alt Hacker NewsYes they are approaching the limits, try asking smaller models niche questions about almost anything, they hallucinate massively because you cannot simply pack in all the raw knowledge from a massive frontier model into something that’s quantified down to 20GB etc.
It breaks fundamental laws of information theory. It’s like saying you can extract 100 joules of energy from 10 joules of energy source. Not possible.
Replies
Sounds like you are describing a quantized model which is a naive form of compression, not a model that is trained more efficiently.
Additionally the information theory angle is for information storage, but a model can access resources and tools to gain information and what we are really seeking to train is reasoning not information retrieval. We reduce the needs to the right capabilities and we don’t get upset if it does not know the lyrics to every song ever written.
It doesn't really matter though. Hardware performance is still growing. The new Mac Studio could just about run this model locally (rather slowly) - something that sits on your desk, that you as a consumer can buy.
Imagine prosumer desktop hardware 10 years from now. The 2036 DGX Spark. For a few thousand dollars you will be able to buy something with hundreds of GB (maybe TB if manufacturers step up) of unified RAM, memory bandwidth in the 10-20TB/s range. Overall AI "compute" will increase 10-20x, while at the same time AI model capability per byte will increase 5-10x.
The hardware would fit today's models, something like Kimi K3, quite comfortably and give performance of maybe 100 tokens/second. So what needs data center hardware today will run on your desk.
But if we also assume the models become more efficient, a 2036 Fable-class model (in terms of intelligence/capabilities, not size) will easily run on this thing at hundreds of tokens per second.
Unfortunately it'll still slow to a crawl with 5 Chrome tabs open, and every Electron app will need at least 200GB of RAM.