Aman Dhesi

The through-line

Two commitments run through everything Aman Dhesi says: a very early and durable fascination with machine intelligence, and a hard-nosed view that a startup only exists once it has found someone to serve. He traces the interest back to learning to program at ten or eleven to make video games, then to a college competition where teams wrote chess bots to play each other, after which "the concept of smart machines was just so fascinating" [1]. That led to AI research as an undergraduate and a summer at Microsoft Research in Seattle in 2009, and he describes AI as sitting at the intersection of the two things he cared about, computers and math [1].

The entrepreneurial half of the through-line is more austere. He has been thinking about businesses since around 2014 and has "tried so many things" and failed at most of them, to the point that failure no longer fazes him, which he treats as constitutive of the work rather than incidental to it [1]. The lesson he draws is the same each time: what startups "eventually boil down to is finding a a a market that you can serve and serving him as a product" [1]. When that did not happen, he shut the idea down and moved on with the lessons.

On what a startup is actually for

His most recent venture began as a machine learning company applied to crypto security, before the vocabulary shifted: what people now call AI he was doing as machine learning, which he defines as the process by which AI is trained [1]. The problem was concrete. If you self-custody crypto, you can press the wrong button or land on the wrong website and lose everything, so the team built what amounted to an antivirus for the wallet [1]. Through Stelo Labs, that took the form of Stelo, a browser extension that simulates a transaction and shows contextual information and a risk assessment before the request reaches the wallet, plus an API exposing that same simulation and risk backend to developers, which the team used to build approvals.xyz, a token approval dashboard [3]. He and his co-founder also discussed an upcoming API-powered interface for safe interaction with arbitrary smart contracts [3].

Two years in, the company wound the idea down. His account is unsentimental: consumer crypto in 2021 and 2022 was "very much like a kind of a like a bubbly phenomenon where hype exceeds the reality of it", the wave receded, and they never found product-market fit [1]. He does not frame this as bad luck or bad timing so much as the ordinary verdict of the only test that matters, and he carries the lessons forward rather than relitigating them [1].

On crypto as two different things

He insists on separating crypto as an asset class from crypto as a developer platform, because the two have behaved completely differently [1]. As an asset class, measured in price, it has gone essentially straight up and to the right since the ETF launch [1]. As a developer platform, it has had repeated ups and downs and is currently in one of the downs, still without real product-market fit [1]. This distinction is what lets him be constructive about crypto's monetary role while being blunt about how little of the builder promise has landed.

On the monetary side he offers two mental models. The first is crypto as a liquidity sponge: "the more liquidity that governments print the more kind of gets absorbed into crypto", partly because people stop trusting government judiciousness about the money supply, producing an exit from the fiat system into a system where monetary policy is rules-based rather than subject to the whims of a central bank [1]. The second is insurance, and specifically insurance against a named risk. Anarchy is a lack of rules, and the opposite of a lack of rules is a protocol whose rules cannot be changed, so he treats it as "an insurance policy against catastrophe and Anarchy" and thinks everyone should hold some [1]. Asked whether an entire national or global economy might one day run on crypto as its financial operating system, he puts the probability at "low but not zero" [1].

On where adoption actually comes from

Adoption, in his reading, tracks political instability rather than technological enthusiasm. The closer a country is to chaos and anarchy, the more likely it is to need something like crypto, which is why several politically unstable South American economies have moved closest to it [1]. China is the mirror image: an extremely regimented system, and therefore one moving away from crypto rather than toward it [1]. He is also careful not to overclaim from headline cases, noting that El Salvador adopting Bitcoin as legal tender does not mean the country's economy actually runs on it [1].

On why decentralized AI training does not work

He is direct about the limits of the crypto-AI intersection, prefacing it by saying he has not researched everything people are attempting. Of what he does know, the popular idea is decentralized GPU networks for training, and he considers it a physics problem rather than a coordination problem: when GPUs in a cluster are not physically next to each other, the time to move data across the wire between them becomes the dominant cost, more so than the actual computation [1]. That is why the companies assembling dense physical GPU clusters, effectively the AWS of GPUs, have seen their stocks double, triple and quadruple, and why he is not bullish on decentralized GPU training [1].

Where he does see a role is the same counterbalancing function he ascribes to crypto in monetary policy. He suspects something similar exists against monolithic AI superpowers and monopoly control of the technology, while being candid that he does not know what it looks like; his most concrete guess is people pooling capital to buy their own GPUs [1].

On having been early to ranked feeds

His 2009 stint at Microsoft Research produced what he describes as the first paper on ranking for social media feeds, written before Facebook ranked posts, on showing content by predicted relevance to the user rather than chronologically [1]. He notes the irony that the work happened inside Microsoft, which in hindsight was not where that idea would be commercialised [1]. The episode is the origin of a wider observation he makes about the field: AI only entered the collective consciousness around 2022, but the ideas and the research were running long before that [1].

Takeaways

• Separate crypto as an asset class, which has run straight up since the ETF launch, from crypto as a developer platform, which still has not found product-market fit [1].
• Crypto functions as a liquidity sponge for printed money and as "an insurance policy against catastrophe and Anarchy", because its rules cannot be changed [1].
• The chance that an entire national or global economy runs on crypto as its financial operating system is "low but not zero"; legal tender status, as in El Salvador, is not the same thing [1].
• Adoption is driven by instability: the closer a country is to chaos, the more it needs crypto, while regimented systems like China move the other way [1].
• Decentralized GPU training is inefficient because data movement between physically separated GPUs dominates the computation, which is why dense single-site clusters win [1].
• Crypto's plausible role in AI is as a counterbalance to monopoly AI superpowers, possibly through pooled capital buying GPUs, not through decentralized training [1].
• Stelo's browser extension simulates a transaction and returns context and a risk assessment before it reaches the wallet, with the same backend exposed via API and used to build approvals.xyz [3].
• Startups reduce to finding "a market that you can serve"; when a two-year effort in consumer crypto security did not find it, the right move was to shut it down [1].