OpenAI's new big model points to something worrying: an AI slowdown
Engineers from the company who have already tested the model, called Orion, seem to be clear that the jump in performance is by no means exceptional. ChatGPT's second birthday is approaching, and even Sam Altman himself echoed it and suggested the arrival of a possible "birthday gift." A recent leak has given us potential details about that gift, and there is good news (great new model) and bad news (it won't be revolutionary). Let's take a look.
Orion . That’s the name of OpenAI’s next big AI model, according to company employees in comments leaked to The Information . The news comes just before the second anniversary of ChatGPT’s launch on November 30, 2022. As reported by TechCrunch , OpenAI denied that it plans to launch a model called Orion this year.
Low expectations
These employees have tested the new model and have discovered something worrying: it performs better than OpenAI's existing models, but the jump is better than that between GPT-3 and GPT-4 or even the flashy GPT-4o .
How they tackle the problem . The relatively “evolutionary” version of ChatGPT seems to have prompted OpenAI to look for alternative ways to improve it. For example, by training Orion with synthetic data produced by their own models, and also by further polishing it in the process immediately after training.
AI slowdown
If this data is confirmed, we would be faced with clear evidence of how the pace of improvements in generative AI models has slowed significantly. The jump from GPT-2 to GPT-3 was colossal, and the jump between GPT-3 and GPT-4 was also very noticeable, but the increase in performance in Orion (GPT-5?) seems like it may not be what many would expect.
Sam Altman and his claims
This would also contrast with the unbridled optimism of OpenAI CEO Sam Altman, who a few weeks ago said that we were "thousands of days away from a superintelligence." His message was logical, since he was looking to close a colossal round of investment , but it was also worrying: if expectations start to turn into unfulfilled promises, investors could withdraw the support they are now giving to the company.
But they are already very good
In fact, this slowdown is reasonable: the models are already really good in many areas, and although they still make mistakes and invent things, they do so less and less and we are also more aware of how much we can trust their responses. In areas such as programming, for example, it seems that Orion will not be especially superior to its predecessor.
What now?
But this slowdown in AI poses other opportunities that we are already seeing. If the models become more polished enough for us to trust them more, future AI agents could be a new impetus for these kinds of functions.
