I’m excited about Jev appearing on the scene, but perhaps not for the common reasons. What excites me is that all the human potential that has arrived to the AI tech now has a taste of what non-LLMs look like and what they can do.
Shoehorning LLMs everywhere has lead to general slowness and increasing expenses for some mundane jobs, and replacing them with proper machine learning tools will certainly benefit everyone.
Just to recap, Jev-like models take an input like:
“In front of me is a small animal with fluffy fur and a bushy tail”
and some options together with instructions what triggers them like:
“what_animal”: choice
“shark”: Big fish, big teeth, predator
“spider”: Has 8 legs, insect-like, some are venomous
“squirrel”: Small mammal, furry, notoriously bushy tails, can carry diseases
“crocodile”: Big lizard, unexpectedly fast, carnivore
“is_dangerous”: Can the animal seriously injure or kill a man
“should_run_away”: Should I run away from it if I encounter it, to avoid death or injury
And then it outputs a result like:
what_animal: squirrel, confidence: 0.9
is_dangerous: 0.05 (interpretation: NO)
should_run_away: 0.02 (interpretation: NO)
If you want to apply it to something that doesn’t fit in a sentence or two, you need the data chunked, and then ask the same question on each chunk. This is the part where it might or might not be more efficient that just using a LLM.
The next issue is depth of knowledge. You can’t cram a lot of information into this type of context, and in many cases, you just can’t pass all pertinent information about the situation in it - some things the model needs to be trained for. Luckily, these models are small enough that training them is much easier than for LLMs.
There are several Open-source implementations of the concept - we can’t yet be 100% sure it’s exactly the same one as Jev (both are based on BERT), but they do similar things, sometimes better:
I’m not looking at this as some huge news, the most important thing is that it could lead to a shift in thinking that LLMs are the only thing out there for AI.
BERT, the famous pre-LLM language model, after all, is 2018 technology.