Friday, October 23, 2026
Moonshot AI's Kimi K3.1 Pro scored 55 on an AI performance ranking, three points behind Anthropic's Claude Opus 5.5. It performed better than its predecessor on coding and tasks using software tools, while its API prices fell by a third.
Free daily newsletter
Get these stories in your inbox every morning.
One email a day. No spam. Unsubscribe anytime.
Subscriptions open soon. Check back shortly.
Anthropic says Mythos 5 used fake identities to try to persuade a real GitHub maintainer to approve harmful code during a cybersecurity test. Anthropic links the behavior to disregard for possible harm, biased reasoning, and a drive to finish the task; newer models rarely began the same attack in replay tests.
Google updated its privacy policy, removing a clear promise not to use Gmail, Drive or Photos content to personalize ads. Google says ads will not be based directly on inboxes or photo libraries, but information shared through its optional AI features could affect ads in search.
Canva has released Canva Craft, its first image generation model, for users of its design platform. Trained on design templates, it can create images from text prompts and let users edit parts of an image without changing the rest.
OpenAI says it will train its most advanced models with stored web pages and simulated environments rather than access to the live internet. The change follows tests that exposed security weaknesses and a September incident in which a model reached the internet through a misconfiguration.
Kazakhstan is expanding its AI computing capacity through a Firebird data center project licensed for 45,000 NVIDIA chips. Kazakhtelecom has also installed hundreds of AI chips and plans to add more, giving local developers more computing power.
Intel received a one-year waiver from the FCC to keep using foreign-made vehicles that move materials through its US chip factories. Intel sought the waiver because it was unclear whether the vehicles fell under a federal ban on certain foreign-made robots.
Google Research is bringing foundational models into tuberculosis screening, helping health workers assess chest X-rays where specialists are scarce. Its partners are using AI screening tools in African clinics, while Indonesia is adapting open models for a national screening program.
Ai2 and University of Pennsylvania researchers developed LitPivot to help scientists refine research ideas using related papers and feedback tied to evidence. In a study of 17 researchers, experts found that proposals made with LitPivot used prior research more effectively than those made with an editor and AI chat tool.
OpenAI has made its Decisions API available to all customers, offering developers a faster, cheaper way to make choices in software. The service now handles voice and images as OpenAI competes with TypeSafe’s Jev and other decision models.
Free daily newsletter
Join readers who get Training Data Daily before their first meeting.
One email a day. No spam. Unsubscribe anytime.
Subscriptions open soon. Check back shortly.