Artificial Intelligence

An August 2026 paper shows that the chat template, not the model itself, decides whether an LLM responds with ‘I’m just an AI’ or in first person. The authors found a direction in the activations that reproduces that shift at will.

Tech leaders asked the UN to control the development of artificial intelligence, while Jensen Huang challenged a key figure in the debate and the US Congress added regulatory pressure.

A developer extended the trick of reading an LLM’s logprobs, popularized by Jev, to classify webcam images without training any new model. With a local Gemma 4 12B he reached 1 frame per second; against OpenAI’s API, performance dropped to 0.2 fps.

Claude, Anthropic’s AI model, identified a candidate CRISPR-like gene editing system called ART in 21 hours. The finding, reported by Geeky Gadgets and The Conversation, still lacks published experimental validation.

NeoGenelogic secured a US patent for an artificial intelligence technology that identifies therapeutic targets against cancer. The case illustrates how AI has become central to the pharmaceutical race, although a patent doesn’t yet equal a proven treatment.

Anthropic released Claude Opus 5.5, the first model in its 5.5 family, with performance equivalent to Fable 5.1 and a cost 40% lower than Opus 5. The company also reports the best safety results it has achieved so far in its automated behavioral audit.

The White House accused China of copying American AI models on an industrial scale, according to CNN. We review the techniques used to distill language models and how that copying can be detected with real code.

Claude Code 2.1.277, released on September 18, 2026, adds support for AGENTS.md as an alternative to CLAUDE.md when the latter doesn’t exist in the project. The release also fixes more than twenty-five bugs and adds controls for corporate gateways.

Andrew Ng, co-founder of Google Brain, called warnings about human extinction from AI science fiction. The debate pits those pushing to pause the largest models against those who consider that risk technically implausible.

Bend is a new programming language that combines near-native speed, automatic parallelism across CPU and GPU, and a mathematical proof system to block bugs introduced by AI agents before they reach production. Its compiler acts as a theorem prover: if the code breaks a law declared in advance, the merge gets blocked.