Sacks to Amodei: “Go ahead. You guys are the frontier.”
David Sacks, the former White House AI and crypto czar, replied on X to Amodei’s “We Must Pace the Frontier”: “Dario has written that we need to ‘pace the frontier,’ and Sam has agreed. People may be surprised by my response: go ahead. You guys are the frontier.” His argument is that the two labs already lead on adoption and model strength, so if they want to slow down they can do it unilaterally, without asking for an antitrust waiver or a new approval regime as the price. He published the same response the same day as a signed op-ed on RealClearPolitics. Yesterday I broke down the structure of Amodei’s plan: of its three steps, only the first (resident evaluators) needs nobody’s cooperation. Sacks left the czar job in March when his 130-day special government employee term ran out (CNBC), but he now co-chairs the President’s Council of Advisors on Science and Technology (White House announcement). He no longer speaks for the administration, and his post announces no government decision; still, my read is that a PCAST co-chair publicly waving off the premise makes the narrow antitrust waiver that step two depends on look unlikely anytime soon. What “go ahead” skips over is the reason Amodei gave for asking for coordination in the first place: in his framing, a lab that slows down alone simply loses the race. The generous-sounding reply declines every part of the plan that needs government to move.
Thinking Machines open-sources Inkling-Small, and the student outscores the teacher
Mira Murati’s Thinking Machines released Inkling-Small with full weights on Hugging Face: a mixture-of-experts model with 276B total and 12B active parameters, text, image and audio input, a context window up to 1M tokens, priced at $1.20 per million output tokens on Tinker versus $4.05 for the flagship Inkling (975B total, 41B active). Post-training of an earlier checkpoint partly used on-policy distillation, followed by further RL on coding tasks: the student generates its own attempts and Inkling, as the teacher, grades every token of them, instead of the student imitating the teacher’s finished outputs. The student now beats the teacher on several reasoning benchmarks: 31.6% vs 29.7% on Humanity’s Last Exam, 80.2% vs 77.6% on SWE-bench Verified, 89.5% vs 87.2% on GPQA Diamond. The announcement is candid about what didn’t transfer: the big model keeps its edge on knowledge coverage and factuality, and AIME is slightly lower. In this run, distillation carried reasoning; it didn’t carry knowledge capacity. For builders, my take: an open-weights model at 80% on SWE-bench Verified with 12B active parameters changes the math on self-hosted coding agents. One caveat: 276B total parameters still puts the weights alone in the hundreds of gigabytes at BF16 (model card), so even quantized this is multi-GPU territory, and “small” describes the inference compute, not the memory footprint.
Obama urges Democrats to bring a “clear plan” for AI safeguards
At a Democratic fundraiser on Thursday, in a conversation with House minority leader Hakeem Jeffries, Barack Obama said AI is “moving very fast in private hands, and if we don’t get on top of it, I think can be dangerous,” and that Democrats should “put together a framework for a very public conversation” once they retake the House. He also pointed at the upside: done right, AI could speed up drug development and disease treatment. The remarks reached the press through a partial transcript from his office, reported by the New York Times and relayed by TechCrunch. Put next to the Sacks post, the week’s contrast is clear: the administration’s former AI czar tells the two leading labs to restrain themselves without new rules for anyone, while the opposition’s most prominent voice wants AI safeguards to be a central part of the Democrats’ agenda, with midterms less than two months away. The practical point for practitioners, as I read it: how much federal AI regulation the US gets now depends in part on who wins in November, so “no federal AI rules” is an election bet, not a constant you can plan around.
Reverse-engineering Claude Web’s sandbox: a Firecracker microVM with careful cross-session hygiene
Independent developer AprilNEA ran strace, strings, and other standard Linux tools inside a Claude Web session and mapped out “Antspace,” Anthropic’s internal sandbox platform. Each session is a Firecracker microVM (4 vCPUs, 16GB RAM) restored from a frozen snapshot rather than booted fresh; PID 1 is a custom Rust init process that supervises everything over WebSocket and HTTP APIs instead of systemd. The most interesting part is the cross-session hygiene: the kernel zeroes freed memory pages (init_on_free=1), the random number generator is reseeded after every snapshot restore, and the image ships without SSH, cron, or logging daemons. To be precise about what this is: not a vulnerability disclosure. The author did not escape the sandbox; the reverse engineering went this deep because one key binary shipped with debug symbols intact, so what leaked is engineering detail, not a security boundary. If you build agent sandboxes, I’d read it as a free worked example of how one platform handles a threat class that is not unique to Anthropic: data bleeding between sessions through leftover memory or predictable randomness.
One line for today: Amodei shipped the mechanism draft yesterday; the replies arrived today. Sacks tells the labs to slow themselves down and ask nothing of government, and Obama wants Democrats to campaign on safeguards. Neither reply moves the two steps of the plan that need governments, which makes the one step that needs nobody’s permission, resident evaluators, the real test of intent.