4 ms·
At the very least, it needs to be able to collate training data, design, code, train, fine tune and "RLHF" a foundational model from scratch, on its own, and ha
by root_axis 1y ago
At the very least, it needs to be able to collate training data, design, code, train, fine tune and "RLHF" a foundational model from scratch, on its own, and have it show improvements over the current SOTA models before we can even begin to have the conversation about whether we're approaching what could be AGI at some point in the future.
- kadushka 1y agoI cannot do all that. Am I not generally intelligent?
- root_axis 1y agoYou could if you were trained to do so. LLMs cannot, even if they are.
- kadushka 1y agoAre LLMs trained to do all that? How do you know they cannot?
- root_axis 1y ago> Are LLMs trained to do all that Yes, they are trained to do all that more or less as much as anything else they are trained to do. > How do you know they cannot? The same way I know they can't run a fortune 500 company. It's clearly apparent.
- kadushka 1y agoLast time I checked they were trained to predict the next word. Why would you expect them to solve research problems?
- root_axis 1y ago> Why would you expect them to solve research problems? I don't.
- kadushka 1y agoGreat, then we established that I (generally intelligent human), and LLMs, both can't perform the task in question without specific training. We still don't know if given that specific training we (a specific human or a specific LLM) would be able to perform it.