Evening brief

Thursday, July 30

1 people · 0 companies · 2 topics

The signal

  1. “Pacing the Frontier”: 1,178 AI lab employees call for US-backed pacing mechanism; OpenAI and Anthropic formally endorse What happened: On July 28, 1,178 verified employees across OpenAI, Anthropic, Google, Meta, and several smaller labs published “Pacing the Frontier” – asking the US government to develop the technical and governance tools needed to deliberately pace automated AI development at some future point, not now. darioamodei and several Anthropic co-founders signed; OpenAI chief scientist Jakub Pachocki signed; Meta’s VP of AI research signed. Both OpenAI and Anthropic formally endorsed on July 29. Mark Zuckerberg published a same-day WSJ op-ed opposing it, framing the danger as AI concentration rather than capability. Why it matters: For the first time, sitting CEOs and chief scientists at both US frontier labs are publicly aligned on a specific future control posture, and both companies are on record backing US-led pacing infrastructure directly ahead of the August 1 TRAINS deadline. Confidence: Confirmed Sources: CNN (independent) · TechTimes (independent) · thezvi (independent) · AI Weekly (independent)

Since last brief

  • Kimi K3 llama.cpp PR #26185: still open as of July 30, no new code owner approvals; community testing extended to 8xRTX PRO 2000 Blackwell with CPU-mmap experts (~4.9 tok/s hybrid); vision tower support proposed in a companion PR by danielhanchen; --cache-reuse 0 still required for multi-turn. PR #26185 (primary)

Moves on the board

Player Move Strategic implication
OpenAI + Anthropic Formally endorsed “Pacing the Frontier” July 29 Both labs now on record backing US-led pacing infrastructure ahead of the Aug 1 TRAINS deadline
Meta / Zuckerberg WSJ op-ed opposing the pacing letter July 28 Frames slowdown as concentration risk; firms up open-weights stance directly before TRAINS

Other

  • EU AI Act enforcement goes live August 2 (2 days): chatbot disclosure, AI-content labeling, deepfake marking, and GPAI model-recall powers all activate; Digital Omnibus pushed high-risk AI obligations to Dec 2027 and Aug 2028 but did not delay Article 50 transparency or GPAI provisions; penalty cap is the greater of EUR 15M or 3% global annual turnover. technology.org (independent) · EU Commission (primary)
  • TRAINS August 1 (tomorrow): Meta still absent; White House pressing Meta directly, citing Llama’s open-weights release model as incompatible with post-release access controls; no formal consequence mechanism for holdouts confirmed. TechTimes (independent) · AI Weekly (independent)

Watchlist

  • Kimi K3 llama.cpp KDA kernel work: PR #26185 still open, no new code owner approvals; vision tower work proposed in companion PRs; --cache-reuse 0 required for multi-turn (opened 2026-07-23)
  • OpenAI ExploitGym / Kill Switch Act: “Pacing the Frontier” letter now company-endorsed by OpenAI and Anthropic; Senate companion to Kill Switch Act still absent; AISI cross-lab cheating investigation still pending (opened 2026-07-23)
  • TRAINS August 1: tomorrow; Meta absent; White House pressing directly; no formal consequence mechanism for holdouts (opened 2026-07-25)
  • EU AI Act August 2: 2 days out; GPAI enforcement + Article 50 transparency go live (opened 2026-07-26)