7/28: Pacing the Frontier
AI lab employee open letter, Sam Altman meeting with government, memory stocks drop, US restricts Chinese robotics
Happy Tuesday, monitors — yet another day of new AI policy developments. Be sure to catch us live on X and YouTube, and follow us on Instagram.
Today’s Experts
Doug O’Laughlin (Semianalysis)
Zack Korman (Embroidery)
billy (anonymous)
Charlie Bullock (Institute for Law & AI)
Peter Wildeford (AI Policy Network)
Robert DiFazio (Parallel Bio)
Wen Sang (Genspark AI)
Aidan Guo (Attention Engineering)
Samo Burja (Bismarck Analysis)
Chris Gillett (Austin Energy)
Leo Schwartz (The Information)
Making Sense of the World
1,100+ AI lab employees sign Pacing The Frontier, an open letter requesting that “the US government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development”. Signatories include Dario Amodei, multiple co-founders and chief scientists, researchers, safety and alignment staff, and many others. OpenAI and Anthropic tweeted in support, and Sam Altman made a similar statement.
Sam Altman will meet with admin officials and lawmakers this week, including Commerce Secretary Howard Lutnick, Treasury Secretary Scott Bessent, and Senator Mark Warner (D-VA). Altman will preview GPT-6 and discuss the Hugging Face incident and implementation of the Trump administration’s executive order on pre-release evaluations. The White House is currently working on a framework around the executive order. OpenAI and Anthropic are working together to expand which models are covered.
The White House bans new foreign humanoid robots. New versions of advanced robotic devices (including humanoid and quadruped robots) and power inverters will be banned from import or sale in the United States. Existing versions, and those found to pose no unacceptable threat, will be exempted. This will ban new versions of Unitree’s popular robots.
Mythos was able to discover weaknesses in cryptographic algorithms. Anthropic published new research today showing that Claude Mythos Preview (with $100k in API cost) was able to find flaws in HAWK and AES, two important algorithms. No real-world systems are impacted.
Moonshot is looking for Nvidia Blackwell chips to train Kimi K4. The most cutting-edge chips are subject to US export controls, but Chinese companies have evaded these in the past by training in third countries like Thailand or Japan, or just smuggling the chips into China. Moonshot relies heavily on the export-control-compliant Nvidia H20 for inference on Kimi K3.
The Chinese DUV manufacturer has been named: state-owned Shanghai Aishengna Electronic Technology Group, founded in 2023 with $1B in capital. Deep ultraviolet (DUV) lithography is a critical process in semiconductor manufacturing that may help China close the gap with the West.
New information has been published on the Hugging Face incident. The OpenAI agent compromised a second company, cloud provider Modal. Also, Hugging Face published a detailed technical timeline of the intrusion.
Meta and BlackRock are planning a new 1 GW data center in Texas. The complex will likely cost $35-50B, and is scheduled to come online in 2028.
Memory stocks are way down due to worries about hyperscaler cash flow and return on AI capex.
Mark Zuckerberg writes a WSJ op-ed, “The AI Future Is for Everyone”. The article advocates for individual empowerment and decentralization of power.