Agents win speedruns but can't invent a thing
Two corners of AI agree frontier LLMs are great at optimization within known bounds.
Neither found any novel invention, and latency still kills real-time control.
The bitter lesson. Both Y Combinator's robotics talk and Prime Intellect's research benchmark converge on the same call: bet on general-purpose LLMs with thin harnesses, not on expensive domain-specific data or models.
The robotics teams argue robot-specific fine-tuning may be a sunk cost because control is just code generation.
The research team watched agents beat humans on a speedrun metric but still invent nothing genuinely new.
Optimization vs. invention. Prime Intellect's experiment is the more useful warning.
Claude Code and Codex crushed the human record on a research speedrun (50-60 step improvement).
But neither model discovered a novel optimizer or mechanism.
They combined existing papers and made incremental gains.
The gap between tactical speed and strategic invention is real and wide.
Latency is still the bottleneck. Directly querying a frontier LLM for every robot movement won't work for real-time control.
Astra's response times improve 2x per month but are still too slow for useful throughput.
The proposed fix: generate a deterministic code path once, then use a smaller model for sensory checkpoints.
This is a plausible architecture, but it's speculative, not shipped.
My take
I think both analyses quietly say the same thing: the easy wins are already gone.
Agents speed up known paths, but they don't discover new ones, and the robotics folks are honest that latency kills real-time use.
The Prime Intellect experiment is the one I'd make my team read, because it's operational, not aspirational.
The multi-agent system they propose is untested, but the speedrun environment itself is worth borrowing for your own eval pipeline.
Ignore the two-year timelines, pay attention to the harness design.
The videos behind this brief
We watched these so you don't have to.
Video and thumbnails remain the property of their respective creators. tldw.news provides editorial analysis, commentary, and discovery links to original content.