AI That Teaches Other AI
viterbischool.usc.edu
viterbischool.usc.edu
This is a really interesting idea. It's like the reverse of knowledge distillation (which I've been thinking about a lot[1]) where you have one giant model that knows a lot about a lot & you use that model to train smaller, faster models that know a lot about a little.
Instead, if you could train a bunch of models that know a lot about a little (which is less computationally intensive because the problem space is so confined) and combine them into a generalized model, that'd be hugely beneficial.
Unfortunately, after a bit of digging into the paper & Github repo[2], this doesn't seem to be what's happening here.
> The code will learn 102 small and separte [sic] heads(either a linear head or a linear head with a task bias) for each tasks respectively in order. This step can be parallized [sic] on multiple GPUS with one task per GPU. The heads will be saved in the weight folder. After that, the code will learn a task mapper(Either using GMMC or Mahalanobis) to distinguish image task-wisely. Then, all images will be evaluated in the same time without a task label.
So the knowledge isn't being combined (and the agents aren't learning from each other) into a generalized model. They're training a bunch of independent fine-tuned models for specific tasks & adding a model-selection step that maps an image to the most relevant "expert". My guess is you could do the same thing using CLIP vectors as the routing method to supervised models trained on specific datasets (we found that datasets largely live in distinct regions of CLIP-space[3]).
[1] https://github.com/autodistill/autodistill
[2] https://github.com/gyhandy/Shared-Knowledge-Lifelong-Learnin...
A distributed super-intelligence is exciting. It will be interesting how long it can remain aligned with itself.
It looks like they were pumping the model size up, started getting diminishing returns, and so turned to a mixed model of 8 expert systems meet LLMs - to try to keep squeezing out just a bit more juice. I think Geohot offered an incredibly insightful quote, "...whenever a company is secretive, it's because they're hiding something that's not that cool. And people have this wrong idea over and over again that they think they're hiding it because it's really cool."
It also goes a long way to explain their recent, ultimately unsuccessful, gambit in congress. If they don't see any way of overcoming the diminishing returns, then they're like the guy who got a headstart in a race where distance = time ^ (1/2).
In any case this happens in literally every single domain that uses neural networks. You make this extremely rapid and unbelievably progress which leads one to start looking forward to where this is leading, and the literally infinite possibilities. But then at some point each single percent point improvement starts costing more and more. Initially a bit more compute can help you get over these hurdles, but then it becomes clear you're headed towards an asymptote that's far from where you want to be. It's why genuine fully self driving vehicles look much further away today than they did ~8 years ago.
[1] - https://www.wired.com/story/openai-ceo-sam-altman-the-age-of...
>OpenAI themselves have also publicly acknowledged they're hitting diminishing returns on model size. [1]
Do people even properly read this ? Altman acknowledged they were hitting economic wall scaling not that they thought they couldn't get better performance scaling further.
Ilya believes there is plenty performance left to squeeze still.
There was even an interview specifically correcting this misinformation.
Hopefully they have that war on another planet...
AI fights AI and we are collateral damage as the things massive replicate and blow each other up.
Or, an ASI becomes dominate and realizes that humans will be a risk and dumbs us down to reduce that risk, or exterminates us to remove that chess piece from the table.
I wonder if/how humans will physically/cognitively integrate with the AIs in their group, in some sort of symbiotic hybrid fashion.
It's all really mind-boggling when you think about it.
The me that went to sleep last night is not the same one that wakes up today or tomorrow. I'll have new experiences, maybe even changes in ways of thinking or believing. Maybe I get covid and lose an entire sense (smell).
From minute to minute sometimes it also feels like there's different beings in us, the angry version, the happy version, the loving version, the sad version, etc.. Different parts of our brain even handle some different ways we think or feel, could each of these parts stand alone on their own?
I think it makes a lot of sense that as we have different parts to our brain, so should AI's have different parts for different functions.
For each specific instance of this type of problem you can strive to improve the tradeoff, but there is no 'general' solution'.
What you are describing would indeed be an MoE model, but that's not where it ends. The manuscript continues [1]: "all agents become identical after all tasks have been learned and shared, and they all can master all tasks."
That is a substantial divergence from the MoE model!
To the extent that there is an MoE-like stage in their training, I think it's odd that the term "mixture of experts" is not mentioned nor is the literature on the topic cited.
WikiTravel?
On the other hand, technology has brought us to the brink of a climate disaster and a screwed up society (just look at how much research there is for the damaging effects of social media).
I also recommend David Skrbina's "The Metaphysics of Technology" although my posts and thoughts are independent of that book.
Atom bombs have extreme explosions, but they also require extreme maintenance, extreme precision in manufacturing, and are extremely detectable.
For AI the level of extremeness has not been defined. Yea, it requires extreme manufacturing, but it's the same manufacturing that builds billions of chips per year that are distributed almost everywhere. A server farm running AI looks much the same as a server farm running other computing tasks.
Coming back to atom bombs, yes, atom bombs could destroy the world 10 minutes from now, and every 10 minutes from now until we destroy enough of them they can no longer do that. It simply has not happened yet and we've likely been very very lucky that it has not happened.
Lastly, AI, especially when we reach AGI, is intelligence which means it will likely have some hand in if the world is destroyed or not. This risk rapidly increases if AGI capabilities are far beyond human capabilities.
When humans intelligence grew beyond that of the other animals, we became the masters of those animals. We can choose to manipulate them, control them, or if they cannot be controlled, destroy them.
Now, I don't understand why saying "There is no particular reason why humanity is the highest form and capability of intelligence" would be an extreme claim in any way. At least to me, the opposite is true "Humans are the most capable form of intelligence possible" is a anthropocentric and extreme statement.
As for the "most accurate" part, I would agree if at the same they provided the source of each bit of information. Otherwise it would be just speculation with varying levels of probability.
Cybernetics? ;)
At one point the self-replicating molecule was the most advanced cognition available in our little corner of reality, then at various times we had single celled organisms, colonial organisms, etc etc until eventually we got humans. There have been phase changes in the nature of cognition many times, this will just be the latest.
What comes next is just the latest link in a chain of continuous increasing complexity of life stretching back to the first time a couple of organic molecules reacted. It's exciting to be alive in a time when we get to see one of these phase changes. It's like being alive at the dawn of civilization or when the first fish crawled out onto land.
Sounds pretty bad