How Autonomous Labs Will Transform Scientific Research: Ginkgo Bioworks’ Jason Kelly
In a Nutshell
Jason Kelly of Ginkgo Bioworks explains how autonomous labs, powered by AI and robotics, will revolutionize biotech by automating high-mix, low-volume experiments—running 24/7, slashing costs from overhead to reagents (90% of spend), and enabling 10x more data per dollar than manual labs. Their OpenAI collaboration optimized cell-free protein synthesis, beating benchmarks by 40% after six rounds through rapid iteration, while platforms like modular racks and cloud services centralize underutilized equipment for remote access. This shift from human-limited science to AI-driven execution promises US leadership in bio-innovation against China, accelerating breakthroughs in therapeutics, consumer biotech (e.g., GLP-1 drugs, longevity), and democratized experimentation.
These notes were generated by AI and may contain inaccuracies.
All previous tech revolutions—internet, social media—have been meaningless to biotechnology and biopharma. They improved communication slightly or handled back-office IT, but did not change fundamentals. Autonomous labs will disrupt how science and big industries like biopharma are done, unlike the last 30 years of tech.
Jason Kelly founded Ginkgo Bioworks in 2008 to make biology programmable. Bootstrapped for 4-5 years without raising capital until 2014, as biotech VCs avoided young founders straight out of grad school not making drugs. Focused full-time, survived on government grants and service business. In summer 2014, Sam Altman blogged about applying Silicon Valley model to deep tech like biotech. Kelly emailed him; joined Y Combinator after meeting in San Francisco.
Mission unchanged: make biology easier to engineer. DNA is code (A, T, C, G), like computers but moves atoms, not just information. Programming cells today is poor. Initial approach: build foundries, centralized automated labs for biotech lab work. Analogy: compiling/debugging DNA code is physical—build DNA via phosphoramidite chemistry, insert into cells, grow, test.
- Make compile/debug cheaper: more experiments faster, less expense.
- Improve design to increase odds of success. Twin activities for 15 years; AI opportunities on both.
Design is software-like (e.g., Chai Bio protein models, Arc Institute's EVO 2 genomic model). Ginkgo shifted focus last two years from design (hard science problem) to testing/execution (engineering problem: cheaper/faster lab experiments). Recent OpenAI project used AI for this.
New companies like Medra use robotic arms. Life science tools industry provides incremental tools without fundamental change. Ginkgo seeks platform shift, like planar semiconductor manufacturing at Intel's start.
Design is science (unpredictable); execution is engineering (solvable). Now resourced to pursue execution.
Cells perform Intel/Nvidia/TSMC-level atomic placement, self-repair, self-replicate—product of 4 billion years evolution. >Half of drugs today produced by biotech; we cure cancer with limited tools. Near-term value from being better than today (e.g., Chai deal with Lilly).
Built autonomous lab to remove humans from bench for efficiency. Experimental science (95% of spending) requires physical testing hypotheses. Gave reasoning model a robotic lab for cell-free protein synthesis (DNA → RNA → protein in test tube, like smallest 3D printer). Optimized Stanford benchmark (Mike Jewett's lab, August paper).
Ran 384-well plates (30,000 experiments total). After 4 rounds, beat state-of-the-art; after 6 rounds, beat by 40%. Model designs logically like scientist: hypothesize, experiment, analyze, iterate. Breakthrough from running more experiments, not superior biology modeling.
AI handles scientific method (hypothesis, testing) + robotic labs. Scientists push back on AI creativity, but advantages:
- Runs 24/7.
- 100 AI scientists pursue hypotheses (e.g., Alzheimer's), share raw daily data across all (vs. 1-2 year papers with distilled info).
- Weekly lab notebook/mini-papers for direction.
NIH indirect rates pay for manual labs (5 days/week use, duplicated equipment). Of $60-80B biopharma/$40B NIH annual spend, <5% on reagents (true cost); rest overhead/people/space. Robotic labs make reagents 90% cost → 10x data per dollar.
AI + robotics win even if dumber than scientists. Past tech revolutions (internet/social) were back-office; this changes science fundamentals, first real tech impact after 15-20 years adjacency. Leaders in biopharma R&D, universities, NIH must act—unlike communication revolutions, this is computation (processing info).
PhDs do manual liquid handling (e.g., Kelly's 5-year MIT bioengineering PhD). High mix, low volume work hard to automate (like Hadrian in manufacturing). Transportation analogy:
- Low mix/high automation: subway (fixed stops).
- High mix/low automation: car (flexible destination).
- Goal: Waymo-like (subway automation + car flexibility).
Since the industrial revolution, everything low mix and low variability has been automated, from the loom onward. AI pushes past the wall on flexibility. Every part of physical infrastructure post-industrial revolution needs reexamination through this lens as variability increases.
Lab automation exists but is like subways: repeating the same experiment. Diagnostic companies like Quest or high-throughput screening in pharma have automation, but it can't handle variability. 99% of lab work remains at the bench, just as 99% of miles traveled are in cars.
Autonomous cars have sensor suite (radar, LIDAR, cameras), software suite (perception, planning, actuation), and tie into vehicle manufacturers. They require simulating gazillion corner cases due to constantly changing physical world.
In labs, the physical environment is controlled—same room every day, nothing changes. The challenge is variability in scientist requests: using different equipment combinations. Key problems:
- Integrating a thousand longtail third-party benchtop lab equipment with poor software into one system.
- Liquid handling: Pipetting is trickier than expected—humans adjust pressure for viscous liquids like syrup based on visual feedback.
Nailing liquid handling and sending samples to equipment solves it. It's tractable.
Ginkgo has it basically working: 50 robots scaling to 100 in Boston lab. Scientists dogfood it, like early Google engineers next to Waymo wheels.
Challenges:
- Scheduling with tight timing across parallel work—algorithmically tricky.
- Equipment reliability for all-day operation vs. sporadic bench use.
Liquid handling solved via existing industry robotics; manage different handlers and liquid classes.
Scientists don't code. Visual programming like LabVIEW is hated. Recent issues: sending sealed plates to pipetting robots, causing failures. Scientists, experts at bench work, made basic mistakes—bad UI.
Solution: Submit written protocols via cloud code or Codex. Model generates code; update skills file for errors (e.g., unsealing plates). Big win for usability.
Currently, 50 scientists submit jobs to one unique robotic setup in Boston.
Labs designed for humans, but better ways exist than bipedal walking: track systems move plates with micron specificity. Humans are the limitation long-term.
Biology is microscopic discipline—humanoids won't disrupt TSMC chip etching.
Future labs smaller, not enormous like data centers. Current setups (e.g., Merck's million sq ft campus) underutilized (<20%), follow scientists, duplicate across sites.
Centralized autonomous labs like a cloud: scientists order via computers remotely. Benefits:
- Eliminate duplication.
- 70% equipment utilization.
- Tighter packing around tracks.
Sold 97 robots to Department of Energy for Genesis mission (AI for science). Denser than equivalent traditional labs; saves on lab space (main cost, not reagents).
Unit of work: Rack—reconfigurable automation cart with benchtop equipment, 6-axis robotic arm (from manufacturing), maglev track. Lego-block them; arms pick/place samples.
Business: Sell racks (service fee + software subscription per box), eventually automation-friendly reagents. Two halves: build autonomous labs (e.g., Pacific Northwest National Lab, Merck); run Boston lab as cloud.
Generates dataset for inference now, but valuable for training. Two levels:
- Edge cases from variable requests (like Waymo miles): physical breaks, liquid classes—requires real execution, not digital twins. Becomes giant skills file.
- Model turning intent into experimental plans; loop results to teach better science (Newton/Einstein level). 10x/100x discovery speed advances species.
Results feed back to update model weights—frontier models prioritizing this. Pushing knowledge frontier most important for intelligence.
From White House OSTP, run by Department of Energy (big science projects: Manhattan, Human Genome). Targets breakthroughs in bio, materials, energy at national labs to double science acceleration.
- Mine existing data with models.
- Autonomous labs generate new data per models.
Ginkgo ribbon-cut first 18 robots with DOE Secretary. Needs Congress funding for scale.
Chaired National Security Commission on Emerging Biotech (now Sen. Young). US had post-Soviet science lead via funding/universities, but China matches: more papers, biotech drugs from 5% to 40% of deals (startups sell to Merck/Pfizer for $2-10B).
China: as many smart scientists, paid less, more hands-on experimentalists → more research per dollar.
Need AI + autonomous labs to boost US intellectual horsepower. DARPA created post-Sputnik to avoid tech surprises.
OpenAI from nonprofit research to $830B valuation shows fundamental research valuable. Economy forgot this amid engineering focus; pharma exception.
Hope industries (e.g., Dow Chemical) prioritize breakthroughs over plants. Likely from researchers rethinking processes (e.g., protein design).
Gut instinct suggests breakthroughs will come from teams taking big swings. The tricky part is getting funded during biotech winter. Proving better discovery engines and candidates is hard due to asymmetric information in biotech—you can't easily tell the difference. Isomorphic, backed by Google, feels closest with deep pockets to prove it out. Bet on vertically integrated research teams taking big swings.
Funding these companies all the way through is the challenge. Unequivocally yes to more breakthroughs and fundamental research as a dividend of the AI wave. Phase one: AI becoming human-level intelligent across categories, doing today's things better, faster, cheaper. Phase two: super intelligent across specific categories one at a time—that's where breakthroughs come. We're in transition to human-level across many things, about to hit super intelligence in multiple areas within months or years, leading to breakthroughs.
Backed a team from Google's AlphaChip project using AI to design chips better than human designers. Expect pockets of super intelligence across industries. Reference to AlphaGo's move 37 as that kind of intuitive breakthrough humans wouldn't make.
Two ways to breakthroughs: super intelligence intuiting novel solutions, or greatly accelerating experimental work via combinatorial approach. Much physical world isn't simulatable, so run the machine faster, feed back real-world data to make it smarter.
If breakthroughs happen, commercialization question arises—what does venture look like? Example: room temperature semiconductor—do you go to big companies or do it yourself?
Discussion on whether venture capital shrinks as production costs drop with efficiency. Analogize to cloud transition: didn't need own data centers, explosion in creation followed, making distribution competitive. Million companies fighting to break through noise. Cost of creation drops, cost of distribution rises with more options.
Language-based foundation models vs. ACTG-native models. ARC's Evo is ACTG-native, trained on trillion bases of DNA—awesome tool for reasoning models. Reasoning models can access AlphaFold, design proteins, synthesize as reagents. Reasoning models can do experimental scientist job—a thousand in a box, no miracle needed.
AI papers on drug discovery, but has pace accelerated? Bio data is often trash. Jurassic Park inspiration for genetic engineering. Knows Colossal (woolly mammoth de-extinction, brought back direwolf, mammoth coming).
Excited for kids to design biology like programming computers. Only working app ecosystem is therapeutics (85% market), agrobiotech (10%), industrial (5% like enzymes in cold water laundry detergent). Embarrassing totality for programmable matter compilers (cells). If computers only for drug discovery, they'd seem painful.
Ginkgo focuses on making biology design easier, faster, cheaper for new functions. Therapeutics have ROI but annoying: long time from inception to money due to regulation (careful sticking things in people). China does phase one trials in 6 months vs. US 2.5 years; Australia faster, FDA will match. Still not like launching phone app.
Other attempts (animal-free meat, nutrition) haven't disrupted industries yet—would accelerate genetic engineering flywheel like computers.
Drug development more expensive year-over-year for 25 years because done manually. Scientists, rent more expensive; slight productivity gains with tools, but manual work doesn't get cheaper. Retrenched after 15 years to solve by automating labs—get people out first.
Fully automated enables crazy stuff, eventually like chips (from vacuum tubes/human-scale electronics to alien tech). First step: drop vacuum tubes, build system without people in middle.
Unpredictable, but near-term excited about GLP-1 drugs (Eli Lilly ~$1T)—consumer product, not disease treatment. Best since iPhone: regain willpower, don't think about food. Biotech today is disease industry; most life not diseased. Consumer apps: weigh less, sleep better, more muscles, feel better.
Biotech adding 2-5 years lifespan worth infinity/50T (Sequoia Capital). Healthcare aims for this; pill/shot version huge. No good regulatory pathway now—FDA oriented to disease. Mahala concept (healthy when no disease) good for industry.
First YC investor Bryan Johnson (pudgy VC then, now longevity-focused) normalized monitoring: $2M/year diagnostics, weekly tests. Oura rings pathetic (heart rate tells nothing); real is molecular. Weekly blood samples for longitudinal readouts to test interventions molecularly—aging is molecular.
Quest tests painful (10 vials); at-home early. Big near-term app alongside consumer therapeutics.
Ginkgo launched cloud lab service: experiments as cheap as $39, run experiment, get data back (no sample). Science is formalized human curiosity—everyone curious. Lab access blocks people: brutal, gatekept, nearly illegal.
If labs accessible via AI-designed experiments, millions could be scientists. 1960s: kids programming seemed insane vs. IBM mainframes. Drop costs, kids/everyone asks original questions—cool market on other side of AI for science.
Keep Sequoia Capital in your library
Save the videos and channels worth coming back to, and find them again in one place.
