Re: Remember "Bit-Slice" Chips ?

On Sat, 21 Dec 2024, 186282@ud0s4.net wrote:

On 12/20/24 12:19 PM, rbowman wrote:

On Fri, 20 Dec 2024 01:31:30 -0500, 186282@ud0s4.net wrote:
>

    It's the 'hand-wave' thing that sunk the first AI paradigm.
    Marv Minsky (who posted on usenet for awhile) and friends saw how
    easily 'decisions' could be done with a transistor or two and assumed
    it would thus be easy to build an AI. AC Clarke used the Minsky
    optimism when fashioning the idea of "HAL".

 Minsky threw a wrench in the works with his 9169 'Perceptrons'. He had
tried to implement B. F. skinner's operant condition with a analog lashup
that sort of worked if the vacuum tubes didn't burn out. Rosenblatt has
built a 'Perceptron' and Minsky pointed out original design couldn't
handle an XOR. That sent research down another rabbit hole.
 By the '80s the original perceptron had evolved into a multilayer network
train by back propagation. When I played around with it 'Parallel
Distributed Processing' by Rumelhart and McClelland was THE book.
 https://direct.mit.edu/books/monograph/4424/Parallel-Distributed-
Processing-Volume
 The ideas were fascinating but the computing power wasn't there. Most of
what I learned then is still relevant to TensorFlow and the other neural
network approaches except now there are the $30,000 Nvidia GPUs to do the
heavy lifting.
 The '80s neural networks weren't practical so the focus shifted to expert
systems until they petered out. The boom and bust cycles led to the term
'AI Winter'
 https://www.techtarget.com/searchenterpriseai/definition/AI-winter
 I think something worthwhile will come from this cycle but ultimately it
won't be the LLMs that are getting all the hype.

>
 With Minsky and friends it was just naive enthusiasm ...
 it was SO EASY to do logic and thus it seemed SO EASY
 to wire bits of it together and get an 'intelligence'.
>
 The same gen also promised us those flying cars and
 luxury Mars living by 1999 .......
>
 IMHO, if we're gonna get anything largely indistinguishable
 from 'sentience' these days it'll be the next few gens of
 LLMs. You can argue it'd be "fake" - but if you fake something
 WELL ENOUGH it's not fake anymore. LLMs and near derivs are
 where the HUGE money is these days.
>
 I did have a few posts with Minsky as his vision was
 falling apart. He did admit that he'd totally underestimated
 the problem. A few transistors did NOT replace 600 million
 years of evolutionary experiments - 'intelligence'/'self'
 was really deep/complex with endless fuzzy processing and
 pattern matching steps between 'I' and 'O'.
>
 However I still keep a copy of his "Society Of Mind"
 as a reminder of yesterday's optimism. He THOUGHT
 about it, TRIED ... and thus eventual failure was
 not really a failure - it just inspired new directions.
 There had to be a foundation to build on.
>
 There was a short-lived UK series about androids
 that eventually came to self-awareness (and the
 hate/fear directed towards them). The idea there
 was that 'self' was a sort of fractal, self-reflective,
 kind of paradigm. I suspect they had something there.
 Chat/LLMs maybe can't achieve that on their own, but
 who says you can't splice on a few more methods ?
 Organic brains seem to have LOTS of layers, lots
 of 'little people' inside that merge into 'Me'.

I've also thought about this approach. We have OCR, we have voice recognition, we have chess, we have diagnosing cancers (OCR), we have parrots (LLMs) etc.
The "only" thing lacking is the integrating "ego"-layer that coordinates all these separate systems we've developed, and provides the AI with volition/initiative/independence.