Let's talk about Artificial Intelligence

[old man emu]

5 hours ago, red750 said:

AI-generated answer. Please verify critical facts.

An OK generalisation, but still not tellingh me anything. How is an algorithm created in code?

[rgmwa]

My understanding is that the neural network uses formally coded algorithms to alter its internal weighting parameters based on whether the model has made a successful prediction or not. If the neural network is big enough and the data it is being trained on is extensive enough its predictions become more accurate to the point where the AI model seems intelligent.

Jerry can probably explain it better than I can. 

[old man emu]

11 hours ago, rgmwa said:

Jerry can probably explain it better than I can. 

How do you create a "formally coded algorith"? That is my question.

[facthunter]

Why expect to find the answer HERE? We are mere Mortals.  Nev

[red750]

2 hours ago, old man emu said:

How do you create a "formally coded algorith"? That is my question.

 

Just ask AI.    https://www.phpcluster.com/how-to-write-an-algorithm/

[rgmwa]

2 hours ago, old man emu said:

How do you create a "formally coded algorith"? That is my question.

Like any other program. The programmer decides what the rules are that govern how the state of the ‘neurons’ in the network should change depending on the outcome of a trial and implements those rules in a programming language like C++ or Python. It’s just a feedback loop. You start by assigning random values, say between 1 and 10 to a node in the network and run a trial that gets it to predict an outcome. If the prediction is correct that node’s value is increased by some amount. If it guesses wrong the value is decreased so its importance in the network is reduced because its opinion is less ‘reliable’.  Although I’ve said ‘a’ node, it’s applied to all nodes in the network. That’s the basic theory as I understand it. If you have billions of nodes being ‘trained’ the network may eventually start to seem to produce intelligent answers to questions although nobody really knows exactly what its internal state of ‘mind’ is because although the rules of the game have been coded the process of learning’ is fundamentally random trial and error. 

 

Edited 17 hours ago by rgmwa

[rgmwa]

I suppose an analogy might be chess. The rules of chess are known (coded) but the process by which each game reaches a conclusion is unknown (the AI model) even though the conclusion is known - it's going to be a win or a draw ( a correct prediction).

 

[facthunter]

There are Many different ways to win at chess and there are Many checkmate Positions. Can't you also have a stalemate?. .The Rules are the  Way the Various Pieces  are Permitted to Move.  Nev

[Marty_d]

In theory it should be easier for a good computer program to win chess against even the best humans.

A computer can look at all possible combinations of moves on both sides for the next what, 6-8 moves or more?

Try doing that yourself... I have difficulty working out the possible combinations for the next 3 moves in just the critical areas.

[nomadpete]

What does A.I. convert human language into, in order to compare the words with the algorithm?

 

Is it really just  language modelling and comparing with databasees of similar language? 

 

If so, then it isn't rationally thinking.

[rgmwa]

15 minutes ago, nomadpete said:

What does A.I. convert human language into, in order to compare the words with the algorithm?

 

Is it really just  language modelling and comparing with databasees of similar language? 

 

If so, then it isn't rationally thinking.

This what it said when I asked it, but don't ask me to explain it:

 

What does A.I. convert human language into, in order to compare the words with the algorithm?

 

AI converts human language into high-dimensional numerical vectors, often called vector embeddings. These are essentially long strings of decimal numbers representing coordinates in a multi-dimensional mathematical "meaning space

How Language is Transformed into Math

To process your words, AI follows a strict pipeline:

1. Tokenization: The algorithm breaks text down into small pieces called tokens (which can be words, prefixes, or characters) and assigns each a unique ID number.
2. Vector Embeddings: Each token is mapped to a dense list of numbers. A single word can be represented by thousands of dimensions (e.g., 1,536 to 3,072 separate numbers). Each number acts as a coordinate tracking a specific micro-feature of meaning, syntax, or context.
3. Contextual Math: Words with similar meanings or grammatical rules are mathematically positioned close together in this coordinate system. Because language is reduced to geometry, the AI can perform equations like:
   Vector("King") - Vector("Man") + Vector("Woman") approximates to  Vector("Queen")

Is it really "just" database matching?

No, it is much more dynamic than a static database lookup. While some AI applications use standard vector databases to fetch similar documents (Retrieval-Augmented Generation or RAG), modern Large Language Models (LLMs) operate on active predictive generation, not simple database matching.

• The Transformer Architecture: Models use Self-Attention mechanisms to dynamically alter a word's vector depending on the words around it. For instance, the word "bank" will be assigned a completely different vector in the phrase "river bank" compared to "investment bank".
• Probabilistic Computation: The algorithm doesn't copy-paste text from a memory bank. Instead, it processes the text through billions of internal mathematical nodes (weights) to calculate a probability distribution for the next logical token.
• Emergent Logic: Stacking these layers of math allows the AI to discover the complex geometric patterns of human logic, enabling it to map out reasoning paths rather than merely retrieving matching records.

The only part of this that seems close to what I was saying earlier is the Probabilistic Computation bit. Clearly AI models have made some progress since I last used finite state machines to unsuccessfully predict next week's Lotto numbers in the previous century. 

[nomadpete]

Gee, thanks for that.

 

Much of that sounds like self agrandising word salad.

 

So, LLM  A.I. is still just using contextual word and phrase comparisons. 

 

But you lose me with predictive aspects. I'd agree that probability might be possible, but predictive results gets too close to seeing into the future. That's impossible.

 

My take on A.I.'s own answer is that it's often limited when it comes to initiative or sound logic.

[rgmwa]

1 hour ago, nomadpete said:

But you lose me with predictive aspects. I'd agree that probability might be possible, but predictive results gets too close to seeing into the future. That's impossible.

Yes, that’s why I never won Lotto. To be fair that wasn’t really a good test of the methodology. Lotto results are ultimately random, so there was no long term pattern to them that the algorithm could have been trained on to predict the next draw. Had there been an underlying bias it’s possible that the algorithm could eventually have improved the odds in my favour by picking a set of numbers that more likely than not contained next week’s winning numbers.

That was my aim at the time. It didn’t work but the programming was fun. While my finite state machine approach was a far less sophisticated AI model than the technology behind ChatGPT or Claude, the principle of using a data set to train an algorithm to mimic human learning by modifying itself by trial and error until it can produce answers to questions that appear logical and intelligent remains the same. Of course it doesn’t mean that the answers are right. 

 

[Jerry_Atrick]

I don't have a lot of time, so. I will try to be quick (given I have only skimmed the posts since OME asked the question). 

 

The algorithm is not predictive in the sense of predicting the future. The predict what you meant by what you typed, and whether or not their response will make sense. For example, if you asked how many legs a dog normally has, it will predict that 4 is the correct answer; not that my some genetic mutation caused by increasing radiation in the atmosphere, they will grow 5 legs in the future.  Well, not unless you ask it, anyway. But even if you did, it would use existing data to try and predict the correct answer based on the data it has, not to perform its own scientific analysis and determine for itself. It may take the factors and apply it to an existing theory to mimic scientific analysis and intelligent conclusion drawing, but it is not doing that itself. 

 

In terms of AI implementation, it is likely to be a lower level language such as C or C++ = not python because it is too slow; the end point AI functions can be done in Python, but that is on a precomputed large dataset - usually. 

 

The algorithm will be surprising small.. usual Bayes theorem or something similar. It is able to be repeated over billions of nodes - or data points. The code will likely be recursive in building the precomputed data set to apply simpler probability maps across the data you provide. And it will likely be recursive in using that map to answer your question. By recursion we mean the same code calls itself.. So, you will be able to crawl heterogenous data sets, over and over using the same code developing deep nested stacks (well, heaps to be accurate) of responses from which it will map as many of the different types of data that can be used to probably represent what ever bit of data it actually is. 

 

Tagging data is old these days.. That is what these maps do. 

 

This of something like if you type the number 4 into your LLM. That is easy, because it is represented by whatever coding mechanism you are using... and there are only a handful of coding mechanisms - ASCII, EBCDIC (not used anymore) UTF-8, UTF-16 and a few others. But, they are translated to the discrete representation of 4 to AI. No thinking required. Now, say you provide it a scanned doc, which contains the number 4 somewhere. It has to read all of the pixels and look at the shape of each character to determine if it is a letter, number, whitespace, part of a diagram, etc. Say it worked out it was a digit. Now it was to work out what digit it is? It has to look through its maps/nodes and get the most likely digit it is.. it predicts what digit is is because it spits out a probability index of what it has been taught what that digit could possibly be, across any digit which it may match. Is it a 4, or is it a H.. 

It also has to take into account font. As you know, a 4 can be represented very differently in different fonts...

 

So, the sort answer to your question is that there will be a set of probability algorithms that will read data, build probability maps, and then apply the same probability algorithms to inputs and compare to what it has learned. Machine learning means it constantly refines its probability maps. 

 

The technical implementation is likely to to be a lower level language (a lot of Python maths/stats libraries are written in C++).. and there will be a lot of recursion.

 

 

[rgmwa]

How does it decide how the probability maps should be altered to get a better result?