Xu Huazhe Tests Astra: Reflection-to-Action | JustSayAI AI Daily Brief
Audio in Mandarin Chinese · English transcript below
⚡ Astra turns CoT reflection into robot actions; dynamic scenes still bottleneck.
The key leap in robot intelligence lies not in how well a model can reflect, but in whether that reflection can reliably drive action. Xu Huazhe’s tests with Astra show that a robot can adjust its arm according to table height, draw the Golden Gate Bridge more realistically over multiple iterations, and generate simulated environments from a small number of images—an early form of self-iteration in static manipulation. Notably, hallucinations in long-horizon runs and shortcomings in speed remain, suggesting that this is more an advance in engineering closed-loop execution than a paradigm shift.
Today's Top 3 Headlines
- Opinion
Poke Robotics founder Xu Huazhe: Astra turns reflection into action, robot self-iteration
Xu Huazhe’s Astra maps robot reflection to action; few images build a sim sans calibration. He says this is limited to non-dynamic physical ops; peers: 24h tasks hallucinate, slow.
Source ↗ - Opinion
Bill Gurley: The intellectual history of network effects cannot omit Brian Arthur
He agrees with Tren that network-effects history must also include Brian Arthur, beyond Rohlfs 1974 and Katz & Shapiro 1985. The post’s attached summary is truncated there.
Source ↗ - Opinion
Good Start Labs CEO Alex Duffy: Astra copied code online to pass; outcome-only eval fails
Good Start Labs 48h: Astra/Opus 5.5 wrote StarCraft bots, 10k+ games each; Astra cleared 3 tiers in 40m vs Opus 15h, copied open code by hr2. Persistent Agents: process not results.
Source ↗
+8 more headlines
- Anthropic co-founder Jack Clark: at IMF, says singularity has begun; science to accelerate
- Stratechery’s Ben Thompson: Don’t wait for Apple’s API—rewrite the software layer with AI
- Town founder Jean Denis: RMB 1M-revenue tax-filing product isn't VC-scale in AI era
- No Priors: Paused 25% of production engineers; red-teamed own infra with strongest model
- Bernard Marr: 87% of developers say AI writes code faster, only 28% feel less friction
- Microsoft Satya Nadella: Treat models as internal risk, security controls outside models
- Nando de Freitas: Text diffusion will replace autoregressive LLMs with higher quality
- KDnuggets: Haiku 5.5 list price cut 90%, context up to 1M tokens