Bulls vs Bears

Let's hear from the opposition on my Massive Uncertainty thesis

My whole schtick so far with AGI Friday is that no one has much clue where things are headed, AI-wise. Everything from Star Trek to The Matrix is on the table, including the possibility that nothing meaningfully changes for many years. But what if I'm wrong about this?? There sure are a lot of highly confident commentators out there. Of course they're highly confident in opposite directions, which makes me all the more certain about the uncertainty.

Here's one small facet of the disagreement:

Will deepfakes reach the point that you can give an AI a picture of your friend and ask it for a fake photo of your friend getting their toenails painted purple by Ryan Gosling in footie pajamas? And have it immediately generate something that actually fools people? Some people are very confident this is less than three years away (some of them were also very confident in 2023 that this would happen in 2024).

The broader disagreement is between what I call the AI Bulls and the AI Bears. The bulls believe, for example, that scale is all you need. That we’re on a trajectory towards AGI and that it will emerge inexorably by scaling up existing neural net architectures. (If you believe that it’s scale plus a handful of additional breakthroughs that the time is ripe for, that still puts you pretty far towards the bull end of the bear-bull spectrum.) The AI bears believe that leaps in capability and seeming-understanding shown by DALL-E and Claude and the like over the last several years are going to hit another wall before reaching human level and that AGI is thus still decades away.

Note that the bear-bull spectrum is orthogonal to the spectrum of beliefs about how hard AI alignment is. For example, an AI bear may agree that if we did somehow hit upon AGI in a few years, that would literally destroy humanity. Likewise, some AI bulls believe that AI doom is nigh and others believe that AI utopia is nigh. (See last week’s post on AI Risk and the Technological Richter Scale which maps out both dimensions of this uncertainty.)

Gary the Bear

I think the best representative for the AI bears is Gary Marcus, who believes generative AI is a dead end. I’ve been (one-sidedly) beefing with Gary Marcus since 2022 when he mocked Scott Alexander for being “snookered” by Google. Before that, I was totally on board with Marcus’s reasoning. Prior to 2022, I would've used simple questions like “what’s bigger, your mom or a french fry?” to immediately expose an AI as having no genuine understanding. Then, starting in 2022 with Google’s PaLM, LLMs have been able to answer questions like that.

Gary Marcus made a ton of sense in 2021 and has since seemed to fail to update his views in light of new evidence and seems to keep being wrong in his predictions (as I was) about what AI will be able to do. He ridiculed Scott Alexander’s credulity but Scott keeps looking gradually more and more correct. In any case, we’re about to find out. I wish Gary Marcus would commit to a line in the sand beyond which he would agree he was wrong. Because so far it looks like he’ll keep moving the goalposts until sentient robots carry him away screaming that this does’t count as intelligence because they’re ultimately just following dumb algorithms.

His latest prediction is that GPT-5 isn’t coming. Of course he predicted the same for for GPT-4. But just because he’s been wrong so far doesn’t mean he won’t ultimately be proven right. If AI plateaus below human-level, we can forgive his errors in predicting exactly where the wall would be (just like we can forgive the bulls who predicted deepfakes of anyone doing anything in 2024, if it happens in 2027 instead). His certainty is what’s less forgivable.

(Interestingly, Gary Marcus has been in favor of pausing training of AI frontier models and regulating AI, not for fear of misaligned AGI but because of short-term risks of deploying unreliable AI.)

Next week: the bulls.

In The News

• Elon Musk’s AI company, xAI, released Grok 3 this week and it... exceeded expectations. Which doesn’t say a lot. But it is generally considered at least near the frontier for LLMs. My prediction is that it will fall behind again. My recommendation for normal people is to stick with Anthropic’s Claude.ai.

• Also this week, Google announced an “AI co-scientist”. Again, I’m skeptical.

• For something I’m bullish about, just today the Beeminder community has a way to connect LLMs directly to Beeminder. This has me excited about my nannybot vision for Ultimate Productivity, but that’s a story for another blog.