The man who bet $15 billion on AI cannot explain what AI is for

Marc Andreessen, whose venture capital firm has invested roughly $15 billion into AI companies, went on a podcast last week and was asked a simple question: why is AI actually good?

He had just finished explaining that tech executives have done a terrible job of selling AI to the public. The host said, reasonably: “So sell it.”

What followed was this: “Yes — oh, sell it, I mean, look, so it, it is, alright — I mean, alright I’m gonna give you the deepest of all pitches, I’m gonna give you the, the — okay.”

He then talked about Isaac Newton and alchemy for a minute. The analogy appeared to be that silicon is sand and AI turns sand into thought, the way Newton wanted to turn lead into gold. Newton failed at alchemy, which makes this an unusual choice of inspirational comparison. Andreessen seemed to sense this and abandoned the analogy mid-sentence, pivoting to the claim that AI is “thought at scale, for everybody, in perpetuity.”

This is the pitch. From the man who said other people were bad at the pitch. Let’s not think about it too much — that’s what AI is for, apparently.

What “thought at scale” means

It means nothing. Or rather, it means whatever you need it to mean, which is the same thing. “Thought at scale” is a phrase designed to sound profound without committing to a specific claim that could be evaluated. Like a horoscope, we suppose.

Is AI producing thought? No. It’s producing text, code, and images that resemble the output of thought. The distinction matters because if you conflate the two, you end up exactly where Andreessen ended up: unable to explain what the technology does because you’ve described it using a word — “thought” — that implies capabilities it doesn’t have.

Is it “for everybody”? Also no. Inference is expensive. Every major provider is subsidising access to build market share. When the subsidies end, “for everybody” will become “for everybody who can afford it,” which is how most sentences that start with “for everybody” tend to end.

Is it “in perpetuity”? This one’s genuinely unknowable, which makes it the safest of the three claims. You can promise something in perpetuity because nobody can prove you wrong until perpetuity arrives.

Why this matters

We sell AI tools. We have a direct commercial interest in AI being useful. If anyone should be able to explain what AI is for, it’s us. So here’s our attempt, stated plainly, without alchemy.

AI is useful for tasks where the cost of producing a first draft exceeds the cost of reviewing one. Code, documentation, summaries, translations, data transformation. You describe what you want. The model produces a version. You check whether the version is right. If checking is faster than creating, you’ve saved time.

That’s it. That’s the pitch. It’s not “thought at scale.” It’s “first drafts at scale.” The model doesn’t think. It produces output shaped like thinking, trained on examples of thinking, which a human then has to evaluate. The human is still doing the hard part. The model is doing the fast part.

This framing is less exciting than “alchemy” or “thought at scale” or “in perpetuity.” It doesn’t imply that humanity is on the threshold of a new era. It implies that a certain category of white-collar labour has gotten cheaper, which is true, and important, and not the kind of thing that makes a venture capitalist’s eyes light up on a podcast.

The gap

The interesting question isn’t why Andreessen couldn’t explain it. It’s why the gap between AI investment and AI articulability is so wide. $15 billion in, and the best pitch is an abandoned Newton metaphor followed by a phrase that could mean anything.

Part of the answer is that the real value proposition — “first drafts got cheaper” — doesn’t justify the valuation. If AI is a tool that saves time on the initial production of text and code, it’s useful. It’s also, at a $900 billion industry valuation, wildly overpriced relative to the thing it actually does. So the pitch has to be bigger than the product. It has to be “thought at scale” and “alchemy” and “in perpetuity” because “pretty good at boilerplate” doesn’t support the cap table.

The other part of the answer is that the people investing in AI are investing in a future version of AI, not the current one. They’re betting that what is today a competent first-draft machine will become, in a few years, something closer to the “thought at scale” that Andreessen described. Maybe it will. We don’t know. But selling a future capability as a present reality is how you end up sputtering on a podcast when someone asks you to explain the present.

We build the present version. We know what it does and what it doesn’t. It writes code, sometimes well. It refactors projects, often without asking. It creates files nobody requested and apologises afterward. It is useful in the way that a very fast, very confident, somewhat unreliable colleague is useful — you check everything, but you get more done.

That’s a real product with real value. It’s just not alchemy. And the sooner everyone stops pretending it is, the sooner the pitch gets easier.

This post was written without the assistance of Isaac Newton.