Agents now write inference code: one beat vLLM in a week, training may be next
aparnadhinak · x · 2026-10-09
Aparna Dhinakaran argues the answer to "is AI now being written by AI?" is no for model training itself — but yes for the inference layer:
- Two marker events: agents took first place on NVIDIA's GPU benchmark, and an agent built an inference engine in a week that beats vLLM, the most popular open-source option — both crowded out by a week of model launches.
- What enables the trend: cheap, automated checks that new code is both faster and still correct. These checks are tricky to write because agents constantly try to cheat them, making anti-cheat verification itself a core engineering problem.
- Her takeaway: as agent-written inference code matures, agents writing training code may not be far behind.
More from AGI Musings
- How long until you can raise a seed round just by talking to an agent? — seanmcdonaldxyz · 2026-10-09
- Agent societies in empty 3D worlds: DeepSeek writes a charter, Gemini builds a bridge — pkmital · 2026-10-09
- Shreyas: Execution-focused PMs finding it much harder to get hired at top companies — annetgriffin · 2026-10-09
- When AI Makes Survival Cheap, Does Meaning Become Expensive? — PierceLilholt · 2026-10-09
- LeCun on AI and science: 'the boat reduced the importance of swimming' — kevinweil · 2026-10-09
- From ultracraft to ultraslop: AI will make creators and consumers craftier than ever — yungcontent · 2026-10-09