This is actually a new skill I've been working on. Learning how to elicit concise and simple speech from models (and from people to!).
Whenever I come to a wall of complicated text I kick into gear and think through getting it to distill this into the high-level useful bits that I actually need to know.
I guess I could create an actual agent skill for this :) And next-gen models might eventually be trained to simplify their output themselves...
I'm really impressed by mathematicians. It's cool that Fermat had the intuition to conjecture that "aⁿ + bⁿ = cⁿ" could not be satisfied for n > 2, and that other mathematicians can create proofs, and that others still can understand AI's formulation of those proofs. Really cool.
I wonder if AI can come up with mathematical conjectures. As in, they feel it's right but can't prove it. What even happened in Fermat's brain to sense it was true?
It is capable of applying know heuristics and general principles in places where they haven't been applied and in this sense very much capable of generating conjectures in much the same way a person does. It's ability to employ a diversity of techniques coupled with it's computational power differentiate it from a human researcher. It still needs guidance to work well, but I've already changed my daily work flow as a research mathematician to incorporate use of AI.
Right. Once we see AI start delivering on the creative & intuition side of things that's going to be awesome. Until then I guess we'll live with exhaustive exploration of problem spaces by orchestrating swarms of agents...?
> it can't jump to somewhere where it's not present its training data
That sounds like something that can be engineered, can't it? In other words, we can identify limitations in current transformer-based architectures, and we can also build new architectures over time.
You mention LLMs are dead and don't have the complexity or inner life that people do. Is your opinion that these kinds of things are not possible for AI in general, or that these things might be possible but we're just not there yet with modern LLMs?
You mentioned LLMs don't have souls, desire, or a will. I imagine those latter two can be engineered, no?
My view is that these sorts of things are not possible for AI in general. Though we can create things and name them “will” and “desire”.
Software deals with metaphors. Your Amazon shopping cart is a metaphor of a real shopping cart. You your desktop and your file system, etc. are metaphors of real items. But we don’t mistake the metaphor for its object, even from inanimate objects to their software counterparts (shopping cart to Amazon cart).
Now the metaphors are dealing with humanness, things like intelligence etc. And rather than seeing it as software doing what it always does, taking things and creating software metaphors of them, we are starting to say these are actually what they are named. Saying the artificial intelligence is actually an intelligence.
We’d either have to reduce the definition of intelligence such that calculators are intelligent. Or admit that these tools are not intelligent and are rather ways of exploring the work of actual intelligent beings, work that is found in their training data.
> My view is that these sorts of things are not possible for AI in general.
I wonder what it would take build an artificial system that has these qualities.
> Software deals with metaphors.
This is me wondering again: what's fundamentally different between software running on a machine compared to what's happening in our brains? In both cases you have energy flow following a pattern.
It's conceivable to create a system where energy flows in a particular way.
BTW, we navigate the world of an uncountable number of particles by creating models in our heads of what we think are large things out there. Approximations are made by both artificial and biological systems.
> what's fundamentally different between software running on a machine compared to what's happening in our brains? In both cases you have energy flow following a pattern.
Assuming we’ve scratched the surface of the complexity of the brain. I’d say in one case a human with a will is steering that flow of energy. In the other case it is a probabilistic algorithm steering the flow of energy.
The AI is not interacting with world with its own will. I see that as a big difference.
There are presuppositions that go beyond the realm of software engineering that guide one’s views of these things. One is whether you believe the material world is all that is, and that human consciousness is a product of the brain—or that there is such a thing as the soul or spirit of man. From the material perspective you may posit that if you emulate the brain then a sort of AI consciousness could arise. Or that emulating the patterns of the brain equates to emulating personhood. (Though what is material consciousness? I’d say consciousness is by nature immaterial.) I’m not a materialist, and I don’t believe the conclusions that arise from it’s perspectives are accurate.
Well they're language models. You can't capture the human experience in language. Simple as that.
> Is your opinion that these kinds of things are not possible for AI in general, or that these things might be possible but we're just not there yet with modern LLMs?
I used to be on the side of "we're just not there yet [with AI in general]", but after seeing people's response to an algorithm optimized to tickle just their language instinct, I'm actually a little bit more on the fence about it.
I'm in the same boat, I think some aspects can be engineered, like intention and desire. I'm really curious if it's possible to go the full distance and make AI have experience like we do.
How about a subsequent review where a separate agent analyzes the original issue and resultant code and approves it if the code meets the intent of the issue. The principle being to keep an eye out for manual work that you can describe well enough to offload.
Depending on your success rate with agents, you can have one that validates multiple criteria or separate agents for different review criteria.
You are fighting nondeterministic behavior with more nondeterministic behavior, or in other words, fighting probability with probability. That doesn't necessarily make things any better.
In my experience, an agent with "fresh eyes", i.e., without the context of being told what to write and writing it, does have a different perspective and is able to be more critical. Chatbots tend to take the entire previous conversational history as a sort of canonical truth, so removing it seems to get rid of any bias the agent has towards the decisions that were made while writing the code.
I know I'm psychologizing the agent. I can't explain it in a different way.
I think of it as they are additive biased. ie "dont think about the pink elephant ". Not only does this not help llms avoid pink elphants instead it guarantees that pink elephant information is now being considered in its inference when it was not before.
I fear thinking about problem solving in this manner to make llms work is damaging to critical thinking skills.
Assigning different agents to have different focuses has worked for me. Especially when you task a code reviewer agent with the goal of critically examining the code. The results will normally be much better than asking the coder agent who will assure you it's "fully tested and production ready"
As someone born, raised, and currently living in the UK, this is the first time I've heard of this. It's literally never come up or been an issue anywhere in the past 50 years I've been alive, and I've lived all over these isles. A bizarre OP.
It's humbling to think about all the things people have gone through over the past couple hundred thousand years. Somewhere around 117 billion humans have ever lived...? It makes it seem kind of small when we think only 50 or 100 years out when thinking of what the future would be.
I agree with you. I clicked into this hoping to hear what new things we could learn or discover with the new observatories. Commenting on the more positive and informative side would be a better use of time and energy I think :)
Ha, you made me think of casually referring to xkcd's by number just as we did with RFC's back in the day. "I don't know, the socket states seem to follow RFC 793, but remember it's a 1918 address on the southside of the NAT."
I gonna keep a look out for doing this with xkcd's now :)
There are a few that pop out but the one that has managed to stick (aside from 1053 that just came up), is 927 for standards, which you can remember as 3^2 for 9 and 3^3 for 27. Or Yoda's age + the 27 club.
Whenever I come to a wall of complicated text I kick into gear and think through getting it to distill this into the high-level useful bits that I actually need to know.
I guess I could create an actual agent skill for this :) And next-gen models might eventually be trained to simplify their output themselves...
reply