Bet it answers quickly though. Good old Groq.

This amused me.

https://www.businessinsider.com/chatgpt-giving-users-unhinged-answers-no-one-knows-why-openai-2024-2

There’s something interesting in the idea of ChatGPT having an off day. Ironically, it humanises the unknowable bot. Maybe there’s a creative idea here about deliberately training in a little craziness.

ChatGPT wrote:

Would it glad your clicklies to grape-turn-tooth over a mind-ocean jello type?

Chatbots be jazz, yo.

LLMs in themselves aren't the powerful part of AI although they are a powerful way of getting info from the internet. The really powerful stuff is custom models built for a particular dataset but it's not general use in the same way as LLMs are for the public. But yeah there's still the problem of not knowing how they reach a decision. 

It'll be interesting to see how Google's multi-modal approach works for general AI but I suspect for a long time AI models will be data specific. That makes them stupid good at a particular task but there's also the danger of trying to get differing models to work well together on differing but linked datasets to understand a more complex problem. Again it depends on what we allow it to do and that'll be country specific.

The optimist in me hopes they can figure out how models reach a certain decision and that'll help humans understand a problem in a deeper way. A lot might depend tho on human interpretation of the AI and political persuasion. It might be a little too easy to interpret the results using confirmation bias.

I can recommend this.

https://sourestdeeds.github.io/pdf/Deep Learning with Python.pdf

It really is the best book imo, both for learning the concepts and building your own neural nets.

I've been coding for over 40 years since I was 15. I can pick up any language pretty quickly. The book I've ordered off amazon arrived today and it's Python code samples too.   I've found most languages are starting to converge now. You get the occasional syntactical quirk but once you know the package/library names for your import statements it's not too bad really.  The concepts are the same no matter the language unless you get into heavyweight declarative stuff.

My main issue now is picking a nice dev environment. I usually just use a text editor with a CLI compiler but I'm sure there's easier ways.    The cloud environments are nice.

Will have a look. I've got some environments already - have one for the phone app dev (React language) and I have Visual Studio but I've only been tinkering. Got a pile of React training material in a humblebundle I need to get back to.

Okey doke. Cheers.

React is not a language [/pedant]

Stuff like that I can only see as noise we eventually learn to identify and filter out.

Buying the item with an obviously hand written description will become attractive.

acemuzzy wrote:

React is not a language [/pedant]

'tis a fair point.

https://amycastor.com/2023/09/12/pivot-to-ai-pay-no-attention-to-the-man-behind-the-curtain/

I do like the Mechanical Turk workers using ChatGPT and other AI tools to process the AI-input tagging/manual-labour tasks Canny use-case and could be another route for the "garbage-in-garbage-out" factors leading to AI-training-on-AI-output-inbreeding-self-degradation-spiral.

I don't regard cannibalisation as a long term problem and the idea you need to clean the data is wrong, as long as you have enough data. You might not always have a lot of data. These things are just a result of the current bad practices the current boom is bringing with it as everyone jumps on the train.

Even if you do have data made by AI that was based on another AI (and so on), the patterns in the chain are still there but deeply buried, and if you get enough data the model can figure this out. The unreasonable effectiveness of data is a thing and it's hard to get experienced coders to appreciate this.

The new rule is "Gazillions of garbage in, useful stuff out". 

That's not to say the article is wrong, and there are many badly trained models out there, but it's just a side effect of early tech hype. Datasets never lie and it's finding the complexity in them. I could randomly generate a bunch of numbers on my PC and use that to train a model. If I had enough data a good model could find how the RNG worked even if I was training it on something else because that info is there, buried in seemingly random garbage. Needs a shit-ton of data though.

But at the mo then yes, it can be a problem if you let it and people are letting it.

Sounds like it’s got confused around words like accessibility. Which in your case should have meant legibility, but in other contexts will be found alongside words like diversity and inclusion.