Re: definitions, problem spaces, methods

Dave
Machine Generated KR would be nice, but it presumes machine has  correct
model of the world
Machine cannot be presumed, as things stand, to have a correct model of the
world
without KR

Nonetheless, some level of trivial KR can easily be automated to some extent
with inference machines (not new)

Solomonoff, R. J. (1957, January). An inductive inference machine. In *IRE
Convention Record, Section on Information Theory* (Vol. 2, pp. 56-62). New
York: Institute of Radio Engineers.

to verify the quality/validity of this machine generated KR you would need
to develop
a more intelligent AI than the AI that generated the KR, and ultimately
some very good systems engineers to take responsibility for the overall
outcomes


I expect a whole lot of public funding will be wasted in the meantime
but no AI or KR can change that

PDM


On Wed, Nov 9, 2022 at 6:16 PM Dave Raggett <dsr@w3.org> wrote:

> Hi Mike,
>
> GPT3’s weaknesses in respect to coherence, etc. is unsurprising and should
> encourage research on moving on from text and image prediction to work on
> reasoning and continuous learning. This in turn motivates work on imprecise
> and imperfect knowledge, where I expect neural networks will shine. That
> includes the role of causal explanations as a basis for learning, and the
> ability to reason about past, present and imagined situations, including
> the beliefs of others. I very much believe that hand authoring of knowledge
> representations will give way to machine generated KR. I hope you can
> respect my view that KR is a subset of AI, as in the Wikipedia definition.
> The media hype around AI is frustrating in that attention grabbing
> headlines detract from a clear exposition of the underlying concepts.
>
> p.s. you may be interested in the talk I’ve prepared for next week’s
> NSF/EU workshop on research priorities.
>
> https://www.w3.org/2022/11/Raggett-AI-Priorities.pdf
>
> Best regards,
>
> Dave Raggett <dsr@w3.org>
>
>
>
>