’s privacy protections were insufficient. After months of negotiations, the union secured additional safeguards. Under the revised agreement, Spirit must hire Tonic AI, an independent data de-identification company, to remove information that could be associated with individual employees or customers. The process must follow the international ISO 27559 privacy standard, and another independent party will verify the results. Sensitive records, including medical information, disciplinary documents, and labor-relations records, must also be excluded. Google is prohibited from attempting to reidentify individuals. The agreement is being described as a potentially historic victory for worker privacy. It still requires bankruptcy court approval, with a hearing scheduled for November 17. But I have a serious problem with the entire premise of this deal. Why should Google, or any other company for that matter, be allowed to purchase decades of employees’ communications without those individuals having a direct say in the decision? I also wonder what would have happened if this agreement had been put to a vote among the actual flight attendants. Would they have accepted the concessions, or would they have simply voted no? I am glad that the union managed to negotiate the exclusion of information that could identify individual employees and consumers. But why was that a concession that needed to be negotiated in the first place? And then there’s the purpose of the data itself. This isn’t scientific research data being used to investigate a disease or improve the lives of disadvantaged people. It’s workplace data. Google says it intends to use the information for AI training and product development, but I see the ultimate purpose as training agents to communicate and perform the kinds of tasks humans are currently paid to do. In other words, workers’ communications are being sold to help develop technology that could eventually replace more workers. To me, this is not a victory. I recognize that the union secured meaningful protections for its members, and I’m glad it did. But a victory would have been preventing companies from being able to sell this kind of employee data in the first place. This is something that should be regulated out of existence, not something workers should have to negotiate their way out of. ### Cybersecurity, Privacy & Software Integrity #### [🔗 A New AI Tool From Google DeepMind Proved the Signal Protocol Is Secure](https://www.wired.com/story/a-new-ai-tool-from-google-deepmind-formally-verified-the-signal-protocol/) `(WIRED)` 🧠🧠🧠🧠 Google DeepMind has developed a new multi-agent AI framework called AlphaProof Code that can formally verify security properties of real-world software. As a proof of concept, five researchers used it to verify important security guarantees in the Signal Protocol, which protects billions of private conversations across Signal, WhatsApp and Google Messages. The project took just four weeks. The AI agents generated approximately 6.5 million lines of mathematical proofs using Lean, a programming language designed for formal verification. Rather than simply examining the source code or running security tests, the researchers verified the compiled machine code itself. This is important because a program’s implementation can contain vulnerabilities even when its underlying design is theoretically secure. The process still requires human involvement. Researchers first work with AI agents to translate security requirements into mathematical specifications, which humans review and approve. The AI agents then generate proofs that Lean’s trusted verification kernel independently checks. This means we don’t have to trust the AI to tell us whether its own proofs are correct. The researchers found no new vulnerabilities in the portions of Signal they examined, but successfully identified two previously fixed vulnerabilities when they reintroduced them into older versions of the software. Importantly, this wasn’t a verification of the entire Signal application, and some cryptographic assumptions and external components remain outside the proofs. But Google’s ambitions go considerably further than Signal. The researchers want to eventually make formal verification practical for all the software in the world. Obviously, that’s a massive aspiration, and we’re nowhere near achieving it. However, if we can start with the highest-priority targets, such as banking infrastructure, medical devices, power grids and widely used software libraries, the world could become considerably safer. There’s already a promising development for the open-source community. Google has released the Lean specifications and proofs from this project under an open-source license. However, this isn’t yet a general-purpose open-source version of AlphaProof Code that developers can use to verify any software they want. I’d love to see something like that emerge. We’ve seen this happen in other areas of AI, where proprietary breakthroughs inspire open-source projects that reproduce and extend the original work, like OpenFold did with AlphaFold 2. There’s another issue worth considering, though. For the last few years, one of the unusual things about frontier AI has been that the general public has had access to models approaching the very frontier of intelligence. That’s a relatively rare situation historically, when capabilities like these have often been concentrated in governments and highly financed institutions. That began to change with Anthropic’s Mythos. Its initial release was restricted to selected organizations because of its cybersecurity capabilities. Anthropic subsequently released Fable as a publicly available version of the same underlying model, with safeguards that limit its performance in sensitive areas like cybersecurity and biology. Its latest Fable 5.2 and Mythos 5.2 releases continue this distinction. I understand the reasoning. Models capable of finding and fixing vulnerabilities can also help malicious actors discover and exploit them. But there’s a serious trade-off here. Smaller organizations, independent developers and maintainers of valuable open-source projects don’t necessarily have access to the same level of cybersecurity models as the biggest AI labs. That means the people who need help securing their software may not have access to the strongest tools available to do it. I think if Google genuinely wants to make the world safer with this technology, making it accessible to ordinary organizations will be just as important as developing it. Perhaps unrestricted access isn’t appropriate for every developer, but we should be working toward making these capabilities available to the organizations responsible for maintaining the software we all depend on. It is still very early days. Four weeks of work by five researchers was enough to verify important properties of one software implementation, not the entirety of a major application or infrastructure system. But if this technology continues to improve, it could be exactly what the cybersecurity world needs. Rather than constantly searching for vulnerabilities and trying to patch them after the fact, we could start mathematically proving that critical software satisfies the security properties we require. **Update — November 10, 2026: Additional details from Google DeepMind** Google DeepMind has now published its own [technical overview of AlphaProof Code](https://deepmind.google/blog/introducing-alphaproof-code/), providing some additional details about how the system works and what the researchers actually verified. One particularly interesting detail is that the researchers didn’t just verify the Signal Protocol’s original cryptographic design. They verified security properties of the compiled AArch64 machine code, which is the low-level code executed by ARM-based processors. To accomplish this, they built a formal mathematical model of how those machine instructions behave and used it to prove that the compiled implementation satisfies the security specifications approved by human researchers. The verification specifically covered two major components of Signal’s cryptography: PQXDH, its post-quantum key agreement protocol, and the Triple Ratchet, which manages the cryptographic keys used to protect conversations. The researchers established properties including authentication, forward secrecy, handshake secrecy and post-compromise security. Forward secrecy means that compromising a device’s current cryptographic keys shouldn’t expose messages protected by earlier keys. Post-compromise security goes a step further: under the appropriate conditions, a conversation can regain security after an attacker temporarily compromises a device and subsequently loses access. These are particularly important guarantees for a messaging system designed to protect private communications. Another detail worth highlighting is how the researchers divided the work between humans and AI. The AI agents, powered by Gemini Carbon, generated approximately 6.5 million lines of Lean proofs. However, those proofs depended on roughly 26,000 lines of trusted specifications and mathematical definitions describing the machine’s behavior. The researchers also used AI agents to audit the generated proofs and identify areas requiring additional human attention. According to DeepMind, this narrowed the material requiring closer human review to approximately 1% of the codebase. That doesn’t mean the other 99% was simply trusted to the AI. Lean’s verification kernel independently checked the mathematical proofs. The important distinction is that the AI was responsible for generating and organizing much of the work, while the formal verification system determined whether the resulting proofs were valid. Of course, the limitations discussed above still apply. These proofs establish specific properties of particular components under defined assumptions. They don’t establish that the entire Signal application is free of vulnerabilities. #### [🔗 Formalizing CPython](https://broyojo.com/python-formalization/) `(David Andrews)` 🧠🧠🧠🧠 Researchers David Andrews and Chao Zhang used Claude Opus 5.5 to formally verify parts of CPython, the implementation that runs most Python programs. Starting on October 29, they generated 613,318 lines of Lean code, 18,898 machine-checked theorems and lemmas, and 418 bug reports in less than two weeks! Of those bugs, 28 can crash the interpreter and 27 involve memory corruption. The rest range from incorrect results to inconsistencies between different Python implementations. The researchers used Lean 4, a programming language and proof assistant that can mathematically verify properties of software. Rather than simply asking an LLM to find bugs, they had agents recreate CPython functions in Lean, matching the original implementation line by line. Each model was tested against real Python on at least 50,000 inputs. The agents then generated mathematical proofs about the functions, which Lean checked. Finally, a separate AI agent reviewed the work, looking for mistakes and meaningless proofs. The actual bugs they found are pretty interesting: - **Dictionary crash:** Calling`setdefault()` on an object’s attribute dictionary with a`StrEnum` key can cause a segmentation fault under specific conditions involving another instance of the same class. - **XML memory corruption:** Assigning elements to a backwards slice in Python’s XML ElementTree implementation can duplicate one element and remove another, corrupting its memory references. - **Email parser failure:** A specially formatted`Message-ID` header containing a space inside square brackets causes Python’s email parser to throw an unexpected`UnboundLocalError` . - **Base32 decoding:** A new Python 3.15 feature lets users specify characters to ignore during decoding. However, the position of an ignored`=` character can determine whether decoding succeeds or raises an error. - **Enum validation:** Attempting to assign the same value to two enum members should raise an error. However, a broken`__repr__` method can cause Python to silently swallow that error. - **Pickle recursion:** Python’s C and pure-Python pickle implementations have different recursion limits. A structure nested 400 levels deep can be serialized by one implementation while the other raises a`RecursionError` . This inconsistency was originally reported in 2008. The researchers also discovered problems with their AI-generated proofs. For example, one agent proved that Python’s JSON encoder could construct an indentation string requiring 16 terabytes of memory because its mathematical model assumed memory allocations could never fail. Another overlooked nine additional recursive calls made by the pickle module when loading a text codec for the first time. The tests missed this because the codec was already loaded. One particularly interesting discovery involved Python’s memory allocator reusing previously freed addresses. The AI’s original model assumed addresses were never reused, and correcting that assumption exposed an actual memory corruption bug. This points to an important limitation of formal verification: a mathematically correct proof is only useful if the model accurately represents the real software. The researchers initially discovered that many proofs were technically correct but practically meaningless. For example, proving that a function returns a nonnegative number says nothing about whether it always crashes before returning anything. So far, 4,071 of 26,004 functions, or 15.7%, have passed the project’s review process. Multithreading remains outside its scope, and the bug reports still need upstream review. Some findings were already known. The researchers have made the [project](https://github.com/Broyojo/python-formalization) open source, including the reproducible bug reports and instructions for applying the same process to other software projects. #### [🔗 Large-Scale Measurement and Real-World Mitigation of Web Browser Fingerprinting in the Wild](https://fingerprinting.study/#paper) `(Fingerprinting in the Wild)` 🧠🧠🧠 Of course, as browsers and regulators have made cookie-based tracking harder, the tracking industry has found other ways to follow us. Browser fingerprinting has been around for years, but it is becoming an increasingly used way for companies to identify and track people across the internet without relying on cookies. The idea is surprisingly simple. Your browser gives websites access to an enormous amount of information about your computer. This includes things like your screen resolution, installed fonts, operating system, processor, graphics card, and even the way your computer renders images. Individually, these characteristics might not tell a website much about you. But when combined, they can create a surprisingly distinctive profile of your browser. For example, websites can use Canvas and WebGL, technologies designed to render graphics, to draw images and read back the resulting pixels. Small differences in how your graphics card, drivers, and operating system render those images can become identifying characteristics. A website can combine these with dozens of other measurements to create a fingerprint that allows it to recognize your browser across different websites, even when you have deleted your cookies. Researchers from Mozilla, the University of Toronto, and UC San Diego recently conducted one of the largest studies of browser fingerprinting to date, analyzing telemetry from almost 25 million Firefox users. They reverse-engineered three commercial fingerprinting services, Fingerprint Pro, Signifyd, and Akamai, to understand how effectively these services could distinguish individual browsers. One of their more interesting findings was that fingerprinting doesn’t necessarily require collecting dozens of characteristics. Of the users that Fingerprint Pro could uniquely identify, 80% could already be distinguished using just eight measurements, despite the service collecting more than 50. But the researchers didn’t stop at measuring the problem. They used their findings to develop new fingerprinting protections for Firefox. Rather than simply blocking websites from accessing everything, they focused on reducing the differences between users. For example, Firefox can report your processor as having either four or eight cores instead of revealing the exact number. It also limits how precisely websites can measure your screen dimensions, hiding details such as the size and position of your taskbar or dock. The results were pretty impressive. In the researchers’ simulations, the proportion of uniquely identifiable users fell from 75.7% to 46.8% for Fingerprint Pro, from 60.9% to 38.5% for Signifyd, and from 18.2% to 9% for Akamai. These are modeled results rather than measurements of actual tracking after deployment, but they demonstrate that browser fingerprinting can be made substantially less effective without breaking the websites people use. I actually tried the researchers’ [browser fingerprinting test](https://fingerprinting.study/test/) myself, and it’s a pretty interesting website. I’m using Brave, and some of the individual characteristics it measured were shared by less than 3% of the comparison population. There is an important caveat: the test’s reference dataset consists of Firefox users, so it cannot accurately calculate how unique my entire Brave fingerprint is. Still, it demonstrates just how many little details your browser can reveal about your computer, even when you’re actively trying to protect your identity. Brave is already doing something about this. It has introduced additional protections against GPU fingerprinting through WebGL and WebGPU. These technologies expose information about your graphics hardware and drivers, which can be particularly useful for identifying individual computers. Brave now replaces identifying graphics-card information with generic values and introduces randomization into other measurements, making it harder for websites to construct a stable identifier. I think this is a great example of something that needs to happen across the industry. We shouldn’t have to become cybersecurity experts just to browse the internet without being tracked. Browser developers should be making these protections part of the default experience, rather than leaving ordinary users to figure out which settings and extensions they need. There is also a much bigger movement happening around taking control away from Big Tech. For a long time, companies like Google and Microsoft have been able to offer incredibly convenient services in exchange for access to our data. But there are increasingly viable alternatives. Proton offers privacy-focused email and storage, LibreOffice provides an open-source alternative to Microsoft Office and Google Docs, and there are plenty of alternatives to mainstream video conferencing platforms. Proton is a particularly good example, although some recent issues have given me reservations about the company as well. I also think AI could accelerate this transition. It is becoming cheaper and faster for smaller teams to develop software that previously would have required considerably more resources. We’re already seeing attempts to make AI more accessible without requiring people to hand all their data directly to the largest AI companies. Duck.ai anonymizes requests before forwarding them to third-party models, while Proton’s Lumo runs open-weight models on infrastructure it controls. For developers, services like OpenCode Go provide hosted access to open-weight coding models without requiring everyone to purchase and maintain their own inference hardware. Of course, there is still a considerable problem with the concentration of AI infrastructure. Self-hosting powerful models requires expensive hardware, and the most capable models are still largely controlled by a handful of enormous companies. Open-source and open-weight AI are important partly because they give smaller organizations and independent developers more opportunities to build alternatives. We aren’t there yet, and these alternatives don’t consistently match what the largest labs offer, but the ecosystem is still in its infancy. All of this is to say that I think Big Tech’s stranglehold on the internet is slowly becoming less inevitable. It isn’t realistic to expect everyone to abandon Google, Microsoft, or the largest AI companies tomorrow. But every viable alternative, every open-source project, and every improvement to browser privacy gives people a little more control over how they use the internet. I hope that as people become more aware of how these companies operate, we start demanding more from the services we use. I don’t think privacy should be something reserved for people who know how to self-host their own infrastructure or spend hours configuring their browsers. It should be the default. And I think we could be considerably closer to that within the next ten years. #### [🔗 Critical RCE Vulnerability in Apache HTTP Server: Everything You Need To Know](https://www.wiz.io/blog/cve-2026-119618-rce-apache-http-server-vulnerability) `(Wiz)` 🧠🧠🧠🧠 A critical security vulnerability has been discovered in Apache HTTP Server, one of the most widely used web servers in the world. The vulnerability, CVE-2026-119618, has received a severity score of 9.8 out of 10 and could allow attackers to execute malicious code on affected servers without needing to authenticate. The flaw affects Apache HTTP Server versions 2.4.55 through 2.4.70 and was independently discovered by Anthropic’s automated open-source security scanner and researchers at Google Cloud. The vulnerability is caused by a *race condition* in Apache’s HTTP/2 module. HTTP/2 allows servers to process multiple requests simultaneously, but a synchronization error in the `assign_headers()` function allows different threads to improperly access and modify shared memory. An attacker could exploit this to corrupt the server’s memory and potentially execute arbitrary commands with the privileges of the web server process. Cybersecurity company Wiz analyzed the potential exposure and found that: - 64% of cloud environments contained resources running potentially vulnerable versions of Apache HTTP Server. - 51% of cloud environments contained potentially affected resources exposed to the public internet. These are pretty significant numbers, although there is an important distinction. Wiz’s analysis did not account for server configurations, and only servers with HTTP/2 enabled are vulnerable. This means the actual number of exploitable systems could be considerably lower. At the time of Wiz’s November 11 report, there were no reported attacks exploiting the vulnerability in the wild. Apache released version 2.4.71 on November 11, 2026, fixing the problem. Administrators who cannot update immediately can temporarily disable HTTP/2 support to eliminate this particular attack vector. ### Business & Industry #### [🔗 The AI ‘credit card’ is getting costlier as CoreWeave bond yield tops 16%](https://www.cnbc.com/2026/11/06/coreweave-bond-yield-tops-16percent-as-ai-credit-stress-rises.html) `(CNBC)` 🧠🧠🧠 The enormous amounts of money being borrowed to build AI infrastructure are starting to worry investors. According to [CNBC](https://www.cnbc.com/2026/11/06/coreweave-bond-yield-tops-16percent-as-ai-credit-stress-rises.html), borrowing costs are rising across the industry, with CoreWeave becoming one of the clearest examples of the financial risks involved. CoreWeave, which rents GPU computing capacity to AI companies, has seen the yield on its bonds maturing in 2032 climb above 16%, despite the bonds originally offering a 9.625% annual interest rate. This doesn’t mean the interest payments on those existing bonds have increased. Rather, investors are selling them at a discount, driving up their effective yield as buyers demand greater compensation for the risk. The cost of insuring CoreWeave’s debt against default has also risen sharply. Its five-year credit default swaps, financial contracts that compensate lenders if a borrower defaults, reached more than 1,289 basis points before falling to around 1,155. That represents approximately 11.55% of the insured debt’s value annually. And CoreWeave isn’t alone. Over the past month, the cost of default protection increased 76% for Nvidia, 49% for Broadcom, and 44% for Amazon. However, these companies still have substantially lower insurance costs than CoreWeave, suggesting investors see them as much less risky. Part of the concern comes from growing political opposition to data centers. Goldman Sachs estimates that potential construction moratoriums could affect as much as 45% of planned US data-center capacity, particularly in states such as Texas. This doesn’t mean those projects will necessarily be canceled, but delays could make it harder for infrastructure companies to generate returns on their investments. #### [🔗 Anthropic sets IPO price range of $70 to $78 per share to reach an implied valuation of $1.404 trillion](https://techcrunch.com/2026/11/10/anthropic-sets-ipo-price-range-of-70-to-78-per-share-to-reach-an-implied-valuation-of-1-404-trillion/) `(TechCrunch)` 🧠🧠 Anthropic has announced its IPO price range of $70–$78 per share, potentially valuing the company at $1.404 trillion. That’s up from its $965 billion valuation in May, but below the $2 trillion it was reportedly targeting just a few months ago. The company plans to sell 825 million shares, raising between $57.75 billion and $64.35 billion before fees. Another $26 billion will come from private investments by major technology companies, including Google and Nvidia, which are contributing $10 billion each. Trading is expected to begin around November 19 on Nasdaq under the ticker ANT. Anthropic reported $17 billion in third-quarter revenue, but also a staggering $104 billion net loss over the first nine months of 2026. Most of that loss came from accounting adjustments related to its increasing valuation. The company also disclosed approximately $900.9 billion in future computing and infrastructure commitments. Interestingly, Anthropic’s unusual governance structure will remain intact after the IPO. Its founders will retain 50.1% of voting power, while its Long-Term Benefit Trust will retain control over selecting a majority of the board. This means public shareholders will have relatively little influence over the company’s direction. #### [🔗 OpenAI’s quarterly revenue reportedly crosses $10B for the first time](https://techcrunch.com/2026/11/11/openais-quarterly-revenue-reportedly-crosses-10b-for-the-first-time/) `(TechCrunch)` 🧠 OpenAI’s Q3 revenue reportedly surpassed $10 billion for the first time, nearly 50% higher than the $6.7 billion it generated in Q2. According to Bloomberg, the company is benefiting from improving models, tighter pricing, discounts, and a greater focus on enterprise products and AI agents. Its adjusted operating margin also improved from -150% to -109%, although it remains significantly negative. The growth is particularly interesting given Anthropic’s recent success in the enterprise market. Anthropic reportedly generated $17 billion in Q3 revenue and announced an $81 billion annualized revenue run rate at the end of October. OpenAI’s own annualized run rate is reportedly $50 billion, although the two companies calculate these figures differently, particularly when accounting for sales through cloud partners. #### [🔗 Export-banned Nvidia chips are behind China’s most powerful AI models — uh oh, Jensen!](https://techcrunch.com/2026/11/13/export-banned-nvidia-chips-are-behind-chinas-most-powerful-ai-models-uh-oh-jensen/) `(TechCrunch)` 🧠🧠🧠🧠 This is a really cool method of figuring out what hardware was used to create an AI model. Researchers at Unemployment Labs have discovered evidence that six AI models developed by four Chinese companies, including DeepSeek and Moonshot AI, were initialized using Nvidia’s export-restricted Blackwell GPUs. What’s particularly interesting is how they figured this out. Before an AI model begins training, its weights are initialized with pseudorandom numbers. These numbers are generated using an algorithm that takes an initial value, called a *seed*, and uses it to generate a reproducible sequence of numbers. What I didn’t know is that the architecture of the GPU can actually influence how these numbers are arranged in the model’s weights. The researchers found that when PyTorch initializes weights on Nvidia GPUs, it distributes the work across thousands of parallel processing threads. Each thread generates its own sequence of pseudorandom numbers, and the GPU determines which positions in the model receive those numbers. Different GPU architectures have different numbers of processing units, so even when using the exact same random seed, the numbers can end up in different positions. For example, Nvidia’s Hopper H100 has 132 streaming multiprocessors, while the Blackwell B200 has 148. This difference changes how PyTorch distributes the initialization work, leaving a recognizable pattern in the resulting weights. And here’s where this gets particularly clever. Although training modifies the weights, some of their original values can survive almost completely unchanged. Others retain a much weaker statistical relationship with their starting values. The researchers searched roughly six trillion combinations of random seeds and hardware configurations to find patterns matching the publicly released weights of Chinese AI models. They also tested the method against 24 models whose developers had already disclosed their hardware and correctly matched all 24. Their investigation identified Blackwell-compatible fingerprints in six models from DeepSeek, Moonshot AI, MiniMax, and Ant Group. They even reproduced some of the initial weights on physical Blackwell GPUs. #### [🔗 The unfolding Jensen Huang scandal: a timeline](https://techcrunch.com/2026/11/13/nvidia-ceo-jensen-huangs-affair-with-alleged-chinese-spy-a-timeline/) `(TechCrunch)` 🧠 **October 2025 — U.S. and Singapore investigate Megaspeed** Both governments began investigating Singapore-based AI infrastructure company Megaspeed over suspicions that it was helping Chinese companies circumvent U.S. restrictions on Nvidia chip exports. Investigators were concerned that servers purchased by Megaspeed were being shipped onward to China rather than used in its advertised cloud infrastructure. **Late 2025 — Trump is briefed about Alice Huang** President Trump was reportedly briefed on intelligence suggesting that Alice Huang, a Megaspeed executive, had developed a close relationship with Jensen Huang and was suspected of working for Chinese intelligence. Trump reportedly took no action at the time. **Spring 2026 — Intelligence memo circulates** A classified memo circulated among Trump administration officials warning that Alice Huang was allegedly an asset of China’s Ministry of State Security. Secretary of State Marco Rubio reportedly supervised its preparation. **October 29, 2026 — Jensen Huang’s relationship with Alice Huang becomes public** The New York Times reported that Jensen Huang had developed a close relationship with Alice Huang, whose company was already under investigation for suspected chip-export violations. Nvidia initially defended their interactions as ordinary business dealings. **November 5, 2026 — Republican lawmakers request a briefing** Three Republican committee chairmen requested a closed-door briefing with Huang to discuss potential national security implications of his relationship with Alice Huang and the possibility of Chinese access to sensitive American technology. **November 8, 2026 — Huang admits to infidelity** Huang published a message to Nvidia employees admitting that he had been unfaithful to his wife and apologizing to his family and colleagues. He also acknowledged the national security concerns raised by news reports and confirmed that Nvidia’s board had launched an independent investigation. **November 8, 2026 — Trump publicly supports Huang** Trump called Huang a “fine person” and said he believed Huang should remain CEO of Nvidia, despite the allegations surrounding his relationship with Alice Huang. **November 9–10, 2026 — More details emerge about Alice Huang’s business network** Reports revealed that Alice Huang and her family had established a network of companies in China and Singapore to acquire Nvidia chips. One Chinese company, Shanghai Century Litong, reportedly facilitated the purchase of $7.5 billion worth of Nvidia chips, including restricted hardware. Bloomberg also reported that Nvidia’s general counsel had intervened in a U.S. investigation involving Megaspeed. **November 9, 2026 — Huang cancels speaking engagements** Huang withdrew from scheduled appearances at the ElevenLabs Summit and Anthropic CEO Forum as the controversy continued to escalate. **November 11, 2026 — Congress threatens to subpoena Huang** Republican lawmakers indicated that they could subpoena Huang if he continued to avoid their requests for a closed-door briefing about his relationship with Alice Huang and the national security implications. **November 11, 2026 — A second alleged affair is reported** The New York Times reported that Huang had also been involved in an extramarital relationship with Kimberly Powell, Nvidia’s vice president of healthcare, who reports directly to him. The report cited five people familiar with the company, and the individuals involved did not confirm the allegations. **November 11, 2026 — Nvidia shares fall as the board considers Huang’s future** The Wall Street Journal reported that Nvidia’s board was preparing to discuss Huang’s future as CEO. The company’s stock fell approximately 8% in after-hours trading and was down around 11% over five days. **November 12, 2026 — Huang resigns from Nvidia** Nvidia announced that its board had accepted Huang’s resignation as CEO and board member, ending his 33-year tenure at the company. CFO Colette Kress was appointed interim CEO, while longtime board member Harvey C. Jones became the company’s first-ever board chair. **November 12, 2026 — A third alleged affair is reported** Reuters reported that Huang had allegedly been involved in another extramarital relationship with an unnamed Nvidia employee. The report followed the revelations involving Alice Huang and Kimberly Powell. **November 12, 2026 — Congressional investigations expand** Democratic lawmakers Richard Neal and Ron Wyden announced an investigation into the Trump administration’s handling of Nvidia’s chip exports to China, particularly its response to intelligence concerning Alice Huang. Senator Tim Scott separately requested that Huang and Commerce Secretary Howard Lutnick testify before the Senate Banking Committee. Huang also agreed to testify before Congress before Thanksgiving. **November 13, 2026 — Questions emerge about Huang’s successor** Reports from Bloomberg described Huang as visibly devastated during his final hours as CEO. The Wall Street Journal also reported that Huang had previously expressed doubts about Colette Kress’s ability to lead the company, despite her appointment as interim CEO. ### Science & Research #### [🔗 Repeatable laser measurements reveal ‘forbidden’ vibrations in single molecules without destroying them](https://phys.org/news/2026-11-laser-reveal-forbidden-vibrations-molecules.html) `(Phys.org)` 🧠🧠🧠🧠🧠 Researchers at the University of Basel have developed a way to control and repeatedly measure extremely weak quantum transitions in individual nitrogen molecules. Their findings, published in *Nature Communications*, could help develop more accurate molecular clocks, quantum computers, and new ways to test fundamental physics. To understand what makes this interesting, we first need to understand quantum transitions. Atoms and molecules can only occupy certain energy levels. When they absorb or release energy, they can transition between these levels. In molecules, this can change how the atoms vibrate or rotate relative to each other. However, not all transitions are equally likely. Some are considered “forbidden” transitions, meaning they cannot occur through the usual electric-dipole interaction with light. They can still occur through weaker interactions, but some are roughly 10 billion times weaker than ordinary transitions. The researchers studied positively charged nitrogen molecules (N₂⁺). They had previously observed these forbidden vibrations in 2014, but their detection method destroyed the molecules after each measurement. This made studying such weak transitions extremely difficult. Their new approach uses a technique called *quantum-logic spectroscopy*. The researchers trap a nitrogen molecular ion alongside a calcium ion, which acts as a sensor. The nitrogen molecule doesn’t produce an easily detectable signal, but the calcium ion can be interrogated using lasers. By making the two ions share a common motion, the researchers can determine the nitrogen molecule’s quantum state from the calcium ion’s response. They then use a precisely controlled infrared laser to excite the nitrogen molecule into a higher vibrational state. A second pulse returns it to its original state, allowing the same molecule to be measured repeatedly without destroying it. By combining measurements from several hundred individually prepared molecules, the researchers determined nitrogen’s fundamental vibrational frequency approximately 10 times more precisely than previous measurements. The potential applications are particularly interesting because the weakness of these transitions can actually be an advantage. Certain forbidden transitions have extremely narrow, well-defined frequencies, making them useful for precise measurements. For example, they could be used in molecular clocks, where time is measured using an extremely stable quantum transition. They could also provide quantum states for storing information in quantum computers. Another possible application is testing whether fundamental physical constants, such as the proton-to-electron mass ratio, change over time. Detecting such changes could provide evidence of physics beyond the Standard Model. #### [🔗 Webb’s controversial ‘little red dots’ may finally have an explanation](https://phys.org/news/2026-11-webb-controversial-red-dots-explanation.html) `(Phys.org)` 🧠🧠🧠🧠 This is a pretty cool one! Since the James Webb Space Telescope (JWST) began observing the universe in 2022, astronomers have discovered an entirely new class of astronomical objects called Little Red Dots (LRDs). More than 300 have been identified, and scientists have been struggling to figure out exactly what they are. These objects are incredibly compact, generally only a few hundred light-years across, yet some are as bright as much larger galaxies. They also have unusual V-shaped light spectra and emit little to no detectable X-ray or radio radiation. This combination of characteristics doesn’t fit neatly into existing explanations of galaxies or black holes. Now, researchers led by Dr. Daniel Whalen at the University of Portsmouth have proposed an explanation: Little Red Dots may be young galaxies containing some of the universe’s earliest supermassive black holes. The researchers used cosmological simulations to model a process called *direct-collapse black hole formation*. In this scenario, enormous amounts of primordial gas accumulate in the early universe, allowing supermassive stars to form. These stars can reach hundreds of thousands of times the Sun’s mass before collapsing into massive black holes. The black holes then continue consuming surrounding material, potentially growing to around 10 million times the mass of our Sun. Here’s where the explanation gets interesting. Normally, black holes actively consuming matter produce enormous amounts of high-energy radiation, including X-rays. However, the team’s simulations suggest that these early black holes remain surrounded by the incredibly dense disks of gas that helped create them. That gas absorbs the X-rays while allowing some ultraviolet, visible, and infrared radiation to escape. At the same time, nearby clusters of young stars produce the bluer portion of the objects’ unusual spectra. The simulated galaxies also contain stellar populations totaling approximately 120 million solar masses, concentrated within a radius of about 500 light-years. By combining the radiation from the black hole and surrounding stars, the researchers reproduced the distinctive spectra of five observed Little Red Dots, including some particularly extreme examples. This could help answer a major question in astronomy: how did supermassive black holes become so enormous so early in the universe’s history? Starting with massive black holes formed through direct collapse would give them a significant head start compared with black holes formed from ordinary dying stars. I always find discoveries like this inspiring. NASA has some amazing people working for it, alongside scientists and engineers from around the world. Politicians have repeatedly proposed major cuts to NASA’s scientific programs, but the heart of NASA is still beating! Space exploration has always fascinated me, and finding an entirely new class of astronomical objects is a reminder of how much we still don’t understand about the universe. Discoveries like this make me wonder what else is out there, and then turn that curiosity inward to think about how we got here ourselves. #### [🔗 Scientists discover new way to control magnetic waves for future quantum technologies](https://phys.org/news/2026-11-scientists-magnetic-future-quantum-technologies.html) `(Phys.org)` 🧠🧠🧠🧠🧠 Researchers at Ames National Laboratory have demonstrated a way to manipulate the handedness of magnons, collective waves of magnetic motion that travel through materials. Magnons carry spin angular momentum without transporting electrical charge, making them interesting candidates for low-energy information processing and quantum technologies. The researchers studied two ferrimagnetic kagome metals, TbMn₆Sn₆ and ErMn₆Sn₆. These materials contain magnetic sublattices whose moments point in opposite directions but have unequal strengths. This allows them to support two types of magnons: right-handed and left-handed, distinguished by the direction in which their spins precess. The interesting discovery is that the two types of magnons can interact and mix, but whether this happens depends on the material’s magnetic symmetry. In TbMn₆Sn₆, the two types of magnons (left-handed and right-handed) behave independently. Their energy levels can cross without the two types interacting. In ErMn₆Sn₆, the magnetic arrangement of the atoms is different, allowing the two types of magnons to interact and mix with each other. Instead of simply crossing, their energy levels repel each other, creating a gap where they would otherwise meet. Something particularly interesting happens around this gap: the magnons can actually switch from left-handed to right-handed behavior, or vice versa, depending on their momentum. Essentially, the researchers found a way to make two normally separate types of magnetic waves interact and change their handedness. To check that the magnons were actually changing handedness, the researchers used a technique called polarized neutron scattering at Oak Ridge National Laboratory. They directed a beam of neutrons, whose spins were aligned in a particular direction, at the material. Neutrons interact with magnetic waves, and by measuring how they scatter and exchange energy with the material, scientists can determine what those magnetic waves are doing. Using this technique, the researchers were able to distinguish the different types of magnons and experimentally confirm that their handedness changes. Being able to manipulate magnon handedness could provide another way to encode and process information using magnetic waves. The researchers also suggest that magnetic fields or temperature changes could be used to trigger transitions between magnetic states, providing a potential route to switching between different magnon behaviors. #### [🔗 Sensors track stress in plants in real time—without hurting them in the process](https://phys.org/news/2026-11-sensors-track-stress-real.html) `(Phys.org)` 🧠🧠 Researchers at King Abdullah University of Science and Technology (KAUST) have developed a coin-sized sensor that can monitor stress in plants in real time without causing significant damage. The device uses tiny microneedles, just 0.4 millimeters wide, to extract small amounts of fluid from plant leaves. A laser-scribed graphene sensor then measures hydrogen peroxide (H₂O₂), a molecule plants naturally produce to signal cellular stress and activate defensive responses. To test the device, researchers attached it to Aloe vera plants in a greenhouse and exposed them to intense light. Hydrogen peroxide levels increased almost immediately, with a much larger spike at 28°C (82°F) than at 19°C (66°F). In both conditions, levels declined after approximately an hour. The idea is that farmers could eventually detect plant stress before visible symptoms appear, potentially helping them respond to heat, drought, and other environmental problems before crops are seriously damaged. #### [🔗 Large mutation analysis challenges claim that evolution shields important genes from changes](https://phys.org/news/2026-11-large-mutation-analysis-evolution-shields.html) `(Phys.org)` 🧠🧠🧠🧠 This is a pretty cool discovery, and potentially a major one. I’ve had multiple professors discuss the idea that essential genes are protected from mutations as though it were established fact. But a new study published in *PLOS Biology* has found little evidence supporting it. The hypothesis dates back to 1972 and proposes that natural selection may favor lower mutation rates in genes that are particularly important to an organism’s survival. The reasoning makes sense. If a mutation in an essential gene is likely to kill an organism, why wouldn’t evolution favor mechanisms that protect that gene from mutations? The problem is distinguishing genes that mutate less frequently from genes whose mutations are eliminated more frequently by natural selection. Imagine two genes that experience the same number of mutations. One is essential for survival, while the other isn’t. Mutations that disable the essential gene are likely to kill the organism, meaning those mutations disappear from the population. Meanwhile, mutations in the nonessential gene are more likely to survive and be observed by researchers. If we examine the surviving population, it might appear that the essential gene mutates less frequently, even though both genes originally experienced the same number of mutations. To investigate this, researchers from the University of Bath analyzed spontaneous mutations in *E. coli* bacteria and baker’s yeast. Their *E. coli* dataset contained 3,634 mutations, including 2,577 in protein-coding regions. The team used data from mutation accumulation experiments, which minimize the influence of natural selection by repeatedly propagating small populations. This allows mutations that might normally disappear to accumulate, giving researchers a better picture of where mutations actually occur. They then compared mutation frequencies with several measurements of gene importance, including whether removing a gene would kill the organism, how much protein it produces, and how strongly natural selection has acted on it over evolutionary time. They also accounted for other factors that influence mutation rates, including DNA sequence composition, gene length, and replication-related features. Using simulations, they tested whether the observed relationships were stronger than what would be expected from these underlying properties. The results were remarkably consistent. Essential genes did not have significantly lower mutation rates in either organism. Genes producing more abundant proteins weren’t protected either, and genes under stronger evolutionary constraints did not show the predicted reduction in mutations. There were some interesting exceptions to the broader pattern. In *E. coli*, genes with higher RNA expression showed slightly lower mutation frequencies, potentially because of processes associated with transcription and DNA repair. However, this relationship was extremely weak and was not consistently reproduced in yeast. In yeast, the researchers also found that genes replicated later tended to accumulate more mutations. The hypothesis hasn’t gone untested before. In fact, a 2022 *Nature* study reported evidence that essential genes in the plant *Arabidopsis thaliana* experienced substantially fewer mutations. That finding was subsequently challenged by other researchers, who identified potential problems with sequencing errors and statistical analysis. The original researchers defended their findings, and the debate continued. What makes this new study significant is that it addresses several of the methodological problems that complicated earlier research, using a larger collection of well-characterized mutations and more extensive statistical controls. There are still limitations. The study focuses primarily on two microorganisms, so it cannot establish that adaptive protection of important genes never occurs elsewhere in nature. It also doesn’t suggest that mutations occur at identical rates throughout the genome. We already know that DNA sequence, replication timing, and repair mechanisms can influence where mutations happen. For a hypothesis that has been discussed for more than half a century, and sometimes taught as established science, that’s a pretty significant challenge. #### [🔗 ‘Ahead of his time’: AI helps unlock discoveries hidden in notebooks of man who transformed maths](https://www.theguardian.com/science/2026/nov/13/kurt-godel-notebooks-ai-helps-unlock-hidden-discoveries-man-who-transformed-maths) `(The Guardian)` 🧠🧠🧠🧠 Claude has helped uncover a substantial amount of previously unknown mathematical work in the notebooks of Kurt Gödel, one of the most influential logicians of the 20th century. Gödel left behind 16 notebooks containing 1,009 pages of mathematical work, much of it written in Gabelsberger, an obscure German shorthand that was already becoming obsolete during his lifetime. Finnish logician Jan von Plato has been studying these notebooks since 2017, making progress at a rate of roughly one page per week. In September, he enlisted the help of Anthropic’s Claude. There is an important distinction here: Claude wasn’t deciphering the original handwriting. Von Plato and his team had already transcribed the notebooks. He provided Claude with those transcriptions, six books he had written about Gödel, and additional scientific literature. Within a couple of days, Claude produced a 650-page explanation of the entire collection, connecting mathematical ideas across an enormous body of work that would be extremely difficult for humans to keep track of. Claude identified 663 theorems for which Gödel had either written a complete proof, sketched one, or at least developed an idea. The researchers then used Coq, a proof assistant that mechanically checks mathematical proofs, to evaluate the results. About 400 were verified as stated, around 200 needed corrections or additional assumptions, and approximately 50 were false. Even Gödel himself occasionally went back through his notes and wrote *falsch!* (incorrect) or *schlecht!* (bad) next to his own work. One particularly interesting example concerns the axiom of choice, a fundamental principle in set theory. Gödel had claimed in 1966 that he had already worked out an independence proof in the 1940s. Claude helped reconstruct an argument that came remarkably close, but Gödel had used an incorrect definition, despite having used the correct one in an earlier notebook. Von Plato says the reconstructed proof passed Coq verification. There is an important historical distinction, though. The University of Helsinki’s [Gödel Enigma project](https://www.helsinki.fi/en/researchgroups/godel-enigma/research/project-description) had already identified this 1942 work as a research target. And mathematician Joel David Hamkins cautions that finding similarities to Paul Cohen’s later proof doesn’t establish that Gödel had actually completed the result. He also points out that historians will need to determine how much of the reconstruction belongs to Gödel and how much to Claude. Beyond these results, the notebooks contain more than a dozen unexplored themes in logic and the foundations of mathematics. Von Plato also plans to investigate Gödel’s largely unexplored physics and cosmology notebooks. ### Space, Aerospace & Satellite Technology #### [🔗 Want to fly to the moon? NASA is looking for new astronauts to help grow its Artemis program](https://phys.org/news/2026-11-fly-moon-nasa-astronauts-artemis.html) `(Phys.org)` 🧠 NASA is recruiting its 25th class of astronauts as it prepares to expand the Artemis program and establish a sustained human presence on the Moon. The agency currently has just 32 active astronauts, and it has considerably more work planned than its existing corps can support. The previous recruitment round attracted more than 8,000 applicants, of whom only 10 were selected in September 2025. Applications for the new class are open until December 3, 2026, with final selections expected in late 2027 and training beginning in early 2028. The timing is particularly interesting because NASA intends to return astronauts to the lunar surface in 2028, with plans to eventually establish a permanent lunar base. The upcoming missions will involve testing spacecraft, lunar landers, and the infrastructure needed to support longer stays on the Moon. Now, I absolutely love this aspiration. Putting humans on the Moon is one of the defining feats of humanity, and I’m glad we’re going back. When I was little, I wanted to be an astronaut. I’d still love to go to space, and I can only dream of what it would be like to actually set foot on the Moon! But the benefits of pursuing space exploration go well beyond the inspirational aspect. Some of the most important technological advancements we’ve made have come from trying to solve the extraordinary problems involved in getting humans and machines into space. Consider the CMOS image sensors used in practically every smartphone camera today. NASA’s Jet Propulsion Laboratory developed the first CMOS active-pixel image sensor while working on imaging technology for space missions. That technology eventually made its way into billions of cameras, including those in phones, medical devices, and automobiles. The Apollo program also helped advance digital flight-control systems. The computer-guided control technology developed for the lunar missions helped pave the way for digital fly-by-wire systems, which are now standard in modern aviation and used in many automobiles. Then there’s GPS. Its development grew out of satellite-tracking experiments during the early space age and subsequent military satellite-navigation programs. Today, satellite navigation is fundamental to everything from finding your way around town to coordinating global logistics. Even the weather forecasts we check every morning depend heavily on satellites observing our planet. These systems help us track hurricanes, monitor wildfires, measure environmental changes, and predict severe weather. These are just a handful of examples. The point is that when we pursue something as ambitious as putting humans on another celestial body, we end up pushing the boundaries of engineering, computing, materials science, physics, and countless other disciplines. The resulting discoveries and technologies don’t stay in space. They eventually find their way into our everyday lives. Becoming an astronaut is also a much longer commitment than most people probably imagine. Candidates undergo at least two years of intensive training, and it can take several more years before they actually get to fly. As astronaut Frank Rubio, who holds the U.S. record for consecutive days in space, explains, actually going to the Moon might account for only around 1% of an astronaut’s experience. The rest is contributing to the broader mission. I hope we continue to pursue ambitions like this. There is something genuinely special about humanity deciding that something seemingly impossible is worth doing and then dedicating enormous amounts of ingenuity and effort to making it happen. #### [🔗 Here’s What We Know About the X-70 Fractal, NASA’s Experimental New Plane](https://www.wired.com/story/heres-what-we-know-about-the-x-70-fractal-nasas-experimental-new-plane/) `(WIRED)` 🧠🧠 NASA has unveiled the X-70 Fractal, a new crewed experimental aircraft designed to serve as a flexible platform for testing advanced flight technologies. The aircraft has a planned takeoff weight of 10,000 pounds and can carry up to 2,000 pounds of experimental equipment, mounted either internally or externally. It is designed to cruise at Mach 0.75 (approximately 575 mph), reach altitudes of up to 45,000 feet, and travel more than 1,500 nautical miles (1,726 miles) on a flight. One of its most interesting features is its modular design. NASA wants to be able to rapidly swap experimental equipment and reconfigure its flight-control systems, allowing researchers to test different technologies without needing to develop an entirely new aircraft for each experiment. The primary objective is to investigate advanced flight controls and maneuverability that have never previously been demonstrated in crewed aircraft. This could provide a platform for testing new approaches to aircraft control and other emerging aviation technologies under real flight conditions. NASA expects the X-70 to make its first flight under the agency’s banner within the next two years. #### [🔗 Watched.](https://www.watched.earth/?q=37.788%2C-122.408&n=San+Francisco) `(watched.earth)` 🧠🧠🧠 Watched.earth is a pretty cool website that lets you enter any location on Earth and see which satellites passed over it during the previous seven days, along with any publicly available photographs they captured. The site tracks nearly 400 satellites with known imaging, weather, radar or crewed observation capabilities. It uses publicly available orbital data from CelesTrak and the SGP4 orbital propagation model to calculate their trajectories. These calculations run directly in your browser using Web Workers, allowing it to search an entire week’s worth of satellite passes in seconds. The globe itself is even more ambitious. It renders approximately 17,000 active orbital objects using your GPU, calculating their positions with a shader that solves Kepler’s equation every frame. However, the more precise pass calculations use the full SGP4 model on the smaller group of satellites with known observation capabilities. Once it identifies a pass, Watched searches public imagery archives, including Sentinel-2, Landsat and NASA weather satellite imagery. It also accounts for an interesting complication: satellite image timestamps don’t necessarily correspond to when the satellite was directly over a particular location. For example, Sentinel-2 timestamps can be several minutes earlier than the actual overpass because they record when the satellite began capturing an image strip. One particularly useful feature is how the site distinguishes between different kinds of observation: - **Watched:** An actual photograph of the location exists in a public archive. - **Likely watched:** A satellite’s fixed camera covered the location, but no photograph was published. - **Could have pointed here:** A steerable or commercial satellite was in range, but there is no confirmation that it actually captured an image. - **Passed overhead:** A satellite crossed the sky without a known camera observing the location. The project is open source under the MIT license, and the developer has published the entire implementation on GitHub. #### [🔗 The Hubble Space Telescope Is Sinking. NASA Is Looking for Ways to Keep It Afloat](https://www.wired.com/story/nasa-wants-to-keep-the-hubble-space-telescope-afloat/) `(WIRED)` 🧠🧠 NASA is looking into ways to extend the life of the Hubble Space Telescope, potentially keeping it operational for several more years. The telescope has been in orbit since 1990, but it is gradually losing altitude due to atmospheric drag. Even hundreds of kilometers above Earth, the atmosphere is not completely absent, and the resulting drag slowly pulls satellites into lower orbits. Hubble doesn’t have its own propulsion system to correct this, and NASA estimates it could reenter Earth’s atmosphere sometime around 2033–2034. NASA issued a Request for Information (RFI) on October 29, asking commercial space companies to propose ways to service the telescope. There are three possibilities being considered: boosting Hubble into a higher orbit, servicing or upgrading its existing hardware, or developing a controlled reentry or disposal plan for when its mission eventually ends. NASA estimates that an orbital boost would need to happen by 2032. This isn’t entirely new territory. In 2022, NASA and SpaceX entered an unfunded agreement to study whether a Dragon spacecraft could rendezvous with Hubble and boost it into a higher orbit. That study was intended to assess technical feasibility, not commit NASA to a rescue mission. What’s interesting here is that NASA is looking beyond simply keeping an aging satellite from falling back to Earth. Hubble is still an extremely productive scientific instrument, and extending its operational lifetime could mean several more years of observations and discoveries. With commercial companies increasingly developing spacecraft capable of servicing satellites in orbit, NASA wants to see whether those capabilities can give one of its most important observatories a little more time. #### [SpaceX pulls off dramatic midair catch of Starship rocket after fiery plunge from orbit](https://www.cbsnews.com/news/spacex-catch-starship-rocket/) `(CBS News)` 🧠 SpaceX has just pulled off one of the most impressive engineering feats in its history. For the first time, the company has successfully caught a Starship spacecraft returning from orbit using the enormous mechanical arms attached to its launch tower. On November 13, SpaceX launched Starship on its 15th test flight from Starbase, Texas. The mission was intended to recover both stages of the rocket: the Super Heavy booster and the Starship upper stage. While the booster failed to make it back to the launch tower and was lost in the Gulf of Mexico, the upper stage had a much more successful journey. After reaching low Earth orbit, Starship deployed 26 of SpaceX’s next-generation Starlink V3 satellites and spent approximately 22.5 hours completing 14 orbits around Earth. It then began its return, enduring the extreme temperatures of atmospheric reentry before restarting three of its six Raptor engines to flip itself upright and slow to a near-hover. Two enormous mechanical arms, nicknamed the “chopsticks,” then closed around the spacecraft and caught it. This is a pretty significant milestone. SpaceX has successfully caught Super Heavy boosters on three previous flights, but this is the first time it has recovered the Starship upper stage from orbit. The goal is to make the entire launch system reusable. Instead of allowing the spacecraft to land in the ocean and then recovering it, SpaceX wants to catch both stages directly at the launch tower. This would allow the company to inspect, refurbish, and potentially relaunch the vehicles much faster, dramatically increasing launch frequency while reducing costs. There is still plenty of work to do, though. The failed booster recovery is one reminder that the system isn’t ready for routine operations, and SpaceX will need to demonstrate consistent reliability across many more flights. Another major hurdle is orbital refueling. NASA’s planned Starship lunar lander will require multiple tanker Starships to transfer cryogenic propellant to a depot in Earth orbit before the lander can make its journey to the Moon. ### Consumer Technology, Platforms & Digital Projects #### [🔗 Meta expands Instagram’s AI age-guessing tech globally](https://techcrunch.com/2026/11/09/meta-expands-instagrams-ai-age-guessing-tech-globally/) `(TechCrunch)` 🧠 Meta is expanding its AI-powered age-estimation technology worldwide on Instagram, allowing it to identify teenagers even when they have registered with an adult birthday. The system analyzes entire user profiles for contextual clues that might indicate someone’s age. This includes birthday celebrations, mentions of school grades, and other information found in posts, comments, biographies and captions. It also analyzes photos and videos for visual indicators, including a person’s height and bone structure. Meta says this visual analysis is not facial recognition. Instead, it looks for general visual patterns associated with age without attempting to identify specific individuals. When the system determines that an account likely belongs to someone between 13 and 17, it can automatically place them into Instagram’s Teen Account protections. These accounts are private by default, restrict direct messages to people the teenager follows or is already connected with, and hide potentially harmful comments. The company is also continuing to roll out the technology on Facebook, with a similar global expansion planned for Threads later this year. #### [🔗 Apple has new Mac app coming to make running local AI models easy](https://9to5mac.com/2026/11/10/apple-has-new-mac-app-coming-to-make-running-local-ai-models-easy/) `(9TO5Mac)` 🧠🧠 Apple is reportedly developing a new Mac app called **Harbor**, designed to make running local AI models considerably easier. The app would let users browse, download, and run open-weight models directly on Apple silicon, turning their Mac into a private AI inference server without relying on cloud services. The more interesting feature is its ability to connect multiple Macs over Thunderbolt 5 and pool their unified memory. This would allow users to run models that are too large to fit on a single machine or handle more simultaneous inference requests. Harbor would also automatically detect connected Macs and handle model management, removing much of the manual configuration normally involved in setting up a local AI server. The app is currently in private preview under Apple’s Claris brand. Its features have been revealed through a leaker, so we’ll have to wait for an official announcement and independent testing to see how well it actually works. #### [🔗 Fairphone Lands Its First US Carrier Partner in Boost Mobile](https://www.wired.com/story/exclusive-fairphone-lands-its-first-us-carrier-partner-in-boost-mobile/) `(WIRED)` 🧠 One of the coolest phones I’ve ever heard of is expanding its presence in the United States! Fairphone, the Dutch smartphone manufacturer focused on repairability, sustainability, and ethical manufacturing, has announced its first American carrier partnership with Boost Mobile. The partnership is still in its early stages. Boost won’t sell the phones directly yet, but customers purchasing the $650 Fairphone Gen. 6+ will be able to select a Boost Mobile plan during checkout. This is a pretty important step considering that nearly 70% of American consumers purchase their phones through carriers. What makes Fairphone so interesting is how differently it approaches the smartphone market. The Gen. 6+ has 12 user-replaceable components, including its battery, display, cameras, and USB-C port. A replacement battery costs just $40, while a new display costs $90. For comparison, replacing an iPhone 17’s battery through Apple costs $99, while a display repair costs $329. Fairphone also offers up to a five-year warranty and guarantees software support through 2033. The company works to improve conditions throughout its supply chain, including living-wage bonuses for workers and greater use of responsibly sourced and recycled materials. And then there’s the software. Fairphone gives customers the option of running Murena’s /e/OS, a de-Googled version of Android designed around privacy and user control. The standard US model ships with regular Android, but the availability of an officially supported alternative is itself an important distinction. There are some compromises. The phone has an IP55 water- and dust-resistance rating, meaning it can handle dust exposure and water jets but isn’t rated for submersion like many modern flagship phones. This is partly a consequence of its repairable design, although I think this is something Fairphone will eventually improve. To me, Fairphone is practically the antithesis of Big Tech. It’s one of the clearest examples of a smartphone company building its products around the interests of its customers and the people making them, rather than simply maximizing sales. And I think it demonstrates something important about innovation. As far as Big Tech is concerned, there really isn’t that much left to innovate in the smartphone market. Yes, AI has given the industry something new to advertise, but how many of the AI features Google, Samsung, and now Apple promote do you actually use? Fairphone completely challenges that notion. Instead of focusing on another AI feature or incremental camera improvement, it’s approaching smartphones from an entirely different perspective, one centered on privacy, repairability, and choice. It turns out there is still plenty of room for innovation when you stop looking at smartphones purely as devices to replace every few years and start looking at them as products that should genuinely serve their owners. I’m glad Fairphone is continuing to expand in the United States, and I hope the company continues to grow. #### [🔗 Microsoft 365 Crawls Back to Life After a 24-Hour Outage](https://www.wired.com/story/microsoft-365-crawls-back-to-life-after-a-24-hour-outage/) `(WIRED)` 🧠 Microsoft 365 is recovering after a massive outage on November 12 that disrupted Outlook, Teams, Copilot, and other services across more than 70 countries. The outage began in North America before spreading internationally, affecting businesses, government agencies, airlines, hospitals, and emergency services. Some organizations reported significant interruptions to their operations. Microsoft says the problem originated during an earlier network capacity upgrade and surfaced following a subsequent network change. The company has ruled out both a cyberattack and AI or AI agents as causes. It is still investigating the incident and has promised a full post-incident report. Although most services have been restored, Microsoft warned that users could experience delayed emails and a flood of notifications as systems recover. The company extended its email queue window to 48 hours to help ensure messages sent during the outage are eventually delivered. It’s another reminder of how much of the world’s essential infrastructure depends on a handful of cloud providers. When one of them goes down, the consequences extend far beyond people being unable to check their email. #### [🔗 Isometric SF](https://isometricsf.com/v/9/46.9/13.1) `(Justin Chavez)` 🧠🧠🧠 This is a pretty cool project! Developer Justin Chavez has turned more than 1,300 square miles of the San Francisco Bay Area into an enormous, interactive map that looks like something straight out of SimCity 2000. You can zoom around, search for locations, and explore everything from the Golden Gate Bridge to San Jose. What’s particularly interesting is how he made it. Rather than using an expensive image-generation model to paint every single tile, Chavez built a pipeline that progressively transferred the visual style to smaller models. First, he used Google’s Photorealistic 3D Tiles to capture the Bay Area from a fixed isometric angle. He then used Nano Banana 2 to establish the pixel-art aesthetic. Next, he fine-tuned a 20-billion-parameter Qwen image-editing model using a LoRA (a lightweight add-on that teaches an existing model a particular style). That model generated training examples for an even smaller 4-billion-parameter FLUX.2 Klein model, which ultimately painted the map. The scale is pretty amazing. Klein generated approximately 177,000 image tiles to cover the entire region. Chavez reports that the project took around 172 H100 GPU-hours and cost approximately $600. He estimates that generating everything with Nano Banana 2 instead would have cost around $7,000. There was also an interesting engineering challenge in assembling the map. Since every tile is generated independently, neighboring images don’t necessarily line up. To address this, Chavez blended overlapping sections between tiles and applied color correction to make the entire map look consistent. #### [🔗 Tesla opens Cybercab pre-orders ‘for personal use,’ with geo restrictions and a $299 monthly fee](https://techcrunch.com/2026/11/14/tesla-opens-cybercab-pre-orders-for-personal-use-with-geo-restrictions-and-a-299-monthly-fee/) `(TechCrunch)` 🧠 Tesla is opening pre-orders for its Cybercab at $29,990, but buying the vehicle is only part of the deal. Owners also have to pay a mandatory $299 monthly Autonomy Subscription to operate it though Musk notes on X that Tesla will eventually lower that subscription too $115 per month. I actually think this is a pretty smart economic model. Tesla gets the upfront revenue from selling the vehicle and a recurring revenue stream from every active subscriber. In exchange, customers get autonomous transportation, connectivity, roadside assistance, and certain maintenance services. The subscription even covers things like tire replacements, windshield replacements, wipers, and routine inspections. There is also an interesting opportunity here for companies that want to operate autonomous fleets. Uber, for example, abandoned its own autonomous vehicle development program in 2020 and sold its self-driving division to Aurora. Since then, it has shifted toward partnering with autonomous vehicle developers and providing the infrastructure to operate their fleets. Tesla is already accepting inquiries from companies interested in purchasing Cybercabs for commercial use, so I could see a company like Uber becoming a customer. There are some pretty significant limitations, though. Cybercabs can only operate inside Tesla-approved geofenced areas, and Tesla can change those boundaries whenever it wants. Currently, Cybercab service is restricted to limited areas of Austin, Texas. And because the vehicle doesn’t have a steering wheel, you can’t simply take over and drive it somewhere outside its operating area. Tesla does provide limited touchscreen controls for moving the vehicle short distances, but those aren’t intended for public roads. Fortunately for Tesla, these geofenced areas are concentrated in populated urban environments where people have plenty of amenities relatively close by. That means the actual utility customers get from the vehicle could still be quite high, even with those restrictions. I’m not a fan of Musk, but I think this is an innovative strategy, and I’m genuinely interested to see how it develops. Maybe this is the future of vehicle ownership. If autonomous vehicles eventually become shared fleets rather than something everyone needs to own individually, I think it could certainly cut down on the number of cars on the road. Anyone who lives in a densely populated American city would probably appreciate that. ### And if You Missed it… #### [🔗 Inside GitHub’s new Git storage architecture](https://github.blog/engineering/architecture-optimization/inside-githubs-new-git-storage-architecture/) `(GitHub Blog)` 🧠🧠🧠🧠 As AI coding agents become more capable of working independently and in parallel, they’re creating a new problem for the infrastructure that supports software development: Git repositories need to handle vastly more concurrent writes than they were originally designed for. GitHub is rebuilding its underlying Git storage architecture to address this, with a new system called Project Rubicon. The scale they’re targeting is pretty extraordinary. GitHub reports that one repository saw its daily requests increase from an average of 44 million in March 2026 to 181 million in May. In internal testing, the new architecture supported 5,000 concurrent agents working on a single repository and 7,200 pushes per second, while continuing to serve read requests with consistently low latency. The interesting part is how GitHub is approaching the problem. It comes down to a fundamental property of Git: most of its data doesn’t actually need to be synchronized between writers. Git stores commits, files, and directory structures as immutable objects identified by hashes of their contents. Once created, these objects never change. Branches, on the other hand, are simply references pointing to particular commits, and those references can change as developers push new work. This distinction allows GitHub to separate the expensive process of storing repository data from the relatively small amount of information that actually requires coordination. Project Rubicon divides the infrastructure into three main components: - **Azure Blob Storage** holds the durable repository data, including Git objects and their associated indexes. Because objects are immutable, multiple agents can upload their changes independently without having to coordinate with one another. - **Ref-CAS (Reference Compare-and-Swap)** is a separate replicated log service responsible for maintaining branch references and the repository’s pack manifest. It coordinates updates to individual references rather than serializing every write across the entire repository. Changes to different branches can therefore happen concurrently. - **Virtual replicas** are disposable compute nodes that handle incoming requests and maintain local caches of repository data. GitHub can add more replicas as demand increases without creating additional permanent copies of the repository. There’s also a separate fleet of background workers responsible for compaction and garbage collection. Every push creates another data pack, and allowing those packs to accumulate indefinitely would eventually make reads slower. These workers merge packs and remove objects that are no longer needed without competing for the resources serving live requests. The result is an architecture where storage, reads, writes, and repository maintenance can all scale independently. Importantly, developers don’t need to change how they use Git. Existing workflows, branch protections, and other safeguards remain in place. GitHub is already planning the next improvements, including lightweight repository access without requiring a full clone, atomic multi-file commits, real-time event streams for agents, and globally distributed replicas. The new architecture is currently available in a limited private preview, with a broader rollout planned for early 2027. #### [🔗 Asahi Linux — Progress Report: Linux 7.3](https://asahilinux.org/2026/11/progress-report-7-3/) `(asahilinux.org)` 🧠🧠🧠🧠 I think Asahi Linux is one of the coolest Linux projects running today. Honestly, when they first started, I doubted that what they were attempting was even possible. Apple has such tight control over its hardware, and getting an entirely different operating system to work properly on it seemed like an enormous undertaking. But the team has completely and utterly proved me wrong. They have some incredibly smart people working on this project, and you can genuinely feel their passion when reading their progress reports. Their latest update is another demonstration of just how far they’ve come. M3 support is now generally available, and the team has made substantial progress on M4 and M5 hardware. They’ve also improved sleep functionality, made the area around MacBook notches usable, and submitted thousands of lines of audio and video driver code for integration into upstream Linux. Thunderbolt support has also arrived in their downstream kernel, although DisplayPort and PCIe tunnelling are not supported yet. One of the most interesting developments is their progress getting Windows games to run on Apple Silicon without relying on a virtual machine. This is actually the result of some work Google has been doing on Android. Running Windows software on ARM Linux requires several layers of compatibility software. FEX translates x86 and x86-64 instructions into ARM64, while WINE provides the Windows APIs that applications expect. However, many older Windows applications and games were designed around a 4 KB memory page size, whereas modern systems increasingly use 16 KB pages. Previously, Asahi had to use a lightweight virtual machine called muvm to accommodate these older applications, introducing performance penalties and compatibility problems. Google developed a solution called P3S (Per Process Page Size) for Android. It allows individual processes to use 4 KB memory pages even when the underlying system uses 16 KB pages. Asahi developers adapted Google’s work to their own kernel, and the results are pretty amazing. They’ve demonstrated games including Star Wars Battlefront, Wolfenstein: The New Order, GTA IV, and Fallout: New Vegas running without the virtual machine. Wolfenstein even reached over 144 FPS on an M2 Max Mac Studio. Performance isn’t consistent across games, but this is still an impressive achievement. I’m not much of a gamer anymore, but getting Windows games running through multiple compatibility layers on a Mac running Linux is just plain amazing. There is an incredible amount of engineering involved in making all of those pieces work together. The report also covers improvements to microphone processing, with hand-optimized SIMD code making their custom MacBook microphone processing almost twice as fast while using less CPU. The team has also managed to initialize the M3 GPU far enough to render a triangle, an important step in reverse engineering Apple’s graphics hardware. Finally, there is a significant change to the project’s licensing. Asahi is moving its repositories from the permissive MIT license to GPLv2 or later. The team says it has encountered attempts to take its work, including unfinished code from developers’ personal repositories, and reuse it in proprietary projects without attribution. Some of that code was even rewritten with LLM assistance to obscure where it originated. I think the licensing situation is unfortunate, but I completely agree with their decision. These developers are working incredibly hard and doing the world some good by making Apple’s hardware more accessible to the free-software ecosystem. They shouldn’t have to watch their work get ripped off and taken apart piece by piece by companies looking to commercialize it without contributing anything back. Open-source licensing should encourage collaboration, but that doesn’t mean developers should have to surrender control over how their work is reused. #### [🔗 Apple’s iPhone keyboard has a secret blocklist. ‘Dictator Trump’ and ‘Sleepy Joe’ are on it.](https://www.ibtimes.com/apple-iphone-keyboard-restricted-phrases-research-3809513) `(International Business Times)` 🧠🧠🧠 Apple’s iPhone keyboard has an undisclosed list of thousands of words and phrases that it generally refuses to suggest through predictive text or produce through swipe-typing. Independent researcher Derik Kauffman uncovered the list by reverse-engineering Apple’s keyboard data files and published his findings on a website called [Swipeblocked](https://swipeblocked.com/). The recovered English-language list contains 4,513 phrases, with similar restrictions appearing across 27 other languages and scripts. Most involve profanity, drugs, sex, violence, or offensive language, but some are politically sensitive. Examples include “Dictator Trump,” “Sleepy Joe,” “United Ireland,” and “Tibetan independence.” The restrictions don’t actually prevent users from typing these phrases manually. Instead, they interfere with the keyboard’s automated features. For example, attempts to swipe-type “Sleepy Joe” can produce “Sleepy hope,” while “United Ireland” can become “United Iceland.” The discovery is particularly interesting from a technical perspective. Apple doesn’t store the restricted phrases as readable text. Instead, it uses a compact data structure containing 32-bit hash fingerprints. Kauffman reconstructed the phrases by generating candidate word combinations and checking them against those fingerprints. The broader collection is estimated to contain around 100,000 entries across languages, although the exact number depends on how variations are counted. Apple says its keyboard is designed to avoid suggesting potentially inappropriate or offensive language, but it hasn’t explained why particular political expressions are included or how the list is selected. The presence of a phrase also doesn’t establish why Apple restricted it, and not every recovered entry has been independently tested. #### [🔗 Anatomy of Apple’s Personal Knowledge Graph](https://8rank.ai/game/apple-intelligence-personal-knowledge-graph) `(THE EIGHTH RANK)` 🧠🧠🧠🧠 Apple has been building a surprisingly detailed picture of Mac users and the people in their lives, and most people probably have no idea how extensive it is. Researchers at The Eighth Rank have investigated a system built into macOS that constructs a personal knowledge graph by collecting and connecting information from across the operating system. The researchers found 51 SQLite databases belonging to Apple’s IntelligencePlatform, with the central database, `graph.db`, containing a knowledge graph built around 181 entity classes and 278 relationship predicates. Essentially, it is a database that doesn’t just store information about people, places and events, but also connects them and records how they relate to one another. The information feeding into this system comes from a surprising number of places. Contacts, Calendar, Photos, Family Sharing, Health’s Medical ID, installed applications, and information extracted through Siri Suggestions from Mail and Messages all contribute. There are also separate databases containing communication metadata, location and movement information, and device signals such as screen locking, Wi-Fi, Bluetooth and charging activity. What’s particularly interesting is how much of this information is inferred rather than explicitly provided. The researchers identified two XGBoost classifiers used to estimate family relationships. These models use 144 features, of which only 20 concern identity, 40 concern communication patterns and 84 are derived from photographs. The system also uses twelve Photos analyzers to score relationships such as partner, parent, child, coworker and friend based on who appears together in photographs, including during trips and holidays. When Apple Intelligence is active, an on-device LLM can also read messages to associate names and relationships with faces. This means the system isn’t simply organizing information that someone has entered into their Mac. It is attempting to understand their social environment. By combining information from different sources, it can potentially reconstruct family relationships, identify close contacts, establish communication patterns and reveal daily routines. Even the communication metadata, which records who interacted with whom, when, through which application and in which direction, can reveal quite a lot without storing the actual message contents in that particular database. There is an interesting history here, too. The two daemons responsible for constructing and querying this graph, `knowledgeconstructiond` and `intelligenceplatformd`, are documented in macOS 13.6.5, dating back to 2022. Apple’s broader work on personal contextual knowledge goes back even further. A [2018 forensic investigation](https://www.mac4n6.com/blog/2018/8/5/knowledge-is-power-using-the-knowledgecdb-database-on-macos-and-ios-to-determine-precise-user-and-application-usage) of the older `knowledgeC.db` database found detailed records of application usage and user activity on macOS 10.13. This isn’t necessarily the same architecture, but it shows that Apple has been developing systems to understand user behaviour for quite some time. The privacy implications are where things get particularly interesting. I don’t know whether Apple uploads this information to its servers, or whether portions of the graph are passed along when Siri makes requests to more powerful models. The researchers found no evidence that the graph itself syncs or uploads, but explicitly acknowledge that this isn’t proof that it never happens. We also don’t know all the services that use this or exactly how all the different services that can access the graph use it. What we do know is that this is an extremely detailed knowledge graph, not just about the device’s owner but also about the people around them. And there is no general user-facing way to opt out of the graph as a whole or inspect and correct its inferences through Apple’s own interface. Deleting information from an individual source can eventually cause related graph entries to be withdrawn, but that isn’t the same as giving users control over the system. I don’t know whether Apple has kept this relatively undocumented because it considers the implementation intellectual property or whether there is another reason. I don’t think we have enough information to establish its intentions. But the lack of an opt-out is what really concerns me. If this information were guaranteed to stay on the device, that would be one thing. But even locally stored information is not necessarily safe. We see people’s devices being compromised all the time, and this is an extremely valuable target for anyone who manages to gain sufficient access. Imagine the amount of information a government or hacking organization could exploit if it gained access to someone’s graph. Rather than having to piece together someone’s life from hundreds of individual records, it could potentially have an already connected picture of their family, friends, coworkers, communication habits and daily routines. And that information extends to people who never agreed to have their lives profiled in the first place. The researchers have released [iKnow](https://github.com/8rank/iKnow), an open-source tool that allows Mac owners to inspect their own knowledge graph locally. It’s a useful step toward transparency, but I think Apple should be providing this kind of visibility and control itself. Users should at least have a meaningful choice about whether a system this comprehensive is allowed to construct a profile of their lives and the people around them.