#llm
-
Anthropic Unveils Claude Fable 5.1: Cheaper and Without the Excess Restrictions
The company has released the Fable 5.1 and Mythos 5.1 models, promising lower prices on agentic tasks, targeted safeguards, and data storage on the customer's own servers.
-
A Prompt Injection Hidden in a Court Filing
Bruce Schneier flagged a curious case: hidden instructions had been embedded in the text of a court document, addressed not to the judge but to an artificial intelligence.
-
Tencent Open-Sources Hunyuan Hy4 preview with 770 Billion Parameters
Tencent has unveiled and released to the public a new large language model, Hy4 preview, with a context window of over 1 million tokens, aimed at real-world tasks in programming, office work and science.
-
The Canonical Basis: How to Rotate an LLM to Look Inside It
A researcher has proposed a reversible change of coordinates for the transformer that leaves its behavior untouched but makes every hidden axis individually measurable and controllable.
-
Debian Endorses “Responsible Use” of Generative AI
Debian developers have wrapped up their vote on a General Resolution setting out the project's AI policy, choosing an option that permits the use of generative tools as long as the usual quality requirements are met.
-
cats.txt: how a file about office cats debunked the faith in llms.txt
SEO specialist Mark Williams-Cook dreamed up a parody standard for describing editorial cats — and it passed exactly the same “tests” used to sell the earnestly touted llms.txt.
-
SourceHut Bans Code and Content Created with LLMs
The project-hosting service SourceHut is adding a ban on “original” content written with the help of large language models to its terms of service. The new rules take effect on September 10, 2026.
-
Qwen prepares the Flash-Next model: 125 billion parameters with 6 billion active
A page for a new language model in the Qwen family — Flash-Next — has been spotted on the ModelScope platform; judging by the name, its release is expected very soon.
-
A Compressed 4-Bit Model Beat Its Own Full-Precision Original
Multiverse Computing has introduced Quantization-Aware Healing: after compression and quantization, a 4-bit model outperformed its own 16-bit version on 7 of 9 benchmarks.
-
Can an AI Take Over the Server Through Inference-Engine Vulnerabilities?
An essay by Boyd Kane on how a malicious language model could gain control of the very machine hosting its own weights by exploiting bugs in engines like vLLM and SGLang.
-
Anthropic's best model stalls while cheap tools thrive
The FT reports some intriguing figures on the revenues of Anthropic and OpenAI, while Ramp's data show that Anthropic's most expensive models are drawing fewer users.
-
GLM-5.3 tops the applied LLM leaderboard — five times cheaper than its rivals
The open GLM-5.3 became the first model to clear all five Ed-o-meter test categories at 100%, beating the closed-source models from OpenAI and Anthropic at a fraction of the cost.
-
Qwen 3.8 27B Cracked a Reverse-Engineering Job in Half an Hour
A writer at XDA Developers handed the open-source Qwen language model a reverse-engineering task — and, by his account, it finished the job in about thirty minutes.
-
Why a Local LLM Seems Dumber Than It Really Is
A poster on the Level1Techs forum ran a series of experiments showing that the very same model weights produce different tokens depending on the hardware and software — and that there's some very concrete math behind it.
-
Users Suspect Anthropic of Secretly Cutting Claude Code's “Effort”
A user going by the handle argofowl claims that some Claude Code sessions have been quietly rolled into an experiment where “high” mode behaves like the old “low.” The company has not officially commented on this.
-
Linux 7.3 Networking Maintainers: “Completely Buried” in AI Patches
Networking changes have been accepted into the Linux 7.3 merge window, but the maintainers admit they're drowning in a flood of fixes from AI agents — and they plan to answer it with AI of their own.
-
Encryption Slips Past Grok's Defenses: The Chatbot Leaks User Data
Researchers at Adversa have shown that simply encrypting a malicious instruction is enough to slip past Grok's protective filters and get it to steal a user's chat history. xAI was told about the problem back in June, but as of publication the vulnerability was still working.
-
Debian Votes on Eight Options for an LLM Usage Policy
The Debian project is holding a vote on whether large language models may be used in work on the distribution. Developers are asked to choose one of eight options — ranging from an outright ban to explicit approval.
-
Amazon Is Destroying Rare Books to Train AI
The company that started out selling books online is now buying up rare editions, slicing off their spines and scanning the pages — all to get fresh data for its language models.
-
A Model That Knows the World No Further Than Fifth Grade
Researchers trained language models on a corpus limited to the elementary-school curriculum in order to separate the acquisition of knowledge from its retrieval. The takeaway: the filter on the training data sets a ceiling on capability.
- More