swyx Burned 20B Tokens Stress-Testing Astra on Real AI Engineering Tasks — All for Under $6/Hour
charliermarsh · x · 2026-09-04
swyx of Latent Space says he got "sucked into extreme LLM psychosis" and declares a new age of AI Engineering. His team spent 20B+ tokens throwing OpenAI's Astra at real AI engineering tasks beyond cute demos, and found it can:
- choose and train models
- label data (and run active learning on your labels)
- keep pipelines saturated
- instrument and read logs
- deploy and debug entire systems in one shot
- fan out and command eval subagents (including agents running other models)
- keep coherence over billions of tokens in a single agent thread
All at under $6/hour. Full reports coming on Latent Space.
Related event: Latent Space Stress-Tests OpenAI Astra with 20B Tokens(2 posts)→
More from coding & agent
- Google clarifies Antigravity terms change only covers Antigravity and Gemini CLI accounts — rseroter · 2026-09-04
- Matt Shumer claims playable AI-built NYC in under 7 months via sub-agent loop — mattshumer_ · 2026-09-04
- Largest open experiment finds dev tools are always mentioned but never chosen by coding agents — ycombinator · 2026-09-04
- Enterprise AI Agent Cohort Kicks Off With Engineers From PlayStation, Salesforce, ServiceNow — hugobowne · 2026-09-04
- Mathematician tests GPT-6 Astra: live Lean proof verification while writing arguments — teortaxesTex · 2026-09-04
- Makepad flow: a Rust single-executable ComfyUI alternative built in one day — anselm · 2026-09-04