Dark Labs and Physical AI
- huangpf
- Mar 29
- 3 min read
While software engineers debate the next breakthrough in large language models, Lars Henrik is automating chemistry labs in Sweden. His dark labs - laboratories where robots design and run experiments with no human oversight - point to a different kind of AI revolution. One where the constraint isn't intelligence, but infrastructure.
The Lab That Doesn't Need Lights
A dark lab gets its name from a simple fact: when no humans need to work inside, you can turn off the lights. Robots don't need to see in the visible spectrum to pipette samples or run mass spectrometry.
Lars walked us through his setup during Friday's session at SQ Collective. Robotic arms move samples between stations. Automated systems prepare solutions, run reactions, and analyse results. The entire workflow happens without human intervention - hence "dark."
This isn't about building smarter AI. It's about removing humans from the bottleneck entirely.
The 1940s Problem
Scientific research still runs on workflows designed in the mid-20th century. Water analysis methods Lars encountered in his previous startup date back to standards written in the 1940s and 1960s. The procedures assume a human chemist will manually prepare each sample, run each test, record each result.
"Something broke along the way in the industrialisation," Lars said. "Factories scaled. Labs didn't."
While automotive plants and electronics manufacturers automated decades ago, most laboratories still operate like artisan workshops. A skilled chemist might run dozens of experiments per week. A dark lab can run thousands per day.
The gap isn't in the sophistication of the work - it's in the parallelisation.
Physical AI Hits Different Bottlenecks
Physical AI faces constraints that software AI doesn't. You can spin up another GPU cluster in the cloud, but you can't instantly materialise another robotic arm or mass spectrometer.
Lars's background spans a water analysis startup and a PhD in scientific instrumentation. He's seen firsthand how physical systems create different scaling challenges than digital ones.
In combinatorial chemistry, researchers need to test thousands of molecular combinations to find promising candidates. Each combination requires precise measurements, controlled reactions, and detailed analysis. The intelligence needed to design these experiments isn't the limiting factor - current AI can already generate reasonable experimental protocols.
The bottleneck is execution. How do you physically prepare 10,000 samples? How do you run 10,000 reactions in parallel? How do you analyse 10,000 results without human error creeping in?
Dark labs solve this through infrastructure, not intelligence.
Precision at Scale
The breakthrough Lars described isn't about smarter robots. It's about precise robots working in parallel.
A human chemist can prepare solutions accurately, but preparation time scales linearly with the number of samples. A robotic system can prepare hundreds of solutions simultaneously with sub-microlitre precision. The robot doesn't get tired. It doesn't make transcription errors. It doesn't accidentally contaminate samples.
This precision compounds. When you're running thousands of experiments, small errors in sample preparation or measurement can invalidate entire datasets. Humans are remarkably good at one-off experiments but terrible at maintaining consistency across thousands of repetitions.
The AI component handles experiment design and result interpretation. The robotic infrastructure handles execution at scale. Neither alone would create the step-change in research throughput that dark labs enable.
The Infrastructure Revolution
Lars's vision extends beyond chemistry labs. Any research field that relies on repetitive experimental procedures could benefit from dark lab approaches.
Materials science. Drug discovery. Synthetic biology. Environmental monitoring. These fields all involve testing large numbers of variables across many conditions. They all currently rely on human researchers to physically execute experiments.
The pattern Lars identified - that laboratories never industrialised the way factories did - suggests a massive opportunity. Not for better AI, but for better infrastructure that lets AI-designed experiments run without human bottlenecks.
This isn't about replacing scientists. It's about letting scientists focus on the creative parts - hypothesis generation, experimental design, result interpretation - while robots handle the repetitive execution.
The most interesting research questions often require testing hundreds or thousands of conditions. Dark labs make those questions answerable.
Related
The room was packed with builders from across the AI spectrum - computer vision engineers, NLP researchers, robotics founders. The questions came fast: How do you handle contamination between experiments? What happens when a robot breaks mid-experiment? How do you validate results without human oversight?
Lars fielded each question with the practical experience of someone who's actually built these systems. The energy shifted from theoretical curiosity to concrete implementation challenges. Several builders started sketching robotic workflows on napkins.
Missed out last week? Don't worry, these conversations happen every Friday at SQ Collective.
Usually over laptops. Sometimes over pizza.
Comments