The Current State and Future of AI

This is sure to be completely uncontroversial

I’d like to talk about where AI is and where it’s likely to go in the future. In some sense this is a fool’s errand: It’s been obvious for many decades that there’s no physical limit preventing technology from surpassing human brains so any guesswork is about when, not if, that happens, and it’s impossible to guess when major technological breakthroughs will occur. But the current boom isn’t about a series of big breakthroughs, it’s one big breakthrough and a lot of scaling up and polish. So I’m going to say what I think the limits of the current technology are and what that means for the future.

(The one big breakthrough was realizing that if you stick with sublinear functions in the middle of a neural network you can back propagate over any depth. There’s another important but less revolutionary insight that you can make the amount of computation in each layer less than quadratic if you use a transformer architecture. I have an idea for another big advance but it’s working within this framework and doesn’t fundamentally change the outlook.)

The state of AI today is comparable to what the internet was like circa 2000: An obviously very promising and important technology in the midst of its hype cycle which has yet to make a meaningful economic contribution. Improvements in AI could come to a screeching halt tomorrow and we’d still see a process over the next ten to twenty years of figuring out how to use it in industry, resulting in meaningful economic gains which show up in GDP and benefiting peoples lives beyond the fun of talking to a chatbot.

One example is in therapy. Right now chatbots are maybe being a good place for mostly mentally healthy people to vent and find companionship but they aren’t trained for treating serious mental illness and are apparently badly aggravating schizophrenia. This is easy to improve on. For treating the symptoms of depression a chatbot needs to be trained to say ‘Tell me what you’re going through. I care about you.’ For anxiety it needs to say ‘The world is a stable place and everything is probably going to be okay. Freaking out doesn’t help. Stay calm and carry on.’ Schizophrenia is more problematic and possibly not something which current LLMs aren’t good for. Just those straightforward improvements could result in much cheaper therapy available in unlimited quantities at any time of day or night for the most common mental health problems.

That said, AI improvements are obviously not coming to a screeching halt tomorrow. But what’s going on now is mostly scaling up: More data and more training. Eventually you run out of data and can’t afford any more training. An adult human has processed less than a gigabyte of linguistic information and is on a completely different level, so there are still some mysterious fundamental improvements to be had in getting training to work well. The LLMs we have today give a very misleading impression of how good they are. They can do things like make up plausible-sounding recipes but if you try following those recipes you’ll find they need a lot of tweaking to get dialed in. And I have to snark that the new Opus 4.5 model is a massive regression for things like recipes and figuring out which actor was referred to by a given pronoun. It’s best to think of LLMs as chatbots. They’re a massive enhancement in search technology and extraordinarily good at language translation and the tedious parts of coding. But they’re still fundamentally collating things from their training data and dumb as a rock.

One thing affecting the optics of the quality of LLMs is that they’re very good at chatting and math. What’s going on here isn’t so much that the LLMs are exceptionally good at these things as that the state of the art prior to them was bizarrely insanely bad. This had long been a mystery. Why can’t we apply simple statistical techniques to at least make a vaguely plausible chatbot which won’t give itself away in literally sentences? The best we could do were things which obfuscated and said vague generic things and hope that the user doesn’t notice that there isn’t much meat in what it’s saying. What we have now are LLMs, which apparently are those simple statistical techniques which can make plausible text. They just happen to require a technique we didn’t know before and require about nine orders of magnitude more data and computation to train than we expected. They also still work in no small part by obfuscating and saying vague generic things and hoping the user doesn’t notice there isn’t much meat in what they’re saying. But they also augment that by agreeing with the user and repeating what they say a lot.

By the way, if you want to bust something as being a chatbot the best approach isn’t to leverage what they’re bad at but what they’re good at. Ask it to play a game where it doesn’t use certain letters, or only speaks in iambic pentameter, or only uses words containing an odd number of letters, and it will immediately give itself away by demonstrating utterly superhuman abilities. It has no idea how to emulate human frailty realistically.

Unrelated to all that, a note about my last post: It turns out that my napkin model missed that supercritical fluid density is highly nonlinear and in particular gets very dense close to the critical temperature so in practice you want the critical temperature to be just barely below the minimum temperature of the cycle you’re using. Carbon Dioxide’s critical temperature of 31.1 Celcius is very good given typical Earth air temperatures. This paper considers the scenario where you have a solar thermal plant out in he desert so the ambient temperature is considerably higher than normal and you want to increase the critical temperature of the working fluid. They suggest doing this by adding Perfluorobenzene. The problems with this approach are that there’s the counterfactual of pumping water underground for cooling or replacing the whole system with photovoltaics. It may be more promising to instead go in the opposite direction: If you have a power plant next to frigid arctic waters which stay near 0 Celcius year round you can lower the critical temperature to around 15 Celcius by adding in about 12% Argon. That’s a boring but low risk modification which is likely to result in a small improvement in efficiency, and any improvement in efficiency of a power plant is a big deal.