Astro Teller

Who he is

Affiliation: Captain of Moonshots and CEO, X (Alphabet's moonshot factory) since 2010 — the longest continuous tenure of anyone on the Moonshots LIVE bill. Born Eric “Astro” Teller, 1970, Cambridge, England; grandson of nuclear physicist Edward Teller. BS Computer Science and MS Symbolic Computation, Stanford; PhD in Artificial Intelligence, Carnegie Mellon, 1998 (advisor Manuela Veloso; Hertz Fellow). Former Stanford faculty. Co-founder and chairman of BodyMedia (wearables) and co-founder of Cerebellum Capital. Novelist — Exegesis (1997), Among These Savage Thoughts (2006) — and co-author of Sacred Cows (2014) with Danielle Teller.
One-line position: Radical innovation is not a talent problem but a process problem — build an institution that makes killing your own best ideas cheap, fast and socially rewarded, and breakthroughs become a throughput question rather than a matter of genius.

Discipline & technical bet

A credentialed AI researcher who stopped doing research to build the machine that does it. The wager is portfolio-and-process over insight: run 100+ candidate moonshots a year, spend almost nothing per candidate, attack each one at its weakest point first, and accept a ~2% hit rate as the design target rather than a failure. His actual product, in his own framing, is “producing new Alphabet entities” — the output is companies, not technologies. He is the only person on the Moonshots bill who runs an idea-to-kill pipeline at institutional scale and publishes its metrics.

Key claims (Says)

• Definitional: a moonshot needs three things — a huge real-world problem, a product or service that would eliminate it, and a breakthrough technology that offers hope of getting there. Miss any one and it isn't a moonshot.
• Process: “If it sounds reasonable, we're not interested” — reasonableness is a disqualifier by definition.
• Process: attack the hardest part first and try to kill the idea cheaply. “I want you to get information about whether this is really a once-in-a-generation opportunity or not, and it's okay if the answer is no.” Ideas that test more crazy get terminated; ideas that test less crazy get incremental funding.
• Empirical (self-reported): ~2% hit rate; 100+ projects started annually; 44% of spending goes to the graduates that become “outrageously good”; 5–6 years to exit for successful projects.
• Cultural: breakthroughs need equal parts audacity and humility — and creativity is relearnable in an environment where failure is genuinely safe. The claim is that most organizations have trained it out, not that they lack it.
• On AI: AI is “a component technology” — it makes systems smarter rather than replacing them. Safety comes from intelligence preventing failure rather than surviving it.
• Contrarian, on robotics: “personal robotics” is not a moonshot because it is not a problem — it's a form factor. Dishwashers and thermostats are already robots. This puts him against the humanoid wave that most of the 2026 bill is long on.
• On Glass: a learning platform mistaken for a consumer product — he uses his own most public failure as the cautionary case, which is rarer than it sounds.

Notable predictions — with falsifiable checks

• (ongoing) The moonshot-factory model is repeatable. The whole claim. Check: name X graduates after Waymo with independent enterprise value above $1B. Waymo, Wing and Verily all originate in the lab's first six years. If the next decade produced no comparable graduate, the model reduces to one extraordinary draw plus a well-told process story. This is the sharpest question you can put to him and it is entirely fair.
• (2016–) Ambient AI in ordinary devices — his coffeemaker illustration: within ~20 years appliances learn preferences rather than take button input. Directionally held and now unremarkable; the interesting part is that he called it a component-technology story rather than a robot story, and that framing aged better than the humanoid one.
• (ongoing) Anti-humanoid position. Falsifiable and currently under maximum pressure. Check: by 2028-12-31, is there a general-purpose humanoid deployed commercially at >10,000 units doing varied unstructured work? If yes, he was wrong in the most public way available. If no, he was the only optimist in the room with the discipline to say so.
• (ongoing) Near-zero delivery cost would sharply increase resource sharing and reduce waste. Untested — Wing's economics are the natural test bed and the unit economics are not public. Check: any published per-delivery cost curve from a drone-logistics operator at scale.
• Structural claim to watch: that the failure database prevents re-litigating dead ideas. Check: has X ever revived a killed project on changed circumstances and shipped it? He claims the database exists for exactly that; a named example would convert a nice idea into evidence.

Revealed behavior (Does)

• Kills things and says so. The most consequential fact about him is a track record of public shutdowns: Makani (airborne wind, closed 2020) and Loon (stratospheric balloons, closed 2021) were both flagship, years-deep, well-loved projects with real technical achievements — Loon had flown multi-month missions and provided live emergency connectivity. He shut them anyway, on unit economics. Very few people who preach the kill discipline have actually killed their own favourites.
• Maintains an institutional memory of failure — a database of 100+ rejected concepts, kept so ideas aren't re-argued from scratch and can be revisited when conditions change. This is _dropped.html as corporate infrastructure.
• Spends asymmetrically. 44% of budget on the winners means the other 56% buys information about what not to build. He treats the losses as the product's cost of goods, not as waste.
• Stayed sixteen years in the same seat in an industry that rewards moving. Whatever else is true, he is not optimizing for personal narrative velocity.
• Teaches the method publicly — Solve for X, the masterclass format, the podcast circuit. The process is deliberately not a trade secret, which is itself a claim: he believes the constraint is institutional will, not knowledge.

Feels

Comfortable with loss in a way that is genuinely unusual and appears to be trained rather than temperamental — the emotional work of his job is making other people's grief over a killed project survivable, and he has built ritual around it. Reads as playful rather than evangelical; the register is closer to an engineer enjoying a puzzle than a founder selling a future. Notably free of the urgency affect that runs through the rest of the Moonshots bill.

Hears

Two audiences that pull opposite ways: Alphabet capital allocators who want graduates and returns, and a public that wants wonder. He has kept both by making the failures part of the story rather than hiding them. What he likely hears least is the outside view — X is inside a company that can absorb a decade of losses, so almost no one in his hearing has to ask whether the model works without a balance sheet like Alphabet's.

Sees

Sees process where others see genius, and sees form factors as a category error. His characteristic move on any proposal is to find the load-bearing assumption and test that first — the same instinct as §13 first principles, operationalized as a budget line. The blind spot is scale-dependence: the method he sees as general may be specific to an environment with unlimited patient capital and no requirement that any individual bet return.

Incentive map

Unusually clean by the standards of this bench — and that is worth stating plainly, because most personas here need a heavy discount. He sells no course, runs no fund, has no book to move, and takes no carry in the projects. He is a salaried executive of Alphabet.

But: his job exists only if moonshots are believed to pay, and X has faced recurring internal scrutiny about whether it earns its budget. Cannot easily say: that the model needs Alphabet-scale capital to work at all; that the graduate record is front-loaded; that a 2% hit rate is indistinguishable from luck at his sample size. The self-reported metrics are also exactly that — self-reported, unaudited, and defined by the person they flatter. Moderate discount. Grade him on named graduates and dated shutdowns, both of which are public.

Theories aligned with

• Portfolio theory applied to R&D — many small bets, asymmetric payoff, losses as cost of information
• Fail-fast / lean experimentation, at capital scale rather than startup scale
• Techno-optimism — but the engineering-discipline variant, not the abundance-rhetoric variant. He is the least evangelical optimist on the bill.
• Anti-humanoid / task-specific robotics — a live minority position
• Institutional memory as infrastructure — the failure database as a first-class asset

What he's reacting against

• Innovation as a talent search — the belief that breakthroughs come from hiring geniuses rather than building a process
• Reasonable proposals — explicitly disqualifying; incrementalism dressed as ambition
• Form-factor thinking — “personal robotics,” “the metaverse,” any category defined by shape rather than by a problem
• Sunk-cost loyalty to beloved projects — the thing he is institutionally built to defeat
• Failure shame — he treats the emotional cost of killing work as the actual bottleneck, which is a more interesting claim than it first appears

Where he overlaps / splits (with Rich)

• The strongest overlap on the entire Moonshots bill — and it is operational, not philosophical. Your Forge runs a promote/kill sweep every 7th run with _decisions as the append-only system of record; your Ideas folder has graduation thresholds (≥6 principles, passive ≥3, authenticity, JTBD); your _dropped.html is a managed reject list. Teller runs precisely this system with a $100M+ budget and a decade of data. Nobody else you could talk to on Sept 25 has that.
• The questions worth asking him are yours already: what fraction of kills get revived and ship? What is the honest false-negative rate — how often did the cheap early test kill something that was actually good? How do you keep a kill sweep from becoming a formality once the team knows the ritual?
• A real split worth sitting with — §22. Teller says try to kill your best ideas, hardest part first. You hold that hypotheses are where the fun is, and that the moment “I wonder if” becomes “I must prove,” the fun dies. These are not compatible instincts, and the tension is load-bearing (§14). His method is right for a portfolio with a budget and a board; yours is right for a one-person bench where the scarce resource is appetite, not capital. Killing your best idea on Tuesday is cheap for Alphabet and expensive for you. Worth naming rather than resolving.
• Overlaps on §5 systems over tasks, §10 judgment over execution, §13 first principles — the “attack the hardest assumption first” rule is §13 with a budget attached.
• Splits with Diamandis on register — same optimism, no urgency rhetoric, no timeline claims, no product to sell you. Useful as a control on the rest of the bill.
• Splits with the humanoid-robotics consensus shared by much of the AI investment world including parts of Blundin's and Wood's theses.
• For the bench: he fills a genuine hole. Your 113 personas are heavy on researchers and economists and near-empty on operators who run selection systems. That is the gap step 13 measured.

Track record

Graduates: Waymo (autonomous driving, now the clearest commercial success in the category), Wing (drone delivery), Verily (life sciences), Google Brain (moved into Google in 2012 and became foundational to Alphabet's AI). Shutdowns: Makani (2020), Loon (2021), Google Glass (released, withdrawn from consumer). Tenure: sixteen years running the same lab.

The honest read: a genuinely remarkable record and a front-loaded one. Waymo, Wing, Verily and Brain all originate in X's first half. The strongest case for him is that he shut Makani and Loon anyway — a leader protecting a legend does not close the balloons. The strongest case against is that a 2% hit rate over ~15 years is a small enough sample that one Waymo carries the entire result.

Empirical vs normative

Empirical (testable, self-reported): 2% hit rate, 100+ projects/year, 44% of spend on graduates, 5–6 year exit. All unaudited and internally defined — treat as directional. Empirical (externally verifiable): the graduate and shutdown list, which is public and is the better grading instrument. Normative: that reasonable ideas are not worth doing; that failure should be socially safe; that form-factor framing is a mistake. Cleanly separated in his own speech — he does not usually dress values as data, which distinguishes him from most of this bench.

Weak spots / open questions

• Sample size. A 2% hit rate on a portfolio this small cannot be distinguished statistically from luck. He presents it as a validated design parameter.
• Capital dependence. The method may not generalize below Alphabet scale, and he has little incentive to test that. For your purposes this is the crux: does the kill discipline survive at n=1 with no balance sheet?
• Front-loaded graduates. The strongest exits are from the lab's early years. Worth asking directly rather than assuming.
• Self-defined metrics. “Graduate” and “outrageously good” are his terms, unaudited.
• False negatives are invisible. A kill-fast culture cannot easily measure what it wrongly killed — the failure database records what was rejected, not what was rejected and would have worked. He has never, as far as this pass found, published a wrong-kill.
• The humanoid position is his live exposure. He is on record against the direction most capital is currently flowing. He may be right; he is definitely testable.

Rich's take

• (your synthesis here)

Delta log

2026-09-05 — created

• New persona, built at v2 depth for Moonshots LIVE 2026-09-25, where he gives the masterclass on building moonshots.
• Verified this pass: biography, degrees and dates (Wikipedia); the three-part moonshot definition, kill process, 2% hit rate, 100+/year, 44% spend, 5–6 year exit, Waymo/Wing (TechCrunch, 2025-10-27); component-technology AI view, anti-humanoid position, failure database, Glass framing (IEEE Spectrum).
• Caveat logged: the IEEE Spectrum interview is dated — it discusses Loon and Makani as active and self-driving as “currently graduating,” placing it around 2016. Its views are used here; its project status is not. Makani (2020) and Loon (2021) shutdown dates are carried from general knowledge and were not re-verified against primary sources this pass. Verify before quoting the dates to anyone.
• Framed the two open questions that matter to Rich: post-Waymo graduate record, and whether the kill discipline survives without a balance sheet.
• Named the §22 tension explicitly rather than smoothing it — kill-your-best-idea vs protect-the-wonder is a genuine conflict, not a misunderstanding.
• Confidence: medium. Raise after checking the post-2018 graduate list against a primary source.

Sources

• Wikipedia — Astro Teller (biography, degrees, companies, books)
• TechCrunch, 2025-10-27 — “CEO of Alphabet's X, Astro Teller, on what makes a moonshot” (definition, process, all quantitative claims)
• IEEE Spectrum — Astro Teller on AI, robots and coffeemakers (views; project status is out of date)
• Moonshots LIVE 2026 — masterclass session
• Not yet consulted: X's own site, the Moonshot Podcast, Alphabet financial disclosures on Other Bets. Next pass.