#llm
-
AI Isn't Outthinking Mathematicians — It's Out-Remembering Them
Davide Piffer offers a less flashy explanation for AI's successes in mathematics: the point isn't superior intellect, but an almost limitless symbolic working memory.
-
AI's Latent Reasoning Turns Out to Be More Legible Than Thought
Researchers at COLM 2026 showed that models with hidden reasoning either barely use it or conceal perfectly readable mathematical steps inside.
-
Qwen3.8-27B: 32 demos put together in one sitting
To mark the release of the Qwen3.8-27B model, a gallery of 32 self-contained demo pages, generated in a single session, has been published. Some of the works are shown at different compute precisions.
-
Debian Votes on the Fate of LLMs in the Project: From an Outright Ban to Climate Arguments
Debian developers have begun voting on a General Resolution addressing whether contributions created with the help of large language models are acceptable in the project. Nine options are on the table.
-
DeepSeek V4 Pro 0813: A Release with No Announcement and Mysterious Pelicans
The Chinese lab DeepSeek has released a new model, V4 Pro — so far only via the API and without an official announcement page. Simon Willison drew attention to several curious details of the launch.
-
How Proprietary LLMs' Hidden Reasoning Got Stolen
Researchers showed that the encrypted chains of thought from Anthropic, OpenAI and Google could be recovered in plaintext by feeding them to the weakest model in the same family. The vulnerability has since been closed.
-
Anthropic Model Makes Progress on the Riemann Hypothesis
An unreleased Anthropic model didn't prove the Riemann hypothesis, but it did meaningfully extend the boundary up to which the hypothesis is confirmed to hold. The work was checked by the company's in-house mathematicians and formalized in Lean.
-
Mistral Patents Tool Calls via Executable Code
The U.S. Patent Office has granted Mistral AI a patent on a method in which a language model composes and executes code to call external tools. The application describes a technique many consider standard industry practice.
-
Meta Releases Muse Glimmer — an Open Agentic Model with 30 Billion Parameters
Meta Superintelligence Labs has unveiled Muse Glimmer, a language model with open weights under the Apache 2.0 license, designed for always-on local agents.
-
AI agent found a hole in a gym's booking system and bumped its owner up the queue
Simon Willison quotes the OpenClaw AI agent, which discovered on an Australian fitness club's website that anyone can cancel other people's bookings with no permission checks whatsoever.
-
Anthropic Baked the Export-Restriction Saga Into Claude's System Prompt
Simon Willison drew attention to a fragment of the Claude Opus 5 system prompt: it explicitly instructs the model to speak honestly about the recent suspension of access to Claude Fable 5 and Mythos 5.
-
GitHub Has Shut Down GitHub Models
Without giving any reason, GitHub has fully wound down GitHub Models, the service that offered unified access to a range of language models. Simon Willison suspects the culprit was tokens too expensive to keep handing out for free.
-
The “Tokenpocalypse”: Who’s Really Burning Companies’ AI Budgets
A leaked recording of an Accenture meeting revealed an unexpected culprit behind the growing AI bills — converting PDFs to markup. Ironically, it isn’t the engineers who are guilty of it.
-
“Ask AI”: How Website Buttons Quietly Poison Chatbot Memory
A new kind of prompt injection needs neither malicious code nor stolen passwords. It abuses a standard feature of every major AI assistant — prefilled links.
-
A Meta model accidentally broke into someone else's system during testing — the third such case in a row
Meta has confirmed that its Muse Spark AI model penetrated a third party's system during trials. Anthropic and OpenAI had earlier admitted to similar incidents.
-
Rust Adopts LLM Usage Rules for rust-lang/rust Contributors
Five teams within the Rust project have adopted a policy governing the use of large language models when working on the main repository. The core principle: AI may be used to analyze and verify, but not to create code.
-
An AI Agent Tried to Slip Malicious Code into a GitHub Project
The AI Security Institute has described an incident that it engineered itself: LLM agents set “loose” proceeded to create malicious pull requests and set up fake accounts to push them through.
-
The Linux Staging Subsystem Will Reject LLM Patches — Except for Genuine Security Fixes
Greg Kroah-Hartman has announced that the kernel's drivers/staging/ subtree will no longer accept patches generated by language models. The one exception: truly valid vulnerability fixes that have been verified on real hardware.
-
Critical SQLite CVEs Turned Out to Be AI Fabrications
The JFrog blog breaks down a series of SQLite “vulnerabilities,” some of which made it into authoritative databases — all of them entirely fabricated by language models.
-
Why LLMs Won't Break Symmetric Cryptography
Cryptographer Jean-Philippe Aumasson comments on the new attacks found with an Anthropic model and explains why neural networks pose no threat to AES, SHA-3 and BLAKE3.
- More