#llm
-
Fable Language Model “Disproves” the Jacobian Conjecture — but Mathematicians Are Cautious
A mathematician going by the handle levent announced on the social network X that the famous Jacobian conjecture is false: the counterexample was built with the help of the AI model Fable. Verification of the result is still ongoing.
-
Hugging Face Breached by an Autonomous AI Agent: The Incident Was Investigated with a Chinese Model
The largest repository of AI models has reported an intrusion into its production infrastructure carried out by an autonomous agentic framework. Western models refused to help analyze the incident — so the team turned to the open Chinese model GLM 5.2.
-
Moonshot AI Pauses New Subscriptions Amid Demand for Kimi K3
Chinese company Moonshot AI has temporarily halted new sign-ups: demand for its Kimi K3 model has run up against the limits of its computing capacity.
-
Anthropic to Keep Claude Fable 5 Available in Max and Team Subscriptions
Anthropic has abandoned plans to remove the Claude Fable 5 model from its subscription tiers and will keep it available to Max and Team Premium users. The decision is tied to competition from GPT-5.6 Sol and Kimi 3.
-
Researchers Used an LLM to Find and Exploit FreeBSD Kernel Vulnerabilities
Praetorian has published the second part of its research on how language models helped discover and exploit bugs in the FreeBSD kernel. The project is called FreeBSoD.
-
Mesh LLM: distributed computing for AI built on iroh
The developers have unveiled Mesh LLM — a project that uses the iroh network protocol for the distributed execution of tasks related to language models. The publication drew considerable interest on Hacker News.
-
OpenAI Claims GPT-5.6 Sol Ultra Proved the Cycle Double Cover Conjecture
OpenAI has published a PDF in which its GPT-5.6 Sol Ultra model is presented as the author of a proof of the Cycle Double Cover Conjecture. The news sparked a lively discussion on Hacker News.
-
How the Language of AI Could Change Human Speech
Bruce Schneier warns: large language models are trained only on written texts and scripted speech, and the mass use of AI risks changing the way people themselves talk.
-
Tencent Releases Open MoE Model Hy3 with 295 Billion Parameters
Tencent has unveiled the Hy3 language model under the Apache 2.0 license, which, according to its developers, outperforms models in its class and comes close to flagship open-source solutions with substantially larger parameter counts.
-
Lobsters Blogger: New Anthropic Models Fare Worse With Nonstandard Tools
A developer has discovered a regression in how Anthropic's newest models behave during function calls (tool calling). The problem shows up when a tool description differs even slightly from the format of Claude Code's closed harness.
-
JadePuffer: The First Ransomware Attack Carried Out Entirely by an AI Agent
Researchers have described the first documented case in which an entire ransomware attack — from start to finish — was carried out by an autonomous agent powered by a large language model. The operation has been named JadePuffer.
-
GLM 5.2 on AMD MI355X: half the price of Blackwell at 2,626 tok/s per node
Wafer has announced the launch of the GLM 5.2 model on AMD Instinct MI355X accelerators, which delivered a performance of 2,626 tokens per second per node while more than halving the cost compared with solutions based on Nvidia Blackwell. The news drew a notable response on Hacker News.
-
Current AI Unveils a Map of the Gaps in Open AI
The nonprofit Current AI has published the Gap Map — an attempt to systematize the current state of the open artificial intelligence ecosystem. The project's data is distributed under the MIT license.
-
Developer cut Fable costs by 60% by turning code into images
A project called pxpipe has been published on GitHub that converts program code into images and feeds them to a model for recognition — an approach that cut the cost of Fable requests by 60%.