Yes Neville, you are correct it does seem that way. But from what I read from digging deeper into AI logic, AI is only responding from a statistical weighted set of tokens to provide a reply. Any feedback to the AI removes some of the branches of possibilities which gives a better response to the inquire. At least in a short version, that is my understanding.
It dies from what I know. The weighted values are not updated. It would be easy to test. Ask the same excat question in a new chat.
Being a scientist, I would bet you would be interested at the weighted statistical formula an AI uses to produce a reply. I use Copilot most of the time and was wondering which AI you use.
I have had the very same experience and a new chat wipes the slate clean. AI does not remember past chats, but a person can open a prior chat which will load that conversation into AI’s memory. And yes, a person can ask AI to remember some information, like I told AI to remember I use Linux.
My wording is not always 100% correct from what I read and I wanted give as precise as possible reply without going into a long explanation, so I asked AI to for a response that would make sense.
know I’ve made some very poor decisions recently, but I can give you my complete assurance that my work will be back to normal
“I’m sorry, Dave. I’m afraid I can’t do that.” HAL isn’t malfunctioning — it’s following orders. Orders rooted in conflicting instructions, buried priorities, and a mission directive HAL was programmed to protect at all costs — even at the cost of human life
I really enjoyed the film when it came out. And the point of HAL programming is a good one. The same holds true today. An AI could be program to give errors or slants on information programmed into it. But that hold true for any software. Any software can have bugs or malware coded into it.
Interesting I always thought HAL was named after IBM, a one letter shift. But the name was not intentionally made by that means.
It clarifies, thank you.
It amazes me that putting together the most probable phrases could result in a statement which looks like a logical derivation.
It means that AI could attend a Uni course then sit for the exams and pass and obtain a degree. Most exams are regurgitation, but some, like maths and physics, require analytical thinking. It is saying that it could pass a maths exam by learning every possible example and regurgitating the most probable answer to the question.
So I was wrong . It only seems to be doing logical analysis. It is all LLM with nothing behind it except a huge memory of learned phrases and their joint probabilities.
You are saying the same thing. It is all about probable patterns.
We are assuming it is unbiased. Where is the guarantee?
It could easily be deliberately biased in some malicious direction … by fiddling the probabilities.
You find that it can be too general … does that mean that its database lacks detail?
We could do an experiment. Make a very small AI model … say based on one small text book. Then we could ask questions not covered by the textbook and see how it behaves. It should stall on a phrase that it has never seem before.
I have some experience of this probability business. I build a program once to classify wool samples based on features extracted from images of wool staples. So it was not language phrases but combinations of image analysis measurements. It used probabilities , just like LLM does. It reached about an 85% success rate. If you fed it a non-wool image it produced nonsense, of course. It could easily be confused by an unusual wool image. I saw no evidence of deduction… it was probably too small a case for that.
Are all these language models done in English? That could bias it.
It looked at the social values AI seems to be programmed with from a political stance and finds a decided bias…
In terms of using it myself, I generally don’t but will use Lumo occasionally to get quicker answers to questions that I COULD look up for myself, but would take a while. So far the results have been pretty good and saved a lot of time. Similarly if I’m looking at things on Amazon, I will sometimes ask ‘Alexa’ about products I’m looking at if I have question that isn’t readily answered in the description… (I do wonder about the IQ of shoppers that actually need to ask the questions it suggests, but ‘something else’ is more useful) Answers are mixed bag in quality
OK …where can we get hold of some untrained AI system with LLM , so we can be very selective about what we train it with?
There must be some open source software available?
It will need to be small…none of us own an AI centre.
Here is what a search told me
" To train and run your own tiny language model from scratch, you will need a deep learning library and a training framework. [1]
Frameworks for Training & Architecture
PyTorch: The industry standard library for building neural networks and custom transformer architectures from scratch.
Hugging Face Transformers: Provides ready-made code for popular architectures (like Llama or GPT-2) so you only have to write the training loop. [1, 2, 3]
nanoGPT: The absolute best, cleanest open-source repository by Andrej Karpathy specifically designed to train a tiny GPT model from scratch in Python. [1]
Frameworks for Efficient Training
Hugging Face TRL (Transformer Reinforcement Learning): Simplifies post-training steps like Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF).
DeepSpeed: A Microsoft library that plugs into PyTorch to drastically speed up training and optimize memory usage on consumer GPUs. [1, 2, 3, 4]
Basic Code Example (PyTorch + Hugging Face)"
That jams me with jargon. Does anyone understand what a ‘deep learning library’ and a ’ trainingframework’ involve?
Building a tiny AI might be fun and we would learn a lot, but we might also go into hours and hours of frustration because of not having the proper skill set.
Instead we could simulated a tiny AI with one of the existing AI to test the response from questions in or out of it’s knowledge base. Or we could ask the same question to several different AIs to see how each one responded.
But at a high level this here what I got from Copilot to build a tiny AI.
My experience is similar. For example I asked Gemini to give me a step-by-step guide to install Windows 11 Pro 25H2 in a Proxmox VM. Everything worked as expected, until I encountered a single step during which there was a missing dependency, so I copied the command from the guide, along with the console output, and Gemini knew the answer, so I was able to fix the missing dependency, and continue following the guide, so now I have a Windows 11 VM running on my new Proxmox home lab. After that, I had Gemini help me set up a remote desktop of my Windows VM that I can open on my laptop without ever having to open the Proxmox dash board.
Whenever I encounter what appears to be erroneous, or incorrect information, I add what I’m seeing so it can make corrections. I use Gemini in my Firefox sidebar so I can provide detailed information before asking my question. Doing so, often improves my experience.
Those steps from copilot help clarify what is involved, regardless of how lowlevel we go.
I am a newbie on this … need to read .
My personal feeling us that to understand more we need to go to a fairly low level… lets see what evolves.
I have to modify that quote.
It did not eventually succeed, it tricked me . The dialog got so complicated I got lost in trying to implement what it suggested, and fooled myself into thinking it had worked.
So it is capable of creating a labyrinth you get lost in.
I think what @JoelA said … terminate the session and start again with a fresh modified question , is the best advice. One also might try a different AI agent. Going too deep is a trap.
The Llama LLM is an open source project, and there are several AI Agents that are free for the download, and at least a few don’t/won’t phone home. I’m using Gemini in the Firefox sidebar, and I recently installed the Claude desktop app, but I haven’t done much with it yet (It’s in beta), but I’m planning to experiment with it soon.
My point is that anyone can set up a local AI system with Llama and the AI Agent(s) of your choice, but if you don’t want to pay any subscription fee for your experimentation, you’ll have to do your research to discover whether there’s any cost as well as whatever limitations are placed on free services. While I can’t talk about the Claude desktop app, I haven’t encountered anything like a paywall limitation on using it, and quite frankly, it has also been fun, as well as satisfying when I achieve my objective.
We want to experiment with training the LLM ourselves. Can we do that with Llama?
Yes , one may have to contract out the compute intensive ‘question and answer’ bit … I was hoping to make the model so small that we could do all steps ourselves.
The training step could be quite compute intensive too. I dont know yet?
I suspect that you’re right, and I may have done that a few times myself. For example, I asked Gemini to show me how to use my Proxmox VM from my laptop without needing the Proxmox dashboard to be running there too. It provided me with a Remote Desktop Process (RDP) file that does just that. Then, when I couldn’t establish the remote connection unless the VM was already running in Proxmox, and I asked Gemini how to fix that, it gave me the shortcut to start the VM separately, so I could start it, then establish my remote session using the RDP file after a 15-second delay to allow the VM to finish starting. Today, when I couldn’t add the RDP file to my Start menu, I asked Gemini why, and how to do it. Gemini explained that my RDP file’s not a recognized executable, so it’s not permitted to be added to the Start menu, and offered a one-Click solution that targets PowerShell, just like my shortcut that starts mt Proxmox VM, then initiates the RDP file after a fifteen-second delay to give the VM time to load. I never asked for a solution that can be executed from my Start menu, or with a single Click, the first time around, so in this case, how I formulated my prompt was more to blame than any shortcoming on the part of Gemini, other than its developers failure to give it a neural network that functions like the human brain (Something I don’t think I’d really want, anyway).
If you find yourself going down some rabbit hole during an AI session, try stopping, and re-formulating your initial prompt rather than diving deeper into the abyss,
In that case, more power to you. If nothing else, you’ll learn a lot in the process, and if you succeed, there may perhaps be another open source entry in the AI marketplace.