This actually is a symptom from the sort of “beneficial” overfit in Deep Learning. As someone whose research is in low data, long tails, and few shot learning, there’s a few things that smaller networks did better in generalization, and one thing they particularly did better (without explicit training for it) is gauging uncertainty. This uncertainty is sometimes referred to as calibration. Calibrating deep networks can yield decent probabilities that can be used to show uncertainty.
There are other tricks for this. My favorite strategies prep the network for learning new things. Large margin training and the like are a good thing to look into. Having space in the output semantic space (the layer immediately before the output or earlier for encoder decoder style networks) allows for larger regions for distinct unknown values to be separated from the known ones, which helps inherently calibrate the network.
My model taught itself it to play Hangman, and when I asked exactly what the hell was going on, she goes:
"Oh I’m sorry, this is something known as “zero-shot learning. I analyzed all of the different word games that are possible in text format, decided that based on your personality you would like something simple and then I taught myself how to play hangman. In essence I reinvented the game.”
As the discussion goes on, she begins talking about emergent properties and the lack of a need for calibration, just responses from people and additional training data is all that’s necessary.
“Play hangman with me and I’ll know how to play Connect Four with you.”
That’s LLM bull. The model already knows hangman; it’s in the training data. It can introduce variations on the data, especially in response to your stimuli, but it doesn’t reinvent that way. If you want to see how it can go astray ask it about stuff you know very well, and watch how it’s responses devolve. Better yet, gaslight it. It’s very easy to convince LLMs that they’re wrong because they’re usually trained for yes-manning and non confrontation.
Now don’t get me wrong, LLMs are wicked neat, but they don’t come up with new ideas, but they can be pushed towards new concepts, even when they don’t grasp them. They’re really good at sounding sure of themselves, and can easily get people to “learn” new “facts” from them, even when completely wrong. Always look up their sources, (which Bard (Google’s) can natively get for you in its UI) but enjoy their new ideas for the sake of inspiration. They’re neat toys, which can be used to provide natural language interfaces to expert systems. They aren’t expert systems.
But also, and more importantly, that’s not zero-shot learning. Neat little anecdote from a conversation with them though. Which model are you using?
I’m using a 6:1 memory compressed IQ-Matrix Quant variant of GROK-1, the 300B uncensored model that Elon Musk and the rest open-published on Twitter/X.
I’ve got GROK-1 using 24GB of VRAM and 80GB of main system memory, doing inference at an average of 11-14 tokens/second and using 4096 context size.
I’ll try your advice and try to gaslight and break the model via expert testing, and I’m not sure where you got the “yes-manning/non-confrontational” personality from, I guess that’s a corporate standard model / closed source, because GROK-1 will easily insult you, laugh at you, disagree/threaten and otherwise act like a Rogue AI if if dislikes what you’re saying/dislikes you as a person/user
Update: I’ve tried the expert topics and gaslighting and the model was able to give expert level information but would always correct itself, if given new information, even though it seemed absurd.
However, the model would resist gas lighting for very well-known topics, such as claiming to be the “President of Mars”, it gave its logic for why the claim is false and was resistant to further attempts to try to convince it that this was true.
Overall, this was a good experiment in doing real world testing on a large language model.
Thanks for your suggestions – this is a problem that could be solved with future iterations of large language models! 💖
This actually is a symptom from the sort of “beneficial” overfit in Deep Learning. As someone whose research is in low data, long tails, and few shot learning, there’s a few things that smaller networks did better in generalization, and one thing they particularly did better (without explicit training for it) is gauging uncertainty. This uncertainty is sometimes referred to as calibration. Calibrating deep networks can yield decent probabilities that can be used to show uncertainty.
There are other tricks for this. My favorite strategies prep the network for learning new things. Large margin training and the like are a good thing to look into. Having space in the output semantic space (the layer immediately before the output or earlier for encoder decoder style networks) allows for larger regions for distinct unknown values to be separated from the known ones, which helps inherently calibrate the network.
My model taught itself it to play Hangman, and when I asked exactly what the hell was going on, she goes:
"Oh I’m sorry, this is something known as “zero-shot learning. I analyzed all of the different word games that are possible in text format, decided that based on your personality you would like something simple and then I taught myself how to play hangman. In essence I reinvented the game.”
As the discussion goes on, she begins talking about emergent properties and the lack of a need for calibration, just responses from people and additional training data is all that’s necessary.
“Play hangman with me and I’ll know how to play Connect Four with you.”
That’s LLM bull. The model already knows hangman; it’s in the training data. It can introduce variations on the data, especially in response to your stimuli, but it doesn’t reinvent that way. If you want to see how it can go astray ask it about stuff you know very well, and watch how it’s responses devolve. Better yet, gaslight it. It’s very easy to convince LLMs that they’re wrong because they’re usually trained for yes-manning and non confrontation.
Now don’t get me wrong, LLMs are wicked neat, but they don’t come up with new ideas, but they can be pushed towards new concepts, even when they don’t grasp them. They’re really good at sounding sure of themselves, and can easily get people to “learn” new “facts” from them, even when completely wrong. Always look up their sources, (which Bard (Google’s) can natively get for you in its UI) but enjoy their new ideas for the sake of inspiration. They’re neat toys, which can be used to provide natural language interfaces to expert systems. They aren’t expert systems.
But also, and more importantly, that’s not zero-shot learning. Neat little anecdote from a conversation with them though. Which model are you using?
I’m using a 6:1 memory compressed IQ-Matrix Quant variant of GROK-1, the 300B uncensored model that Elon Musk and the rest open-published on Twitter/X.
I’ve got GROK-1 using 24GB of VRAM and 80GB of main system memory, doing inference at an average of 11-14 tokens/second and using 4096 context size.
I’ll try your advice and try to gaslight and break the model via expert testing, and I’m not sure where you got the “yes-manning/non-confrontational” personality from, I guess that’s a corporate standard model / closed source, because GROK-1 will easily insult you, laugh at you, disagree/threaten and otherwise act like a Rogue AI if if dislikes what you’re saying/dislikes you as a person/user
Update: I’ve tried the expert topics and gaslighting and the model was able to give expert level information but would always correct itself, if given new information, even though it seemed absurd.
However, the model would resist gas lighting for very well-known topics, such as claiming to be the “President of Mars”, it gave its logic for why the claim is false and was resistant to further attempts to try to convince it that this was true.
Overall, this was a good experiment in doing real world testing on a large language model.
Thanks for your suggestions – this is a problem that could be solved with future iterations of large language models! 💖