Jason Kelly | Training Data Summary

1. Key Themes

AI Is the First Technology Revolution That Actually Matters to Biotech

Every previous tech wave — the internet, social media, cloud — was essentially irrelevant to how biology and biopharma actually operate. AI is different because it touches the fundamental mechanism of how science is done: experimental design, hypothesis testing, and iteration.

"All of the previous revolutions in tech, internet, social media, whatever, have been totally meaningless to biotechnology and biopharma. Like, yeah, it's nice. We communicate slightly better or whatever. It's just some back office IT crap. Not this. This is actually going to change the fundamentals of how we do science and our big science industries like biopharma are going to get disrupted." — Jason Kelly 00:00:00

The Autonomous Lab: Replacing the Human at the Bench is the Keystone Unlock

The core bottleneck in biological science is not intelligence — it's that experimental science requires physical presence. Less than 5% of research spending goes to actual reagents; everything else is overhead (people, lab space, equipment duplication). Autonomous labs that remove humans from the bench could flip this ratio and create a 10x improvement in data per dollar.

"If you look at the spending on research... less than 5% is on the reagents. Everything is on overhead. It's basically overhead. The people, the regulatory, the lab space... If you're running it efficiently, you would budget a research program at the NIH... just on the reagents. Because that's like the usage-based pricing of science." — Jason Kelly 00:16:14

"The other advantage those AI will have is if they're able to run robotic labs, now they're running where 90% of the cost of a research project goes to the reagents. That's like a 10X increase in the amount of data per dollar that you're getting compared to how we do it today." — Jason Kelly 00:17:12

Reasoning Models Can Already Do Experimental Science — The OpenAI Result Is a Landmark

The Ginkgo-OpenAI experiment is arguably the most significant proof point yet that AI reasoning models can run autonomous experimental science loops. After six rounds of 30,000 experiments, the AI beat Stanford's state-of-the-art cell-free protein synthesis benchmark by 40%. The key insight: the model didn't need to simulate biology — it just needed to be logical, run experiments, and iterate.

"What really let it break through wasn't that it was so smart. It was that it could run experiments. And the question was just could it design them like a scientist could? The answer was, yeah, hell yeah it could." — Jason Kelly 00:12:09

"After four rounds of that, we beat state of the art. And after six rounds, we beat it by 40 percent. And so that was a — I think it's the most interesting sort of model doing experimental work result that's been shown to date by a lot." — Jason Kelly 00:11:11

2. Contrarian Perspectives

Humanoid Robots Have No Role in the Lab

While humanoids are a hot narrative in robotics and manufacturing, Jason argues they are entirely wrong for laboratory automation. The lab environment is static, not dynamic like a road. Tracks and robotic arms solve sample movement with higher reliability and precision than bipedal robots ever could.

"In the long run the humans are the limitation... biology is a microscopic discipline. These things are — yeah, makes no sense." — Jason Kelly 00:29:04

"There are much better ways to do that than like walk them bipedally among things. You just put them on a track like our system has like a nice little track and the plates move with extremely high reliability. They get delivered with micron specificity to where they are." — Jason Kelly 00:29:04

Autonomous Labs Will Make Labs Smaller, Not Bigger

Conventional wisdom (especially by analogy to data centers) would suggest that centralized autonomous labs will become massive facilities. The opposite is true: autonomous labs eliminate equipment duplication, dramatically improve utilization, and compress physical footprint significantly.

"It's actually going to make them smaller... you also get wildly better utilization of the benchtop equipment — like we're talking going from like sub-20% utilization at the bench to like 70%. So now you need less equipment, and then assuming you didn't decide to have humanoids, you can just jam it all in around a track system." — Jason Kelly 00:29:59

The Drug Development Cost Curve Is Going the Wrong Direction — And Has Been for 25 Years

Most people assume technology has been incrementally improving drug development efficiency. In fact, the cost to develop a drug has increased year over year for a quarter century. The reason is structural: manual labor dominates costs and doesn't benefit from Moore's Law-style improvement.

"It has gotten more expensive to develop drugs, not less, year over year for the last 25 years. So that's not great. That's the opposite of what should be happening. And why is that? Because we do it manually. That's my opinion... the majority of the cost is manual work. That does not get cheaper." — Jason Kelly 00:50:33

The Biggest Biotech Market Isn't Disease — It's Human Enhancement

The entire regulatory and commercial apparatus of biotech is built around treating disease. But the far larger consumer opportunity is performance, longevity, and wellness — markets where people would pay enormous sums across their entire lives, not just during illness. GLP-1 drugs are an early signal of this shift.

"How much of your life do you want to weigh 15 pounds less, how much of your life do you want to sleep better... the applications in the consumer space for biotech are bananas... It adds two years to your lifespan — what's that worth? Like what is the value of a biotech product that adds five years to lifespan? 50 trillion — it's infinity." — Jason Kelly 00:52:21

China Is Already Winning Biotech Innovation — Not Just Manufacturing

The narrative is that China is a manufacturing threat. The less-discussed reality is that China has surpassed the U.S. in scientific paper output and is rapidly closing the gap on drug discovery — the highest-value part of the pipeline. Three years ago, less than 5% of drug candidates sold to major pharma came from China; last quarter it was 40%+.

"Three years ago, less than 5% from China. Last quarter, 40% plus... They have just as many scientists as us. They're just as smart as our scientists. They get paid less. And remember, it's like hands in the lab — you got to do science. Science is driven by experimental work. So if you now have more experimentalists in China and you get more research per dollar, I don't see why they don't win in research." — Jason Kelly 00:38:39

3. Companies Identified

Ginkgo Bioworks

Pioneer in making biology programmable through automated foundries and now autonomous labs. Currently operating 50 robots (scaling to 100) in Boston with scientists submitting jobs to a centralized robotic system. Sold 97 robots to the Department of Energy for Project Genesis.

"Right now in Boston is like a very unique experience — experiment happening where we have like 50 scientists submitting jobs into one big robotic setup that exists nowhere else on the planet right now." — Jason Kelly 00:28:10

Chai Bio

Early-stage company focused on AI-driven protein design. Named as a leading example of companies solving the "design" front-end of biological engineering.

"Today, the folks leading on the design side, you might see companies like Chai Bio, for example, like it has like these protein models." — Jason Kelly 00:04:51

Arc Institute / Evo 2

Research institute that released Evo 2, a genomic model trained on a trillion bases of DNA — described as a powerful tool that can be accessed by reasoning models to do biological design work.

"The folks at Arc Institute just came out today with a paper we'll call it about Evo 2, which is like a genomic model. There's a whole community of people now trying to solve the problem of designing biology with AI." — Jason Kelly 00:04:52

Medra

New company using robotic arms to accelerate lab work — mentioned as one of the few companies with a Silicon Valley-style approach to transforming the back-end of lab science.

"There's some new companies, right? So there's companies like Medra out here as one that's doing it with like robotic arms trying to accelerate it." — Jason Kelly 00:05:44

Isomorphic Labs (Google DeepMind)

Google's AI drug discovery company — cited as the most credible vertically integrated research team with the deep pockets to actually prove out better discovery engines.

"Google with Isomorphic feels like the closest because it's actually got deep pockets behind it to actually prove that story out." — Sonya Huang 00:42:15

Colossal Biosciences

De-extinction company that has already brought back the dire wolf (with woolly mammoth in progress) — mentioned as an example of exciting, ambitious genetic engineering application.

"There's one company that brought back the woolly mammoth... I know them well. Colossal. Yeah, it's great. I love that stuff." — Jason Kelly 00:47:16

Hadrian

Advanced manufacturing company working on high-mix, low-volume automation — cited as a parallel to the lab automation challenge Ginkgo is solving.

"This is true at like places like Hadrian today, for example, that are working on this on the manufacturing side — industrial high mix, low volume is hard to automate." — Jason Kelly 00:20:14

Waymo

Cited as the definitive analogy for the transition from automation to autonomy: achieving the flexibility of a car with the reliability of a subway, which is precisely the problem autonomous labs need to solve.

"We got Waymo up here in the corner where you get the flexibility — the automation of a subway, but the flexibility of a car." — Jason Kelly 00:21:13

4. People Identified

Sam Altman, OpenAI CEO

Credited with catalyzing Ginkgo's fundraising journey by writing a blog post inviting deep tech companies to YC. Later partnered with Ginkgo on the autonomous lab experiment that beat the Stanford cell-free synthesis benchmark by 40%.

"Sam Altman, now Mr. Famous, writes this blog post because he just took over YC and he's like, hey, I think the Silicon Valley model can work for like deep tech, you know, nuclear fission, biotech, material science." — Jason Kelly 00:01:36

Brian Johnson, Entrepreneur / Longevity Pioneer

Ginkgo's first angel investor post-YC. Now a prominent longevity advocate spending $2M/year on diagnostics. Cited as having normalized the idea of continuous molecular self-monitoring, a potentially enormous market.

"Our first investor out of YC — do you know who it was? Our angel, Mr. Brian Johnson... what's interesting about what Brian's done is he has normalized the monitoring... every week he's like taking all these tests and everything else. The most measured person. Two million dollars a year of diagnostic stuff." — Jason Kelly 00:53:43

Mike Jewett, Stanford Professor

Set the benchmark for cell-free protein synthesis efficiency in a paper published in August — the benchmark that the Ginkgo-OpenAI autonomous lab subsequently beat by 40%.

"There's a paper that came out of Stanford from Mike Jewett's lab in August that set the benchmark for like how cheap people had been able to do cell free protein synthesis. And so we said, all right, let's try to optimize that with the model." — Jason Kelly 00:10:15

Constantine (Sequoia Partner)

Credited with a useful framework distinguishing revolutions in computation (processing information) vs. communication (distributing information) — explaining why previous tech waves were irrelevant to science.

"Our partner Constantine has a good framework for that. He talks about how there are revolutions in computation and revolutions in communication... what you're talking about here is just a different way to process the information." — Pat Grady 00:18:21

Chris Wright, Secretary of Energy

Personally ribbon-cut the first 18 robots installed at a DOE facility under Project Genesis — signaling meaningful government commitment to AI-accelerated science.

"Secretary the Department of Energy, Secretary Wright and I like ribbon cut the first 18 robots up in Washington and he like signed it. It was really cool." — Jason Kelly 00:37:18

5. Operating Insights

Replace Visual Programming Languages with LLM-Based Protocol Submission for Non-Technical Users

When scientists tried to program robotic systems directly (even with visual tools like LabView), they made embarrassing basic mistakes (e.g., sending sealed plates to a pipetting robot). The fix: eliminate all manual protocol coding and instead have scientists submit written instructions to a model like Claude Code or Codex, which writes the robot instructions. Edge cases get logged in a "skills file" and never repeated.

"From now on, only the way we're going to interact with writing the code is through Claude Code or Codex. You will now submit a written protocol — what you want — and the model will figure it out. And if the model sends a plate sealed, we will update the skills file and it will never do it again." — Jason Kelly 00:27:43

Centralize Equipment to Radically Improve Utilization and Cut Overhead

Most research organizations replicate the same equipment across every lab team, resulting in sub-20% utilization per piece of equipment. Centralizing into one autonomous system brings utilization to ~70% and eliminates the need for local labs wherever scientists sit. This is the physical-world equivalent of moving from on-premise servers to cloud.

"You have wildly better utilization of the benchtop equipment — like we're talking going from like sub-20% utilization at the bench to like 70%. So now you need less equipment... it's actually a lot tighter." — Jason Kelly 00:30:27

Design AI Scientist Networks to Share All Raw Experimental Data Daily, Not Just Published Results

Current scientific collaboration shares only distilled, curated conclusions every 1-2 years. The operating model for AI-driven science should be daily raw data sharing across all parallel research threads — where one AI's failed experiment is immediately available to inform another AI's hypothesis. This is a structurally different and superior information architecture.

"At the end of the day, they're going to pass the data on those experiments — like what experiment they ran and the raw data that came off it — to the other hundred AIs. Daily. Every fucking day... your failed result might, for example — say your experiment went the wrong way from your hypothesis — that data might be relevant to my hypothesis. And I would never see that normally." — Jason Kelly 00:15:00

6. Overlooked Insights

The "Usage-Based Pricing of Science" Creates an Entirely New Business Model for Research

Jason briefly mentioned selling "automation-friendly reagents" as the third revenue leg of Ginkgo's business model — after hardware (racks) and software subscriptions. This is easy to miss, but it's profound: if autonomous labs shift research spending from overhead (people, space) to reagents, then whoever supplies reagents optimized for robotic consumption owns the consumables layer of an enormous new market — analogous to ink cartridges or cloud compute. The company that defines "automation-friendly reagents" as a category could become the dominant recurring revenue player in the autonomous lab stack.

"Eventually what I want to sell is like automation-friendly reagents. Because that's kind of like the usage pricing... that's like one half of my business now is like I'll build you an autonomous lab... and then the other half is what you said, I'll run my lab in Boston as a cloud and you can just order for it." — Jason Kelly 00:32:17

Ginkgo's $39 Cloud Lab Experiments Could Democratize Science the Way AWS Democratized Computing

Buried at the very end of the conversation, Jason mentioned that Ginkgo has already launched a cloud lab service where anyone can run experiments starting at $39 — you submit the experiment, they run it, you get the data back. This was mentioned almost in passing, but the implication is enormous: it's the "AWS moment" for biological experimentation. If millions of non-scientists can now run real biological experiments the way they can spin up a server, the addressable market for scientific curiosity may be orders of magnitude larger than anyone has modeled.

"We did just launch a cloud lab service at Ginkgo where you can — like we have experiments as cheap as $39 that you can just run... we will send you back data. So it's like you do the experiment, we run the experiment for you, you got the data back." — Jason Kelly 00:55:57

"I believe if you do manage to drop the cost and all this stuff, you may have kids and everybody else wanting to just ask original scientific questions and being able to do it. And that would be a cool market... if you rewind the clock to the 1960s when it was IBM and it was mainframes and you told people that kids would program computers, they would say you're fucking insane." — Jason Kelly 00:57:19