GPT 5.2 is the first HUMAN LABOR replacement

Can you imagine, at one point we won't be smart enough to tell is something has gotten better in this space....

the "and wells fargo is on the list" comment killed me looool 10:10

People often criticize predictive language modeling by saying it doesn’t “know” answers and is therefore “just making things up.” That criticism misunderstands how both reasoning and knowledge actually work. When a computer answers “2 + 2 = 4,” it isn’t recalling a memorized fact in the human sense. It applies formal rules to arrive at a result. When a child who can’t yet read fluently sounds out a word letter by letter, they are not guessing randomly—they are applying learned patterns of language to reach the correct word. A skilled reader recognizes the word instantly, but the underlying goal is the same. Large language models operate in a similar way. They do not invent answers arbitrarily. They generate responses by applying learned statistical and structural patterns derived from vast amounts of real language, facts, and reasoning examples. The predictions are constrained by those patterns, just as arithmetic is constrained by mathematical rules and reading is constrained by phonetics and grammar. Prediction, in this context, does not mean randomness. It means selecting the most likely correct continuation based on structure, consistency, and prior knowledge embedded in the model. When an LLM gives a correct answer, it has followed a disciplined process grounded in learned relationships—not imagination. The key point is this: methods don’t matter—results do. Whether an answer is reached through calculation, phonetic decoding, memorization, or predictive modeling, the value lies in the correctness of the outcome. Prediction is not a weakness; it is simply another valid mechanism for reasoning toward truth. If the answer is right, the path taken to get there is irrelevant.

WOW the zip file result is CRAZY it's like a person sending you their work for approval.

"Wells Fargo is on the list." - I got that one, Wes.

Launch your site for free at https://framer.link/WesRoth Use code WESROTH for a free month on Framer Pro. Build a site that looks hand-coded. Without hiring a developer.

21:35 Philosophy enthusiast here. In common parlance (no pun intended), "common" can suggest a majority occurrence, eg: "people commonly do X" can imply: "tend to more than not". So the phrase "not uncommon" is used to signal that an event is moderately frequent rather than rare, but avoids the suggestion that it might be "more often than not" that can be implied by the word common.

Yep, I'm convinced GDPval is the most important benchmark around, in terms of imminent real-world impact.

What “shot” means in AI academia: In mainstream ML/LLM usage, a “shot” means an example provided in the model’s input context (the prompt), as a demonstration of the task. Zero-shot: no task examples in the prompt (just instructions, maybe constraints). One-shot: exactly one example (input → output) in the prompt. Few-shot: a small number of examples in the prompt (often 2–20ish; not a strict boundary).

Wow great video Wes! Thanks for keeping tabs on all these new developments across all these models. It’s gotta be a full time job. Between the latest versions of Claude, Gemini, Grok, and now Chat, well they each have their specialties, but then that keeps changing every several weeks. Any way you can compile your personal top 3-10 models and what you like most of each? Of course, that runs stale after a month or two, but it would be very helpful to gain your insights. Happy Holidays!