[GoLUG] LLMs can't reason

[o1bigtenor]

On Wed, Jun 18, 2025 at 9:07 PM Steve Litt <slitt at 444domains.com> wrote:
>
>
> On Wed, 18 Jun 2025 23:07:45 +0100 (GMT+01:00)
> kc-golug at chadwicks.me.uk wrote:
>
> > 18 Jun 2025 22:26:18 Steve Litt <slitt at 444domains.com>:
> >
> > > The real problem is that, with programs designing programs
> > > designing programs, and more and more AI machines and robots are
> > > closely networked together,
> >
> > Except they aren't designing programs. They are just spitting out
> > error probe chunks of code and the bigger the code base gets the
> > worse it gets. The data even shows that humans using LLMs as tools
> > which is the only way they can work results in significantly worse
> > code quality. Any attempts of true AI (completely separate field to
> > LLMs and "AI" of today) where programs are self modifying fail
> > spectacularly. And as I have said innovation and reasoning from a
> > position of understanding is not computable. All LLMs do is
> > probabilistically exhibit what already exists whilst introducing
> > errors and other quality issues.
>
> The preceding paragraph is true today. But five years ago, I think 1
> out of 100 people believed LLMs would get as smart as they have. LLMs'
> genius startled Bill Gates, for gosh sakes. Now extrapolate the last 5
> years of LLM progress into the next 5 or the next 20, and at some point
> they'll write better code than most humans. Computers are already
> beating chess masters, for gosh sakes.

Not doing that yet for bridge - - - grin!