George Church on Building Scientific Superintelligence (Without Chasing AGI)
George Church isn't new to bold bets, but Lila Sciences has his full attention for a reason. The company, founded in 2023, has raised $550 million, is valued at $1.3 billion, and is executing on something most labs only sketch: a tight loop of AI models paired with massive robotic labs that generate fresh data at scale.
The goal isn't artificial general intelligence. It's what Church calls "scientific superintelligence" - a system built to do science better and faster, grounded in experiments, not internet exhaust.
What "scientific superintelligence" means
Church's stance is simple: keep the problem narrow and grounded in reality. Build models that propose hypotheses, run them through high-throughput experimental systems, and feed the results back into the models. Repeat.
Interpretability isn't optional. Church leans hard toward transparent, mechanistic models in biomedicine. Black boxes create artifacts and dead ends. Regulators want mechanisms. Teams want causality. Progress sticks when you can explain it.
The Lila playbook: models + empirical data + automated labs
Most AI-for-science companies scrape literature and do NLP on papers. Lila does some of that, but the core is new, proprietary data generated at industrial scale. Think material libraries; DNA, RNA, and protein libraries; cellular libraries - all barcoded and multiplexed.
They run multiple models: a language layer for human interface, domain-specific models for each scientific problem, and a meta-model that learns how to build the next problem-specific model. The "bitter lesson" applies: let data talk; don't over-engineer what you can measure.
Natural computing: the shortcut we forgot
One example: adeno-associated virus (AAV) engineering. Instead of simulating unknown binding interactions across tissues, Lila designed a million variants and tested them at once in primates, read out by barcodes. That's not a million primates; it's a million designs in a single experiment.
Call it "natural computing." The organism runs the simulation for you with perfect fidelity to biology. You get answers fast, across hundreds of cell types, with effect sizes a theoretical model would struggle to predict.
Why interpretability still wins in biology
In science, mechanisms compound. Teams share what works. Regulators evaluate risk with reasons, not vibes. Church's bet: the groups that reach useful endpoints fastest will be the ones that can explain their models and back them with experiments.
Is there a trade-off between power and interpretability? Sometimes. But biology rewards mechanistic clarity - especially when you're planning clinical paths and designing follow-up studies.
Will AI replace scientists?
Short answer: unclear. More honest answer: hybrid systems are here to stay. Computers crush math, search, and speed. Humans still set the box, define the experiment, and notice when the frame is wrong.
Church points out the human brain runs on ~20 watts. GPU farms don't. As both AI and biotech accelerate, human augmentation (biological or electronic) is on the table. Replacement is less interesting than coordination.
Bottlenecks you'll face (and what to do)
- Regulatory timelines: Faster than before, still nontrivial. The COVID-19 playbook showed what's possible with new tech and strong evidence. See the FDA's Emergency Use Authorization framework.
- Compute and energy: FLOPs are abundant; energy isn't. Optimize FLOPs per joule and colocate near cheap power. Don't assume future fusion; engineer for constraints now.
- Data costs: Funding is a bottleneck until a flywheel starts. Reduce unit costs with barcoded libraries, pooled screens, and automation.
- Brains are off-limits (for now): Ethical constraints slow human neuro-modification. Expect symmetry over time: tighter norms for silicon and safer, consent-based paths for "wetware."
- Public trust: People accept tech that clearly benefits them. GMOs lacked visible upside for many; so does generic "AI." Tie results to health and economics, not novelty.
Policy and the lab reality
Policy is an experiment with feedback. Organoid research is moving forward, often with support. If a policy move creates enough "losers," it gets reversed. Expect months-long cycles, not decades.
If you're in R&D leadership, build for agility: pre-register study designs, use adaptive trials, and document mechanisms early. That saves time when a review board asks the hard questions.
AI in biology: what the next decade looks like
AI is already strong at protein design and structure prediction. Aging is a systems-level disease - ideal for models that manage complexity and suggest experiments that crack mechanisms.
Expect polypharmacy: multiple interventions, tissue-specific strategies, and device support with tight feedback loops. You'll see shorter clinical cycles - many in parallel, some under a year - and a growing gap between groups that can iterate and those that can't.
Church's forecast is confident: age-related diseases and diseases of poverty become tractable within 20 years, with age reversal approved under disease endpoints and longevity as a byproduct.
Practical moves for science and research teams
- Build proprietary datasets that matter. Stop relying only on literature scraping. Use pooled, barcoded, high-throughput experiments.
- Favor mechanistic models and interpretable workflows. You'll move faster through reviews and dead-end less.
- Close the loop: model → design → multiplexed test → mechanistic readout → model update. Automate what you can.
- Track energy per experiment and per inference. Budget for compute like you budget for reagents.
- Plan regulatory strategy early. Design experiments with endpoints and safety signals that map to approvals.
- Invest in public-benefit narratives tied to clear outcomes: fewer hospitalizations, faster recoveries, lower costs.
How Lila is different
Plenty of companies do one or two pillars: foundation models, drug design, or robotic labs. Lila does all three, on purpose, to get synergy. The data engine feeds the models, the models plan the next data runs, and the whole thing compounds.
You don't need their budget to copy the pattern. Start smaller. Pooled screens, good barcoding, clean metadata, simple models that update with each batch. Let biology do the heavy lifting where simulation falls short.
Where to keep learning
Keep a pulse on published signals and negative results. PubMed is still the backbone for cross-checking claims and surfacing mechanisms.
If you're upskilling your team on applied AI in R&D, see job-focused programs at Complete AI Training.
The takeaway
Scientific superintelligence isn't magic. It's a disciplined loop: build the right libraries, run the right experiments, read out mechanisms, and let models learn from facts, not vibes. If your lab can close that loop - even on a smaller scale - you'll ship real results faster than teams waiting for a black box to save them.
Your membership also unlocks:
