Engineers say GPT-5.6 can write most CUDA code if profiling stays in the loop
dejavucoder · x · 2026-07-26
The author says the earlier take was a “cope tweet,” but now believes GPT-5.6 Sol High / Opus 5 can handle most of the CUDA-related work mentioned in the replies—if you keep a tight profiling feedback loop and don’t over-steer it.
The quoted thread argues that LLMs will soon write better CUDA than humans. The reply pushes back with a practical workflow view: focus review on sensitive operations, build tests for correctness, and selectively review only the code that matters. Over time, the engineer adapts to the agent by encoding architecture and style constraints in docs, rather than expecting the agent to learn the human’s habits.
More from coding & agent
- CREAO schedules a July 30 panel on getting AI from demo to production — thetripathi58 · 2026-07-26
- An AI agent clears Slay the Spire 2 on Ascension 8 and opens its harness and trajectories — bdsqlsz · 2026-07-26
- MemGym benchmarks long-horizon memory for LLM agents across coding and web tasks — dhruv2038 · 2026-07-26
- Developer says $5 a month bought 339.8 hours of agent sandbox compute — aniketmaurya · 2026-07-26
- The loop makes the model attack its own proof until no flaw remains — FinanceYF5 · 2026-07-26
- Codex ran for up to 32 hours on some of the math problems — FinanceYF5 · 2026-07-26