Lab automation complexity keeps AI from accelerating science alone
Nicholas Larus-Stone argues that AI won't shortcut scientific discovery because both the models and the surrounding systems are bottlenecks — and physical-world data generation can't be bypassed by smarter reasoning alone.
Score breakdown
The conversation grounds the limits of AI in science not in vague model capability gaps but in a concrete, structural problem: the physical world generates data too slowly and requires too much specialized tacit knowledge for AI reasoning alone to bypass it.
- 01Both model limitations and system limitations are blockers to AI-accelerated scientific research, according to Larus-Stone.
- 02Lab automation — programming robots to replicate hand-run experiments — is cited as a domain where LLMs currently perform poorly.
- 03The difficulty stems from highly specialized tacit knowledge about specific robots, liquid handling tolerances, and physical constraints.
On a clip from Max Agency, a podcast about how AI agents are actually being built, Harrison Chase poses a pointed question to Nicholas Larus-Stone: is the blocker to AI-accelerated science the models themselves, or the systems built around them? Larus-Stone's answer is both. On the model side, he points to lab automation as a revealing case study. While translating a hand-run lab experiment into robot-executable code sounds like a straightforward programming task, it requires deep tacit knowledge — understanding that a specific robot has particular liquid-handling tolerances, or that the liquid class needed differs subtly from what a scientist would use manually. This kind of knowledge is rarely captured in public training data, making it a domain where LLMs currently fall short. A whole professional role — lab automation engineers — exists specifically to bridge this gap.
On the systems side, Larus-Stone argues that even substantially better models won't be able to simply "think harder" to produce scientific breakthroughs.
On the systems side, Larus-Stone argues that even substantially better models won't be able to simply "think harder" to produce scientific breakthroughs. Science still requires generating new experimental data, because humanity hasn't yet produced enough data to fully understand the underlying biology and chemistry. Robust, multi-purpose lab automation remains very hard to build, and as a result, Larus-Stone concludes that humans and agents will need to work in concert in scientific settings for quite some time — making the timeline for AI-driven scientific acceleration slower than many expect.
Key facts
- 01Both model limitations and system limitations are blockers to AI-accelerated scientific research, according to Larus-Stone.
- 02Lab automation — programming robots to replicate hand-run experiments — is cited as a domain where LLMs currently perform poorly.
- 03The difficulty stems from highly specialized tacit knowledge about specific robots, liquid handling tolerances, and physical constraints.
- 04Public training data on lab automation is scarce, limiting LLM capability in this area.
- 05A dedicated professional role, lab automation engineers, exists to handle the translation from manual experiments to automated ones.
- 06Even improved models will still need to interface with the real world to generate experimental data that doesn't yet exist.
- 07Larus-Stone expects humans and agents to work in concert in scientific settings for a long time.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 16, 2026 · 23:11 UTC. How this works →