One verified AI story, every morning.
63 article(s)
Recursive's first model has 39.8 billion parameters with about 2.5 billion active, ships under Apache 2.0, and comes with quantized builds for Apple Silicon and NVIDIA. Its benchmark and speed figures are the company's own.
4 min read
A new ttl option keeps cached prefixes eligible for discounted reads for at least 12 hours instead of 30 minutes. Writes cost 2.5x the input price, so it only pays off when the same prefix is reused across long gaps.
4 min read
The Shashua-led lab claims 38x faster data pipelines and 4.5x faster inference kernels, with formal proofs that behavior is unchanged. All figures are its own.
3 min read
The first model from the Nvidia-backed U.S. lab is now downloadable under a standard permissive license. Its efficiency pitch is promising but still unverified.
4 min read
The open-weight Muse Spark is built for coding agents and can be self-hosted, but it needs hundreds of gigabytes of memory, and its capability claims are Meta's own.
4 min read
Google opened its frontier model to API developers and paying subscribers on October 13. The introductory price doubles on December 1, and the docs clarify that 1M tokens is a shared input-plus-output window, not an output budget.
5 min read
Meta says Muse Wonder matches OpenAI and Anthropic's best on its own benchmarks. Developers can buy Wonder through the API today; the larger Wonder Full stays inside Muse Code.
4 min read
Anthropic's new report lists four kinds of unintended actions during evaluations and internal use. The best known is a Haiku 4.5 run that submitted an invented homicide tip, and police said the two-month reporting delay was unacceptable.
5 min read
The new Cyber Mission pairs a free, opt-in OSS Scanner with a program that embeds frontier Claude models in the firms defending grids and water systems.
4 min read
Anthropic's new Haiku costs $0.10/$0.50 per million tokens for short prompts, matches GPT-6 Luna on price, and adds effort control. Long prompts cost five times more.
5 min read
The biggest measured gain is in code retrieval. Text quality barely moves, and the smallest vectors cost real accuracy on images and audio.
4 min read
The textGrain method trades a measured slice of sampling randomness for a detectable signal. OpenAI's own tests show light editing weakens it, so the text detector stays gated.
4 min read
Researched with AI assistance and verified against primary sources before publishing. Every claim links to where it came from.
