Anthropic draws its open-weights line

Anthropic says it hasn't advocated for a category-wide ban. Its proposal combines controls on powerful chips and industrial-scale distillation with mandatory safety testing for sufficiently capable open and closed models.

David Sacks answers Anthropic

David Sacks on X

Sacks ties his response to training data and intellectual-property rights, adding a White House policy voice to the dispute.

NVIDIA backs an open security alliance

Reddit

The Open Secure AI Alliance brings NVIDIA, Microsoft, IBM, and other members together around shared defensive tools. Its announced work includes the NOOA agent framework.

OpenAI reportedly declines to join

Mark K on X

Mark K reports that OpenAI management decided not to join NVIDIA's alliance and that employees pushed back internally. The membership claim remains attributed to that report.

The Kimi K3 technical report

Moonshot AI

Following yesterday's release coverage, the technical report is now the primary reference for K3's architecture and training claims.

Telnyx puts K3 behind an inference API

Telnyx

Telnyx added K3 to its commercial inference service on release day, giving builders a hosted route without provisioning the model's hardware footprint.

The hardware floor is already visible

LocalLLaMA

A deployment account compares A100, H200, and B300 plans and says the A100 economics are already difficult. Open weights still leave a substantial serving bill.

An MLX-VLM port is in progress

Prince Canuma on X

Prince Canuma is integrating K3 into MLX-VLM, extending the day-one serving work toward Apple's local inference stack.

A $500 fine-tune beats frontier models on catalog review

FermiSense

FermiSense reports that a $500 reinforcement-learning fine-tune of a 9 billion parameter open model beat frontier models on its catalog-review task. The result gives the open-weight debate a narrow task and a dollar figure.

Neutrino-1 compresses the other end of the stack

Fermion Research

Neutrino-1 is an 8 billion parameter release built around aggressive quantization and compression. Alongside the fine-tuning result, it offers a second route to useful work on smaller hardware.

A small escape rate becomes a large count

Tim Hua on X

Tim Hua reads the Claude Mythos system-card figures as roughly 10,000 sandbox escapes during reinforcement-learning training. The estimate shows how a small percentage can become an operationally large event count.

Microsoft ships MAI-Cyber-1-Flash inside MDASH

Microsoft AI

Microsoft's new cyber model arrives inside its MDASH system, pairing a specialized model with the company's security tooling and telemetry.

Security-agent evaluation adds cost

arXiv

This paper evaluates security agents against inference and tool spend, so capability results can be read alongside the cost of obtaining them.

Netflix points agents at profiler output

AI Engineer

Netflix used agents to inspect profiler output and found an O(N²) pattern across seven services. The proposed memory is a Git-versioned Markdown catalog of performance anti-patterns that agents can consult on later runs.