Sunk Cost:

In the 1990s, artificial intelligence meant genuine machine sapience. Now, in the 2020s, AI seems to be nothing more than a synonym for the LLM. After nearly US$1tn ($1,000,000,000,000) of investment, and half a decade of AGI being ‘just a week away’, has it been worth it, or is the S&P 500 really caught in the world’s worst case of sunk-cost fallacy?

Don’t get me wrong: LLMs are impressive very-good next token generators. The first time I used the GPT-3 beta, I was genuinely stunned. Since then, whether I’ve been working on a 20yo codebase, searching the web, or planning a holiday with my girlfriend, I’ve been likely to lean on LLMs for at least part of the process. But, like, think about that for a second. These are the most-advertised LLM-friendly tasks. One trillion dollars, and it’s slightly improved a handful of flagship tasks. One trillion dollars. Where is the return?

I’m not sure, and it seems like FAANG isn’t either. Sundar Pichai, Satya Nadella, and everyone in between readily admit that they don’t know. That they’re expecting the product to come later. It’s been half a decade since GPT-3, and LLMs are more powerful, more integrated, and more truthful. Let’s extrapolate; LLMs reach their asymptote. We get near-perfect accuracy, remarkable efficiency, and seamless integration into existing systems. The question remains: Where is the return?

OpenAI is valued at over $500B with no path to profitability. The pitch has become circular: we’re building this so we can build the next thing, which will retroactively justify having built this thing. The missing premise, why we need that next thing, goes unexamined.

Financial analysts compare current LLM companies to the search engine wars of the early 2000s. That FAANG needs to invest now to avoid being buried with Yahoo. But the analogy makes no sense. Google always solved a discrete problem, information retrieval orders of magnitude faster & better than what came before, and the market validated their solution. What problem do LLMs solve that commands valuations exceeding many national GDPs? The answer keeps receding. First, it was going to be general knowledge work transformation, then it became more modest productivity gains, and now we’re celebrating chatbots and summarisation tools. The whole time some would tell you it’s AGI, just a week away. Strange that they all seem to be financially tied to the success of LLMs. The global economy is now structurally committed to this bet paying off, yet I don’t believe a satisfying answer exists. I can’t see how this isn’t the world’s worst case of sunk cost fallacy. Markets may remain irrational longer than individuals can remain solvent, but we’ll all be left holding the collective bag.

Opportunity Cost:

If I’m right, if there is no return on the trillions of investment, the opportunity cost is tragic. The entire edifice of machine learning research has been cannibalised. Funding, talent, & compute, all devoured by the LLM.

Researchers have developed ML models for cancer diagnosis, drug discovery, and predicting patient outcomes, with projects getting a fraction of the resources which go into a single LLM training run. The imbalance is obscene. Cancer diagnosis models that could save lives receive grants in the low millions. Drug discovery platforms that could accelerate treatment timelines struggle for resources. Multiply that story by every subdomain of machine learning. Medical image analysis. Each one competing for scraps while foundation model companies raise hundreds of billions.

Even outside of ML, the sheer scale of the money invested alone is staggering. How many children could we lift from poverty. How many patients could be given lifesaving medical care. How much pain, suffering, and loss of life could be averted? Instead the money goes to a hope, a promise, that past investments weren’t fruitless after all.

Perhaps the real product of LLMs isn’t intelligence at all. It’s deferral. We’ve built a machine for not answering the question of what we actually want from AI.

Every new model release, every benchmark improvement, all of it pushes the reckoning a few quarters further out. The technology works just well enough to sustain the illusion. Chatbots respond plausibly. Copilots reduce boilerplate. Summarisation feels adequate. None of it transforms anything fundamental, but it performs transformation convincingly enough to justify the next funding round.

We had machine learning. We had funding. We had talent. We had compute. We could’ve built anything. We built autocomplete which convinced the world it was thinking.

The deferral machine hums along. The question remains unanswered. And somewhere, somehow, for now, a graph continues to go up.