I think it's inarguable that OpenAI have at least at times over the last 3 years been well ahead of other companies. Whether that's true now is open to debate, but it has been true.

This suggests they have either: made substantial breakthroughs, that are not open, or that the better abilities of OpenAI products are due to non-substantial tweaks (more training, better prompting, etc).

I'm not sure either of these options is great for the original mission of OpenAI, although given their direction to "Closed-AI" I guess the former would be better for them.

I left pretty soon after a Google engineer decided the internal chat bot was sentient but before ChatGPT 3.5 came out. So I missed the entire period where Google was trying to catch up.

But it seemed to me before I left that they were struggling to productize the bot and keep it from saying things that damage the brand. That's definitely something OpenAI figured out first.

I got the feeling that maybe Microsoft's Tay experience cast a large shadow on Google's willingness to take its chat bot public.

The way I understand it, the key difference is that when training o1, they were going beyond simply "think step-by-step" in that they were feeding the "step-by-step" reasoning patterns that ended up with a correct answer back into the training set, meaning the model was not so much trained to find the correct answer directly, but rather to reason using patterns that would generally lead to a correct answer.

Furthermore, o1 is able to ignore (or even leverage) previous reasoning steps that do NOT lead to the correct answer to narrow down the search space, and then try again at inference time until it finds an answer that it's confident is correct.

This (probably combined with some secret sauce to make this process more efficient) allows it to optimize how it navigates the search space of logical problems, basically the same way AlphaZero navigated to search space of games like Go and Chess.

This has the potential to teach it to reason in ways that go beyond just creating a perfect fit to the training set. If the reasoning process itself becomes good enough, it may become capable of solving reasoning problems that are beyond most or even all humans, and in a fraction of the time.

It still seems that o1 still has a way to go when it comes to it's World Model. That part may require more work on video/text/sound/embodiement (real or virtual). But for abstract problems, o1 may indeed be a very significant breakthrough, taking it beyond what we typically think of as an LLM.

Got it! Super cool and very helpful thanks!

Totally agree. It took me a full week before I realized that the Strawberry/o1 model was the mysterious Q* Sam Altman has been hyping up for almost a full year since the openai coup, which... is pretty underwhelming tbh. It's an impressive incremental advancement for sure! But it's really not the paradigm shifting gpt-5 worthy launch we were promised.

Personal opinion: I think this means we've probably exhausted all the low hanging fruit in LLM land. This was the last thing I was reserving judgement for. When the most hyped up big idea openai has rn is basically "we're just gonna have the model dump out a massive wall of semi-optimized chain of thought every time and not send it over the wire" we're officially out of big ideas. Like I mean it obviously works... but that's more or less what we've _been_ doing for years now! Barring a total rethinking of LLM architecture, I think all improvements going forward will be baby steps for a while, basically moving at the same pace we've been going since gpt-4 launched. I don't think this is the path to AGI in the near term, but there's still plenty of headroom for minor incremental change.

By analogy, i feel like gpt-4 was basically the same quantum leap we got with the iphone 4: all the basic functionality and peripherals were there by the time we got iphone 4 (multitasking, facetime, the app store, various sensors, etc.), and everything since then has just been minor improvements. The current iPhone 16 is obviously faster, bigger, thinner, and "better" than the 4, but for the most part it doesn't really do anything extra that the 4 wasn't already capable of at some level with the right app. Similarly, I think gpt-4 was pretty much "good enough". LLMs are about as they're gonna get for the next little while, though they might get a little cheaper, faster, and more "aligned" (however we wanna define that). They might get slightly less stupid, but i don't think they're gonna get a whole lot smarter any time soon. Whatever we see in the next few years is probably not going to be much better than using gpt-4 with the right prompt, tool use, RAG, etc. on top of it. We'll only see improvements at the margins.

Ilya was the Google researcher..

Wasn't he at OpenAI when transformers and Google's pretrained transformer BERT came out?

Oh, oops, the piece I was missing was Radford et al. (2018) and probably some others. That's perhaps what you were referring to?