AI software still needs the human touch

[Carey Brown]

Article from The Register. Talks about some of the issues surrounding AI and LLMs and some of how LLMs are good for software developers.

[Claude Moore]

Thanks for sharing it !

[Lou Hamers]

Yeah that says some of what I've been thinking myself lately. This stuff might reduce the barrier to entry for some hard to break into fields, but it won't eliminate those jobs.

[Tim Holloway]

It's what those third hands are touching that make me hesitant. And the funny fingers. We're still learning about machine learning.

[Claude Moore]

Tim Holloway wrote:It's what those third hands are touching that make me hesitant. And the funny fingers. We're still learning about machine learning.

ML is still a research field, and despite the fact that deep learning has reached unforeseable goals, we're still far from a General AI.  IMHO what AI based programmers' assistant tools (and the word 'tool' isn't casual here) is to help programmers write basilar code, no more no less like a calculator helps anyone to make, well, calculations. But Math itself is something by far more complex than calculate sums and products.

[Lou Hamers]

Claude Moore wrote:
ML is still a research field, and despite the fact that deep learning has reached unforeseable goals, we're still far from a General AI.  IMHO what AI based programmers' assistant tools (and the word 'tool' isn't casual here) is to help programmers write basilar code, no more no less like a calculator helps anyone to make, well, calculations. But Math itself is something by far more complex than calculate sums and products.

Yes that's why I like to call it ML instead of AI. But I still end up using the "AI" term sometimes, because I guess we already lost that battle. EVERYONE is calling it AI and most people don't know what "ML" is. Intelligence means thinking, to me, and this software isn't doing that.

My main worry about this stuff currently is that it ends up too centralized and controlled. I'd much rather see it open source, accessible to all for a low cost, and hosted locally without needing to rely on some corporate spy company. Theoretically that would help level the playing field for the little guys against increasingly powerful and wealthy organizations/people.

I'd start up a thread here asking for ML tools/software you can set up and run locally in a homelab etc., without relying on servers owned by someone else, but I'm afraid the result would be rather depressing.

[Claude Moore]

Lou Hamers wrote:

Yes that's why I like to call it ML instead of AI. But I still end up using the "AI" term sometimes, because I guess we already lost that battle. EVERYONE is calling it AI and most people don't know what "ML" is. Intelligence means thinking, to me, and this software isn't doing that.

Right, ML is a field of AI, but are often threated as synonims. And that's conceptually wrong.

Lou Hamers wrote:
My main worry about this stuff currently is that it ends up too centralized and controlled. I'd much rather see it open source, accessible to all for a low cost, and hosted locally without needing to rely on some corporate spy company. Theoretically that would help level the playing field for the little guys against increasingly powerful and wealthy organizations/people.

Please consider that training you own model at a state of the art level  would imply spending A LOT of money per day, something that only Big IT Firms can really afford.  If you want to play with a local version of ChatGPT, you may have a look at GPT4All : it's a tool that lets you to download a free LLM model and run it locally. Of course, we're far away from GPT4, but it could gave you a glimpse of what you can do with AI powered bots in your own business.

[Lou Hamers]

I have to wonder, can we be competitive with much less spending? Is it possible to get 80-90% of the way to "state of the art" with only 10-20% of the spending? If so, I'd say that's probably good enough. I've also seen some articles about how the cost to do this stuff is dropping fast, which is awesome if true.

My hope is that this stuff will follow the same path that other software has taken. Open source software has closed the distance in quality behind commercial software to the point where you really don't need the commercial stuff anymore. The only commercial stuff I use now is dev tools, but they're very much optional.

I think I found GPT4ALL before, but forgot about it. Maybe I'll make some time to play around with that one. I've got a 'homelab' server with way too many idle CPUs and memory...

[Tim Holloway]

The main worry about ML is that I can see a parallel with the infamous problem where people created advanced Excel spreadsheets containing small but significant errors and it became Garbage-In/Gospel-Out. I can't blame The Great Recession on that, since I was unwittingly part of the cause, and the fragile valuations we computed for Mortgage-Backed Securites were not done using any sort of AI or spreadsheet, but I've heard rumors enough on what a bad spreadsheet could do to a major corporation.

And the fatal flaw with ML is that it cannot extend beyond what it knows. I've failed at least one "programming aptitude test" because I knew more than the test creators and knew more than one "correct" answer while knowing the weaknesses of what they considered "correct".

The thing to realize about ML is that it's less about math (exepting perhaps statistics) and more about data. There are plenty of publicly-available ML engines out there. In fact, in front of my I have a PCB that has TensorFlow Lite built into it and it cost me about $16. Someday I plan to train it to send me an alert when its onboard microphones hear the "victory songs" of my laundry-room appliances.

What's really valuable and proprietary are the data sets. A good ML system needs LOTS and LOTS of data and the holders of that data would like to be paid for the effort of collecting it. Or, in cases like financial trading sets, want to have a competive secret for their machinations.