Oren Etzioni to AI Startups: Iterate Fast, Scale Slowly

Oren Etzioni urges founders to distinguish between rapid experimentation and difficult-to-reverse commitments. He highlights the value of assets that speed cannot buy — data and relationships — and cites a study attributing a large portion of failures to premature expansion.
Irreversible decisions require time, not haste
In 2011, researchers conducted a study on over 3,200 high-growth tech startups and attributed about 74% of internet startup failures to overly rapid expansion. Jeff Bezos distinguishes, in his 2015 letter to Amazon shareholders, between decisions likened to one-way doors, which are nearly irreversible and must be made slowly and thoughtfully, and two-way doors, which are reversible and can be made quickly. According to this distinction, a startup's experiments fall under two-way doors, while major commitments such as large fundraising rounds, long-term contracts, or hiring frenzies are not reversible in the same way.
Defensibility comes from unique data and relationships
For labs, information such as clinical records, data from industrial sensors, or proprietary transactions, which remain difficult to obtain, represent significant advantages. Acquiring this data often depends on the gradual building of trust-based relationships and sustainable partnerships. Herbert Simon recommended choosing important problems for which one has a "secret weapon" — exclusive knowledge or methods. Translated to startups, this refers to data and relationships that a faster competitor cannot replicate. A data agreement with a hospital network or fraud labels from a payment stream are cited as examples. Etzioni illustrates this point with Farecast, founded in 2003, which predicted airfare and, by 2007, extracted over 175 billion fares; each prediction was then compared to the actual fare, enriching a labeled dataset every day.
What speed really brings: iterating, not scaling
Etzioni distinguishes between speed of iteration and speed of scaling. Rapid iteration involves inexpensive and reversible experiments, such as a prototype, a pilot, or a feature tested with 20 users, and can help achieve product-market fit sooner. Rapid scaling involves heavy commitments, such as hiring before revenue, multi-year contracts, IT resource purchases, or high-valuation fundraising. Iteration should happen as quickly as possible since each cycle is easy to cancel, while scaling should occur more slowly because each step is hard to undo. Etzioni clarifies that speed favors iteration but has little effect on acuity and judgment, which are essential for identifying the right problem.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
When speed misleads: model windows and clones
Every model release opens a gap that leading labs close in one or two cycles, limiting the advantage of moving quickly in these areas. AI makes building inexpensive, and clones can appear within weeks, weakening velocity as a defensible barrier. Early users do generate data, but each new generation of models tends to reset the value of previous fine-tuning datasets. Peter Thiel argues that it is better to be the last entrant to benefit from sustainable positions, and Y Combinator looks for companies that become more powerful with scale and are hard to copy.
Capital and talent: momentum attracts but can trap
Momentum attracts capital and talent. However, raising funds based on growth enthusiasm can trap a company in the narrative that justified its valuation. Conversely, raising funds at a more reasonable pace preserves the ability to pivot. Etzioni recommends financing the trajectory at a pace that allows for pivoting when the next generation of models arrives.
Proposed roadmap: clear direction, rapid experiments, slow commitments
Etzioni describes a context where many founders prioritize speed without a clear direction, driven by FOMO, while some features are now developed over a weekend. He asserts that hope is not a strategy and that speed alone is not either, reminding that a startup resembles a marathon more than a perpetual sprint. Sri Chandrasekar, at AI House, expects founders to find the right problem through judgment, that low-cost building allows for rapid and manual experiments, and that the machine should only be built once it is proven to work. The advocated roadmap includes a direction where labs will not go, rapid experiments, and slow commitments, in the spirit of Jim Collins' "bullets before cannonballs," the quest for data and relationships that cannot be bought, and the ability to name one's secret weapon in one sentence. The final metaphor contrasts rushing with walking to avoid scaring away the target; Etzioni believes that many are rushing today and that discernment will distinguish those who ultimately win the prize. Absolute speed slogans, popularized elsewhere, are mentioned, but Etzioni distances himself from them.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.