We are quickly approaching two years of the generative AI revolution, triggered by the launch of ChatGPT in November 2022 by OpenAI. So far, the results have been mixed.
OpenAI recently announced that it had crossed 200 million weekly active users – nothing to sneeze at, but it got its first 100 million within two months of launch. A recent YouGov study found that including AI in a product is so likely to turn off a potential buyer as much as getting them to hand over their money.
However, money continues to flow into the sector and advances keep coming. OpenAI is looking for investors to fund future development that would allow the company to valued at 150 billion dollarsThat would be Put it on par with Cisco, Shell and McDonalds. And last week, it unveiled its latest model, called o1, which it has touted as a game-changer in the development of generative AI.
The o1 model, formerly called Strawberry, is designed to reason about decisions, much like humans do. The latest version of the model underpinning ChatGPT is actually a step back in terms of output speed and model size, which is smaller for the moment. Think of it as GPT-4.5, rather than the next major iteration, GPT-5, which is rumored to still be in development.
Mission: Impossible?
While on paper o1 is a failure, it does something that Alex had previously highlighted in this newsletter as a problem with LLM-based chatbots, and which he called the “Tom Cruise problem.” The problem was that researchers could ask ChatGPT a question one way, but when it was asked a question that directly related to the initial one (e.g. who is Tom Cruise’s mother? (Answer: Mary Lee Pfeiffer) and then asked who is Mary Lee Pfeiffer’s son (Answer: Tom Cruise), it balked.
If you ask o1 those pair of questions, the response is excellent. It even provides clues about how it arrives at the answer (something OpenAI has done cleverly and imprecisely because AI models don’t have brains), called “thoughts.” (If you want to know why anthropomorphizing AI models is a problem, see This story (I wrote in February.) When I was asked the second question, o1 “thought” for four seconds, including tracing family connections and confirming details.
So far, so good. OpenAI says that o1 can reasonMany aren’t so sure about such a declarative statement, but let them have it for marketing purposes. That would mean a significant shift in how generative AI can be used: instead of regurgitating facts from its training data or producing answers it statistically deems most likely to please users, it could consider the information and respond.
The key word, though, is “could.” We still largely don’t know how these things work, and “we” includes the developers of such tools. OpenAI has said that this ability to reason is a very big deal; the company has even made a questionable claim that o1 is its most dangerous model yet (see here (Because sometimes this is more of a marketing strategy than anything else.) Those who have tried to investigate the limits of the o1 model seem to agree with its argument about reasoning, but not so much with the danger part.
Pay no attention to that man behind the curtain!
Well, sort of. Because research can only go so far. To try to understand the chain of thought processes that underpins o1, if you want a good introduction, Simon Willison is always reliable – Users who want to dig deeper have tried to get a little more detail on what exactly o1’s “thought” process is. The information currently shown to users is a brief summary of each step in the chain of thought.
And so, they’ve been asking the model itself how it arrives at its answers, although… I have also received emails from OpenAI asking them to stop doing so or their accounts will be suspended.
After the newsletter promotion
All of this means that we are left a bit in the dark. This seems like a radical change in the world of AI and something that could take the tool from being a tool whose results should be viewed with suspicion to a tool that must be used.
What is particularly interesting is that OpenAI’s dominance has effectively wiped out the coverage of any and all competitors in recent times. Mistral, the much-hyped French competitor, launched its first multimodal model last weekThe Pixtral 12B adds image recognition to text generation. It should have received high praise, but OpenAI and o1 stole the show.
All of this means, however, that AI is still advancing and is finally starting to deliver on its promises. Whether those who tried ChatGPT in the early days and found it lacking will be able to try out the fantastic new models again is another question.