On Thursday, August 10, 2023 at 10:33:57 PM UTC-4, rbowman wrote:
> On Thu, 10 Aug 2023 21:08:29 GMT, Scott Lurndal wrote:
>
> > The term "AI" has been misused by media and most non-computer
> > scientists. The current crop "AI" tools (e.g. chatGPT) are not
> > artificial intelligence, but rather simple statistical algorithms based
> > on a huge volume of pre-processed data.
> Not quite...
>
> https://blog.dataiku.com/large-language-model-chatgpt
>
> I played around with neural networks in the '80s. It was going to be the
> Next Big Thing. The approach was an attempt to quantify the biological
> neuron model and the relationship of axons and dendrites.
>
> https://en.wikipedia.org/wiki/Biological_neuron_model
>
> There was one major problem: the computing power wasn't there. Fast
> forward 40 years and the availability of GPUs. Google calls their
> proprietary units TPUs, or tensor processing units, which is more
> accurate. That's the linear algebra tensor, not the physics tensor. While
> they are certainly related the terminology changes a bit between
> disciplines.
>
> These aren't quite the GPUs in your gaming PC:
>
> https://beincrypto.com/chatgpt-spurs-nvidia-deep-learning-gpu-demand-post-
> crypto-mining-decline/
>
> For training a GPT you need a lot of them -- and a lot of power. They make
> the crypto miners look good.
>
> The dirty little secret is after you've trained your model with the
> training dataset, validated it with the validation data, and tweaked the
> parameters for minimal error you don't really know what's going on under
> the hood.
>
> https://towardsdatascience.com/text-generation-with-markov-chains-an-
> introduction-to-using-markovify-742e6680dc33
>
> Markov chains are relatively simple.

I developed ANN classifiers in the mid-late 80s for explosive recognition in suitcases and as hardware plant observers for control loops. The size of the training set is one of the issues...but..... the web has a gazillion images that training sets are harvested. I know of a few local self-driving car companies that do exactly that.

Yes, verification of the ANN is and still is a big issue.
J