StephenMiller.eth ♨️

My take from GPT-4.5 is that humanity has designed an AGI architecture - it is just prohibitively expensive. This model is not great, because training a $1 billion transformer only gives us a 12.5% improvement over a $100 million one, in a paradigm where, apparently, utility scales logarithmically with training cost... That also means that a dense GPT-5 would be only ~11% better than GPT-4.5, for the cost of $10 billion. Similarly, to get a jump as big as the one we've seen from GPT-2 to GPT-4, we'd need to train a GPT-7 (*not* a GPT-6), and that would cost about $100 trillion, i.e., the world's entire GDP. So, that's the wall: we saturated humanity's capacity to scale. Or, to be more specific, we'd need 1,000,000x more compute than GPT-4, to see that sort of jump again. Some argue that reasoning breaks this wall, but I feel like it only weakens it. If test-time compute laws hold, then, we'd need a GPT-4 scale model to "think for 100 million tokens per output token" to emulate a GPT-7. Except it would take days to produce each token. That's not viable. So, unless we make 1,000,000 clones of planet Earth, we could be stuck at roughly this capacity for several decades, and never see a jump as big as the one from GPT-2 to GPT-4 again. Unless, of course, new ways to improve the efficiency of these systems are discovered. AGI has become an optimization problem. I, for one, suspect that GPTs are embarrassingly sub-optimal, and that these big matrix multiplications are merely emulating an underlying "learning algorithm" with a massive overhead. Now, it isn't hard to see that, at this scale, a single attention (i.e., "neural dict") pass takes easily more than 1,000,000x the compute than a dict lookup. If that is true, it wouldn't be surprising if the first team to break the "matmul wall" would be able to train a model equivalent to GPT-4 for as little as $100. Of course, attention is doing much more than a dict lookup; but we don't know what it is doing that leads to reasoning capacities. And, once we figure that out, we may be able to have GPT-7 for the cost of GPT-4, and not for the world's entire GDP. That said, this would require a complete redesign. Gradient descent and matmuls have to be replaced by something entirely different - and nobody knows what that would be. It took us decades to go from neural nets to transformers, so, it could take us a decade to figure this out. Or someone could be stuck with a rush of inspiration and it would happen overnight... Anyway sorry if I got some napkin math wrong, and all the respect for OpenAI for this release. Publishing a result that isn't a complete success is great science. Now I just want to understand what transformers are emulating, and how we can do the same, for less. I have many ideas, and I have many experiments to run... I'll try not to disappear completely but excuse me if I do