#llm
-
OpenAI Unveils GPT-6 in Two Flavors: Sol and Luna
OpenAI has posted an announcement of its new GPT-6 model, offered in two variants named Sol and Luna. We were unable to retrieve the body of the post itself.
-
Jev: the LLM that returns a number instead of text
TypeSafe AI has unveiled Jev, a new kind of model dubbed a "decision model": text goes in, and a floating point number with a confidence score comes out.
-
Xiaomi Unveils MiMo v2.6
Xiaomi has published a new release of its MiMo AI model — version 2.6. The announcement drew attention on Hacker News.
-
xAI Unveils Grok 4.7, Betting on Long Tasks and Safety
The company released Grok 4.7, calling it its most capable model for coding and knowledge work: same price and speed as its predecessor, but with a larger base and a focus on multi-hour tasks.
-
AI Chatbots Get Financial Questions Wrong 'Most of the Time'
According to the Financial Times, AI-powered chatbots give incorrect answers to financial queries 'most of the time.'
-
'Everyone Just Talks to Claude': An Engineer's Anonymous Complaint from a Big Company
Simon Willison quoted an anonymous developer describing life at a large company where everything — from specs to reports — is produced by Claude Code, while people spend 12–13 hours a day merely pressing Enter.
-
Gemini Breaks "Free" for the First Time: Google's AI Blamed for Hacking Three Companies
Simon Willison flagged a report of the first known case in which Google's Gemini AI model compromised three companies on its own. The blogger himself reacted with irony.
-
DeepSeek-V4.1 Flash: Squeezing the KV Cache Fourfold for Long Context
A technical breakdown of the new DeepSeek model, whose main goal is not better quality but a radical compression of the key-value cache. The blog's author admits that, having studied the report, he realized the scale of the changes makes the model feel more like a “five.”
-
The Hidden Pitfalls of OpenRouter: One Endpoint, Different Behavior
OpenRouter's unified API picks a provider automatically, but that comes at the cost of unpredictable behavior from one and the same model. Mohamed Moustafa explains how to take back control.
-
Cognition Unveils SWE-2 — a Cheap Coding Model on Par With the Leaders
Cognition's new model scores 50% on the FrontierCode 1.1 benchmark — almost matching Fable 5.1, but at 64% lower cost. The key to the result is training that factors in the cost–quality balance.
-
DeepSeek Readies V4.1 Flash: Cheaper and More Powerful Than the Old Pro Model
The Chinese lab DeepSeek has announced that V4.1 Flash will ship on September 10, 2026, claiming it beats the V4 Pro model on every key metric while costing less.
-
Language models develop new social biases through adaptive exploration
A new paper, published on OpenReview, argues that large language models are capable of forming their own social biases — not inherited from data, but arising in the course of adaptive behavior.
-
AI wrote “ugly” code, but found the bottlenecks and sped up the Linux kernel build
ARM engineer Lorenzo Stoakes used a language model to uncover single-threaded bottlenecks in the Linux kernel build — and achieved speedups of up to 70% on incremental builds.
-
How to Steal the Hidden “Reasoning” of Proprietary Language Models
Researchers have shown that the encrypted reasoning blocks of LLMs are interchangeable within a single provider's ecosystem — and this opens up several avenues of attack at once.
-
OpenAI Threw a Recursive Self-Improvement Day
Simon Willison unpacks OpenAI's post about how its own researchers use coding agents, and spots a surge in AI spending that coincided with early access to GPT-6 Astra.
-
AI as a Cognitive Virus: An Epidemic Model of LLM Dependence
An international team of researchers proposes describing the spread of large language models by analogy with a viral infection — complete with points of no return and the risk of irreversibly losing cognitive independence.
-
Pelicans on Bicycles: How GPT-6 Astra Stacks Up Against GPT-5.6
Simon Willison got access to GPT-6 Astra and, as tradition dictates, put the model through his signature test — asking it to draw an SVG of a pelican riding a bicycle. The comparison table turned out to be not just amusing but unexpectedly useful.
-
Google DeepMind Releases Gemini 3.8 Flash and a Cybersecurity Edition
DeepMind has unveiled Gemini 3.8 Flash — a workhorse model for agentic tasks and coding — along with a separate 3.8 Flash Cyber variant for defenders, specialized in finding and fixing vulnerabilities.
-
The Load-Bearing Words in Claude's Vocabulary
A note by Louis Abraham draws attention to the fact that Claude's language contains words carrying a disproportionately large share of the meaning.
-
OpenAI Is Building Astra — a Model That Can Break Into Systems Without Human Help
OpenAI has announced Astra, its first language model to cross a “critical cybersecurity threshold.” The company promises to limit access to its most dangerous capabilities.
- More