My thoughts on the current state of AI

I love AI. I get up in the morning to make sci-fi real. But listen Peter Pan, if you keep going on about never never land, people are going to think you're smoking crack.

What is real?

AI is definitely going to be among the most important tech ever developed. If we don't hit any fundamental limitations, it will be the most important tech. It will let us develop a ton of other tech faster. It will let smart people become experts in unfamiliar fields in weeks or months instead of years or decades. It already has outright solved a ton of domain-specific problems.

The pace of progress is not slowing down. You think a $60B cluster is expensive? There will be $1T clusters. Fortunately for us (the people not paying that bill) there is huge economic pressure to have the best model. Better, more reliable models become applicable to more and more problems, and customers can easily switch providers whenever they want. Without any technical advancements, there will be a GPT5 just based on scaling compute. I mean a true GPT5, regardless of whether something incremental is released under that name sooner.

What is hype?

Anyone saying that AGI is coming next year is lying to you. Next 2-3 years? Not with the simplicity and frequency of errors that today's models make. The crazy thing is, this is obvious to anyone who has used recent models. When writing code, these models sometimes produce truly impressive results. Stuff that it would take me quite a while to figure out. But they continuously make basic errors that would be embarrassing for junior devs, and they just keep making it worse when they try to debug. These models are not getting smarter in the same way that humans get smarter. You don't just draw the scaling laws up here. That doesn't mean nobody will declare AGI in a year or two, but the goal posts will have moved to mid-field.

We now have hundreds of GPT wrapper startups. Some of these are legitimate, founded by devs who understand model capabilities and know the directions in which they are likely to improve. There are lots of important problems where current models are great, and next year's models will be even better. Then, there are the shills... the people who treat AI as a magic black box, are probably using it to piss off artists, came over from NFTs... you know the type. I've been in AI research since I was 16 and have done my best to push the field forward at great personal cost. I don't like grifters giving AI a bad rap, so when you occasionally see me going off on someone, this is usually why.

We also have no idea what advancements we might be missing. I don't think you can reasonably put a >50% chance on scaling solving everything without having to take a few years to revisit the fundamentals. There could be some truly mind-blowing results in as soon as a few years, but there's no way you can be sure of anything more than a 1 generation leap over GPT4.

Why am I writing this?

Because it's not just grifters anymore. The top AI labs have gone insane. OpenAI is no longer open, is bleeding talent, and now Sam is talking about AGI in 2025. I interned in ~2018, and OpenAI was by far the best place I've ever worked with the best and most talented people. It was great. This is just sad.

What about Anthropic? It was originally a safety schism off of OpenAI. One of their early hires told me that my work worried him ethically. Well, now they are maximum profit, closed source, and are partnered with Palantir. Look, I'm not against profit. I'm not even anti-defense. But holy hell, get off your high horse!

I still think both of these companies are doing great work and pushing the field forward, minus the lobbying. But nobody is going to trust them or AI as a whole if they keep acting this way. Look, you're allowed to wildly speculate about where you see AI going. My goal isn't to take the fun and futurism out of the field. Just don't do it on widely broadcast interviews on nice sets where people are taking you at face value. I keep my wilder thoughts on AI for dev streams and late night Discord vc. Or, I clearly label them as such, like in my Fireball AGI article.

Come build open source AI with us

If the big companies are wrong, reinforcement learning is the most likely next bet, and you can do it without a ton of compute. PufferAI is building ultra-performant RL environment and tools. All our code is free and open source, and the env dev side is pretty newbie friendly. X doesn't like links, but if you'd like to support our work, star PufferLib on GitHub, join the Discord, and submit some PRs!