E256｜Agent、机器人与制药，谁先被GPT-6 Astra改写？/ Agents, robots, pharma: who does Astra rewrite first?
B · Useful, limited Three practitioner-researchers (Tsinghua robotics, Pokee.ai, University of Washington biomedicine) give first-hand use reports on GPT-6 Astra; informed and specific, but anecdotal, with several second-hand figures.
Summary based on the full transcript·Original: about 80 minutes to listen
In short
An 80-minute Mandarin episode of the Silicon Valley 101 podcast (published 20 October 2026, recorded 29 September) on how GPT-6 Astra affects agents, robotics and drug discovery. Guests: Xu Huazhe (Tsinghua assistant professor, founder of a robotics start-up), Bill Zhu (founder and CEO of Pokee.ai, formerly applied RL at Meta) and Wang Sheng (University of Washington assistant professor). Their shared view: Astra is a large step in long tasks and planning, but speed, cost, refusals and lab validation set the pace of real use.
Key numbers
- 8-24 hourslength of overnight Astra runs in a University of Washington biomedical lab
- ~30%share of a usable wet-lab experiment design that models complete, by the lab lead's estimate
- 60-70%share of tokens Pokee.ai aims to generate on local hardware with ~30B models
- 16-24xhow much slower than real time Astra runs on video in robot control, per the Tsinghua guest
As reported by the source.
Chinese·Video·opens youtube.com
What it reports
- Robotics (Xu): Astra adapts to unfamiliar robot bodies and turns reasoning into motion, but runs 16-24x slower than real time on video, has no intuitive physics, and spends similar thinking time on easy and hard tasks. Robots need an on-board model under 10B; cloud APIs only suit near-static tasks. Repeated tasks plateau (1 hour, then 45 minutes, then 45 minutes). Astra cannot fold clothes.
- Robotics industry: real-to-sim improves appearance but not physics, so he calls the simulation revival hype. A Stanford group connected Astra to a Unitree robot in about two weeks.
- Agents (Zhu): consumer agents depend on browser use because shopping and travel apps do not open APIs; Amazon now blocks Meta's Muse agent. He expects identity tied to the user's own hardware. Pokee targets ~30B models on local hardware producing 60-70% of tokens.
- Enterprise lag: people he knows in legal and healthcare are still "stuck in the Microsoft Copilot swamp" and cannot freely use Codex or Claude Code.
- Biomedicine (Wang): his lab moved internal tasks to Astra for computer use on old domain software and 8-24-hour overnight runs. Gains come from coding, literature search and long tasks, not domain knowledge. Frontier models increasingly refuse biomedical tasks, so the lab uses Kimi K3 or Muse Spark 1.3 internally.
- Drug discovery: models complete about 30% of a usable wet-lab experiment design and ignore the lab's actual equipment. Validation is the bottleneck: mouse studies take weeks, primate studies about four months.
- Second-hand figures: Meta reportedly gives large free weekly token allowances per account; a third-party estimate puts an OpenAI Navier-Stokes attempt at about $22M in retail token cost.
How they got it
A recorded panel. Guests describe their own labs' and companies' use; some figures are relayed from others and flagged as such by the speakers. The digest is based on uploader captions; the last 22 minutes were reviewed only in outline.
Limits
- Anecdotal and partly second-hand; no measured comparisons between models.
- Recorded three weeks before publication in a fast-moving period.
- Guests have commercial interests (robotics start-up, agent start-up, AI drug-discovery firm).
Why it matters for companies
Three useful operating notes: long agent runs are moving to overnight batch work with faster models used by day; frontier models' refusal policies can push regulated domains toward other models, which is a governance question; and validation steps in the physical world, not model capability, set the pace in science and robotics. The remark about enterprises stuck on a single sanctioned assistant describes a real gap between firms whose staff can use coding agents and those whose staff cannot.
Key passages in translation
Chinese original
所以我们其实内部用的时候都是用Kimi K3 或者Muse Spark 1.3
So internally we actually use Kimi K3 or Muse Spark 1.3.
Chinese original
我觉得完成度可能只有30%吧很多时候是需要人类的专家去来给它修改
I think the completion rate is maybe only 30%; much of the time a human expert has to revise it.
Chinese original
现在还纠结于怎么能够脱离Microsoft Copilot的泥潭甚至都不能自由地使用Codex和Claude Code
They are still struggling to get out of the Microsoft Copilot swamp and cannot even freely use Codex and Claude Code.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com
- Tags
- Robotics