Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
AI #167: The Prior Restraint Era Begins
TL;DR
The White House is openly floating FDA-style preapproval for frontier AI releases — Kevin Hassett said future models might need to be “proven safe” before release “just like an FDA drug,” and the administration has already reportedly blocked wider access to Anthropic’s Mythos.
Anthropic solved an immediate compute crunch by leasing SpaceX’s Colossus 1 — the deal adds over 300 megawatts and 220,000+ Nvidia GPUs within a month, letting Anthropic double Claude Code 5-hour limits and massively raise Opus API rate limits.
xAI looks less like a frontier lab and more like an infrastructure company now — Grok 4.3 landed with little excitement, xAI is being folded into SpaceX, and the speaker’s read is that Musk “fumbled the model” but built a highly competitive neocloud.
Anthropic’s business is exploding even while compute-constrained — SemiAnalysis numbers cited here put ARR at over $44 billion by mid-2026, up from $9 billion late in 2025, with inference gross margins improving from 38% to over 70%.
Musk and Anthropic abruptly went from enemies to partners — after meeting Anthropic’s leadership, Musk said “no one set off my evil detector,” approved leasing Colossus 1, and framed future compute access as available to firms taking the “right steps” for humanity.
The video’s core warning is that ad hoc AI restraint is replacing rule-of-law governance — the host argues the U.S. may be entering a “prior restraint era” where the executive branch pressures labs before release without clear legal authority, while also exploring coordination with China on model access restrictions.
The Breakdown
Cheap AI therapy, Go cheating, and the weirdly mundane reality
The video opens with a grounded point: AI is already useful in boring but real ways. A GPT-4.1-mini mental health app improved outcomes for depressed Mexican women by 0.3 standard deviations over six months and even made users more likely to seek human help. Then the mood flips: in Go, AI mostly fuels cheating and shallow learning, with players memorizing openings instead of learning to think — “a giant pile of slop” is the vibe.
Product churn: GPT-5.5 Instant, Gemma 4 speedups, and Grok’s shrug of a release
OpenAI ships GPT-5.5 Instant, Google makes Gemma 4 roughly 3x faster with multi-token prediction, and OpenAI adds opt-in advanced account security. Meanwhile Grok 4.3 technically exists, but the host is openly unimpressed: priced at $1.25/$2.50, scoring 53 on Artificial Analysis, and barely registering as a frontier release. The harsher subtext is that xAI’s ecosystem is unstable, including sunsetting earlier models with only two weeks’ notice.
The compute story of the week: Anthropic rents Musk’s GPU empire
After Anthropic committed $200 billion over five years to Google Cloud, it still needed more compute now, not next year. Enter the obvious deal: leasing all of SpaceX’s Colossus 1, adding 300+ MW and 220,000+ Nvidia GPUs, which immediately unlocked higher Claude and API limits. The host’s read is blunt and memorable: Musk may not have won at frontier models, but he is exceptional at “known hard things” like standing up massive infrastructure fast.
Benchmarks, lawsuits, slop books, and other signs of the times
Program Bench arrives and every tested model scores 0%, with Opus 4.7 leading at just under 3%, which the host treats as both funny and telling. Meta gets hit with another copyright lawsuit over training on pirated books, while a paper claims LLMs tripled new book releases from 2022 to 2025, pushed AI-authored books above half of 2025 releases, and increased consumer surplus by 7% — a conclusion the host clearly thinks depends on unrealistically perfect filtering. It’s the same theme all over again: more output, worse search, shakier quality.
Musk v. OpenAI: the courtroom version of the same old drama
The trial mostly rehashes known history, but under oath, which gives it more weight. A few details still pop: Musk admitted xAI is distilling OpenAI models, the judge swatted away extinction-risk grandstanding as beside the point, and testimony reinforced that Sam Altman’s ouster was about management and trust, not primarily safety. The line that sticks is Helen Toner on Mira Murati: she “didn’t realize that she was the wind.”
Anthropic’s revenue curve goes vertical — and Musk suddenly likes them
The financial section is wild: Anthropic’s ARR allegedly rockets from about $100 million in early 2024 to over $44 billion by mid-2026, and the host keeps stressing this is with compute constraints. That makes the détente with Musk feel less philosophical than strategic, though Musk still adds a human flourish, saying nobody at Anthropic triggered his “evil detector.” The whole segment has the energy of rivals realizing they need each other more than they need the feud.
The real thesis: the “prior restraint era” may have started
The back half is the heart of the video. With David Sacks out, the host says the White House is drifting toward executive-branch control over frontier model releases, especially after Mythos triggered cyber fears. Kevin Hassett invoking the FDA as the model is presented as the nightmare scenario — not thoughtful safety policy, but innovation-chilling prior restraint with unclear legal basis and huge room for arbitrary pressure.
If access controls are coming, they may become international
The closing twist is that U.S.-China talks on AI restrictions may be easier to stabilize than broad agreements to slow development. The host argues a “we won’t release if you won’t release” access pact is much more game-theoretically feasible than any true pause, even if it only buys time. That’s the unsettling final note: the world may not coordinate on alignment, but it might coordinate on controlling who gets to touch the models.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
