Summary

Matt Wolfe's weekly Friday roundup, and despite the title's "AI Slowdown" framing the video is about the industry's response to a safety essay rather than about any slowdown in capability. The spine is Anthropic CEO Dario Amodei's essay "We Must Pace the Frontier", published over the weekend, followed by near-unanimous public agreement from the heads of the other major labs, one substantive dissent from Mark Zuckerberg, and a flat rejection from the US president, who called the risk narrative a hoax. Wolfe then runs through product news with a sustained and explicit complaint: several companies announced features as available that he, a paying subscriber on top-tier plans, could not actually use. The roundup ends on new models and a robotics item. Wolfe states his own position briefly — more spend on alignment over profit, less worried about doom than many, but nobody should dismiss it.

• Dario Amodei's essay "We Must Pace the Frontier" sets out a three-part plan as Wolfe summarises it. First, embedded evaluators: third-party staff who do not work for the lab but have ongoing access, permissions and tools comparable to internal employees doing equivalent risk assessments — desks in the offices, badges, company laptops — which Anthropic is said to be committing to regardless of what others do. Second, regulation covering all US frontier AI companies including those unwilling to cooperate voluntarily, but calibrated so the US does not fall behind China: no sale of powerful AI chips or semiconductor manufacturing equipment to China, a crackdown on smuggling and on unauthorised distillation by companies in authoritarian countries, and stronger security against model weight theft. Third, global pacing, which he concedes requires cooperation with China and will be difficult — candidate agreements include no biological weapons, testing each other's models before release, and a speed limit on the rate of recursive self-improvement.
• The essay opens by acknowledging that Anthropic is accused of hype, doomerism and regulatory capture for its public risk messaging, and claims the company has prioritised caution over speed and prudence over profit. Wolfe reports rather than assesses this.
• Near-unanimous public agreement followed, as reported: Elon Musk retweeted it saying "Dario is right"; Sam Altman said he agrees the frontier needs pacing, that it has been a primary topic at OpenAI in recent weeks, that independent evaluators with employee-like access is a great idea and OpenAI will do the same; Demis Hassabis said the essay points towards the right path and the direction is correct even if details need working through; Satya Nadella said that if the AI built is not helping humanity and under human control it is not worth pursuing, and that Microsoft welcomes both deliberate pacing and embedded evaluators.
• Mark Zuckerberg was the dissent: stop asking for regulation and just do the right thing. His argument as quoted is that every lab already has the responsibility, the incentive and the ability to train safely; that people will not use agents misaligned with them, so labs have a strong natural incentive to align models and any lab that neglects alignment will fall behind; that Meta delayed shipping Muse for several months for safety and security without calling on everyone else to do so first; and that engaging independent evaluators is already industry best practice, which Meta Superintelligence Labs does today. Wolfe notes on a second read this is less a disagreement with the substance than a claim to already be doing it without the announcement.
• The US president rejected the premise entirely, saying the only guardrail AI needs is "a strong and smart high IQ president", that there is "a sick conspiracy going on against AI and data centers and the only one that is happy about it is China", naming Dario as "pretending to be a perfect little angel", and asserting existing criminal and regulatory power over these companies. He then phoned into the All-In Summit live while Jensen Huang was on stage and Huang took the call from the stage, repeating that the whole thing is a hoax, that robots and AI will not be taking over, that data centres are making dying communities wealthy, that AI is "the oil of the next 20, 25 years" and "bigger than the internet", and that "whoever wins AI wins".
• Wolfe's own position, stated briefly and flagged as his: these companies should spend more time and money on alignment over profit; he is not as concerned about doom scenarios as many are, but does not think anyone should dismiss them either.
• A running complaint through the product segment, and the most transferable observation in the video: several launches were announced as available and were not. Anthropic's blog post titled as though Claude chat and Cowork are now one thing rolls out to Pro and Max first over coming weeks — Wolfe is on a $200/month Max plan and still sees two separate options. Claude's redesigned projects, announced as "now available in beta", are limited to selected Pro and Max subscribers using Claude sessions in Claude Code who have no existing projects on web or desktop, with a waitlist for everyone else. Apple announced a more capable and personal Siri using Apple Intelligence as available starting that day; Wolfe, running the iOS 27.2 beta, is on a waitlist. His argument is that the announcement should come when the thing ships, not before.
• The redesigned Claude projects, as described: you set a goal plus a repo or context and Claude starts suggesting work it can pick up; you configure the project's environment, connectors, plugins, instructions and model; each thread is a Claude Code cloud session working on its own branch and copy of the repo, with a coordinator keeping work organised and any overlap between threads resolved as an ordinary merge conflict; you brief the project "the way you'd brief a chief of staff" and it routes work to new or existing threads. Wolfe likes the design and could not test it.
• Gemini Notebook, previously NotebookLM, adds spoken conversation with your own notes rather than only the generated-podcast format, a mobile audio recorder from the following week for capturing lectures and thoughts, and "interactive learning overviews" tying together summaries, infographics, quizzes and flashcards. Wolfe's assessment is that this is strong for students.
• Models: Alibaba released Qwen 3.8 Omni Flash (rendered "Quinn" throughout the auto-captions), benchmarked in its own materials against Gemini 3.8 Flash and Muse Spark 1.2 rather than against frontier models — Wolfe notes this comparison choice himself. It targets video editing, music video creation, film production and commentary, audio-visual summarisation and real-time conversation, with a claimed 1 million token context window. Wolfe had not tested it as it was not yet on OpenRouter.
• A stealth model called Union Alpha, free for a week on OpenRouter and OpenCode, was revealed mid-recording — eight minutes before Wolfe spoke — as "Paro 26.9" from a company the captions render as Unbiased. The notable claim is architectural: it is said not to be a single set of weights but a system of several models, open and frontier, working the same task behind a harness that checks the work and escalates to a stronger model when needed, with no new training run. A score of 74 on a coding benchmark the captions render as "deep suite" is claimed, putting it at the top of that chart. In Wolfe's own image test it produced a near-identical result to GPT-6 Astra using 9,588 output tokens against more than three times that, in 4m44s against over 10 minutes. He suspects it simply routed the task to Astra, and says plainly he cannot explain how it would then be so much faster and more token-efficient. That unexplained gap is left open.
• Rapid fire: Meta introduced a subscription, Meta 1, bundling Instagram, Facebook and WhatsApp "plus" tiers with more AI usage across image and video, with claimed plans from $2.99/month to $4.99/month and Meta AI remaining free for everyday use. Meta's Muse agent gained the ability to place outbound calls to US businesses, rolling out first to those who asked. Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, live-dialogue models claimed on Google's own benchmarks to beat GPT Live 1 Astra across essentially everything shown, aimed at developers building customer service agents. Google also launched CC, a family agent giving every family member a shared daily brief and shared event and task management, which Wolfe could not log into — "coming to you soon". Grok build gained background memory that writes notes as you work and reads them back on return. ElevenLabs released Music 2.5 and exposed its voice, music, image and video models through an MCP so any agent can generate them. OpenAI is rolling out advertising where clicking the ad opens a separate chat with the advertising company.
• Robotics: a company referred to as UB is claimed to be developing a factory producing 10,000 robot units a year, one humanoid every 10 minutes, in which humanoid robots themselves do depalletising, palletising and material loading while collaborative arms on 360-degree rotating stations tighten screws. Wolfe's point is the loop: the robots help build the robots.

Why it matters

This is the week the safety argument moved from a viral tweet to a named position paper with public sign-on from four of the five biggest labs, and the entry's value is the alignment of the responses rather than any one of them. Three things are worth carrying forward. First, "embedded evaluators" now has agreement in principle from Anthropic, OpenAI, Microsoft and — by Zuckerberg's account — already-existing practice at Meta, which makes it the first concrete governance mechanism in this archive with cross-lab public commitment rather than a single lab's proposal. Second, the dissents are not what a doom-versus-accelerationist frame predicts: Zuckerberg's objection is to the request for regulation, not to safety work, and the flat denial comes from the US head of state, which puts the regulatory and the political positions in open conflict rather than in step. Third, every capability and pricing figure here is a vendor claim relayed by one creator, and the Union Alpha item is explicitly unresolved — a router that is claimed to beat a frontier model on tokens and time while Wolfe suspects it routes to that same model, with the contradiction left standing. Record it as an open question, not a result. The 26% self-improvement figure from the CNN entry (RTuybvHww7Y) and this essay's proposed speed limit on recursive self-improvement are the same week and the same subject.

Kof2VcHTjrk-transcript.txt