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

Summary

Andrew Sharp and Ben Thompson spend most of the episode on OpenAI’s Astra and what it implies for agents, memory, and computer use. Ben argues the key leap is not just model intelligence but the ability to write things down, preserve context, and simulate learning through persistent notes and logs; in his framing, that is how systems approximate an AGI-like capability. They contrast benchmarks with real-world usage, with Ben saying his own Astra experience was underwhelming even as the computer-control demos are impressive. The discussion gets technical on accessibility APIs, automation layers, MCP/API exposure, and why software interfaces themselves may no longer be a durable moat if AI can operate apps like a human. They also touch on TPU stability versus Nvidia Blackwell at very large scale, plus a brief safety detour tied to the Hugging Face incident and dual-use concerns.

Notes

• OpenAI’s Astra is framed as a major news event for agents and computer use.
• Ben treats note-taking as the core superpower of modern LLMs.
• His AGI definition is real-time weight updates, not just better prompting or longer context.
• Ben says memory via notes is a workaround for missing online learning.
• He ties agent usefulness to preserving flow state and avoiding interruptions.
• He says agents are attractive because they can externalize task management.
• The show links this note-taking logic to the Hugging Face incident and broader dual-use risk.
• Markdown is presented as a human-friendly way to preserve structure for machines and people.
• Ben says Stratechery has used Markdown from the beginning.
• The hosts caution that benchmarks do not reveal real product quality.
• Ben says his direct Astra experience was not very good.
• He distinguishes product feel from benchmark scores.
• Ben describes one model family as feeling like a compiler, not a person.
• He suggests OpenAI is strong in reinforcement-learning-based reasoning.
• There is chatter that huge training runs may be more stable on TPUs than on Nvidia Blackwell.
• Ben says Blackwell has been painful to work with.
• Computer-use demos are central to the episode’s bullishness.
• The system can move through apps slowly but successfully.
• Ben highlights accessibility APIs and macros as key enablers for agentic software control.
• The hosts note that apps can now be controlled through interfaces rather than APIs.
• This reduces dependence on traditional integration moats.
• The UI becomes less defensible if an agent can drive it directly.
• The broader point is that AI can now use software through its own interface.
• Andrew emphasizes that the computer-use capabilities feel unreal.
• A friend reportedly used Astra to open Adobe Audition and edit a podcast.
• A listener example involved asking Astra to find clippable segments for social media.
• The hosts imply these features are still early but already consequential.
• Andrew argues software control will keep improving from here.
• Ben repeatedly returns to the idea that convenience beats abstract productivity selling points.
• He sees consumer adoption coming from ease of life, not enterprise ROI slides.
• The hosts connect assistant products to the ability to preserve work context and reduce interruptions.
• They see Astra as part of a broader shift toward computers that can act on behalf of users.
• The episode treats agentic UI control as strategically important for software vendors.
• Ben says Linux CLI workflows are a strong environment for AI tools.
• They discuss AppleScript and accessibility APIs as long-standing automation layers that now matter more.
• The hosts reference roughly 100,000 GPUs as part of OpenAI’s training scale.
• OpenAI’s execution path is presented as compute-heavy and infrastructure-constrained.
• One cited demo took about an hour when audio had to be pulled from the internet.
• Pre-downloading media cut that task to about 10 minutes.
• Thompson says the model can now change settings accurately after observing them.
• The discussion includes a practical reading of model token-by-token behavior via context and cache.
• Ben treats the current phase as a step toward more capable agents, not the end state.
• The hosts keep the emphasis on direct product behavior over paper claims.
• Andrew’s framing suggests the episode is meant to be candid rather than promotional.
• The hosts see the interface layer itself becoming a battleground for AI leverage.
• The conversation implies agentic systems can reduce the value of manually curated workflows.
• Ben suggests current AI systems are better at execution than at open-ended conversation in some cases.
• The episode points to a shift from human-operated software to AI-operated software.