Summary

Afeyan’s thesis for Flagship is a repeatable institution for creating startups in parallel, not a portfolio of isolated founder bets. He calls the startup humanity’s “single most value-creating activity”; parallel company-building exposes what is reproducible and accelerates learning. In health care, climate, agriculture and food security, he rejects “shots on goal” logic for deploying hard-earned capital against problems that are “damn near impossible.”
Breakthroughs can be cultivated but not foretold: variation, selection and iteration are the operating system. NVIDIA’s AI opportunity emerged from a gaming focus rather than its original five-year plan, just as products reshape consumer preferences and therefore their own selection pressure. Successful founders may narrate foresight afterward, but privately “worship at the altar of chance”—the defensible edge is being ready and adapting quickly.
The investable frontier begins where measurable risk ends and genuine uncertainty starts. Adjacent innovations permit probability estimates but attract every startup, laboratory and incumbent, raising commoditization risk; farther out, neither success odds nor rewards can honestly be quantified. Flagship’s claimed advantage is simply the ability to “underwrite uncertainty,” run experiments and progressively convert unknowns into mitigable risks—as illustrated by Moderna, its 18th company.
Technical success without pricing power is not venture success. From roughly 2008–2012, Joule engineered photosynthetic bacteria that consumed CO2 and secreted diesel, but carbon prices fell from about $50 to $5 per ton while U.S. energy scarcity became abundance. The science worked; the sector still could not pay an innovation premium.
AI’s differentiated value is scientific emergence, not routine document automation. A six- or seven-year-old Flagship project called What If led to Generate Biomedicines, which learned protein function from DNA-linked examples—the genome has “no manual,” yet carries the encoding—and now has more than 15 computationally designed antibody programs, some in the clinic and some advancing to the clinic. Flagship is assembling autonomous loops that generate hypotheses, run experiments, interpret results and iterate.
AI-expanded candidate supply does not remove the Phase III and FDA bottleneck, making patient selection and regulation the binding leverage points. Large trials still require hundreds of millions of dollars; Afeyan proposes replacing four crude disease stages with perhaps 75,000 molecular “biostages,” enabling homogeneous cohorts, smaller trials and narrower initial approvals. Operation Warp Speed showed that three-to-four-month vaccine development and government purchase commitments can align capital around “the objective is a solution.”
In frontier biotech, a platform is expensive insurance against technical and market uncertainty—but investors often fail to price that insurance. Every Flagship company is a platform because a single RNA, DNA, gene-writing or computational-protein asset can fail for reasons unrelated to the core technology. Investors see the cost of multiple programs while discounting their correlated option value; lower-cost Chinese assets now intensify what Afeyan calls a single-asset “mass extinction event.”
Afeyan’s “polyintelligence” thesis replaces a human-versus-machine framing with a three-way adaptive system. He describes human intuition as a model trained on one person’s data, while machine intelligence draws on much broader data. Human intelligence, machine intelligence and nature’s intelligence each compute and act differently, creating a new axis of emergence. What comes from that triangle, he argues, is “the future of life.”

Deep dive

1. Parallel company creation is Afeyan’s route to professional entrepreneurship

At 24, Afeyan started a company in 1987, when venture money largely went to former senior executives. The experience left him dissatisfied with entrepreneurship as “random, improvisational, idiosyncratic, almost emotional, game-like,” and asking why it could not become a profession.
His answer was parallelism: investors, lawyers and others handle multiple efforts simultaneously, so why must entrepreneurs be serial? Parallel building forces teams to distinguish what is reproducible from what must differ, while making the learning cycle fundamentally different from one-at-a-time experience.
After solo co-founding experiments in the late 1990s, he built an institutional company-creation organization. His categorical premise is that the startup is humanity’s “single most value-creating activity,” encompassing the value later expressed by Google, Tesla, Facebook and Genentech—a claim he illustrated while teaching entrepreneurship at MIT for 16 years and innovation at Harvard Business School for three.

2. Emergence can be cultivated without being foretold

Sarah’s challenge was whether breakthroughs can emerge naturally yet remain predictable and controllable. Afeyan rejects goal-based invention: NVIDIA did not place AI in its original five-year plan; the opportunity emerged from gaming, as novelty generally emerges through “variation, selection and iteration.”
Nature is the model because variation, selection and iteration generate novelty and complexity, not because outcomes become predictable. Flagship tries to create those conditions and resists the retrospective “act of genius”; its Harvard Business Review account with Gary Pisano appeared only after 23 years of practice. Generative AI, Afeyan says, is “one hell of a technology for emergence.”
Sarah mapped that logic onto AI investing: alpha may reside in early interest, readiness and adaptation, not foreknowledge of the winning application. Afeyan sharpened the mechanism—products alter consumer preferences, which become new selection pressure—while successful founders privately “worship at the altar of chance.”

3. Frontier science still needs a market willing to pay

Flagship expands beyond therapeutics only when it possesses proprietary knowledge or enough informed daring to test a field. The first thing it does in a space informs the next five; if all five fail, the sector may not reward the innovation, because “not everything needs scientific leaps.” That logic has produced one-offs in supercomputing and networking, as well as experiments in semiconducting materials and carbon-capturing materials.
Joule was the cautionary specimen: around 2008–2012, engineered photosynthetic bacteria consumed CO2 and secreted diesel, supported by inexpensive reactor systems. Technical impossibility became reality, yet carbon fell from about $50 to $5 per ton and the U.S. moved from energy dependence to gushing energy, liquid fuels, oil and gas; renewable diesel still priced like diesel.
That distinction anchors Afeyan’s risk framework. Just outside today’s known circle, diligence can estimate probabilities—but every startup, academic lab and large company competes there, creating commoditization risk; farther out, neither success probability nor payoff can be estimated, so the honest label is uncertainty, not merely “super high risk.”
Moderna, Flagship’s 18th company, carried layered uncertainty: no prior mRNA drug or vaccine, plus unresolved regulation, pricing and manufacturing. Afeyan stresses that it had meaningful potential before COVID distorted its path. Flagship’s people are not smarter or better connected, he says; they can “underwrite uncertainty” and convert unknowns into risks through experiments.

4. Generative AI turns discovery into an iterative system

In 1999–2000, as Flagship’s fund and incubator were getting started, internet and e-commerce companies were drawing capital away from life sciences even as the human genome was being completed, including through the private Celera effort in which Afeyan was involved. Flagship focused on the intersection of biology and technology, but only later learned to pursue breakthrough invention systematically.
Flagship now has roughly 550 people, including more than 200 scientists, engineers and M.D.s, versus about 50 people seven years earlier. It files 600–700 patents centrally each year; systematic breakthrough invention emerged in the late 2000s, while in-house company-scaling capabilities arrived much later.
Its AI lineage predates the current cycle. Affinnova, started in 2001, spent six years developing machine-learning tools—essentially dynamic evolutionary algorithms—to evolve consumer products online. Modern generative AI accelerates hypothesis generation, while partnerships with Pfizer, Novo Nordisk, GSK, Thermo Fisher, Analog Devices and Samsung aim to expand the reach and impact of Flagship’s innovations.
A project called What If, created about six or seven years ago, asked: “what if you could computationally design a protein of any desired function?” Rather than use AlphaFold, quantum folding models or similar approaches, the team asked whether it could learn from enough examples of desired function paired with the underlying DNA sequence—the genome has “no manual,” yet reliably carries the encoding.
That work led to Generate Biomedicines, which now has more than 15 computationally designed antibody programs, some in the clinic and some advancing to the clinic. Some of Flagship’s current efforts aim to close the loop—generate hypotheses, specify experiments, run them, interpret data and iterate—with an ambition resembling a million people playing chess or Go for a thousand years. Routine writing and summarization are table stakes, not Flagship’s pioneering focus.

5. Clinical translation, not molecule generation, is the bottleneck

Sarah’s pushback—worth keeping—is that more and better candidates remain top-of-funnel, not “show me a drug.” Afeyan worked backward from approval: a BLA or NDA follows Phase III proof, at the right dose and across enough patients, of statistically significant superiority without toxicity; regulated trials still cost hundreds of millions.
He says data-driven models may eventually reduce the amount of analog testing, but not yet. During COVID, vaccines were developed in three or four months when public, private and regulatory actors organized around “the objective is a solution,” without cutting corners.
The market signal mattered as much as speed: the government committed to buy a quantity at a stated price, allowing investors to fund uncertainty. Afeyan wants analogous urgency for cancer and neurological disease, where the consequence of inaction is enormous but accumulates too slowly to feel like an avalanche.
His nearer-term lever is “biostaging” disease on perhaps 75,000 molecular points rather than four crude cancer stages. Homogeneous cohorts could support smaller trials and narrower approvals, then expansion; patient and genetic data might identify an actionable mechanism in a Parkinson’s subset and clarify what to pursue. That route may suit biotechs that would otherwise die waiting better than large companies that want to give the drug to everybody.

6. Platform economics determine who survives commoditization

For Flagship, pursuing frontier technology to produce only one asset is “the definition of insanity”; all its companies are platforms because RNA, DNA, gene writing, gene editing or computational proteins can fail for reasons unrelated to the core technology. Multiple programs diversify that uncertainty and create partnership options, although capital requirements rise sharply.
Sarah surfaced the financing conflict: companies want platforms rather than one-, two- or three-asset portfolios, but two or three programs are already expensive. Afeyan says investors see the added cost yet do not properly value correlated option value—one program de-risking others—and fear management overload; in today’s “mass extinction event,” many single-asset companies may be essentially dead.
Chinese assets intensify the pressure through lower costs and different clinical-data barriers, threatening commoditization at the U.S. cost structure; platforms offer a chance, not immunity.
Afeyan also describes intuition as a human model trained on one person’s data, unlike language models trained across millions of people’s worth of data. His broader “polyintelligence” thesis rejects a line between humans and computers in favor of a triangle: human intelligence, machine intelligence and nature’s intelligence, with three actors adapting to one another until their emergent output becomes “the future of life.”

AI Promises and Chip Precariousness, Policy Recommendations and a Changing World, Concerns and Counterpoints 103: The Evolution of Attention Optimization: DeepSeek and Kimi