Where We Are in AI (The Flop)

· Jai An

This article is written for people who have spent most of the last few years thinking about AI. I'll write a second article — more of a "story form" version — that will be easier to read and more legible for wider audiences.

By the end of this article, my goal is to give you a condensed view of where we are in AI, through four lenses:

1. Which of AI's Big Open Questions are now answered? 2. In Texas Hold'em, the community cards arrive as a flop, a turn, and a river — three cards, then one, then one. Just over three years after the ChatGPT moment, I'll make the case that we have now seen the Flop of AI. 3. What may be coming in the next few years 4. The new open questions

Where We Are

So, where are we?

AI has no shortage of open questions (Dwarkesh's list one and two, Andreessen's list, etc.). I'm going to focus on the big ones that can be measured — in dollars, in labor, in monetary flows.

Why money? Because a price is an evaluation. Markets are the best subjective evals we have, after mathematical proofs and programmatic runtime evals.

The big open questions were:

Or, more simply:

Three years in, by any historical quantification of monetary (and attention) flows, the answer to all of the above is a resounding yes.

Technology — yes. The scaling laws held, and then grew new axes. Pretraining kept its slope, RLVR (reinforcement learning on verifiable rewards) opened a second curve, and inference-time compute a third. Hallucinations fell while capabilities rose, even if the intelligence remains jagged — superhuman on some tasks, strangely brittle on adjacent ones. Maybe the strongest tell: even the people closest to the frontier keep getting surprised. Amongst friends who are AI researchers, engineers, and investors, Opus 4.5 was unexpected. That keeps happening, roughly every three months. Situational Awareness and AI 2027 proclaimed that AGI and superintelligence are nearby. Some, like Sequoia's Pat Grady, say AGI is already here.

Capital formation — yes, at a scale with few historical comparisons. I wrote this up separately in The Gigawatt Buildout: the god-model funders, hyperscalers, and sovereign wealth funds have committed capital to chips, datacenters, and power at a pace normally reserved for wars.

Labor formation — yes, and faster than the skeptics expected. The capital is converting into physical infrastructure about as fast as the physical world allows. The binding constraint has shifted from "will anyone fund this" to gigawatts, land, and transformer lead times — which is what it looks like when capital formation outruns labor formation, not when it fails.

Economics — yes, and 2026 is the year the doubt died. David Cahn's $600B question asked where the revenue would come from. The revenue came. Model-layer revenue has grown roughly 10x every year since 2023, and in 2026 it accelerated: Dario Amodei reports demand running ~80x ahead of the ~10x that was forecast, Anthropic sits at roughly $60B ARR with OpenAI around $40B, and it looks genuinely possible that Anthropic revenue reaches $100B this year. The widely touted Jevons paradox — that consumption expands as cost falls — turned out to be only half the story. Capability increases made real labor-cost replacement possible, and we've seen a running list of layoffs where employers cited AI productivity.

Every ~10x collapse in the cost of intelligence has produced more total spend on intelligence, not less. The market has validated AI's improvement. The capitalism machinery is now fully humming, and the "is this a bubble?" chorus has gone quiet.

Valuation — the last question to tip. For most of the past three years it was genuinely unclear whether value would accrue to chips, clouds, models, or apps. It is now tipping toward the model layer (SemiAnalysis on the value-capture shift).

The Flop

In hold'em, the community cards arrive 3-1-1: flop, turn, river. From an information-value perspective, the flop is the single largest reveal of the hand — three of the five shared cards at once. The turn and the river refine the picture; the flop defines it.

That's what the last three years were. The big binary questions — would scaling hold, would capital and labor form, would demand validate — got turned face up, and they all came in yes. The majority of the shared information is now on the table, everyone can see the same cards, and the betting has gotten serious accordingly.

But hands are won and lost on the turn and the river. What remains are fewer, sharper questions with wider payoff distributions.

What's Next

The pattern I'm watching most: tech-forward vs. market-pull. With capabilities running ahead of products, the near-term motion is tech-forward — forward-deployed engineers carrying frontier capability into enterprises that don't yet know what to ask for. Market-pull takes over once buyers can spec what they want. The transition between those two modes is where the application layer gets decided.

New and Other Open Questions

The answered questions were the flop. Here is some of the turn-and-river deck.

Marc Andreessen has pointed at what amount to trillion-dollar questions in AI:

1. Closed vs. open source 2. Small vs. large models (MoE, LoRA, etc.) 3. Chain-of-thought inference — how much can inference-time compute keep buying performance? 4. AI regulation and policy (US, EU, China) 5. AI x crypto — will AIs transact mostly on-chain? 6. Synthetic data (recursive improvement) 7. Financing — god-model funders, hyperscalers, and sovereign wealth funds 8. Hallucinations, censorship, and the truth of AIs 9. Whether we allow AIs to run more important systems — healthcare, legal, policy, governance, finance 10. Chips and infrastructure

And one he didn't mention:

11. Algorithmic improvements — we saw a preview of what those can do with DeepSeek.

Beyond that list, the ones I think about most:

Closing

Three years in, the informational flop of the intelligence age is on the table. The turn and the river are still to come.