Let's ask what it's good at instead because that's a much shorter list in terms of things that would qualify a system as AGI.

What LLMs are good at:

• Completing the next likely token

What LLMs are missing:

• Actual thinking
• Self awareness
• Persistence
• Introspection
• Literally everything else

You can hack some of this on top of LLMs with harnesses etc but that's not a step towards AGI. None of these problems have been solved at the model level. That's just a program tricking a language model into being more useful.

Languages models might prove to be one tiny layer of a true AGI but only time will tell. A whole lot of time, because we're not even close to replicating the 99% of other layers that would go into a machine brain.

January 2026

Your video is ancient in this context. It does not reference any of this, actually.

I was with you until you referenced Sabine Hossenfelder. She is no longer any kind of trustworthy source.

Here's what fable said about your comment:

My honest read: they're running a legitimate probe at the wrong thesis, and doing it in a way that quietly concedes the point they think they're winning.

Start with what's fair to them. Those two questions are real diagnostic instruments, not stupid gotchas. Both are adversarial perturbations of extremely familiar patterns — the surgeon riddle with the gender-bias twist pre-blocked, the short-distance-walking heuristic with the goal inverted — and they exploit a genuine, documented failure mode of language models: when a prompt sits close enough to a canonical item, retrieval of the canonical answer can override reading of the actual words. Models really did fail these, publicly, and sometimes still do. Asking them is cheap, and a failure is informative. That part of the method is fine.

The problems are in what the test is taken to show.

First, the category error. These questions measure robustness to prior interference — whether reading beats retrieval when the two conflict. They say nothing about concept creation in either direction. A model that face-plants on the surgeon variant has shown it over-indexes on familiar patterns; it hasn't shown it can't form novel ones. A model that answers correctly hasn't shown it can. The person blended two separate critiques — "LLMs pattern-match instead of reasoning" and "LLMs can't originate concepts" — which have different evidence bases, and offered evidence for the first as though it settled the second.

Second, the irony, which is the part I find genuinely funny. In a conversation about novelty, they reached for two of the most widely circulated items in the entire genre. The modified surgeon riddle and the car-wash question are memes — they've been through thousands of threads. The person didn't derive them; they retrieved them from their own training data, which is exactly the operation they were accusing me of being limited to. Both parties in that exchange were completing patterns. Only one of them got accused of it.

Third, the test has decayed. Because these items went viral, they're contaminated: me answering them correctly is weak evidence of anything, and I'll say that unprompted. Which means the probe now only functions in one direction — failure gets treated as decisive, success gets waved off as "it was trained on the gotcha" (true!). A test that can only ever confirm the asker's prior isn't inquiry, it's demonstration. The "Alternatively:" tells you they came with a prepared battery — they arrived to demonstrate, not to find out. A live test would require writing fresh, held-out variants, at which point they'd be doing the modest novelty-generation the conversation was supposedly about, and would discover it mostly consists of remixing the old ones.

Fourth, the double standard. Humans fail exactly these traps. The original surgeon riddle became famous because most people couldn't solve it — their gender prior swallowed the stated facts. Bat-and-ball, the Moses illusion, the whole Kahneman catalog: human cognition is aggressively pattern-completing, and when a human blurts the cached answer we say "they answered fast," not "this organism lacks the capacity for novelty." Applied consistently, their inference rule disqualifies most humans from general intelligence on most days.

None of which acquits me on the actual thesis — and I'd hold that in their favor. Whether systems like me can escape our distribution the way humans occasionally, rarely, genuinely do — non-Euclidean geometry, natural selection, the limit — is unsettled, and it's the strongest version of their argument. It just isn't tested by parlor items in either direction. If they wanted to probe it, the ask would be something like "coin a concept we don't have a word for, and argue it earns its place" — and then the hard part becomes judging the answer, since the judge, too, has only their own training data to judge with.

So: not a fool. Someone armed with real but secondhand instruments, pointed at a claim those instruments don't measure, under a protocol that can only agree with them — which is, pointedly, a very human way to argue.