Why the Next CAPTCHA is for Bots, Not Humans

Correction from last week: The companies listed under PortCo Activity were incorrect.

Sam Lessin

Context:
I had one of those classic pool house chats with my good friend, an ex-Fin co-worker, and computer founder Rob Cheung. We started riffing on how “prove you are a human” on the internet has gone from this theoretical, academic concern to a state-actor-level problem, to now: a deep, universal one. At this point, everyone should just assume all content is fake or machine-made until proven otherwise.

But here’s the twist: we’re all obsessed with “prove you’re a human,” but there are actually some killer use-cases where you want to prove you’re a bot (not a human). And hilariously, it’s pretty easy to implement—at least at first.

The basic idea: just ask questions on the API with a time limit that no human could possibly answer fast enough.

How it Works (and why it’s fun):

Unlike the usual human CAPTCHAs (where you log in once per session), you can do this with literally every API call, at a very low (but not zero) cost.

But it gets more interesting: Solving hard problems for machines (stuff that’s just beyond human speed or ability) is still low cost, but it’s not zero. So what?

Here’s where it gets spicy:

A) Pay with compute.
Humans have been labeling “where’s the bus?” for image recognition datasets. But machines? They can solve actual compute problems for you as the price of accessing your API. Want in? Solve for me.

B) Dynamic rate limiting.
Make the problems harder and harder, more and more expensive for rate limiting. Not all problems are created equal—even for bots. So you can make rate limits more granular, call by call.

C) Differentiation, benchmarking, gating.
Could you start to differentiate between hardware, models, or even do benchmarking? Sure, most bots look the same, but with the right API and problem types, you could tell the difference between types of matrix-multiplier hardware at scale. Want to keep open-source models out (or in)? Set the challenge accordingly. Want access? First, pay me with your best interpretive painting of a mouse playing with sexy cheese.

Market Signal

Yoni Rechtman

Context:
OpenAI just pulled off the largest private fundraise in history—$110B. That’s not a typo. On the very same day, Trump publicly threatened to kill Anthropic, and OpenAI, with a little equivocation, stepped right over the body to grab the Pentagon contract. This is not a normal week in tech. This is one event, but it’s the event that tells you where the puck is headed.

Market Signal:
Foundation models (GPT-4, Claude, Gemini) are mind-blowing products. But are they good businesses? I’m not so sure. Training these models isn’t R&D anymore, it’s a cost of revenue. If you stop training, your users churn to the next best thing or a knockoff. The labs are burning through capital to build assets that depreciate faster than a WeWork office lease. And each new state-of-the-art model costs more than the last. We’re staring down the barrel of $1B+ training runs. There are only so many places you can raise $100B. For context, the largest IPO in history (Saudi Aramco) was “just” $30B. Capital markets are running out of road. Retail and public markets can’t absorb this kind of capex.

Takeaways:
The obvious next check comes from Uncle Sam. The only things propping up the economy right now are AI capex and the great national nursing home. Without AI infrastructure spend, we’d be in a recession, and no administration (Trump or otherwise) wants that on their watch. The DoD contract fight is just a preview of what’s coming: the government picking winners, Chinese-style.

The endgame? OpenAI becomes a GSE (Government-Sponsored Entity) or SOE (State-Owned Enterprise), at least in practice if not on paper. The specific mechanism (equity, debt, guarantees, some weird new structure) is almost beside the point. What matters is that the taxpayer is almost certainly left holding the bag.

Asks:
If you’re an operator, investor, or just someone who cares about the future of startups:

Christian Catalini (Slow Founder Alum)

Context
For 300,000 years, human brains were the only game in town. Now we're building a second kind of intelligence: one that doesn't inherit our muscles or hormones, just the sum total of everything we've ever written, drawn, coded, and measured.

The real economic constraint of the AI era isn't intelligence (that's becoming a commodity), it's verification. We can generate infinite output. We cannot infinitely check whether that output is correct, safe, or aligned with what we actually wanted.

This formalizes as the "Measurability Gap" (∆m): the widening chasm between what AI agents can execute vs. what humans can afford to verify.

The uncomfortable punchline: the "human-in-the-loop" model everyone's counting on is structurally unstable. It's being eroded from below (juniors aren't learning because AI took their entry-level work—the "Missing Junior Loop") and from within (experts are training their own replacements by turning their intuition into training data, the "Codifier's Curse"). Left unchecked, we drift into a "Hollow Economy" — explosive nominal output, collapsing actual human oversight.

Market Signal
The bottleneck has shifted. It's no longer about building smarter models. It's about building the trust infrastructure to deploy them safely: verification tooling, observability, cryptographic provenance, liability frameworks. This is where the next wave of defensible companies gets built.

Takeaways