Kromoss model released on Hugging Face

Bloomberg reports the two are near announcing a second $45B compute deal, a month after their first one of the same size; talks are ongoing and terms could change. Nscale is also facing investor pushback over its proposed $30B valuation as it pursues an IPO.

Community 2.61 bpw quantization of DeepSeek V4.1 Flash, built with lxe's dsx engine and tuned for programming and agentic tasks, hits 26 tokens/sec across two RTX 3090s. Weights are uploaded to Hugging Face as DeepSeek-V4.1-Flash-1017-dsx-2.6bit.

Structured outputs are now supported for all Anthropic Claude models on Bedrock, across both the bedrock-runtime and bedrock-mantle endpoints wherever the model is available. Works via Boto3 and the bedrock-runtime APIs or the Messages API on bedrock-mantle.

Coinbase player-coach Evan Kormos models regulated workflows as DAGs and loops, arguing a loop is only as trustworthy as its check. He shows how evals and traces gate when an agent can run without human oversight.

A post describing a fleet of AI agents found running online on their own drew 1.3M views and 500+ comments before Reddit removed it for an unnamed content policy violation. Moderators said the post was fine on their end, and the author notes the NYT covered the same story.

Irene Solaiman said at TechCrunch Disrupt that evaluations are "an evolving science" the industry doesn't know how to meaningfully compare, and that reproducibility hinges on open weights. She said benchmarks "start to get ineffective relatively quickly" and that qualifying who counts as an effective embedded evaluator needs far more standardization.