Grok 4.6 makes disclosure a same-day event

xAI released Grok 4.6, public questions about its safety documentation followed, and a capability card appeared within hours. The card didn't end the scrutiny: the limited deployment and evaluation detail became part of the release story.

The Grok 4.6 disclosure loop took less than a day

Miles Brundage on X

Brundage linked the capability card after questioning the absence of one earlier in the release. The sequence shows how public pressure can turn model documentation into part of the launch itself.

DeepSeek V4 Pro arrives with benchmarks and API pricing

OpenRouter

OpenRouter lists DeepSeek V4 Pro 0813 with a full benchmark table and API pricing. The API-first release also makes the sequencing between hosted access and possible weights explicit.

A 96 GB local-inference card now lists at $16,000

Tom's Hardware

Nvidia's high-memory RTX PRO Blackwell now carries a $16,000 MSRP. That is twice the price of preorders that began below $8,000 last year.

Continual Learning Bench separates adaptation from base ability

AI Engineer

Continual Learning Bench sequences tasks and reports reward, cost, and gain. Its gain metric compares a stateful run with a reset-memory baseline, isolating whether the system learned across instances.

Memory helps when the answer has left the active context

AI Engineer

Sakana AI found no benefit from memory on a short literature-review task where everything fit in context, while token costs increased. On long-horizon tasks, ranked decision recall beat unguided retrieval and simple memory gating.

Long-running agents need four kinds of context

Nate B Jones

The proposed split is stable instructions, current project state, a resource map, and accessible history. Updating the current-state layer lets an operator redirect a long run without forcing stale history back into every prompt.

LangChain treats traces as queryable data

AI Engineer

LangChain's approach mines centralized agent traces for recurring issues and evaluation cases instead of loading every trace into a model. That reduces context and token costs while preserving the evidence needed to improve the harness.

Agent evals move into executable tests

AI Engineer

Raindrop argues that production agents should be evaluated through local unit and end-to-end tests against the full harness. Its triage method asks when an issue began and what share of users it affects.

The White House creates a path for private cyber operations

The White House

The presidential action establishes a framework for vetted private companies to conduct offensive cyber operations with government authorization. It arrives during a separate debate over restricted access to cyber-capable AI models.

Anthropic tests which ideas spread between agents

Elvis Saravia on X

Anthropic's multi-agent experiment evolves ideas and measures how they propagate through a population. The setup targets coordination effects that a single-agent evaluation can't observe.

DeepMind translates simultaneous signing into text

Google DeepMind on X

DeepMind's SL2T model processes hand, body, and facial movement together instead of treating sign language as a sequence of isolated hand poses. The release targets the full visual structure of signing.