OpenAI is moving towards a world model with all of physics and intricacies of the world compressed into latent space. Not enough people are seeing this as a big leap towards AGI but remember that OpenAI will not be working on anything that they don’t believe helps them achieve

The more I think about ways to allow AI models to generate original thought, the more I realise that I don't understand how they form in my own mind. Time for some crucial reflection and brain toy problems #AGI

Finally got to see the #aurora and got some stunning photos considering they are from an iPhone

LLMs today have a super-human ability to recall information and process large volumes of data. If we assume that the missing ingredient for AGI is their lack of long term planning capability then wouldn’t a system like GPT-4o with human-like planning simply be an ASI?

Makes sense to release the next evolution in model capability after the US elections to keep the heat off at the sensitive time when everything gets blown way out of proportion. Even if this is not quite an AGI precursor, having any planning capability would scare loads of people

With majority of scientists using ChatGPT in some stage of their research and OAI collecting their data to train future models, GPT-5 will probably be more familiar with the state of the art in every field of science than its experts

That is one big series B .. assuming that Musk will keep the team lean, half of the cash will go to Jensen and the rest to whoever is willing to sell data to them

I envy how much less crap US based founders have to put up with

Schedule-Free paper is up! arxiv.org/abs/2405.15682 Joint work with collaborators Alice Yang Harsh Mehta Konstantin Mishchenko Ahmed Khaled Ashok Cutkosky We have some strong small-scale experiments on Transformers, comparing to chinchilla-style cosine 10x reduction schedules.

Seems to imply that GPT-5 will be GPT-4 with something extra. Is it possible that the 4o voice model we were shown is actually what will be released as GPT-5? It seemed a lot smarter in their demos than what we're seeing in 4o

Really nice paper. I’ll be testing it out very shortly!

roon I thought this was interesting when it first came out. Timelines getting shorter since this released in January?

Looks like Sam Altman is quietly the king of e/acc doing the good work in plain sight

1000 years of progress in 5-10 years would surely require a crazy level of manufacturing and laboratory work ... unless the AI will be able to simulate the results 🤔