Wired has been publishing a series of predictions about the coming state of the world. Unsurprisingly, most concern artificial intelligence — how it might impact health, music, our choices, the climate, and more. It is an issue of the magazine Wired describes as its “annual trends briefing”, but it also kind of like a hundred-page op-ed section. It is a mixed bag.
A.I. critic Gary Marcus contributed a short piece about what he sees as the pointlessness of generative A.I. — and it is weak. That is not necessarily because of any specific argument, but because of how unfocused it is despite its brevity. It opens with a short history of OpenAI’s models, with Marcus writing “Generative A.I. doesn’t actually work that well, and maybe it never will”. Thesis established, he begins building the case:
Fundamentally, the engine of generative AI is fill-in-the-blanks, or what I like to call “autocomplete on steroids.” Such systems are great at predicting what might sound good or plausible in a given context, but not at understanding at a deeper level what they are saying; an AI is constitutionally incapable of fact-checking its own work. This has led to massive problems with “hallucination,” in which the system asserts, without qualification, things that aren’t true, while inserting boneheaded errors on everything from arithmetic to science. As they say in the military: “frequently wrong, never in doubt.”
Systems that are frequently wrong and never in doubt make for fabulous demos, but are often lousy products in themselves. If 2023 was the year of AI hype, 2024 has been the year of AI disillusionment. Something that I argued in August 2023, to initial skepticism, has been felt more frequently: generative AI might turn out to be a dud. The profits aren’t there — estimates suggest that OpenAI’s 2024 operating loss may be $5 billion — and the valuation of more than $80 billion doesn’t line up with the lack of profits. Meanwhile, many customers seem disappointed with what they can actually do with ChatGPT, relative to the extraordinarily high initial expectations that had become commonplace.
Marcus’ financial figures here are bizarre and incorrect. He quotes a Yahoo News-syndicated copy of a PC Gamer article, which references a Windows Central repackaging of a paywalled report by the Information — a wholly unnecessary game of telephone when the New York Times obtained financial documents with the same conclusion, and which were confirmed by CNBC. The summary of that Information article — that OpenAI “may run out of cash in 12 months, unless they raise more [money]”, as Marcus wrote on X — is somewhat irrelevant now after OpenAI proceeded to raise $6.6 billion at a staggering valuation of $157 billion.
I will leave analysis of these financials to MBA types. Maybe OpenAI is like Amazon, which took eight years to turn its first profit, or Uber, which took fourteen years. Maybe it is unlike either and there is no way to make this enterprise profitable.
None of that actually matters, though, when considering Marcus’ actual argument. He posits that OpenAI is financially unsound as-is, and that Meta’s language models are free. Unless OpenAI “come outs [sic] with some major advance worthy of the name of GPT-5 before the end of 2025″, the company will be in a perilous state and, “since it is the poster child for the whole field, the entire thing may well soon go bust”. But hold on: we have gone from ChatGPT is disappointing “many customers” — no citation provided — to the entire concept of generative A.I. being a dead end. None of this adds up.
The most obvious problem is that generative A.I. is not just ChatGPT or other similar chat bots; it is an entire genre of features. I wrote earlier this month about some of the features I use regularly, like Generative Remove in Adobe Lightroom Classic. As far as I know, this is no different than something like OpenAI’s Dall-E in concept: it has been trained on a large library of images to generate something new. Instead of responding to a text-based prompt, it predicts how it should replicate textures and objects in an arbitrary image. It is far from perfect, but it is dramatically better than the healing brush tool before it, and clone stamping before that.
There are other examples of generative A.I. as features of creative tools. It can extend images and replace backgrounds pretty well. The technology may be mediocre at making video on its own terms, but it is capable of improving the quality of interpolated slow motion. In the technology industry, it is good at helping developers debug their work and generate new code.
Yet, if you take Marcus at his word, these things and everything else generative A.I. “might turn out to be a dud”. Why? Marcus does not say. He does, however, keep underscoring how shaky he finds OpenAI’s business situation. But this Wired article is ostensibly about generative A.I.’s usefulness — or, in Marcus’ framing, its lack thereof — which is completely irrelevant to this one company’s financials. Unless, that is, you believe the reason OpenAI will lose five billion dollars this year is because people are unhappy with it, which is not the case. It simply costs a fortune to train and run.
The one thing Marcus keeps coming back to is the lack of a “moat” around generative A.I., which is not an original position. Even if this is true, I do not see this as evidence of a generative A.I. bubble bursting — at least, not in the sense of how many products it is included in or what capabilities it will be trained on.
What this looks like, to me, is commoditization. If there is a financial bubble, this might mean it bursts, but it does not mean the field is wiped out. Adobe is not disappearing; neither are Google or Meta or Microsoft. While I have doubts about whether chat-like interfaces will continue to be a way we interact with generative A.I., it continues to find relevance in things many of us do every day.