Independent AI safety researcher
Thank you, it's been fixed.
In terms of LLM architecture, do transformer-based LLMs have the ability to invent new, genuinely useful concepts?
So I'm not sure how well the word "invent" fits here, but I think it's safe to say LLMs have concepts that we do not.
Recently @Joseph Bloom was showing me Neuronpedia which catalogues features found in GPT-2 by sparse autoencoders, and there were many features which were semantically coherent, but I couldn't find a word in any of the languages I spoke that could point to these concepts exactly. It felt a little bit like how human languages often have words that don't translate, and this made us wonder whether we could learn useful abstractions about the world (e.g. that we actually import into English) by identifying the features being used by LLMs.
You might enjoy this post which approaches this topic of "closing the loop," but with an active inference lens: https://www.lesswrong.com/posts/YEioD8YLgxih3ydxP/why-simulator-ais-want-to-be-active-inference-ais
A main motivation of this enterprise is to assess whether interventions in the realm of Cooperative AI, that increase collaboration or reduce costly conflict, can seem like an optimal marginal allocation of resources.
After reading the first three paragraphs, I had basically no idea what interventions you were aiming to evaluate. Later on in the text, I gather you are talking about coordination between AI singletons, but I still feel like I'm missing something about what problem exactly you are aiming to solve with this. I could have definitely used a longer, more explain-like-I'm-five level introduction.
That sounds right intuitively. One thing worth noting though is that most notes get very few ratings, and most users rate very few notes, so it might be trickier than it sounds. Also if I were them I might worry about some drastic changes in note rankings as a result of switching models. Currently, just as notes can become helpful by reaching a threshold of 0.4, they can lose this status by dropping below 0.39. They may also have to manually pick new thresholds, as well as maybe redesign the algorithm slightly (since it seems that a lot of this algorithm was built via trial and error, rather than clear principles).
"Note: for now, to avoid overfitting on our very small dataset, we only use 1-dimensional factors. We expect to increase this dimensionality as our dataset size grows significantly."
This was the reason given from the documentation.
Thanks for pointing that out. I've added some clarification.
Some thoughts:
First, it sounds like you might be interested the idea of d/acc from this Vitalik Buterin post, which advocates for building a "defense favoring" world. There are a lot of great examples of things we can do now to make the world more defense favoring, but when it comes to strongly superhuman AI I get the sense that things get a lot harder.
Second, there doesn't seem like a clear "boundaries good" or "boundaries bad" story to me. Keeping a boundary secure tends to impose some serious costs on the bandwidth of what can be shared across it. Pre-industrial Japan maintained a very strict boundary with the outside world to prevent foreign influence, and the cost was falling behind the rest of the world technologically.
My left and right hemispheres are able to work so well together because they don't have to spend resources protecting themselves from each other. Good cooperative thinking among people also relies on trust making it possible to loosen boundaries of thought. Weakening borders between countries can massively increase trade, and also relies on trust between the participant countries. The problem with AI is that we can't give it that level of trust, and so we need to build boundaries, but the ultimate cost seems to be that we eventually get left behind. Creating the perfect boundary that only lets in the good and never the bad, and doesn't incur a massive cost, seems like a really massive challenge and I'm not sure what that would look like.
Finally, when I think of Cyborgism, I'm usually thinking of it in terms of taking control over the "cyborg period" of certain skills, or the period of time where human+AI teams still outperform either humans or AIs on their own. In this frame, if we reach a point where AIs broadly outperform human+AI teams, then baring some kind of coordination, humans won't have the power to protect themselves from all the non-human agency out there (and it's up to us to make good use of the cyborg period before then!)
In that frame, I could see "protecting boundaries" intersecting with cyborgism, for example in that AI could help humans perform better oversight and guard against disempowerment around the end of some critical cyborg period. Developing a cyborgism that scales to strongly superhuman AI has both practical challenges (like the kind neuralink seeks to overcome), as well as requiring you to solve it's own particular version of alignment problem (e.g. how can you trust the AI you are merging with won't just eat your mind).