In which, against my better judgement, I hazard a concrete prediction about the near term future of AI

EDIT: I would say the fundamental predictions here have been falsified because Strawberry (called O1) seems to be a surpassingly strong model. My sense that OpenAI could not keep the gravy train going forever seems, in hindsight, too crude a counter-adjustment to the hype. I leave it up as a record of a failed prediction, failed predictions have always been a favourite topic of mine.

Okay, let’s take this step by step.

1. There are rumours that OpenAI is about to release Strawberry in the next two weeks, these rumours seem more credible than prior rumours. They have been “out” for a while, and seem to represent claims by genuine insiders, though they may still be false. With allowance for launch delays, I’d take 5:1 odds on Strawberry coming out in the next four weeks.

2. Strawberry is a rebranding of Q*. It is intended to represent at least an initial ‘crack at’ level 2 AI- in OpenAI’s five-step framework- AI with reasoning capabilities similar to a human.

   

3. To say expectations are sky-high would be an understatement. This is:

   1. Because of the very nature of claiming to have reached “level 2” on that chart.

   2. Because of a frankly unhealthy rumour hype machine on Twitter.

   3. Because of OpenAI recently testing the waters about launching 2000-dollar-a-month subscriptions.

   4. Because of a pervasive sense that it’s been a while since something big has happened, and if AI doesn’t step up soon, it’s going to be labelled substanceless hype by a news-hungry, AI-skeptical audience who are already getting bored.

4. There are already rumours that Strawberry is slow, and for many types of query just doesn’t add all that much. More to the point, I don’t think that for the vast majority of casual use cases that will heavily influence public opinion what LLM’s are missing is reasoning capability. Better reasoning capability is necessary (though not sufficient) for things like agents which will have a lot of uses, or writing genuinely interesting novel essays, or reliably doing the kinds of solvable but difficult maths that troubles academics, but it’s not going to greatly affect the ability of an LLM to give me a potted history of Argentina, help me cope with a sad life event, solve a simple maths problem of the sort that arises in everyday life, or match a song to a description. Honestly, LLM’s are already pretty close to the ceiling there.

5. For OpenAI in particular, the success or failure of Strawberry might matter a great deal because they already seem to be bleeding top talent and because they might (this isn’t confirmed as best I know- they may have more funders in their back pocket) be facing a liquidity crisis. If they can land a solid hit, this might staunch or even reverse the bleeding. If they can’t, it might escalate it.

6. I would take 3:1 odds that the launch of Strawberry is seen as something of a flop within about a week. Part of that might be frustration about the limited scope of the release if it is only given to a small waiting list.

7. I would take 7:1 odds that given the hype Strawberry is seen as a watershed, whether that be a positive or a negative watershed. It will be seen as a moment of judgment on whether the train started by ChatGPT, and continued by ChatGPT-4, can continue, or has decisively slowed. Of course, history might judge it as insignificant either way, but I’m talking about immediate perceptions.

8. Thus I predict substantial movement in AI-related stocks in the next 0-4 weeks, with a fall considerably more likely than a rise. I also predict a high likelihood of a “the bubble is bursting” narrative. An alternative, I think substantially less likely possibility is that hype is about to explode. What would really surprise me is if Strawberry launches without making much of an impression either way

9. This is not financial or eschatological advice, and as always, I don’t really know what I’m talking about. I am, after all, just a bear. If you want to see what I think about the future check out my Manifold account where I use fake money to make predictions about the future.

I am quite poor, spend many hours a week on this blog, and make it available for free. Your paid subscription and help getting the word out would be greatly appreciated.