No PriorsNo Priors Ep. 34 | With Ginkgo Bioworks Co-Founder and CEO Jason Kelly
CHAPTERS
- 0:00 – 2:26
Synthetic biology as “DNA is code” — and where the analogy breaks
Elad frames biology as entering a digital revolution and asks Jason to define synthetic biology. Jason explains the core “DNA is code” idea while emphasizing that cells are physical, messy, and not human-designed like computers—making engineering fundamentally different.
- •DNA’s A/T/C/G sequence is code-like, but execution happens in a physical, stochastic environment
- •Proteins are the cell’s nanoscale machinery; biology isn’t cleanly isolatable like chips
- •Synbio is about porting useful engineering concepts into biology—selectively
- •AI is the latest major tool being attempted in this long arc of synbio engineering
- 2:26 – 4:21
From deterministic software to unpredictable wet labs: the Tom Knight story
Jason tells the origin story of Tom Knight, a legendary computer architect who moved into DNA programming and opened a wet lab inside MIT’s computer science building. The anecdote illustrates why many computer scientists bounce off biology: identical steps can yield different results and may never be fully explainable.
- •Tom Knight’s shift from computer architecture to programming DNA catalyzed early synbio at MIT
- •Wet-lab “compilation” is physical (pipettes, cloning) and subject to variability
- •Biology can defy debugging instincts: repeatability and root-cause analysis aren’t guaranteed
- •The unpredictability is a core cultural barrier for software-native builders
- 4:21 – 5:23
Why engineers still choose biology—and the parallel to neural nets
Jason argues that despite frustration and opacity, biology is worth it because of its powerful substrate: self-replication and self-assembly. He draws an explicit parallel to AI systems, where we accept limited interpretability because the models are powerful.
- •Biology’s advantages: self-replication, self-assembly, unmatched physical capabilities
- •Neural nets share a biology-like opacity: understanding ‘how it works’ is often hard
- •Motivation parallels: tolerate unpredictability when payoff is large
- •Different personalities prefer predictability vs “magic,” shaping who loves bio/AI
- 5:23 – 6:37
Self-evolving neural nets and ‘biology-like’ AI systems
Elad extends the analogy: as AI systems begin to code themselves, they may become evolutionary and non-designed. The discussion highlights how evolution optimizes utility in messy ways, creating redundancy and unexpected interactions—foreshadowing similar dynamics in future AI.
- •Evolutionary systems reuse parts, create redundancy, and propagate perturbations unpredictably
- •Self-evolving neural networks could arrive soon and change how we reason about AI
- •Complexity grows when systems aren’t human-designed end-to-end
- •Anticipating emergent behavior becomes central in both bio and AI
- 6:37 – 8:24
Ginkgo’s core strategy: abstraction layers and the bio “foundry” model
Jason explains Ginkgo’s attempt to import a key computing concept—abstraction—into biology. Ginkgo splits scaled, automated lab execution (the “foundry”) from DNA programmers, enabling scale economics and repeatable high-throughput experimentation.
- •Abstraction in computing separated electrical engineering from programming; Ginkgo aims for a similar split
- •The foundry automates and scales lab work, reducing reliance on manual bench science
- •Cultural friction: scientists often resist having others run their experiments
- •Scale enables testing many genetic designs—critical for engineering outcomes and AI
- 8:24 – 9:46
Customer interface and business model: specs, timelines, licensing, royalties
Sarah asks how the abstraction boundary works in practice. Jason outlines Ginkgo’s “prop shop” approach: agree on a target cell behavior, develop it on a timeline, then license the organism/solution to the customer with payments and royalties.
- •Customers include major pharma and ag (e.g., Merck, Novo Nordisk, Bayer) plus startups
- •Work begins with a clear functional spec: “here’s what I want the cell to do”
- •Ginkgo develops the organism and licenses it; revenue via milestones/fees plus royalties
- •Direct customer access to Ginkgo infrastructure is a future vision but ‘too early’ today
- 9:46 – 12:48
How AI enters the loop: protein engineering, iterative design, and foundation models
Jason describes the practical workflow of protein engineering: pick a production host, generate many sequence variants, measure performance, and iterate—now increasingly with neural nets. He then introduces the next step: protein foundation models, including a newly announced partnership with Google.
- •Protein engineering example: improve an enzyme (activity + yield) for industrial use (e.g., detergents)
- •Iterative loop: propose ~1,000 designs, test in the lab, learn from results, repeat
- •Existing engineered strains function like reusable ‘software libraries’ for new projects
- •Foundation model ambition: build a model that broadly ‘speaks protein,’ then fine-tune for tasks
- 12:48 – 14:42
Why bio may be the first domain where AI truly outperforms humans
Jason contrasts AI competing in human-native domains like law versus in biology’s “foreign language.” Because biology is sequential code we didn’t design or understand, he expects AI to surpass human intuition faster here than in English-centric work.
- •Legal/English work is optimized for humans (training pipelines, language, conventions)
- •Biology is code-like but not human-invented; we don’t naturally read/write it
- •Foundation models + specialized fine-tuning may be especially powerful in protein/DNA domains
- •Prediction: bio could see ‘split-the-atom’ scale disruption sooner than many white-collar tasks
- 14:42 – 18:36
Markets and industry structure: pharma vs industrial bio, and the lack of platforms
Elad asks about market sizing across pharma versus industrial/ag catalysts. Jason frames Ginkgo as an AWS-like horizontal provider but notes biology’s current reality: heavy vertical integration and few shared “operating systems,” which shapes how platforms can emerge.
- •Ginkgo aims to support cell engineering across domains, like cloud compute across workloads
- •Economics differ: pharma has high R&D/clinical costs; other markets may monetize molecules faster
- •Royalties vs fees vary by product timelines and go-to-market speed
- •Bio is unusually vertically integrated; shared horizontal platforms are scarce today
- 18:36 – 19:47
Why we haven’t seen many AI-discovered drugs yet: data is the bottleneck
Sarah asks why AI-discovered drugs haven’t materialized despite years of hype. Jason argues the breakthroughs are recent and the key limiting factor is not model size but high-quality, task-relevant data—where Ginkgo’s foundry-generated datasets become an advantage.
- •Perceived ‘long time’ may be misleading; major AI capability jumps are recent
- •Many groups train on the same public datasets, limiting differentiation
- •Data—not compute—is often the real constraint in bio modeling
- •Ginkgo’s large robotic lab generates proprietary experimental data via customer projects
- 19:47 – 23:24
Building a rational pandemic defense: monitoring, rapid response, and ‘bio-radar’
Shifting to infectious disease, Sarah asks if we’re prepared for another pandemic. Jason lays out a practical program: rapid vaccine response plus persistent monitoring (e.g., wastewater sequencing from aircraft) to detect outbreaks early—modeled after cybersecurity and hurricane radar.
- •COVID showed modern healthcare doesn’t automatically prevent pandemic-scale disruption
- •Two pillars: rapid countermeasures (e.g., mRNA vaccine speed) and persistent surveillance
- •Programs: wastewater sampling from inbound flights, pathogen/variant sequencing with CDC and abroad
- •Core logic: pathogens replicate—early detection and response can ‘snuff it out’
- 23:24 – 26:31
AI biosecurity fears vs reality—and why preparedness matters regardless
Elad raises concerns about lone actors using AI to design deadly pathogens. Jason and Elad argue the near-term risk is often overstated because designing for outcomes like fatality requires hard-to-obtain data and deep expertise; nonetheless, society remains unacceptably exposed and should build robust defenses against natural and engineered threats.
- •Designing highly effective pathogens isn’t push-button; data on lethality is difficult to gather
- •The ‘lone actor’ narrative can be internally inconsistent (low expertise yet high capability)
- •Even if AI-enabled bioterror is unlikely today, nature will generate new threats again
- •Goal state: background, always-on bio defense akin to modern cybersecurity infrastructure
- 26:31 – 32:34
Ginkgo’s governance experiment: employee super-voting and platform stewardship
Elad asks about Ginkgo’s unusual corporate decisions, including employee super-voting shares. Jason explains the motivation: powerful bio platforms need thoughtful governance, and Ginkgo chose employee-controlled voting (that expires when leaving) as a human-centered alternative to founder-only or purely capital-market control.
- •Ginkgo was initially ‘unfundable’ (not a drug company; required wet labs) and grew via grants + YC
- •Platform technologies in biology have direct real-world consequences; governance is central
- •Structure: employees receive 10x-vote shares while employed; collective ownership threshold ensures control
- •Tradeoff discussion: governance vs popularity dynamics; long-tenure shareholders may weight tough decisions
- 32:34 – 37:00
Energy, architecture, and learning from evolution (not just brains)
In the closing, Sarah asks about AI’s energy costs and whether future architectures will be more biologically inspired. Jason and Elad explore alternative directions—up to and including “biological compute”—and emphasize that the deepest lesson may come from evolution’s mechanisms for creating evolvable, layered systems.
- •Compute/energy requirements for frontier AI raise economic and physical constraints
- •Neural nets were an escape from purely logical software; future architectures may get ‘crazier’
- •Elad highlights evolution’s messy but hyper-optimizing dynamics once self-replication enters the loop
- •Jason recommends focusing on evolution’s principles (evolvability, layering) and cites ‘The Plausibility of Life’
