It's all fucking flame, and LJL at anyone who doesn't get that by the midpoint of 2026.

(http://www.autoadmit.com/thread.php?thread_id=5868844&forum_id=2#49899270)

You could think of it as the ability to reason, rather than to just complete tasks in response to prompts. It would include things like having a world model, a theory of mind, etc. It's what literally every "AI" company professes to be trying to develop.

(http://www.autoadmit.com/thread.php?thread_id=5868844&forum_id=2#49899364)

as penis says "agi" is a meaningless concept for nerds to jack off to in the anthropic bathroom

no one ever thought the capex was for anything other than code writing / drug discovery / 60 second animal cuck cartoons made by indians. that's far from flame and flame was never needed to sell people on the capex.

(http://www.autoadmit.com/thread.php?thread_id=5868844&forum_id=2#49899312)

humans are just meat robots running connectionist learning algorithms. once you accept that fact, the increasingly general capabilities of modern AI from burning compute cycles is unsurprising and strongly implies near-term AGI. there's simply too much compute and data available, so even garbage learning algorithms with crude training objectives (next token training on trillions of random tokens) leads to increasingly powerful systems.

(http://www.autoadmit.com/thread.php?thread_id=5868844&forum_id=2#49899291)

i think your "poking around and creating a new conceptual understanding" claim is most vulnerable here. If a model independently finds a construction that humans had not used to settle a longstanding conjecture, then by ordinary behavioral standards it has done something very close to “poking around” and “creating a new conceptual understanding.” Especially when it instantiates a new cross-domain solution path that no human ever saw.

maybe it didn't do it with humanlike curiosity, it didn't have the phenomenological experience of sitting with the problem, feeling the sense of strain that something was off, and didn't reach the felt moment of insight. Maybe it needed a specific prompt, and AI grading and flagging then later human verification, but I wouldn't consider those things to fall under the category of "intelligence".

another deeper issue is it seems you are smuggling in an origin criterion here. it seems you are saying that because the model was trained on vast data, RL, and synthetic thought traces that it's solution is "in-distribution", but that does not settle the question. humans are also trained on vast cultural data, examples, heuristics, and feedback. whether the system was trained doesn't seem to be the relevant question. the question should be whether it can generate valid, non-trivial solutions to real problems outside of memorized cases and it clearly can. i agree with you that it is still much weaker in terms of full research autonomy and economic subsitutability (where you can drop it in a firm and completely replace roles under deadlines etc.), but to me that is a different question from "general intelligence".

edit: actually I just thought of another angle on one of your points that I agree with now. the fact that humans can update their competence at any point, learn new things from lived interaction while current base models mostly can't change their weights during deployment, it needs "new training" probably for certain novel activities or for things the weights or guardrails bias against. I think that is a genuine aspect of "intelligence" humans have that these systems are lacking in, because we do say someone is more intelligent due to their ability to learn new things

(http://www.autoadmit.com/thread.php?thread_id=5868844&forum_id=2#49899337)

(http://www.autoadmit.com/thread.php?thread_id=5868844&forum_id=2#49899382)