(I wrote this quickly for my own edification. It isn't original thinking; it condenses takes I’ve read on X over the past few days. Where relevant I’ve linked tweets. I hope it’s helpful for people.)
Watching an OpenAI launch now feels like watching an Apple keynote in the early 2010s. As a kid, I was giddy about which new, unaffordable iPhone would drop. I’ll probably look back on my twenties as the years I felt the same mix of excitement and perhaps fear about whatever OpenAI released next.
At first glance on X, before I had access to the model, the mood was mostly disappointment. Some people pushed their timelines out; a few even talked about the end of the pre-training paradigm. The live presentation had a couple of funny slip-ups.
I’m not a technical person, and my naive expectation was that GPT-5 would feel as big a jump as GPT-3 to GPT-4. I now think two things are true at the same time:
a) GPT-5 is not a big step up from GPT-4.
b) That doesn’t mean AI progress has stalled.
a) does not entail b).
GPT-5 isn’t a brand-new base model or a push to a new state of the art. It’s more of a product release: a family of models plus a router. You get a general GPT-5 for most tasks, a “thinking” variant for harder ones, and the system picks which to use for your prompt.
On compute, the big jumps have come from scale up in pre-training compute. GPT-3 → GPT-4 was roughly a hundred-fold; GPT-4 → 4.5 about an order of magnitude. Epoch think GPT-5 isn’t a major scale-up over 4.5, which also fits the lower cost and faster outputs. That doesn’t mean pre-training is over; it more likely the case that the next big scale up in pre-training will happen further down the line.
Where the push may go next though is RL. RL sits after pre-training and shapes behaviour. If we want broader, task-agnostic capability, we probably need far better, and far bigger, RL environments and rewards. Mechanize call this “replication training”: thousands of diverse, auto-graded tasks that let you scale RL the way we once scaled pre-training.
Okay, so if a) is true, what about b)?
I think the tweet below does a good job of summarising this, but to me, the ‘progress has stalled’ feeling is mostly expectation inflation, ours and OpenAI’s. Calling it “GPT-5” and hyping the launch, perhaps to keep pace with competitors, seems to have backfired. Still, things look on trend for long-horizon tasks, and we are certainly not out of the weeds on making sure the next big jump goes well.
Thanks for the post.
I think many of these excerpted comments are missing the point. They say 'GPT-5 is a product release and not a much larger, newer model'. Granted. But they don't ask why it is not a much larger, newer model. My answer is that OpenAI has tried and does not yet have the ability to build anything much bigger and more capable relative to GPT-4, despite two years and untold billions of investment. What we are seeing is massive dimishing returns relative to investment. The fact that this is the best OpenAI can do to warrant the GPT-5 label after all this time and money warrants a significant update.
I'm not sure this is true. Two key points are made in the Sam Hammond tweet:
GPT-5 is an enormous step up from GPT-4 in capabilities, to be sure, setting aside training compute.