For coding AI, it could make sense to specialize models on architecture, functional/array split from loopy solutions, or just asking 4 separate small models, and then using a judge model to pick the best parts of each.
I haven’t watched the video yet, but I have to say: running a personal LLM on my computer using products like gpt4all produces some really awesome results im very happy with.
I can totally envision a future where everyone can easily run their own local AI in the next ten years.
I’m not going to watch the video — I like most context in text rather than video form — but while I will very well believe that:
-
It’s possible to optimize LLMs to make smaller models more effective than they are today. It would be very surprising if they were already optimal, given that the field is immature.
-
It’s possible to do a series of smaller, specialized models and keep models not-relevant to the current context unloaded from VRAM — I believe that the “splitting into smaller specialized networks” approach is referred to as Mixture of Experts. This should improve memory efficiency for many problems.
…this is countered by the fact that once you free up resources, I also suspect that you can then go use those now-available resources to improve the model by shoveling more data into the model. And while there might be diminishing returns, I very much doubt that there is a hard cap on which one can get better results by throwing more knowledge at a problem.
MoE
This optimization is about smaller matrix multiplications. Experts will specialize on input token types, and while it is better at being split accross resources (GPUs), it is not really specialization on “output domain” (type of work). All experts need to be in memory.
Deepseek made a 7b math focused LLM that beat all other models on math benchmarks, even 540b math specialist LLMs. More than any internal speed/structure “tricks”, they achieved this through highly curated training data.
The small models we get now tend to just be pruned from larger generalist models. Paper/video is suggesting smaller models that are “large tuned” post trained to be domain specialists. Large models could select from domain specialist models and only load those in memory or act as a judge in combining outputs of “sub models”
Where an LLM is a giant probabilistic classifier, there are much faster/accurate/less compute intensive deterministic classifiers (expert/rule systems). Where SLMs have advantages, using even cheaper classification steps is going in the same direction. A smaller LLM is automatically a faster classifier, as a hammer to bang on everything alternative.
Yeah that tracks from what I’ve seen. There were some very interesting new approaches that could improve the base framework of all generative AIs but at this time MoE is the one important improvement that Deepseek pioneeredfor LLMs. I wonder if throwing knowledge at the problem might actually net us a bit more of an elegant solution but MoE is kind of the only thing that helps us scale LLMs.
-
