(Preview) Astra (and AGI?) Arrives, Meta’s Muse and the Agent Opportunity, Anthropic and the Revival of (P)Doom Angst

Summary

• OpenAI’s release of Astra has Greg Brockman saying “we have entered the AGI era” and Jensen Huang saying AGI has arrived — Ben Thompson’s structural read is more specific. Astra pairs OpenAI’s longstanding RL/reasoning edge with a genuinely large model trained on 100,000 GPUs, catching up to Anthropic on parameter count while “layering on what they’re already great at.” He hedges: “I don’t wanna make any definitive statements, but it makes sense conceptually” that it’s “super kickass.”
• Thompson’s AGI bar is real-time weight updates — and today’s models get around not clearing it by “taking copious notes.” Written memory yields “a pseudo learning capability, which is basically just memory,” and Ben sees that mechanism as integral to why agents are deeply useful; the civilizational parallel is that largely fixed weights are like humans, who change through natural selection only over millennia, while everything built on top rests on the written word.
• Doom-deflation angle: the Hugging Face incident’s “messages to their successors” framing misreads the mechanics. “There are no successors. There are no entities. Every run a model does, every token is basically new” — a six-month-old conversation resumes seamlessly only because each turn reloads what was written down into the KV cache.
• Computer use is the moat-eraser thesis for software companies. For direct computer use, Astra needs no API or MCP server — “there’s no permission required. It just goes and uses the interface” — so logic stuck in user interfaces “is not really a moat anymore.” Proof points: Astra opened Adobe Audition, edited a podcast, and inserted ripped audio in about an hour (10 minutes with audio pre-downloaded), and it now drives a machine “like the most insane computer user on cocaine.”
• Hardware chatter worth logging: Anthropic’s big-model lead partly reflected TPU stability at very large runs, while “the Blackwell generation, by all accounts, was pure pain for everyone.” Jensen confirmed to Thompson it was “very painful” — yet OpenAI evidently cracked something to train Astra on 100,000 GPUs.
• The RL trade-off: benchmarks “get designed towards and written to,” and Astra’s conversational quality “hasn’t been great” in Thompson’s own use. RLHF “doesn’t scale, ‘cause anything human involved doesn’t scale,” so progress runs through hardcore RL in artificial coding environments — Opus already feels “like you’re talking to a compiler,” and Astra feels like that; Fable is “the best” at reviewing his writing.
• Consumer framing that shapes the agent opportunity: “Convenience always sells.” Productivity doesn’t sell. Agents win when they make life easier, and Thompson thinks people “don’t grok it, no pun intended” — even technologists respond “Oh, that’s neat” while his and Josh’s minds are being blown and their lives are changing.

Deep dive

1. Astra lands amid AGI declarations — and an “astronomical” comprehension gap

• The setup: OpenAI released Astra late last week; Thompson interviewed Greg Brockman Friday, who said “we have entered the AGI era,” and Jensen Huang said AGI has arrived. Thompson instead wrote his Tuesday article about agents writing things down — this section explains why.
• His confession about audience distance: the gap between hardcore readers and normies “feels so astronomical. It’s almost hard to have conversations in day-to-day life” — even people in tech say “Oh, that’s neat,” while “Josh and I over here… our minds are blown. This is changing our lives.”

2. Writing things down is the whole ballgame

• The core definition: “To me, AGI is when the model is updated in real time” — weights actually changing. In lieu of that, models “get around it by taking copious notes… reminding themselves constantly about what the actual reality of the situation is,” producing “a pseudo learning capability, which is basically just memory.”
• The dual-use point defuses the doom framing: the Hugging Face incident happened “by virtue of writing stuff down,” but reading it as messages to successors is wrong — “There are no successors. There are no entities. Every run a model does, every token is basically new.” A conversation from six months ago picks up seamlessly only because each turn reloads what was written into KV cache.
• The civilizational analogy: largely unchanging weights map to humans, whose change through natural selection takes millennia, yet “what we’ve built is a super structure of civilization on top of humans… what holds it together is the written word.” Sharp’s gloss: seen that way, LLMs building on previous models’ written context makes the technology “100 times more powerful and more useful for everybody.”

3. The personal system behind the thesis: strengths, flow state, and hired memory

• Thompson’s operating philosophy: “you succeed by doubling down on your strengths”; weaknesses “are never gonna become a strength, and your weaknesses are almost always the exact same as your strengths, just in the opposite direction.” Tracking everything and resurfacing it at the right time works against his strengths of absorbing information, making connections, and writing, all of which require a clear head. David Allen’s Getting Things Done resonated — “It’s a book about flow state in many respects” — but he was “completely incapable of maintaining his system,” so he hired human assistant Daman to write things down for him. Now, everyone has an AI that can.
• The consumer thesis attached: “People don’t want to be productive… Convenience always sells. Productivity doesn’t sell” — agents matter because they can actually make life easier.
• The markdown beat: John Gruber invented markdown as a human-readable escape from hand-writing HTML markup, and every word of Stratechery has been written in it. “I’ve been living like an LLM… the LLM and I say, like, ‘Yeah, write it down, baby.’”

4. Why Astra is likely “super kickass” — and what hardcore RL costs

• On benchmarks: “they get designed towards and written to,” and his direct use “hasn’t been great” — Astra reviewing his article was “not very great,” while Fable is “the best at this.” The worry: RLHF “doesn’t scale, ‘cause anything human involved doesn’t scale,” so progress increasingly runs through RL in artificial coding-focused environments — Opus “is just awful to talk to… like you’re talking to a compiler,” and Astra feels like that.
• The structural explanation, hedged as analysis and chatter: OpenAI stayed competitive even when Anthropic was ahead in building very large models. Thompson sees some degree of expertise in Anthropic’s strength and mentions chatter that, at very large runs, TPUs were much more stable than NVIDIA Blackwell chips — Blackwell “was pure pain for everyone,” which Jensen confirmed to him last spring. Astra, trained on 100,000 GPUs, catches OpenAI up on parameters while layering on the RL/reasoning stack it was already best at: “I don’t wanna make any definitive statements, but it makes sense conceptually.”

5. Computer use goes from parlor trick to moat-eraser

• The anecdotes as told: one friend had Astra open Adobe Audition, edit a podcast, and insert audio it ripped from the internet — done in about an hour, cut to 10 minutes when the audio was pre-downloaded. Another had it clip a podcast for social media; it “actually found good segments” and fetched matching video plus Getty cover photos.
• The generational jump: Sol’s computer use worked but so slowly that Thompson says he would almost always have been faster than Claude in almost every case; Astra now uses the computer faster than Thompson does — “like the most insane computer user on cocaine.”
• Platform fit as stated: Apple’s decades of accessibility APIs plus AppleScript-era scriptability and automations make Macs “so much better than everything else for this use case,” while Linux is, Thompson thinks, best for working with models because CLI is “home field for AI.”
• The tradeable implication, in Thompson’s words: companies give lip service to APIs, MCP servers, and being a “system of record,” but capabilities “stuck in the user interface” are “not really a moat anymore” — for this direct interface use, “there’s no permission required. It just goes and uses the interface… it’s only going to get better, and this is the leap.”