AI 2027

Though I think it is probably mostly science-fiction, this is one of the more chillingly thorough descriptions of potential AGI takeoff scenarios that I've seen. I think part of the problem is that the world you get if you go with the "Slowdown"/somewhat more aligned world is still pretty rough for humans: What's the point of our existence if we have no way to meaningfully contribute to our own world?

I hope we're wrong about a lot of this, and AGI turns out to either be impossible, or much less useful than we think it will be. I hope we end up in a world where humans' value increases, instead of decreasing. At a minimum, if AGI is possible, I hope we can imbue it with ethics that allow it to make decisions that value other sentient life.

Do I think this will actually happen in two years, let alone five or ten or fifty? Not really. I think it is wildly optimistic to assume we can get there from here - where "here" is LLM technology, mostly. But five years ago, I thought the idea of LLMs themselves working as well as they do at speaking conversational English was essentially fiction - so really, anything is possible, or at least worth considering.

"May you live in interesting times" is a curse for a reason.

I think LLM or no LLM the emergence of intelligence appears to be closely related to the number of synapses in a network whether a biological or a digital one. If my hypothesis is roughly true it means we are several orders of magnitude away from AGI. At least the kind of AGI that can be embodied in a fully functional robot with the sensory apparatus that rivals the human body. In order to build circuits of this density it's likely to take decades. Most probably transistor based, silicon based substrate can't be pushed that far.

I think generally the expectation is that there are around 100T synapses in the brain, and of course it's probably not a 1:1 correspondence with neural networks, but it doesn't seem infeasible at all to me that a dense-equivalent 100T parameter model would be able to rival the best humans if trained properly.

If basically a transformer, that means it needs at inference time ~200T flops per token. The paper assumes humans "think" at ~15 tokens/second which is about 10 words, similar to the reading speed of a college graduate. So that would be ~3 petaflops of compute per second.

Assuming that's fp8, an H100 could do ~4 petaflops, and the authors of AI 2027 guesstimate that purpose wafer scale inference chips circa late 2027 should be able to do ~400petaflops for inference, ~100 H100s worth, for ~$600k each for fabrication and installation into a datacenter.

Rounding that basically means ~$6k would buy you the compute to "think" at 10 words/second. Generally speaking that'd probably work out to maybe $3k/yr after depreciation and electricity costs, or ~30-50¢/hr of "human thought equivalent" 10 words/second. Running an AI at 50x human speed 24/7 would cost ~$23k/yr, so 1 OpenBrain researcher's salary could give them a team of ~10-20 such AIs running flat out all the time. Even if you think the AI would need an "extra" 10 or even 100x in terms of tokens/second to match humans, that still puts you at genius level AIs in principle runnable at human speed for 0.1 to 1x the median US income.

There's an open question whether training such a model is feasible in a few years, but the raw compute capability at the chip level to plausibly run a model that large at enormous speed at low cost is already existent (at the street price of B200's it'd cost ~$2-4/hr-human-equivalent).