Graham Neubig Underwhelmed by Astra for Coding, Sees GUI Computer Use as the Real Leap
May_F1_ · x · 2026-09-08
- NLP researcher Graham Neubig says he's somewhat underwhelmed by Astra for coding use cases.
- He acknowledges Astra seems very good, but notes "5.6 sol" is also very good, and neither is close to perfect — suggesting no clear generational jump in coding.
- He speculates the big step function may lie in GUI-based computer use and asks for examples of people genuinely wowed by it.
- A firsthand assessment from the creator of SWE-bench, reflecting a broader community sentiment that the new flagship model isn't a breakthrough for coding.
More from Models
- Qwen3.5 35B-A3B sampling differs between Tinker API and Alibaba Cloud — philhchen · 2026-09-08
- GPT-6 Astra scores 95% on robot control task — but critics say demos are gamed — GaryMarcus · 2026-09-08
- Power user burns through Pro limits on Astra, verdict: not AGI, just a bigger code monkey — GaryMarcus · 2026-09-08
- Epoch: Astra beats most humans at card game but shows little continual learning — teortaxesTex · 2026-09-08
- GPT-6 Astra usage limits have been reset for paid users — xiaohu · 2026-09-08
- Astra's computer use is strikingly fast, likely planning steps in batch — shekitup · 2026-09-08