Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Replay Doctor, a locally run diagnostic tool, reads Claude Code transcripts to identify which turn caused prompt cache expiry or other failures and how many tokens were re-billed at write prices. The project attributes r…
GPT-5.6 Luna found 69 verified bugs across 50 public benchmark pull requests at a total cost of $0.20, while GPT-6 Astra found 92 at $5.66, according to an evaluation published by the AI-Code-Review-Evals project. Luna's…
Apple released iOS 27, macOS 27 Golden Gate, watchOS 27, visionOS 27, and tvOS 27, headlined by Siri AI, a large language model-based overhaul of the company's voice and text assistant. The new Siri runs primarily on App…
Apple released Siri AI to the public as part of iOS 27, adding a new opt-in setup prompt that asks users whether to "Share Audio and Text" or "Not Now" to improve Siri and Apple Intelligence. Apple said the data collecte…
Maxime De Greve, a designer on GitHub's Copilot team, published a browser experiment called Octoglow on September 14th that replaces QR codes with an animated particle cloud encoding a six-digit code for screen-to-screen…
Cohere CEO Aidan Gomez warned that advanced AI models are the "most potent cyber weapon" ever created, citing their ability to find and exploit vulnerabilities at scale, and said companies must deploy AI defensively to f…
Developers can cut monthly LLM API bills by up to 60% by trimming conversation history, using prompt caching, and routing simple tasks to cheaper models such as GPT-4o-mini or Haiku, according to a first-person guide. Th…
Victor Vieira, CTO and co-founder of Sigilo Profissional, detailed a privacy-first architecture for the B2B workplace-survey and whistleblowing platform, combining PostgreSQL row-level security with application-level ten…
Microsoft CEO Satya Nadella said any pursuit of superintelligence must be grounded in the principle that AI not helping humanity and under human control is not worth pursuing, and that AI benefits must be diffused broadl…
Agentic workloads run roughly double the context of chat conversations by turn 10 and require a median of 15,146 prefill tokens and 501 decode tokens per turn, according to an analysis of a custom harness running SWE-ben…
Anthropic CEO Dario Amodei published an essay this weekend calling for a brake on the pace of development of large language models, citing looming dangers from the technology. The essay marks a doomer turn in the AI indu…
A holiday-let operator running 23 properties built AI agents into daily operations and found that giving agents more context often degrades their decisions, because large context blobs bury relevant information. The oper…
Apple shipped iOS 27 today with two undocumented mechanisms in the release candidate code that would let third-party AI models such as Claude and ChatGPT replace Siri's AI backend, according to researcher @itspdfu. The M…
Andriy Mulyar, founder and CEO of Nomic, alleged in a four-post thread on X on September 14th that startups tied to OpenAI, Anthropic and Y Combinator are reselling frontier-model traces to unidentified buyers, including…
An anonymous email attributed to a machine learning engineer at a frontier AI lab claims the lab has hit limits on scaling models for programming tasks and that frontier labs lack a competitive moat, predicting an open-w…
Richard Socher's new venture Recursive raised a $4.65 billion seed round to build a "Eureka Machine" aimed at automating AI research itself, with Socher claiming a Recursive AI research system outperformed humans and the…
Elon Musk said on September 14, 2026 that Grok 4.8, a 2.5T model trained with a new C++ software stack, will finish training that week and start RL. The post from Musk's account drew 2.8 million views. The claim concerns…
A developer documented building a fully offline conversational chatbot named Daisy using Ollama, LangChain, and Streamlit on a CPU-only laptop with 16GB of RAM and no dedicated GPU. The project runs Meta's llama3.2 3B mo…
Cursor outperforms GitHub Copilot for Python projects with deep codebases because it indexes the entire local project, letting it reference installed library versions and types across 20+ modules, while Copilot is limite…
Tencent Yuanbao, Tencent's free consumer AI assistant built on the Hunyuan family, has been added to the Uprouter directory, which rates it as medium risk. The current version serves the Hunyuan Hy4 preview with upgraded…