[atleastoptimal]

Far more is being done than simply throwing more GPU's at the problem.

GPT-5 required less compute to train than GPT-4.5. Data, RL, architectural improvements, etc. all contribute to the rate of improvement we're seeing now.

[4gotunameagain]

The very idea that AGI will arise from LLMs is ridiculous at best.

Computer science hubris at its finest.

[atleastoptimal]

Why is it ridiculous that an LLM or a system similar to or built off of an LLM could reach AGI?

[4gotunameagain]

Because intelligence is so much more than stochastically repeating stuff you've been trained on.

It needs to learn new information, create novel connections, be creative.. We are utterly clueless as to how the brain works and how intelligence is created.

We just took one cell, a neuron, made the simplest possible model of it, made some copies of it and you think it will suddenly spark into life by throwing GPUs at it ?

[atleastoptimal]

>It needs to learn new information, create novel connections, be creative.

LLM's can do all those things

[4gotunameagain]

Nope. Can't learn anything after the training data, only within the very narrow context window.

Any novel connections are through randomness, hence hallucinations instead of useful connections with background knowledge of involved systems or concepts.

About creativity, see my previous point. If I spit out words that go next to eachother, it won't be creativity. Creativity implies a goal, a purpose, or sometimes by chance, but utilising systematic thinking with understanding of the world.

[atleastoptimal]

I was considering refuting this point by point, but it seems your mind is already made up.

I feel that many people who deny the current utility and abilities of large-language models will continue to do so far after they've exceeded human intelligence, because the perception that they are fundamentally limited, regardless of whether they actually are or if their criticisms make any sense, is necessary for some load-bearing part of their sanity.

[leptons]

If AGI is built from LLMs, how could we trust it? It's going to "hallucinate", so I'm not sure that this AGI future people are clamoring for is going to really be all that good if it is built on LLMs.

[NuclearPM]

Humans are wrong all the time too. Intentionally and unintentionally.

[saubeidl]

Because LLMs are just stochastic parrots and don't do any thinking.

[atleastoptimal]

Humans who repeatedly deny LLM capabilities despite the numerous milestones they've surpassed seem more like stochastic parrots.

The same arguments are always brought up, often short pithy one-liners without much clarification. It seems silly that despite this argument first emerging when LLM's could barely write functional code, now that LLM's have reached gold-medal performance on the IMO, it is still being made with little interrogation into its potential faults, or clarification on the precise boundary of intelligence LLM's will never be able to cross.

[saubeidl]

Which novel idea have LLMs brought forward so far?

[atleastoptimal]

> Claim: gpt-5-pro can prove new interesting mathematics.

>Proof: I took a convex optimization paper with a clean open problem in it and asked gpt-5-pro to work on it. It proved a better bound than what is in the paper, and I checked the proof it's correct.

https://x.com/SebastienBubeck/status/1958198661139009862

(please excuse the x link)

[leptons]

Call me back when LLMs stop "hallucinating" constantly.