I imagine it could be easier to make sense of the 'biological' patterns that way? like, having bottlenecks or spatially-related challenges might have to be simulated too, to make sense of the ingested 'biological' information.
I've only started using LLMs for code recently, and I already tend to mentally translate what I want to something that I imagine is 'more commonly done and well represented in the training data'.
But especially the ability to just see some of the stuff it produces, and now to see its thought process, is incredibly useful to me already. I do have autism and possibly ADD though.
Yes, code is a key training component. Open-Llama explicitly used programming data as one of seven training components. However, newer models like Llama 3.1 405B have shifted to using synthetic data instead. Code helps develop structured reasoning patterns but isn't the sole foundation - models combine it with general web text, books, etc.
It doesn't have to be deterministic. Anyone who wants assurances would simply get a full plan of actions to take, and could edit the individual tasks before confirming their execution.
Just like with a real assistant, one could set very clear boundaries as to how much the assistant could spend, on what, etc., or even specify that more money can be spent in, say, a particular period and with a few extra passes (using different models?) just to make sure.
It does really feel like the same animal, because to the degree that I remember the discussions, a lot of it was about a perception of some kind of dichotomy (control or not, being able to 'touch' the product or not, etc.) that really doesn't need to exist.
There are cases where this could be true. I could see, say, asking it to do some research for you or something. Or maybe grab some restaurant recommendations. Really anything that brings you options that you can then make a decision on. If I do end up using this tech, thats how I would feel comfortable using it.
When it comes to things like making purchases and such on your behalf, thats where I disagree. Even with boundaries it's just not deterministic enough for most peoples risk appetite IMO. And when I say "most people" I am not talking about HN folks.
And just to be clear, I do use this technology. I just am very aware that its not this magic bullet that the AI bros want us to believe it is. I used it this morning in fact to help me write up some code I felt too lazy to deal with. It very quickly and efficiently wrote what I asked. I was pleasantly surprised and _almost_ missed the very bad bug that it dropped right into the middle of it all. I'm not giving my credit card to that lol.
I do agree with that. What I'm seeing (and building for myself) really boils down to 'for anything I am remotely uncomfortable trusting the AIs with, I trust them enough to conveniently present me the action or actions they want to take, and I trust myself enough to build a UI that doesn't make it too easy for me to lazily or accidentally click yes on a hallucinated 5000 dollar expense.
The fact that such an AI, with sufficient development, even given what we have now, could present everything on a silver platter up to the execution of the /exact/ commands, and this alone is incredibly useful and can save a lot of time/expense/effort.
The crucial bit is that with, say, a human personal assistant, even if they say they will do exactly what you tell them to do, it's still ultimately inexact. But with an AI, it's trivial to 'decorate' a structured 'command object' (as json, xml, whatever) and from that moment let the deterministic system take over.
If anything, we sometimes want an actual human to /not/ deterministically execute /exactly/ what we tell them to, because perhaps they might know we were angry, drunk, or they just heard something on the news that should invalidate our request.
Either way, my point is that this distinction between 'doing research' and 'doing a thing' is not so dichotomous. In practice, I suspect all but the most autonomously-minded people, and especially non-IT folk, given enough sense of control over the confirmation of potentially 'fuzzy' actions, are happy with an AI doing a ton of stuff for them.
It's good enough for me on an M1, 16Gb, and slow but good enough as a background job on my older intel mbp with 16Gb. I somehow expected it to not work on intel macs at all, so that's a freebie.
I mean that's just confabulating the next token with extra steps... ime it does get those wrong sometimes. I imagine there's an extra internal step to validate the syntax there.
I'm not arguing for or against anything specifically, I just want to note that in practice I assume that to the LLM it's just a bunch of repeating prompts with the entire convo, and after outputting special 'signifier' tokens, the llm just suddenly gets a prompt that has the results of the program that was executed in an environment. for all we know various prompts were involved in setting up that environment too, but I suspect not.
reply