site banner
magic9mushroom  If you're going to downvote me, and nobody's already voiced your objection, please reply and tell me  1yr ago (thezvi.wordpress.com)   2440 thread views

Danger, AI Scientist, Danger

thezvi.wordpress.com

Zvi Mowshowitz reporting on an LLM exhibiting unprompted instrumental convergence. Figured this might be an update to some Mottizens.

10
Top Bottom Old Controversial Comments
Jump in the discussion.

No email address required.

they aren't good at synthesizing information from two fields in ways that haven't been done before

They are not good at that yet. But there are already indicators that they could become so.

  1. Midjourney will blend concepts from different places. Not just style transfer. A nice example I saw was an image of 'boy with a hedgehog'. The boy was holding a hedgehog, but also his hair was a bit spiky, like the hedgehogs. I think it is most unlikely MidJourney had ever seen a composition/juxtaposition of that kind before, and both the hedgehog and the hair were modified to make the composition work.
  2. DeepMind have AlphaZero, which plays chess, shogi and go. It plays better than human, i.e. not just based on play it has seen before, and one can argue it is crossing between different genres, not confined to one field.
  3. The often cited example of finding an analogy between compost heap and nuclear fission, again an example of crossing field boundaries.

So to say that machine learning can't synthesise information from two fields in ways that have not been done before needs more qualification, to be defensible.

I was talking about (transformer-based generative) LLMs specifically. I am not a sufficiently good mathematician to feel confident in this answer, but LLMs and diffusion models are very different in structure and training, and I don't think that you can generalise from one to the other. Midjourney is basically a diffusion model, unscrambling random noise to 'denoise' the image that it thinks is there. The body with spiky hair seems like the model alternatively interpreting the same blurry patch of pixels as 'spikes' because 'hedgehog' and 'hair' because 'boy'. Which I think is very different from a predictive LLM realising that concept A has implications when combined with concept B that generates previously unknown information C.

DeepMind have AlphaZero, which plays chess, shogi and go. It plays better than human, i.e. not just based on play it has seen before, and one can argue it is crossing between different genres, not confined to one field.

I haven't kept up to date on RL, but I don't think this is relevant. Firstly because the concept of self-play is not really relevant to text generation, and secondly because I don't suppose the ability to play chess is being applied to go. Indeed, I don't really see how it could be, because the state and action space is different for each game. It seems more likely to me that the same huge set of parameters can store state-action-reward correlations for multiply games simultaneously without that information interacting in any significant way.

The often cited example of finding an analogy between compost heap and nuclear fission, again an example of crossing field boundaries.

I'm not aware of this. Can you give some more info?

Diffusion models work for text too. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10909201/

The blending of concepts that we see in MidJourney is probably less to do with the diffusion per se as with CLIP - a building block within diffusion. CLIP aligns a language model with an image model. Moving concepts between different representations helps with concept generation. There's a lot being done with 'MultiModal models' to make the integration between different modalities work better.

'Self play' is relevant for text generation. There is a substantial cottage industry in using LLMs to evaluate the output of LLMs and learn from the feedback. It can be easier to evaluate whether text 'is good' than it is to generate good text. So multiple attempts and variations can lead to feedback and improvement. Mostly self play to improve LLMs is done at the level of optimising prompts. However the outputs improved by that method can be used as training examples, and so can be used to update the underlying weights.

https://topologychat.com is a commercial example of using LLMs in a way inspired by chess programming (Leela, Stockfish). It does a form of self play on inputs that have been given to it, building up and prioritising different lines. It then uses these results to update weights in a mixture of experts model.

Here's the quote from Geoffrey Hinton:

"why is a compost heap like an atom bomb? And GPT-4 says, well, the timescales and the energy scales are very different. That’s the first thing but the second thing is the idea of a chain reaction.

So in an atom bomb, the more neutrons around it, the more it produces, and in a compost heap, the hotter it gets, the faster it produces heat and GPT-4 understands that. My belief is when I first asked it that question, that wasn’t anywhere on the web. I searched, but it wasn’t anywhere on the web that I could find. It’s very good at seeing analogies because it has these features. What’s more, it knows thousands of times more than we do. So it’s gonna be able to see analogies between things in different fields that no one person had ever known before.

That may be this sort of 20 different phenomena in 20 different fields that all have something in common. GPT-4 will be able to see that and we won’t. It’s gonna be the same in medicine. If you have a family doctor who’s seen a hundred million patients, they’re gonna start noticing things that a normal family doctor won’t notice."

From transcript at https://medium.com/@jalilnkh/geoffrey-hinton-will-digital-intelligence-replace-biological-intelligence-fc23feb83cfb of the video.

Diffusion models work for text too.

Last I checked, diffusion models work at all for text but they don't work very well. More specifically, text diffusion models remind me quite strongly of the classic-style Markov chain text generators that used to be popular for generating amusing locally-coherent-globally-word-salad text. Here's the best concrete example of this I can point to (italicized text is the prompt, normal text is continuation, examples sourced from this JDP tweet, whole tweet is excellent but somewhat OT here):

Diffusion model:

This whole dream seems to be part of someone else's experiment.

In the end, I think the only way to get the full benefit of the dream is to get up and walk around.

A lot of the other strands of the theory of dreams, especially the one called yncrest are known to the person who writes this blog. Consequently, I've tested his reactions via a series of brief, televised interviews with the British Prime Minister, Margaret Thatcher.

“I don't care what anybody says. I'm not going to waste my time with that. The study was conducted in the high-flow, low-flow and quiescent periods of a butterfly species and thus provides evidence for the first three of the four hypotheses listed in the original paper. The only one not covered is the "need to change the rate of capture", where we now know that this is largely a policy instrument - and particularly relevant to Sudanese geese.

GPT-2:

This whole dream seems to be part of someone else's experiment. This is usually hard to see, since you're conscious that mental shortcuts are normal. But you're definitely moments away from being normal again, as it's obvious that you are being coached to think like that by another person and doing our own thing. Why are we having such thoughts when we're paying no attention because that's already taking place?

I'm totally appalled by what these key players were doing, yet I explain it in a very large way. After 2 minutes, I make my way through high teaching circles, recognizing that there is more to learning than just to learn.

Few other teachers would air this incessantly their students' struggles recount the stories of their own memories and teachers'. Whatever it takes is the big fat slut shaming.

All these legendary trainees made the same point:

Cognitive functional aspects: Bayesian modeling, learn science way. And the most important part is: Like the coming of a hurricane, the book is mucking between science and morals.

Twitter Mentions of the first book: Kent

Now obviously in the limit as computational power and training data volume go to infinity, diffusion models and transformer models will generate the same text, since in the limit they're pulling from the same distribution with minimal error. But in the very finite regime we find ourselves in, diffusion models "spend" their accuracy on making the text locally coherent (so if you take a random 10 token sequence, it looks very typical of 10 token sequences within the training set), while transformer LLMs "spend" their accuracy on global coherence (so if you take two 10 token sequences a few hundred tokens apart in the same generated output, you would say that those two sequences look like they came from the same document in the training set).

The blending of concepts that we see in MidJourney is probably less to do with the diffusion per se as with CLIP

Agreed. Obvious once you point it out but I hadn't thought about it that way before, so thanks.

'Self play' is relevant for text generation. There is a substantial cottage industry in using LLMs to evaluate the output of LLMs and learn from the feedback.

Notably, Anthropic's Constitutional AI (i.e. the process by which Anthropic turned a base LLM into the "helpful, honest, harmless" Claude) process used RLAIF, which is self play by another name. And that's one big cottage.

What is this place?

This website is a place for people who want to move past shady thinking and test their ideas in a court of people who don't all share the same biases. Our goal is to optimize for light, not heat; this is a group effort, and all commentators are asked to do their part.

The weekly Culture War threads host the most controversial topics and are the most visible aspect of The Motte. However, many other topics are appropriate here. We encourage people to post anything related to science, politics, or philosophy; if in doubt, post!

Check out The Vault for an archive of old quality posts. You are encouraged to crosspost these elsewhere.


Why are you called The Motte?

A motte is a stone keep on a raised earthwork common in early medieval fortifications. More pertinently, it's an element in a rhetorical move called a "Motte-and-Bailey", originally identified by philosopher Nicholas Shackel. It describes the tendency in discourse for people to move from a controversial but high value claim to a defensible but less exciting one upon any resistance to the former. He likens this to the medieval fortification, where a desirable land (the bailey) is abandoned when in danger for the more easily defended motte. In Shackel's words, "The Motte represents the defensible but undesired propositions to which one retreats when hard pressed."

On The Motte, always attempt to remain inside your defensible territory, even if you are not being pressed.


New post guidelines

If you're posting something that isn't related to the culture war, we encourage you to post a thread for it. A submission statement is highly appreciated, but isn't necessary for text posts or links to largely-text posts such as blogs or news articles; if we're unsure of the value of your post, we might remove it until you add a submission statement. A submission statement is required for non-text sources (videos, podcasts, images).

Culture war posts go in the culture war thread; all links must either include a submission statement or significant commentary. Bare links without those will be removed.

If in doubt, please post it!


Rules


Recommended Posts And Communities

Recommended Realtime Chats