3 ms·

I think the "self-play" path is where the scary-powerful AI solutions will emerge. This implies persistence of state and logic that lives external to the LLM. T

I think the "self-play" path is where the scary-powerful AI solutions will emerge. This implies persistence of state and logic that lives external to the LLM. The language model is just one tool. AGI/ASI/whatever will be a system of tools, of which the LLM might be the least complicated one to worry about. In my view, domain modeling, managing state, knowing when to transition between states, techniques for final decision making, consideration for the time domain, and prompt engineering are the real challenges.

lern_too_spel 3y ago
It's not necessary for the author's purpose of providing more data. We're only training on one kind of input so far, text, from which these models have built some understanding of the world. Humans train on more inputs, and the data to provide those inputs for training a model is readily available, in far larger quantities than individual human brains consume. Data is not the issue.
- Xelynega 3y ago
  We're training on text because that's what we're making the model do. It's a fact of neural networks that to train them supervised you need the training data in the expected input for(vector of n thousand preceding tokens for LLMs) with the expected output(the next token for LLMs). "Training them on video" would mean converting the video to a format we can train the llm with, then training the LLM with that info. This would probably be a 1 OOM increase at maximum, if the video transcripts aren't already a part of the training data for gpt.
  - lern_too_spel 3y ago
    > This would probably be a 1 OOM increase at maximum, if the video transcripts aren't already a part of the training data for gpt. Human brains aren't trained on video transcripts, which leave out a lot of information from the video that human brains have a shared understanding of due to training. You would train on video embeddings and predict next embeddings, thus learning physics and other properties of the world and the objects within it and how they interact. This is many more than 5 orders of magnitude more data than the text that today's LLMs are trained on.

lern_too_spel 3y ago

It's not necessary for the author's purpose of providing more data. We're only training on one kind of input so far, text, from which these models have built some understanding of the world. Humans train on more inputs, and the data to provide those inputs for training a model is readily available, in far larger quantities than individual human brains consume. Data is not the issue.

Xelynega 3y ago
We're training on text because that's what we're making the model do. It's a fact of neural networks that to train them supervised you need the training data in the expected input for(vector of n thousand preceding tokens for LLMs) with the expected output(the next token for LLMs). "Training them on video" would mean converting the video to a format we can train the llm with, then training the LLM with that info. This would probably be a 1 OOM increase at maximum, if the video transcripts aren't already a part of the training data for gpt.
- lern_too_spel 3y ago
  > This would probably be a 1 OOM increase at maximum, if the video transcripts aren't already a part of the training data for gpt. Human brains aren't trained on video transcripts, which leave out a lot of information from the video that human brains have a shared understanding of due to training. You would train on video embeddings and predict next embeddings, thus learning physics and other properties of the world and the objects within it and how they interact. This is many more than 5 orders of magnitude more data than the text that today's LLMs are trained on.