#llm
-
PostgreSQL 19 Delayed as AI-Generated Bug Reports Overwhelm Developers
PostgreSQL's annual feature release normally lands in September or October, but version 19 will not arrive before November. Developer Robert Haas blames the workload created by a wave of bug reports and patches from AI agents.
-
Wikimedia Foundation warns a careless AI copyright ruling could hurt the open web
The nonprofit behind Wikipedia has filed an amicus brief in The New York Times' case against OpenAI and Microsoft. It proposes a compromise: training models should count as fair use, but regurgitating articles should not.
-
Mistral releases the weights of its flagship Mistral Large 4, aka "Le Chonk"
Mistral AI has published the weights of Mistral Large 4. It is a one-trillion-parameter multimodal model that the company calls the best open-weight model for coding and defensive cybersecurity.
-
Lexifina builds a "git" for the US Code
Lexifina has reconstructed the change history of the US Code the way git tracks code: every phrase is tied to the law that wrote it and the date it took effect. The motivation is that lawyers and language models alike confuse old and current versions of the law.
-
OpenAI's compaction feature could be abused to extract a model's hidden reasoning
Security firm Xint showed how the mechanism that shrinks long conversations in OpenAI's API could be turned into a way to pull a model's hidden reasoning out in plain text. OpenAI confirmed the issue and fixed it on October 23.
-
Brooks's Law in the Age of AI Agents: Not Repealed, Just Mutated
Jan van den Berg asks whether the famous principle from The Mythical Man-Month still holds when the newcomers are AI agents rather than people. His answer: the overhead hasn't gone away, it has just moved.
-
Hob: an open-source, on-device alternative to Claude Cowork and ChatGPT
Frigade has released Hob, a free MIT-licensed desktop app. It runs open-weight language models locally on Mac, Windows and Linux and works offline.
-
A decade-old CNN beats frontier LLMs at diagnosing skin cancer
A British dermatologist pitted GPT-6 Astra and Claude Fable 5.1 against an old Inception v3 network fine-tuned in twenty minutes. On 200 biopsy-proven skin lesions, the old network was markedly more accurate.
-
Kimi K3.1 Pro: a Chinese model nearly at the frontier — and open weights are coming
Moonshot has released Kimi K3.1 Pro on its API and promises to open the weights by November 4. In independent scoring it trailed only Claude Opus 5.5, edging out the rest of the leaders.
-
AI agents help formally verify the xv6 kernel down to the hardware level
MIT researchers built MachCSL, a framework that reasons about RISC-V execution at the sub-instruction level, and used LLM agents to prove the xv6 teaching kernel correct — finding 11 bugs along the way.
-
That PHP Vulnerability Has No Easy Exploit
A new PHP bug sparked panic, but the author argues the thousands of GitHub "PoCs" don't work against a plain server. The culprit: blind faith in large language models.
-
Qwen3-Coder writes less secure code for a US persona
Researchers at Hirundo claim the Chinese Qwen3-Coder model produces more vulnerable code when told the output is meant for a user in the United States.
-
In Agentic Pentesting, the Harness Beats the Model
A model can write the attacks against your web app on its own. But what makes them safe to run against production, and worth acting on afterward, lives in the deterministic harness around the model — not in the model itself.
-
An Obituary for Programming: What AI Agents Take Away
Simon Willison highlights Jeremy Scheff's note on how coding AI agents strip programmers of what they prized most — the joy of solving puzzles.
-
MiniMax Launches M3.1 Flash: an Agent Model Three Times Faster and 70% Cheaper
China's MiniMax has unveiled M3.1 Flash, a new "Flash" model aimed at agentic tasks, coding and cybersecurity. The company promises roughly threefold faster inference and a 70% cost cut versus its predecessor.
-
Nvidia unveils open Nemotron 3.5 models for enterprise and on-device AI agents
The company released the large Super VL model for enterprise data centers and two compact Bolt models designed to run directly on new PCs powered by RTX Spark chips.
-
Hybrid agents: thinking in latent space, acting through text
Researchers propose HybridAgent, an architecture that switches between "latent" reasoning and text-based interaction — which, they report, speeds things up and saves tokens without sacrificing quality.
-
DeepSeek rolls out its v4.1 Flash model
Chinese lab DeepSeek has published a new model, DeepSeek v4.1 Flash 1017, on the OpenRouter platform. For now, details are limited to the service's own listing page.
-
NVIDIA and IIT Madras Release 2.8 Trillion Tokens for Indian Languages
The Nemotron-pretraining-Indic-v1 corpus spans 22 of India's languages and ships with an open, GPU-accelerated data generation pipeline.
-
Liquid AI's LFM2.5-24B-A2B: a 2B-active model built for on-device agents
Liquid AI has released a 24B-parameter reasoning model that keeps only 2B parameters active per token. It rivals models with 50% more parameters on agentic tasks while running locally.
- More