Welcome Message

Welcome to the inaugural post of my new weekly publication where we talk all things AI stack. For format, the plan is to lead with a medium-form thought piece and then share some articles/tweets/news I found interesting. Any and all feedback is welcome (email me!) — now, let’s get onto the show.

Introducing the AI stack

Software is eating ate the world. The software stack has dominated the past 20+ years, killing its predecessors and creating many of the most valuable companies today including Google, Facebook, Microsoft, Apple, Amazon, and Stripe. But no era lasts forever. The software stack is on its last legs, and its successor is rapidly gaining market share.

We’re in the early innings of transitioning from the software stack to what I refer to as the AI stack (see above). In the AI stack, consumer agents eat traditional consumer apps by enabling a more feature-rich and generative UX, vertical-specific & enterprise agents out-compete SaaS incumbents by offering more powerful products with far higher gross margins, models replace large swathes of human mental labor, and crypto rails power this natively digital agent economy via programmable money and internet capital markets.

This isn’t a prediction — we’re already starting to see this play out: OpenAI is doing $10B in ARR, Google has begun to cannibalize its search business in favor of AI summaries, Cursor is the fastest growing business of all time hitting $500MM ARR in <18 months, 300+ YC startups are building agentic services and products, and the adoption of crypto primitives into traditional finance is only accelerating.

Caveat: very little of what exists today are what I’d call actual *agents*. Agents by definition should have true autonomy on how they execute and aren’t limited to operating as a “read-only” answering service. But as models gain higher precision with lower risk of hallucination, this evolution feels inevitable.

So where does crypto fit in?

The notion that AI will eat the world is fairly consensus in SF and tech circles. However, crypto today isn’t being seen as part of that story. My bet is this changes — agents will need to send value to each other or to other humans, and crypto rails operating as programmable, digitally-native money infra is good PMF to solve that challenge. But counterintuitively, agents won’t initially use crypto rails for payments or to trade assets. Rather, agent devs will initially find their way onchain to solve their most pressing problem: creating and scaling their own models.

In order for agents to become durable businesses, they must offer product differentiation. In my view, this will come from agents vertically integrating down the stack and owning presence at the model layer such that the UX they offer is customized to the use cases they support. However, doing anything at the model layer today is incredibly cost-prohibitive with the cost of AI talent, hardware, compute, datasets, etc — and as a result, largely untenable for the vast majority of teams. Crypto offers an alternate solution by enabling a crowdsourced approach via aligned incentives where users can make contributions to earn equity-like upside in the company assuming their contribution improves model efficacy against pre-determined benchmarks. Initially, this might look like crowdsourced LoRA adapters that are used on top of closed-source frontier models, but over time, I suspect we’ll see full-scale distributed post-training environments where contributors of all shapes and sizes can contribute models via various RL techniques, datasets, reasoning traces, evals, adapters, and more.

The mission for agent-specific models is not to replace foundational models, but rather to complement them by tackling specialized use cases in a unique, weight-sensitive way. My view is distributed post-training is the “gateway drug” for agents to come onchain. Once they do, it opens up Pandora’s box for the broader agent economy (payments, capital formation, trading, finance, etc) to happen over crypto rails.

Other interesting things I read this week

1. Fully agree with this take. Products, not models, will win.

2. The advances in distributed training land are crazy. So far, many of them are focused on the pre-training side towards OS/OW frontier models, but my view is that post-training will be the area that finds more immediate PMF with agent devs.

3. Enjoyed this read from Suhail on AI agents becoming billionaires.