I think you are conflating capabilities with context. Yes you may have a superpowerful AGI but you do need to giveit the relevant context for a specific job, which is essentially turning it into a specialised AI. Wether that context is given to it through training or prompts or something else, you are still sepcialising it. That’s why agentic AI is so powerful: you take a powerful base model and make it more useful by giving it the context. No matter how powerufl your AI gets, it will still be more effective when given specific context.
chasmani
Did you measure how this affects model performance on other metrics? Specifically, can you train model capabilities that are more advanced than the model you use to make the pretraining labels? It seems like if you are conditioning the model to replicate the <assistant> tagged text then you get not only the moral stance of that text but also its other capabilities. If so, training frontier models like this would lead to less capable models.
I had the thing you’re talking about! I did meditation, then started to feel energy, then chakras started making a whole bunch of sense. My working model is that the mind maps emotional states onto certain parts of the body, kind of like a synesthesia; through meditation you can strengthen (or become more aware of) that association. The benefit is that it gives you a hack into managing your emotional states (and unlocking what feel like entirely new and powerful emotional states) just by focussing on parts of the body.
These ideas appear in existing frameworks. See ecological rationality, resource-rationality, Marr. etc. They assume that cognitive processes are adapted to solve recurrent problems in the environment. Marr in particular claims that representations (abstractions) will mirror the problem being solved. Resource rationality builds on Marr and would agree with the prediction of efficient representations, I.e. compression.
Thanks for the comment. I guess this is about different frames for understanding the same thing.
The evolution frame is useful in understanding the connection between fitness (or profit) landscapes and the behaviour that you see. An agent can have whatever goals you want, but it will only survive and propagate if those goals allow it to gather enough resources to survive and continue.
I agree that the mechanisms with organisations are very different to genetic evolution, but there is a level of analogy. In organisations the information is not encoded in digital form in DNA but is something more like Dawkins’ memes. The success of the memeplex will influence whether it will survive, and whether parts of the memeplex get copied or not. I agree that organisations often do not want identical memeplexes to pop up but the fact this is a threat to them shows that an evolutionary copying function is at play, even if they don’t want it to be. And of course the meme of “stopping similar competitors from arising” is itself a response to the limited resources in the fitness landscape.
Love it! Agentic AI creates another transmission pathway: through the md files etc that tell agents how to use LLMs. These are perhaps quicker
Darwin still applies. Models (and memes within models) that work well and are popular are more likely to replicate via the companies that make those models gaining more resources and choosing to use similar mechanisms and data to train the next models.
Gradient descent all that are just extra steps.
If the economy is so easily understood then why do we have high inflation, a cost of living crisis, rising inequality?
The thing that is not understood is why these things are happening and how we can change things so that normal people are better off.
The fact that some people have some coherent theories for some aspects of the economy is not equivalent to us understanding the economy.
I feel like the entire framing here is the problem. You cannot see “The Thing” because you are looking at it from a perspective where The Thing isn’t apparent.
What is The Thing? It is having a partnership that you are both committed to. At its best this partnership becomes an aspect of your self, and your partner. The frame to see this in is that the partnership is an entity in its own right and is a part of the “I” that each partner identifies with. In this frame the question “what am I getting out of this relationship” is no longer entirely focussed on the individual “I” but also on the partnership “I”. When you frame the question in terms of what the individual “I” gets out of it then you are entirely missing the point and are unable to see the real value proposition.
If we dip our toe back into the individualistic frame: you could describe the value of relationships being in the partial dissolution of self in the partnership, which not only feels amazing but also gives a deeper level of meaning to your life.
I would say that a part of compassion, and empathy, is to recognise that indeed those narratives are valid, or else there is some valid reason that people are as they are. Also, not everyone shares the moral value of optimising themselves or making themselves good at something. Disgust implies judgement that implies a lack of compassion.
Since you seem to be motivated at making yourself better, which I agree is a good motivation, why don’t you challenge yourself to increase your compassion and humility?
Do you have compassion for yourself? What are you bad at that you are unable to make yourself good at? Do you feel disgust for yourself in those situations? Compassion begins with humility, which is something that you might want to work on
I’m not sure I agree that is is easy for humans to robustly understand proofs. I think it takes really a lot of training to get humans to that point.
There’s the argument that increasing access to information creates competition for attention, which drives language to be more concise and readable, e.g. https://www.nature.com/articles/s44271-024-00117-1
In a post-scarcity world you probably want a lot of personal freedom.
Fun read. So, so many possible covariates. The causal web is very complicated here. Birth order affects lots and lots of other things, which can also affect the chance you become a cardinal. There are also lots of things that would affect the birth rate in a family and also affect the chance the children become cardinals.
I have a meta-view on this that you might think falls into the bucket of “feels intuitive based on the progress so far”. To counter that, this isn’t pure intuition. As a side note I don’t believe that intuitions should be dismissed and should be at least a part of our belief updating process.
I can’t tell you the fine details of what will happen and I’m suspicious of anyone who can because a) this is a very complex system b) no-one really knows how LLMs work, how human cognition works, or what is required for an intelligence takeoff.
However, I can say that for the last decade or so most predictions of AI progress have been on consistently longer timescales than what has happened. Things are happening quicker than the experts believe they will happen. Things are accelerating.
I also believe that there are many paths to AGI, and that given the amount of resources currently being put into the search for one of those paths, they will be found sooner rather than later.
The intelligence takeoff is already happening.
I agree with your point in general of efficiency vs rationality, but I don’t see the direct connection to the article. Can you explain? It seems to me that a representation along correlated values is more efficient, but I don’t see how it is any less rational.
I would describe this as a human-AI system. You are doing at least some of the cognitive work with the scaffolding you put in place through prompt engineering etc, which doesn’t generalise to novel types of problems.
You seem to make a strong assumption that consciousness emerges from matter. This is uncertain. The mind body problem is not solved.
Here is one place:
Ultimately all intelligence may be absorbed into one super universal singleton intelligence.
If there are any kind of meaningful resource constraints, and of course there will be because we live in a limited universe, then you will need to give different contexts to your AI to solve different problems. By giving it different contexts you are effectively making lots of specialised intelligences, not one universal singleton intelligence.