Re: Today's talk on defeasible reasoning and AGI

pá 9. 2. 2024 v 10:09 odesílatel Dave Raggett <dsr@w3.org> napsal:

>
>
> On 8 Feb 2024, at 21:13, Melvin Carvalho <melvincarvalho@gmail.com> wrote:
>
> Comments/questions
>
> 1. I know what chain of thought is, but what is type 2?
>
>
> This is explained on slides 17 and 18.
>

Ah, you mean Kahneman type-2 rather than chain of thought type-2.  Got it,
thanks.


>
> 2. Any thoughts on orchestration of all these agents
>
>
> You may want to expand on what you mean by that.
>

Scaling agents to work together requires some kind of cooperation or
command and control when you start to deal with multiple agents.  I think
that's a common challenge in multi agent systems?

Similar to the demo of the ants finding food, but at web scale.


>
> To be good co-workers with humans, agents will need to be sociable, have a
> good grasp of the theory of mind, and the ability to learn and apply
> behavioural norms for interaction.  I helped lead a workshop on behavioural
> norms last year at a Dagstuhl Seminar 23081, and see also Dagstuhl Seminar
> 23151.
>
> 3. Minor: "More like alchemy than science – but early days yet!" this
> comment caught my eye.  I assume it was tongue in cheek, but would be
> intrigued if you were inclined to expand on that.
>
>
> Others have said this before me. We still don’t have a deep understanding
> of how large language models are able to represent and manipulate knowledge
> and provide the results they do.  The output of a large language model is
> entirely determined by the output from the encoding block. How can the
> richness of the semantics for a given response be represented in a single
> vector?
>
> Bill Gates chatted with Sam Altman (Open AI CEO) in a recent podcast, and
> they both agreed that a better (mathematical) understanding would enable
> smaller more effective models. They didn’t talk about the details though.
>

Makes sense.  LLMs are not an exact science.


>
>
> Dave Raggett <dsr@w3.org>
>
>
>
>