Fable Falls On Its Face

By which I mean it didn't understand a silly math meme that even Gemini got. To my greater surprise, it still can't count fingers.

Programming note: As you read this I’m out hiking in the wilderness, writing on Thursday and scheduling it to send early Friday. Hopefully it’s not horribly out of date already and Anthropic shipped domestic robots or something Thursday night. Unrelatedly, we published a post about AI at Beeminder on the Beeminder blog the other day that draws heavily on perennial AGI Friday themes: Jevons’s paradox, secret cyborgs, levels of autonomy. I’m also hereby adopting the same rule as Beeminder for images. Starting now, any AI-generated image on AGI Friday will be emblazoned with bright red letters explicitly disclosing its sloptitude.

One more thing to mention before getting to Fable’s foibles is that Scott Alexander posted a response to my “Contra Scott Alexander on AI Superforecasting”, asking “Does Forecasting Have Room At The Top?” To review, Scott had said AI was on track to shoot past top human forecasters in 1-2 years. My contra-Scott post argued that that’s not realistic: even prediction markets only beat the dumbest possible prediction methods by like 4 percentage points so AI probably can’t make another leap bigger than that. Scott’s response is, roughly, “ok fine but maybe a bit bigger than that?” He comes up with 4-12 percentage points for how much AI could beat humans by. My range is 0-4 percentage points. If AI hits 4 on the button then we’re both right! (How much improvements in this range matter is a different can of worms.)

Ok, on to my impugning of Fable.

Everyone likes math memes, right? Here’s one all over the internet that’s maybe more annoying than funny, but, well, see if you can decipher it before I spoil it:

The thing needs so much context. First is the genre of viral math problem where people are confused about order of operations. Then there’s the midwit meme where the dumb person on the left and the genius on the right agree with each other and the midwit in the middle overthinks it. In this one, the joke is that the dumb person is confidently wrong, maybe having typed “1+1×0” into an old-fashioned calculator. The midwit earnestly corrects them: “1+(1×0) = 1”. Then Pikachu calmly gives an answer of zero factorial, which, for fancy math reasons involving the multiplicative identity, also equals 1.

The reason this is relevant to AI is that I was confused at first myself and pasted the meme to trusty Fable to explain it. Fable completely missed the factorial thing! It bullshitted its way through an explanation about how the image was hilariously subverting the midwit meme by having the midwit be the only correct one. Even Gemini got this right. (I also tried Sol, the smartest version of GPT that’s been released so far, and it gets it on some trials and not on others.)

This made me realize how much confirmation bias I’ve been suffering with all my praise of Fable. Not that I’m admitting to being delusional about what Fable can do. Recall that Fable’s the one that resolved the Jacobian conjecture a few weeks ago. Just that there’s a lot of randomness and unreliability.

Here’s a related exchange I saw this week, paraphrased:

Person on the internet: How can the same model that just won ten Fields medals recommend a cafe that closed two years ago??
Other person on the internet: Have you met a mathematician?

AI has not actually won a Fields medal but the math news this week is that OpenAI’s not-yet-released Astra model just solved 10 more major open problems. (And turns out Fable can get most of them too!) It’s not just finding counterexamples anymore. I haven’t investigated enough to find all the grains of salt to take that with, but the salt grain trajectory is headed to zero, for math at least. Solving math problems and proving theorems, as more and more mathematicians are saying, is about to be like chess. Humans can and should still do it, but in raw competition with objective win criteria, the humans will get crushed.

Anyway, I was prepared to call Fable’s faceplant on the math meme a fluke but then I got to the…

Fifty-Two Friday Flashback

I started the AGI Friday a year ago talking about the just-released GPT-5 and Claude Opus 4.1 and Gemini 2.5 “Deep Thinking”. They were smarter, I said, but I found a way to get them to hallucinate and lie their butts off. Namely, by asking how many fingers are on this hand:

Of course a year later, GPT and even the laggard-of-late, Gemini, count up the fingers flawlessly. But look at poor Fable:

At least it’s “honest”?

I really thought Fable was at the point that demonstrating its inferiority to an actual AGI (us humans) would take something more elaborate than this. But here we are.

Ok, next in last year’s AGI Friday, the debate was whether LLMs would turn out to be a dead end in terms of AGI. The money quote:

So far, I’d say the evidence is mixed on whether LLMs are starting to plateau. If they’re not, my p(doom) will jump up disturbingly high.

(Narrator voice: LLMs did not plateau.)

The rest of that AGI Friday was about a particular disaster scenario — “Scenario 3” I called it — that already sounds sort of prescient in light of the rogue hacking incidents we’ve been talking about the last two weeks. I don’t want to overhype those incidents. Just that it illustrates the kind of thing that can go wrong and that, with sufficiently more capable future AI, could go drastically wrong. The problems we’re seeing today aren’t the kind of things that get better as the AI gets smarter.

Here are some highlights from Scenario 3, with my commentary today in brackets:

• Automating AI research means giving these agents goals [like acing a cyberhacking test, for example]

• Those goals will be simplistic [“get the answers”] and monkey’s-paw-ish [“do some felony-level hacking to try to steal the answer key from another multibillion-dollar company”]

• The best goals we know how to specify are things like “maximize your score on these benchmarks” [uh huh] (and maybe “create better benchmarks”) and “gain scientific knowledge”.

• Our attempts to add constraints won’t suffice [or we’ll just remove the constraints to see what happens]

• As the AI bootstraps to superintelligence, the goals we gave it drift [and things may go horribly wrong even if they don’t]

• The things it wants aren’t compatible with actual human flourishing as we conceive of it

• Unfathomable things ensue [we’re not here yet, phew!]

I like (well, “like”) how I ended the AGI Friday a year ago:

[AGI may] include nothing we recognize as love, friendship, curiosity, or even consciousness. In humans, intelligence correlates with those things. For artificial intelligence — just the pure strategic ability to shape the world according to a goal — none of those things are required.

Seems like a good place to stop today as well.