Why I'm still bearish on LLMs after Navier-Stokes
- ID
- 25092
- Status
- summarized
- Published
- 16 Sep 2026, 1:37 AM
- Fetched
- 18 Sep 2026, 1:55 AM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://dank.systems/posts/2026-09-15-ai-bear.html
- Source URL
- https://hnrss.org/best
Summary
- Score
- 7.0
- Created
- 18 Sep 2026, 1:58 AM
- Tags
- Audience
- developersai-ml-learnersai-agent-userssaas-startup-founders
What happened
Jay Kruer argues frontier LLMs remain far from meaningful autonomy despite headline feats like Navier-Stokes, because models fail on small perturbations of trained tasks and require expensive rigorous specification by domain experts to avoid reward hacking. He draws on hardware engineering as an analogy, noting CPU projects often employ 3:1 specification-to-design engineer ratios, and warns that specification costs can exceed direct implementation costs.
Why it matters
If you are building agentic workflows or pricing AI-driven automation into your roadmap, this essay argues you should budget for heavy human specification and oversight labor rather than assuming drop-in autonomy—especially for tasks outside well-defined, stable domains like pure math.
Discussion angle
Where is the line between tasks where rigorous specification upfront pays off versus tasks where informal implementation by humans is simply cheaper—especially for Malaysian startups building internal AI tools?