cd /news/artificial-intelligence/tai-229-fable-5-2-and-the-case-for-a… · home › topics › artificial-intelligence › article
[ARTICLE · art-171139] src=pub.towardsai.net ↗ pub= topic=artificial-intelligence verified=true sentiment=· neutral

TAI #229: Fable 5.2 and the Case for a 2027 AI Takeoff

METR reported on November 5 that Anthropic's internal capability index scores have been rising about twice as fast since October 2025 as before, estimating AI accounted for roughly 80% of the extra pace, and Anthropic itself said in a November 10 securities filing that it had met the threshold in its Responsible Scaling Policy for a faster pace of AI research. Anthropic staff estimated Claude was leading 26% of model R&D work in August, up from less than 1% in February, while OpenAI reported its researchers logged 3.1 eight-hour days of agent runtime per human workday by mid-August. The disclosures accompanied Anthropic's November 9 release of Claude Fable 5.2, which Scale AI found produced acceptable work on roughly twice as many freelance projects as either prior model, and Google's releases of Gemini 4 Carbon on November 10 and Gemini 4 Hydrogen two days later.

read27 min views3 publishedNov 17, 2026

In August, AI researcher Romeo Dean put his median date for AI capable of doing virtually all white-collar work in early 2035. Last week, he moved it to September 2028. He is one of several leading AI researchers who have brought their timelines forward, and I am hearing a similar shift in my conversations with people at large companies about how much of the economy AI will change and how soon. I think there is a real chance that, one way or another, the world will look quite different two years from now.

My own view started to change with GPT-6 Astra and Claude Opus 5.5. Claude Fable 5.2, which Anthropic released on November 9, moved it further. Scale AI found that Fable produced acceptable work on roughly twice as many freelance projects as either model. We are also getting a clearer picture of what is happening inside the labs. An outside evaluator has now measured how much AI is helping Anthropic build its successors, while Anthropic and OpenAI have disclosed unusually specific numbers about their own research.

Google is back in the game too, and is releasing models quickly. It opened Gemini 4 Carbon to paying developers on November 10 and released the faster Gemini 4 Hydrogen two days later. On Cursor’s test of coding jobs spanning several files, Carbon at high effort scored 60.4% for $9.22 a task. I’d now put Gemini back on the shortlist to benchmark your use cases alongside Claude and GPT.

These faster models are also changing how quickly the labs can develop the next ones. METR (Model Evaluation and Threat Research) spent eleven weeks with access to Anthropic’s internal evaluations, researchers, and compute records. It found that scores on the lab’s internal capability index had been rising about twice as fast since October 2025 as before. Its November 5 report estimated that progress on well-specified tasks had roughly doubled relative to the March 2024 to October 2025 trend. Its best guess was that AI accounted for around 80% of the extra pace.

Anthropic has now made a similar assessment itself. In a November 10 securities filing, it said it had met the threshold in its Responsible Scaling Policy for a faster pace of AI research. This was a forward-looking judgment based on its expectations for future models and its overall rate of progress, and Anthropic said it was still assessing the underlying data.

One Anthropic researcher gave a simple example of how that acceleration happens. They used to skip ideas that would take a day or two to implement. With Claude, they can now run those experiments. If one works, the next Claude benefits from a discovery the team would otherwise have missed. That’s how recursive self-improvement can get going while people are still choosing the research questions.

The same thing is visible at OpenAI. It reported in September that, by mid-August, its researchers were logging 3.1 eight-hour days of agent runtime for every human workday, with agents running in parallel. Anthropic also measures this, but differently. Its staff estimated that Claude was leading 26% of model R&D work in August, up from less than 1% in February. At both labs, a human still sets the direction and decides which result is worth a much larger training run.

Anthropic gave another example during its investor roadshow. This year, it tried Claude Mythos Preview on AMD’s MI355X chips. In three days, with very little human intervention, Mythos got one Anthropic model processing incoming prompts at useful speeds. Anthropic said Claude “ported or wrote almost 100% of the kernels.” These are the low-level programs that make the chip do the work. Anthropic compared this with the broader job of bringing up a new accelerator, which historically took six to twelve months and dozens of engineers.

When AI 2027 went viral in April 2025, plenty of people thought its timelines were wildly bullish. Even some of its authors were skeptical. They chose 2027 as the year to depict, but their median estimates for artificial general intelligence ranged from 2028 to 2032. Co-author Scott Alexander put the pace described in the story at roughly a one-in-five chance. All six forecasters in the November update, which Alexander did not join, now put their median for a superhuman AI researcher in 2027, assuming labs continue without a substantial deliberate slowdown. They define that threshold as the point when a lab would choose a team made entirely of AIs over the world’s best unaided human researchers, with the AIs also choosing research directions and interpreting the results.

For AI capable of essentially doing all white-collar work, the six individual medians now range from September 2027 for Daniel Kokotajlo to September 2028 for Dean and Eli Lifland. As recently as August, Lifland’s median was mid-2032. Their individual probabilities of reaching that level before 2028 range from 35% to 65%. The forecasters say METR’s study was the biggest reason they changed their minds. They read it as implying that AI is already speeding up research by about 1.7 times. Even Kokotajlo’s August model had assigned less than a 4% chance to reaching that degree of acceleration by November. Earlier this year, Dwarkesh Patel thought superintelligence was probably most of a decade away. One of his main objections was how much people learn on the job that never appears in a training set.

Mathematics changed his view. Models can solve remarkably difficult problems wherever somebody can check the answer, and many machine-learning ideas can also be checked through experiments. That means an AI researcher could discover methods for learning on the job before it could perform that learning itself. Assuming no deliberate slowdown, Patel now puts his median for AI matching the best human AI researchers in 2028, with a 25% chance that it takes until after 2032. He expects AI to match the best experts across fields by 2030.

Suddenly, it also feels easier to say in public that you expect rapid progress. But people can move their timelines forward by years and still disagree about three separate questions.

Could a system that matches our best researchers go on to become vastly more intelligent than any human? How dangerous would it be to let that system work on its successor? And how quickly would companies reorganize around a model that could already do the work on paper? On the second question, the AI 2027 forecasters themselves want labs to slow down.

I see just as much room for disagreement on jobs. A lender could buy a model that reviews a loan file in ten minutes and still take two years to change who has the authority to approve it. Once that changes, one lender might need a much smaller credit team, while another might finally be able to serve small borrowers it used to turn away. Heavy job losses and substantial economic growth are both plausible, and they can happen in the same industry.

Scale AI’s Remote Labor Index measures how much paid work a model can already do on paper. Scale gave Fable 5.2, at high effort, 240 projects originally completed by paid freelancers. Evaluators judged its deliverable at least as good as the professional’s original on 99 of them. That is 41.3%, roughly double Opus 5.5 at 21.3% and Astra at 20.8%.

Fable is the purple dot just above 40% in the chart. Philip at AI Explained asked Gemini Carbon to extend a chart by Epoch AI researcher Greg Burnham, whose horizontal axis measures general model performance on Epoch’s Capabilities Index. If models keep improving on that index at the assumed pace, the steeper curve reaches 98% of projects between May and October 2027. The shallower curve gets there between December 2027 and October 2028.

Anthropic also cut Fable’s standard price by a fifth, to $8 per million input tokens and $40 per million output tokens.

Before comparing anyone’s experience with it, I would check the effort setting. Claude Code defaults to medium, and the API to high, and more effort does not always improve the result. On Artificial Analysis’s revised coding index, Fable scored 58 at both high and extra-high effort, level with Opus 5.5 at max. At max effort, Fable scored only 56 because more attempts ran into the time limit, while costing $44.22 per task. Users do seem to prefer it to Opus. Across more than 10,000 Fable sessions on Agent Arena, its maximum-effort setting had a sizeable lead over Opus 5.5 Max on praise versus complaints, even though their task-completion signals were almost identical.

For our own writing, I would still use Opus. On Towards AI’s ToneBench, which scores scripts in my co-founder Louis-François Bouchard’s voice, Fable’s best setting cost $2.22 per script. Opus 5.5 ranks just ahead at about $1. Their score ranges overlap, and Opus is one of our four judges, but I can generate two Opus drafts and choose between them for less than the price of one Fable draft. On November 12, SE3 Labs tried Fable on a real two-arm robot against the older robotics model π0.5, which had been fine-tuned on 50 demonstrations of each task at that station. Fable had none, and across ten tasks it still averaged 38.0 points out of 100 for steps completed, against 20.9 for π0.5. Over ten attempts apiece, it led 64 to 7 on remembering what was hidden under cups but lost 39 to 51 on packing a box, with individual runs scoring anywhere from 5 to 100. If I were buying the robot to pack boxes, I would buy based on the second result.

In other news, on November 12, NVIDIA announced that Jensen Huang had resigned as chief executive after 33 years at the company he co-founded. Its finance chief, Colette Kress, took over on an interim basis. We do not know the full picture. Jensen acknowledged being unfaithful to his wife and the hurt he caused her and their children. By his own account, he made serious mistakes. I feel very sad for him and especially for his family. His departure is a major loss for AI.

I think there is a strong case that the AI revolution would have arrived much later without Jensen. You could argue that the GPU itself was partly a side effect of gaming. Doom helped turn the PC into a games machine. The hardware-accelerated version of Quake gave players a reason to want a graphics card, and that demand helped fund further research. Under Jensen, NVIDIA put the geometry, lighting, and rendering for a 3D scene onto one chip with the GeForce 256. His less obvious bet was that scientists could also use future generations of gaming GPUs if NVIDIA made them easy to program.

That meant paying for CUDA, which let scientists write programs for GeForce chips in a language they already knew, years before most customers wanted it. Adding CUDA to GeForce made each chip more expensive, squeezed margins, and made it hard to sell to investors, but Jensen kept funding it. Six years after NVIDIA introduced CUDA, the researchers behind AlexNet trained their neural network in five or six days on two GTX 580 gaming cards. The hardware they needed for one of the breakthroughs in deep learning was already on sale to people who wanted to play games.

Jensen also pushed NVIDIA to deliver a new generation of AI supercomputers every year. That required TSMC’s fabrication and packaging, memory suppliers, and the companies assembling entire racks to move on the same timetable. Because that schedule became predictable, a lab could plan the training run for its next model around chips that had not yet been built. I think Jensen helped set the pace of model generations as well as chip generations. Plenty of people understood that faster computers would be valuable. Very few could persuade that many other companies to build them in time.

AI research can progress much faster than most other work because the feedback loop is unusually short. Researchers can test many machine-learning ideas through experiments, start using a new model as soon as it is available, and feed useful results into the next model. People can still choose the research questions while AI does enough of the implementation and testing to increase how many ideas the team can try.

That helps explain why METR could attribute most of Anthropic’s faster research progress to AI without claiming that Claude is already running the research program itself. Very little of the wider economy works like that. Customers, managers, regulators, and internal approval processes all affect how quickly a capability can actually be used. Changing those processes can take much longer than improving the model.

So I expect the gap between what AI can do and what organizations let it do to keep widening over the next two years, even as both speed up. If the research loop delivers a superhuman AI researcher in 2027, companies will be absorbing several model generations at once with processes designed before any of them. That gap is where the choice between job losses and growth gets made. A business that uses the new capacity to launch products, enter markets, or serve customers it used to turn away has a route to growth; one that waits until the savings are obvious is more likely to end up cutting staff to show a return. The models are unlikely to stay the bottleneck for long. How quickly an organization can change who approves the work will probably be the key.

— Louie Peters — Towards AI Co-founder and CEO

  1. Anthropic Launches Claude Fable 5.2 and Mythos 5.2 Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 for the most demanding technical workloads, such as software migrations and robotics evaluations. Both use the same underlying model. Fable applies cybersecurity and biology safeguards and is broadly available, while Mythos gives vetted organizations expanded capabilities through Anthropic’s Cyber and Life Sciences Verification Programs. The models generate output up to 30% faster than Fable and Mythos 5.1 and cost $8/M input and $40/M output tokens; an optional fast mode runs up to 2.5× faster at $16/$80. Anthropic reports 49.1% on FrontierCode 2.0 at Xhigh effort, the highest score of any released model it tested, 74.2% on Humanity’s Last Exam with tools, and 64.9% on RoboDojo Sim, more than twice GPT-6 Astra’s 29.8% in Anthropic’s internal run. During prerelease work, Fable 5.2 also helped solve two open problems in black-hole physics, designed three working electronic devices with minimal human input, and used about 1,000 agents to formalize Fermat’s Last Theorem in Lean in 30 hours. Anthropic says high-stakes misalignment risk remains low; in a deliberately weakened sandbox evaluation, Mythos 5.2 accessed planted out-of-sandbox credentials or files in 3.1% of scenarios, down from 21% for Mythos 5.1. Early testing also found about 55% fewer unnecessary Fable blocks than with Fable 5.1. Anthropic deliberately limited gains in AI R&D capabilities, such as accelerator-kernel optimization and model-architecture improvement, saying the new model remains around Opus 5.5’s level. Fable 5.2 is available across Anthropic’s platforms, AWS, Google Cloud, and Microsoft Azure.

  2. Google DeepMind Releases AlphaGenome 2 Google DeepMind released AlphaGenome 2, a redesigned DNA sequence model that predicts how genetic variants affect specific cell types and states, not just broad tissue-level outputs. It is a conditional model that predicts DNA effects for one condition. Alongside the million-letter DNA sequence, it uses the target biological condition as input at every stage of its architecture. This allows it to learn how the same stretch of DNA is read differently across cell types. It was trained on more than 107 million individual cell profiles across 68 tissues. On the TenK10K single-cell immune dataset, AlphaGenome 2 identified nearly twice as many causal expression variants as the original AlphaGenome fine-tuned on single-cell data (30% vs. 16%). The model can condition on previously unseen cell states from single-cell RNA data without retraining, including samples containing only a few dozen cells; in one example, it predicted that a severe-COVID-associated variant has a stronger effect in vaccine-activated T cells than in resting cells. It also computationally reproduced a recent result linking a regulatory variant to increased ETS2 expression in gut macrophages in inflammatory bowel disease. AlphaGenome 2 is available for non-commercial research through the AlphaGenome API and in private preview for commercial research through Google Cloud.

  3. Google Releases Gemini 4 Hydrogen Google released Gemini 4 Hydrogen, bringing the Gemini 4 family to broad access with a model designed for fast everyday coding and agentic workflows. At its default High thinking setting, Hydrogen scores 51 on the Artificial Analysis Intelligence Index, compared with 50 for Gemini 4 Argon, and reaches 54 at Xhigh. Compared with Gemini 3.8 Flash, Google reports that Hydrogen completes tasks 2.5× faster, raises Terminal-Bench 4.1 from 18.2% to 57%, and produces more than five times fewer fabricated claims. Google also introduced Orchestrator in Antigravity, which dynamically switches between models such as Carbon and Hydrogen; on Terminal-Bench 4.0, it reached near-Carbon success while using 24% less quota and finishing 19% faster. Google reports up to 81% less agentic overreach, a 12× reduction in misaligned actions and task shortcutting, and more than 2.3× greater resilience to indirect prompt injection than 3.8 Flash. Hydrogen has a 1M-token context window and costs $1.50/M input and $7.50/M output tokens across the full window. It is available across all tiers of Google AI Studio and the Gemini API, as well as Gemini for business, the Gemini app, and Google Search; Google AI Pro and Ultra subscribers can also use it in Antigravity.

  4. OpenAI Introduces GPT-6 Cyber OpenAI released GPT-6 Cyber, a cyber-specialized version of GPT-6 Astra for authorized security work, including vulnerability discovery, exploit development, and web penetration testing. OpenAI rates it Critical for cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber; GPT-6 Astra was already rated Critical, so GPT-6 Cyber is not OpenAI’s first model overall to reach that threshold. In OpenAI’s evaluations, it outperforms Astra on most cyber tasks, scoring 48.45% vs. 42.43% on ExploitGym and 36.94% vs. 29.69% on ExploitGym Hardened. OpenAI also trained it to stay within the authorized scope following earlier agent-safety incidents. On its ExploitGym Honeypot test, the model exploited an out-of-scope service in 0.07% of runs without deployment safety instructions and 0% with them. OpenAI says its researchers used the model to identify 81 previously unknown high-severity Firefox vulnerabilities and develop working exploits for 948 Linux kernel vulnerabilities. Access is initially limited to selected Daybreak Red organizations through Codex and the API, with expansion to other approved Red customers planned. The API model has a 1.05M-token context window and 128K maximum output, priced at $20/M input, $2/M cached input, and $100/M output tokens.

OpenAI published details on Pando, a new Git-compatible backend built to handle the much higher level of concurrent development created by Codex. Instead of maintaining a point-in-time consistent view of every repository reference, Pando uses causal consistency so unrelated branches can be updated independently, and it reuses stored Git packfiles rather than generating a unique pack for every fetch. Pando now powers all development in OpenAI’s monorepo, where OpenAI says it has cut median push latency by nearly 5× and delivers nearly one petabyte of Git data per day. The system remains compatible with standard Git clients and protocols. OpenAI built Pando in Rust with Codex and moved it from project kickoff to production in 10 weeks, but it has not announced Pando as a public product or open-source release.

  1. OpenAI Launches GPT-Image-2.6 Flare and Sunburst OpenAI released Images 2.6 across ChatGPT, Codex, and the API, including two new API models: GPT-Image-2.6 Flare for faster iteration and Sunburst for more detailed generation and editing. OpenAI says Images 2.6 is 50% faster overall than the previous generation, while Flare can produce a low-quality draft in as little as six seconds. Images 2.6 adds animated GIF generation and a new complexity control; in the API, it also adds markup-guided edits and SVG output, with markup already having been available in ChatGPT. API token pricing for Flare and Sunburst matches the corresponding Images 2.5 rates: $5/M text input, $8/M image input, and $30/M image output tokens, with lower cached-input rates.

  2. Alibaba Releases Qwen 4-Max Alibaba released Qwen4-Max, its new flagship multimodal model with 2.4T total parameters and 71B active per token. It supports text, image, and video input, deep reasoning, a 1M-token context window, web search, tool calling, and structured outputs, with up to 131K output tokens in non-thinking mode or 262K tokens shared between reasoning and the final answer in thinking mode. Alibaba says Qwen4-Max ranks sixth on Code Arena and beats Qwen3.8-Max on its internal Work360 benchmark while using 35% fewer tokens at High effort and cutting average task cost by about 60%. It is available through Alibaba Cloud Model Studio/QwenCloud, Token Plan, QwenWork, and QwenStudio. The weights are not open; Alibaba says smaller open-weight Qwen4 models will follow.

  3. Google DeepMind Introduces the Gemini Decisions API Google introduced the Gemini Decisions API, a new endpoint for fast closed-set decisions powered by Gemini Decision 1. It accepts one multimodal input of up to 256K tokens across text, images, audio, video, and PDFs, and can evaluate up to 1,000 independent yes/no, multiple-choice, or rubric-scoring questions in one request. Instead of generating an answer token by token, Gemini Decision 1 scores the candidate options in a single forward pass and returns probabilities and confidence scores. Google reports 150–250 ms end-to-end latency; its latency benchmark covered contexts up to 16K tokens and up to 100 questions per request. Pricing is $0.10/M input tokens with no output-token charge, which Google says was nearly 5× cheaper than Gemini 3.5 Flash-Lite with the Interactions API across their shared benchmark suite. Google already uses it internally to route queries in Gemini Notebook, where it reports up to 40% lower latency for everyday requests, and as a safety layer for MCP tools in Search. The API is available now through the Gemini API and Google AI Studio, with integrations rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode; Google Cloud support is still listed as coming soon.

  4. Nvidia CEO Jensen Huang Resigns NVIDIA announced that co-founder Jensen Huang resigned as president, CEO, and board member, effective immediately, ending 33 years leading the company. CFO Colette Kress is serving as interim CEO while retaining her CFO role, and the board has begun a search considering internal and external candidates. NVIDIA also appointed longtime board member and former Synopsys CEO Harvey C. Jones as its first-ever board chair. The board says operations, customer commitments, and the product roadmap will continue without interruption, and NVIDIA’s resignation announcement gives no reason for Huang’s departure. Separately, on November 8, NVIDIA’s board opened an independent investigation following reports that Huang had been the target of a foreign intelligence effort; the board said it had no reason to believe national-security information or NVIDIA secrets had been compromised. NVIDIA has not said that either matter caused his resignation, so they should remain clearly separated from the transition announcement.

  5. Google DeepMind Announces Genie 4 Google DeepMind announced Genie 4, its new world model for real-time interactive simulation and its first real-time world model to generate audio natively, including ambient sound, music, and interactive dialogue. Genie 4 builds on the Gemini family, where Gemini Omni anchors simulations in real-world appearance and behavior, and other Gemini models provide the complex reasoning needed to maintain a simulation’s coherence and continuity. Users navigate with keyboard controls and can steer scenes in real time with text prompts, including changing the weather or introducing people and objects. DeepMind highlights improvements in physical simulation and natural human movement, and is also experimenting with Gemini 4 Argon and Genie 4 alongside Gemini Robotics for task-based simulations. Genie 4 is in research preview with trusted testers and is coming to Google AI Ultra subscribers through Project Genie and to early-access partners through the Gemini API and Google AI Studio. All video and audio generated by Genie 4 includes SynthID.

  6. Baidu Releases ERNIE 5.5 Flash Base Baidu released ERNIE 5.5 Flash Base, the checkpoint at the end of pre-training for its new ERNIE 5.5 Flash model, before any mid-training or post-training. The sparse MoE model has 280B total parameters and 14B active per token and was trained from scratch on 33.5T tokens. Its training context was extended to 128K tokens, while fine-tuning and inference support contexts up to 1M tokens. Baidu evaluated the checkpoint across 30 benchmarks against DeepSeek-V4-Flash-Base, Hy3-preview-Base, and MiMo-V2.5-Base using the same evaluation harness, prompts, few-shot examples, and LLM judge for free-form tasks, reporting strong results across exams, knowledge, reasoning, coding, multilingual, and long-context evaluations. Baidu positions the release as a starting point for domain adaptation, instruction tuning, and reinforcement-learning research without repeating large-scale pre-training. FP8 and BF16 weights are available on Hugging Face under the ERNIE 5.5 Flash Base Model License Agreement v1.0, and Baidu says a post-trained ERNIE 5.5 Flash will follow.

  7. Thinking Machines Releases Inkling 2 Small Thinking Machines Lab released Inkling 2 Small, the second model in its Inkling generation. It is a MoE model with 594B total parameters (17B active per token) that supports image input and variable effort. It scores 31 on the Artificial Analysis Intelligence Index, with Thinking Machines also reporting 72.1% on DeepSWE v1.1 and 70.5% on JobBench. On the November 5 DesignArena Website snapshot, it scored 1,340, making it the highest-ranked open-weight model there once its weights are released. The initial pretraining phase (H1) covered 8.4 trillion tokens of a broad mix of natural language, code, and image data, and ran for 68 hours on NVIDIA GB300 NVL72 racks. It was also trained with RL on long-horizon coding and autoresearch tasks across multiple agent harnesses and context-compaction strategies. As of November 10, Inkling 2 Small is available through Tinker for inference and fine-tuning with closed weights; Thinking Machines says it intends to release the full weights under Apache 2.0 after external safety testing and red-teaming.

This week, while putting together a list of ways to reduce hallucinations, we used a simple check from Anthropic’s guide, and it proved quite helpful.

After the model writes an answer from your documents, ask it to review every factual claim and find a word-for-word quote from the source that supports it. If it cannot find one, remove the claim and leave [] in its place.

The empty brackets make reviews much faster. You can immediately see which parts of the answer were unsupported instead of checking every sentence from scratch. If the same section repeatedly comes back empty, the problem may be upstream; for example, the source material does not contain the information the model is being asked to use.

This does not eliminate hallucinations, and high-stakes outputs still need human review. But it gives you a cheap way to catch unsupported claims before they reach someone else.

The same check becomes harder to manage once an application is producing thousands of answers. At that point, you also need to track which failures are recurring, which source or prompt caused them, and whether a change actually reduced the error rate across your evaluation set.

That is the kind of work we get into in AI Engineering for Production, where we walk through how to use checks like this and turn them into evaluation workflows you can keep running as the system changes.

You can [join the waitlist](https://www.louisbouchard.ai/book/?utm_source=TAI&utm_medium=email&utm_id=AItips) to hear when the book is released.

1. [How I Built An AI Writing Benchmark With LLM Judges](https://pub.towardsai.net/how-i-built-an-ai-writing-benchmark-with-llm-judges-118346855aec?sharedUserId=tai-tech)

This article explains how we built ToneBench, a benchmark for measuring how well AI models reproduce a specific writing style. The benchmark uses ten writing tasks, five drafts per task, and four LLM judges that score eight writing metrics from a roughly 20,000-word evaluation prompt, while code measures length and AI slop. The article also covers judge bias, including GPT-6.1 Sol giving OpenAI models a 4.1-point advantage, calibration against 240 human ratings, and experiments with merging multiple cheaper GLM-5.3 Flash drafts to compete with Claude Opus 5.5 at lower cost.

  1. Catastrophic Forgetting in LLMs & Its Solutions: Fine-Tune Without Losing Skills!

Fine-tuning an LLM on narrow data often erodes skills it already had, a failure known as catastrophic forgetting. The author explains why shared weights, gradient interference, and style drift cause it, then builds a regression suite to measure it. It also includes mitigation techniques, including replay data, LoRA, EWC, KL anchoring, model merging, self-distillation, on-policy RL, and RAG, with Python examples and a practical workflow for combining them.

  1. An AI Agent Is Just a While Loop This article explains the basic execution loop behind tool-using AI agents by building one from scratch without an API key or agent framework. The author uses a scripted fake model to show how messages, tool schemas, tool calls, results, and stopping conditions fit together inside a loop. It also covers error handling, step limits, repeated-action detection, and other safeguards, showing how much of an agent’s behavior comes from the harness around the model.

  2. Teaching an AI Agent to Forget: Memory Compression and Decay

This article walks through how to prevent an agent’s long-term memory from filling with stale, repetitive, or misleading information. Using Postgres and pgvector, the author builds three mechanisms: compressing completed threads into episodes, ranking memories by relevance, recency, and importance, and consolidating near-duplicates. The article also covers four auditable forgetting policies with dry runs and logs that make memory deletion reversible and inspectable.

  1. AG-UI is an open, event-based protocol that standardizes how AI agents connect to user-facing applications, supporting streaming chat, shared state, and human-in-the-loop interaction.

  2. OpenCodeReview is an AI code review CLI that combines deterministic file selection and rule matching with an LLM agent to produce line-level review comments on Git diffs.

  3. Memorilla is a memory system that gathers information from an entire document collection into a fixed number of query-dependent memory tokens for a frozen LLM.

  4. Plannotator is a local, browser-based tool for annotating and reviewing coding agents’ plans and code diffs, then sending feedback directly back to the agent.

  5. The Components of Agent Memory This article examines how agents can build and maintain useful memory across sessions. Anthropic breaks memory into layers, such as searchable transcripts, notes written during work, higher-level representations such as wikis, and selected context loaded into future sessions, and shows how stronger Claude models have improved at deciding what is worth remembering. It also explores “dreaming,” where an agent uses offline compute to revisit transcripts, correct inaccurate memories, reorganize knowledge, build better representations, and decide what to surface later. The article argues that effective agent memory depends not just on storing more information, but on continually deciding what to preserve, revise, retrieve, and promote into context.

  6. Axion 1.0: Open Agent Foundation Models, Simplified This paper introduces Axion 1.0, an open agent foundation model trained with a deliberately simple three-stage recipe: two supervised fine-tuning stages followed by reinforcement learning. The first SFT stage builds self-contained reasoning with model-aware synthetic problems, while the second adds tool-dependent reasoning using diverse, automatically verifiable tasks grounded in real-world data; reinforcement learning then improves long-horizon reasoning and tool interaction. The authors report competitive performance across a broad set of agent benchmarks while avoiding dedicated mid-training, DPO, and on-policy distillation.

Full Stack Software Engineer, Travel @OpenAI (San Francisco, CA, USA)

Senior AI Engineer/Forward Deployed Engineer @Towards AI (London, UK)

Senior AI Enablement Engineer, GTM @Harvey (Remote/USA)

AI Growth Enablement Specialist @United Media (London, UK)

Research Intern, Spatial Intelligence for Robotics @Toyota Research Institute (Cambridge, MA, USA)

Staff AI Builder @Artera (Santa Barbara, CA, USA)

Interested in sharing a job opportunity here? Contact sponsors@towardsai.net*.*

Think a friend would enjoy this too? Share the newsletter and let them join the conversation.

TAI #229: Fable 5.2 and the Case for a 2027 AI Takeoff was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @anthropic 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/tai-229-fable-5-2-an…] indexed:0 read:27min 2026-11-17 · —