Re:Behind the curtain: what it feels like to work in AI righ...
by kajecounterhack on Friday April 07, 2023 @02:58AM
(#35478918)
> I can see that these models are significantly better than anything before, but I can’t see yet the “killer app”.
I get what you mean and think there are a lot of early explorations using models that are far from the terminal state (most of them are text only). Just keep in mind that we just got multi-modal GPT; what's going to happen when we have large models that can ingest YouTube? Folks are overly focused on the current state of LLMs and their textual nature here and not thinking about what's going to happen when more multi-modal models find their uses.
I don't know about killer app but a big use case for large models is for them to become "the mother of models." You could train big models that learn stuff like "how to drive a car" -- a big model could take image (+sound? +map & other metadata?) inputs, and output what it would do in different situations / describe scenes. You could then distill this understanding into models that actually go on a self-driving car. All the work self-driving companies have put into evaluation infra can be repurposed to serve this scheme.
You could do something similar with spam: you could refine a site's content policy and train a large model to understand deeply what might constitute spam or a rule violation, and even point back to rules and understand nuanced policy. Sure, such a model is a horrible idea if used alone, but what if you took the infra you built for ML models used in that way today, and essentially used the large model as a helpful representation teacher / discovery mechanism? The ML models used in such systems today are already subject to bias and whatever mitigations we're using for them today can be used with large models in many cases and maybe even improved.
The things that excite me personally are:
1. RLHF is active learning, so these models can _get better fast_. This threatens any ML system today that's suffering from their inability to retrain/incorporate new information quickly, or suffering from a lack of supervised training data.
2. They learn unprecedented representations for things and can draw connections between things to make implementing nuanced programs easier than ever. Did you think it was impressive that a cat could be clustered close to a tiger with paragraph vectors? What if you could show a large model an image of a road full of cones and ask it to find the drivable lanes? What if a model could introspect and get better at that task with even less human involvement? What if large models' deep visual understanding starts to unlock things like manipulating soft objects with robotic e.g. make it easier to communicate goal states or instructions?
3. End users don't ever have to interact with the raw models for them to be useful and in fact a great way to use large models is to have them replace existing ML systems which are now ubiquitous in your life. They learn better representations and can be distilled to replace large models, or provide more metadata to build heuristics on than existing models today. And while they are expensive to train, they could get cheaper in several ways, including that they already reduce the need for large labeled datasets which used to be a big bottleneck.
The limitations of a chat interface, the limitations of the models being big, the limitations of how expensive these are to train -- they all feel like red herrings vs the true promise of universal representations that can hydrate tons of uses and that should become more accessible over time.
I get what you mean and think there are a lot of early explorations using models that are far from the terminal state (most of them are text only). Just keep in mind that we just got multi-modal GPT; what's going to happen when we have large models that can ingest YouTube? Folks are overly focused on the current state of LLMs and their textual nature here and not thinking about what's going to happen when more multi-modal models find their uses.
I don't know about killer app but a big use case for large models is for them to become "the mother of models." You could train big models that learn stuff like "how to drive a car" -- a big model could take image (+sound? +map & other metadata?) inputs, and output what it would do in different situations / describe scenes. You could then distill this understanding into models that actually go on a self-driving car. All the work self-driving companies have put into evaluation infra can be repurposed to serve this scheme.
You could do something similar with spam: you could refine a site's content policy and train a large model to understand deeply what might constitute spam or a rule violation, and even point back to rules and understand nuanced policy. Sure, such a model is a horrible idea if used alone, but what if you took the infra you built for ML models used in that way today, and essentially used the large model as a helpful representation teacher / discovery mechanism? The ML models used in such systems today are already subject to bias and whatever mitigations we're using for them today can be used with large models in many cases and maybe even improved.
The things that excite me personally are:
1. RLHF is active learning, so these models can _get better fast_. This threatens any ML system today that's suffering from their inability to retrain/incorporate new information quickly, or suffering from a lack of supervised training data.
2. They learn unprecedented representations for things and can draw connections between things to make implementing nuanced programs easier than ever. Did you think it was impressive that a cat could be clustered close to a tiger with paragraph vectors? What if you could show a large model an image of a road full of cones and ask it to find the drivable lanes? What if a model could introspect and get better at that task with even less human involvement? What if large models' deep visual understanding starts to unlock things like manipulating soft objects with robotic e.g. make it easier to communicate goal states or instructions?
3. End users don't ever have to interact with the raw models for them to be useful and in fact a great way to use large models is to have them replace existing ML systems which are now ubiquitous in your life. They learn better representations and can be distilled to replace large models, or provide more metadata to build heuristics on than existing models today. And while they are expensive to train, they could get cheaper in several ways, including that they already reduce the need for large labeled datasets which used to be a big bottleneck.
The limitations of a chat interface, the limitations of the models being big, the limitations of how expensive these are to train -- they all feel like red herrings vs the true promise of universal representations that can hydrate tons of uses and that should become more accessible over time.