Wiki Contributions

Comments

Another similar result was that AlphaFold was trained on its own high-confidence predictions for protein sequences with unknown structures:

The AlphaFold architecture is able to train to high accuracy using only supervised learning on PDB data, but we are able to enhance accuracy (Fig. 4a) using an approach similar to noisy student self-distillation35. In this procedure, we use a trained network to predict the structure of around 350,000 diverse sequences from Uniclust3036 and make a new dataset of predicted structures filtered to a high-confidence subset. We then train the same architecture again from scratch using a mixture of PDB data and this new dataset of predicted structures as the training data, in which the various training data augmentations such as cropping and MSA subsampling make it challenging for the network to recapitulate the previously predicted structures. This self-distillation procedure makes effective use of the unlabelled sequence data and considerably improves the accuracy of the resulting network.

I'm also dealing with chronic illness and can relate to everything you listed. I've been thinking that a discord server specifically for people with chronic illness in the rationality community might be helpful to make it easier for us to share notes and help each other. There are different discord servers for various conditions unaffiliated with the rationality community, but they tend to not have great epistemic standards and generally have a different approach than what I'm looking for. Do you have any interest in a discord server?

By forward do you mean sooner (shorter timelines) or later (longer, slower timelines)?

I tried giving this to GPT-3 and at first it would only give the tautological "pawns become more powerful" example, then I expanded the prompt to explain why that is not a valid answer, and it gave a much better response.

I believe this response is the same as your fourth bullet point example of a good answer.

 

 

Here's the prompt in copy/pastable format for anyone who wants to try playing with it:

 

Consider a new variant of chess, in which each pawn can move up to six squares forward on its first move, instead of being limited to one or two squares. All other rules remain intact. Explain how game balance and strategy is changed with this new variant of chess.

Your response should share something not immediately obvious about this variant and provide a plausible justification for why it might be true. Some responses that would not succeed would be

The pawns become more powerful. (Too simple, close to a tautology.)
New strategies will need to be developed. (Too vague.)
Bishops become more valuable. (Needs a justification for why we should expect this.)

Response:

Agreed that it would be insanely impressive. It would probably convince me that a fast takeoff is very likely coming within the next 5 years. Yet I can't really say I'm more than 90% confident that GPT-4 won't be able to do it. Maybe 95%.

I'm not sure about that. See page 8 of the LamDA paper where they gave it access to a "toolset" including things like a calculator. I wouldn't be surprised if they gave GPT-4 access to similar tools including a way to access the current date.

There are probably less impressive things than this that it won't be able to do, but here's one prediction in which I am about 90% confident:

If you invent or find a board game of similar complexity to chess that GPT-4 has never seen before and explain the rules using only text (and, if GPT-4 is multimodal, also images), GPT-4 will not be able to perform as well at the game as an average human who has never seen the game before and is learning it for the first time in the same way. I say "perform as well as" rather than "beat" because, depending on the specific rules of the game, I expect GPT-4 will likely not output legal moves.

This prediction applies to whatever OpenAI reveals with the name GPT-4 or that is clearly the GPT-3 successor regardless of size, assuming it's revealed in the next two years and is not trained on specifically that task (e.g. by generating tens of thousands of unique synthetic board game examples with synthetic dialogues where they're explained and played over text).

I ran into another similar example in the last section of this post.

As another example, I remember somebody (maybe EY or Gwern?) suggest that prompting people with "how would somebody smarter than you solve this problem?" can actually be effective at soliciting better solutions than just asking them to solve it directly.

I don't remember where I saw this so if anybody has a link feel free to share.

Load More