Search
10 stories found
Cognition has released SWE-2 Flash, a smaller, faster coding model inside Devin that it says runs about 3x faster than SWE-2 at a fifth of the cost and is free for paid users through November 30. Its own benchmarks show where the trade-off bites.
A misconfigured DNS resolver, chained zero-days and a covert channel built from CPU cache timing — OpenAI's own red team found the paths. Here's what happened and why it matters beyond OpenAI.
With the 27.3 betas, Apple opened a Siri AI Developer Preview that lets any app expose actions, content and multi-step "skills" to its new assistant. You can build and test now. You can't ship to the App Store until spring, iPhone users in the EU can't use Siri AI yet, and server-side usage has daily limits Apple hasn't quantified.
Carbon tops the Artificial Analysis Intelligence Index and posts big security benchmark numbers. Google is rolling it out to vetted defenders before anyone else. Here's what's in it and what it costs.
Thinking Machines Lab released Inkling 2 Small, a 594B-total, 17B-active mixture-of-experts model with image input, on its Tinker platform. Weights are due under Apache 2.0 after external safety testing, and every benchmark so far is self-reported.
Meta has acquired Conductor, the Mac app for running several AI coding agents in parallel. Sign-ups have stopped, the hosted Cloud goes read-only on November 13 and is deleted from November 20, and the desktop app winds down in early 2027.
DeepSeek's Harness team published an MIT-licensed research prototype that watches a GitHub repo, takes authorized requests and answers with pull requests. It passed 11,000 stars within days — but DeepSeek itself says it is not production-ready and runs with full user access.
An independent METR assessment, built on Anthropic's internal data, estimates the lab is now improving its models on well-specified tasks at roughly twice its pre-October-2025 pace, mostly because of its own AI tools. Anthropic redacted most of the evidence METR found most informative, so outsiders can't check the numbers.
A new arXiv preprint claims a theorem showing that JEPA models trained on different modalities converge to the same representation. It reports that a 109M-parameter text model out-aligns 72B LLMs. The results are striking, but they are benchmark-specific and not yet peer-reviewed.
The MIT-licensed AiSOC ingests security telemetry, runs thousands of detection rules, has LLM agents investigate alerts and keeps a human in charge of any response. It runs on a laptop-class Docker setup with a small local model — here's what it does, what it needs and where it falls short.
