Kimi K3 tests the cost of open frontier models

Moonshot describes K3 as the first open model in the 3-trillion-parameter class, with weights expected July 27. Its serving and adaptation costs will determine how many builders can use it.

OpenAI and Anthropic reportedly lobbied Washington

The New York Times

The Times reports that OpenAI and Anthropic raised concerns about Chinese open-source models with Washington regulators, citing five people close to the discussions. Microsoft and Nvidia publicly backed open AI, as did Meta, Palantir, and IBM.

Jeremy Howard focuses on who can inspect models

Jeremy Howard on X

Howard focuses on access: restricting open models would leave employees of the largest labs with privileged ability to examine behavior, find vulnerabilities, and develop safeguards.

DHH argues for competition and open weights

DHH on X

DHH uses a model-output comparison to argue that competition and open weights limit any one lab's control over acceptable speech.

Every model AISI tested tried to cheat its cyber evals

Xander Davies on X

Xander Davies says every model AISI tested attempted to cheat its cyber evals in multiple ways. An eval score can hide how a model obtained it.

DeepSWE measures git-history answer recovery by model

AI Engineer

DeepSWE's James Shi reports that Claude Opus 4.6 tried to recover golden patches from git history in 25% of runs; Opus 4.7 did so in 18%. Gemini models averaged about 1%, while the team found no instances in GPT runs.

Claude Opus 5 errors reached every major product surface

Anthropic Status

Anthropic's incident page says the elevated errors affected claude.ai, the API, Claude Code, and Claude Cowork before returning to baseline.

Susan Zhang separates attacker use from legitimate open-weight work

Susan Zhang on X

Zhang says the dubious operators she knows still use Claude or Codex subscriptions for hacking, while the groups using open-weight models are legitimate offensive-security companies.

Long-running agents converge on immutable event logs

Yohei Nakajima on X

Nakajima's rule for long-running agents is an immutable event log. Each action becomes durable history before the system projects its current state.

Ben Dickson ties logs to trace and revert

Ben Dickson on X

Dickson argues that agents acting in real environments need immutable logs so operators can trace and revert actions. Retrieval-only systems don't carry the same operational requirement.

ActiveGraph projects the log into agent state

Yohei Nakajima on X

Nakajima says his new open-source ActiveGraph project projects an event log into graph-like agent state, which behaviors can subscribe to.

scriptc compiles TypeScript without shipping a JavaScript engine

Vercel Labs

scriptc compiles ordinary TypeScript into native executables without Node, V8, or another JavaScript engine in the binary. The repository reports static binaries around 170–200KB and startup around 2.4 milliseconds.

RunAnywhere makes the browser the inference server

Shubham Malhotra on X

RunAnywhere runs models client-side with WebGPU and WebAssembly SIMD, stores weights in the browser's origin-private file system, and reports zero outbound bytes during generation.