Expecting internal coherence from language models means asking the wrong question, according to an Anthropic researcher.
"Why does page five of a book say that the best food is pizza and page 17 says the best food is pasta? What does the book really think? And you're like: 'It's a book!'", explains Josh Batson, research scientist at Anthropic, in MIT Technology Review.
The analogy comes from experiments on how AI models process facts internally. Anthropic discovered that Claude uses different mechanisms to know that bananas are yellow versus confirming that the statement "Bananas are yellow" is true. These mechanisms aren't connected to each other. When a model gives contradictory answers, it's drawing on different parts of itself - without any central authority coordinating them. "It might be like, you're talking to Claude and then it wanders off," says Batson. "And now you're not talking to Claude but something else."
The takeaway: Assuming language models have mental coherence like humans might be a fundamental category error.
Claude Code inventor Boris Cherny says his tool wrote nearly all the code for Claude Cowork. | Screenshot via X
Product Manager Felix Rieseberg says the app came together in just a sprint and a half, roughly one and a half weeks. The team had already built some prototypes and explored ideas beforehand, though, and the current release is still a research preview with a few rough edges, Rieseberg says. Claude Code also provided an extensive foundation to build on; Rieseberg is likely referring mainly to the front-end work.
UK startup turns planetary biodiversity into AI-generated drug candidates
UK company Basecamp Research has developed AI models together with researchers from Nvidia and Microsoft that generate potential new therapies against cancer and multidrug-resistant bacteria from a database of over one million species.
OpenAI is buying health app Torch for around 100 million dollars. The deal includes 60 million upfront and the rest in retention shares, The Information reports. Torch unifies scattered health records into what the founders call a "medical memory for AI", "a context engine that helps you see the full picture, connect the dots, and make sure nothing important gets lost." The app runs on OpenAI models. All four employees, including CEO Ilya Abyzov, are joining OpenAI.
Apple will use Google's Gemini models for its AI features, including a revamped version of Siri. The multi-year partnership means Apple will rely on Google's Gemini and cloud technology for its upcoming products, according to CNBC. The new features are expected to roll out later this year.
The move comes as Apple continues to struggle with Siri's underlying architecture. Internal reports describe Siri as a technically fragmented system built from old rule-based components and newer generative models - a combination that makes updates difficult and leads to frequent errors. Apple is also working on an entirely new in-house LLM architecture and a model with roughly one trillion parameters, aiming to eventually break free from external providers. Google faced similar challenges early on keeping pace with OpenAI's rapid progress but managed to catch up.
British media regulator Ofcom has opened an investigation into X over the AI chatbot Grok. The probe follows reports in recent weeks that Elon Musk's chatbot and social media platform were increasingly being used to create and share non-consensual intimate images and even sexualized images of children.
Ofcom is now examining whether X violated the UK's Online Safety Act. The regulator contacted X on January 5, 2025, demanding a response by January 9. The investigation aims to determine whether X took adequate steps to protect British users from illegal content. Violations could result in fines of up to 18 million pounds or 10 percent of global revenue. In severe cases, a court could even order X blocked in the UK.
Ofcom is also looking into whether xAI, the AI company behind Grok, broke any regulations. Last week, the EU Commission ordered X to preserve all internal documents and data related to the Grok AI chatbot through the end of 2026.
Google brings personalized discounts to AI search and launches open commerce protocol
The search engine company introduces personalized discount ads in its AI mode and presents an open protocol to bind retailers more closely to the Google ecosystem.
