Don't mistake LLM with AGI ...

When it comes to AI algorithms, it’s very important to understand what they can do and what they cannot.

ChatGPT has created a lot of buzz lately and wrongly some people even thought it was something like an omniscient AGI (Artificial General Intelligence) that could answer any problem you might ask. And it is true, it can provide an answer… but the answer is not always the solution, especially when it is wrong. And as a LLM (Large Language Model), basically any Generative Pre-Trained (GPT) model, is going to provide an answer to your request… but will not provide any reliability indicator to decide whether the answer is appropriate, or trustworthy. Sometimes based on its training dataset, it will even “hallucinate”, or “confabulate”, i.e. tell you things for the conversation’s sake, but things that are pure “invention”.

Such behaviors can be partially fixed through setting an appropriate context and explicit rules to “ground” the LLM to a specific usage/context.

But one should always remember that LLMs are “Language Models” and that ChatGPT is a conversational model… not something able to generate complex reasoning to extract the “truth” out of ambiguous issue settings.

Lately, as I am a lazy boy who needed to help one student in her bachelor of Maths, I gave a try to ChatGPT+ (with GPT4 model).

Setting the context (I don’t know really if it’s change something to the behavior of the LLM):

“I want you to behave as if you were an assistant professor in mathematics bachelor. You have to assist students during tutorials. you must provide them with clear answers, breaking down the stages of reasoning. You must also provide them with reliable answers. The topic is about matrixes and linear applications”.

ChatGPT answered it had understood my request and kindly proposed its assistance in that “context”.

Initially, it behaves appropriately as we were only manipulating "context"/concepts...

First issue was to enter matrixes in the text interface. However, by saying “let’s say A is a 3x3 matrix, with first line | a b c |, second line |d e f | and third line |g h i|” chatGPT was able to guess out how A does look like.

Then the question was about finding the eigen values of the matrix A and associated eigenvectors.

ChatGPT was able to determine that to do so, we needed to calculate the roots of its characteristic polynomial and it was also able to explain the reasons for this.

The characteristic polynomial is det(A - λI) and we need to find its roots. So far, so good…

But then, it comes to the calculations themselves…it was pure hallucination

ChatGPT identified a polynomial that was wrong and despite I asked to have a step by step calculation, it just provided me with a false polynomial, something like det(A - λI) = -λ(λ^2 - 14λ + 50)

OK why not (at that time I did not check whether the polynomial was the right one or not)… let’s move it forward.

And then ChatGPT did totally hallucinate identifying 3 roots to my surprise:

λ1 ≈ 2

λ2 ≈ 7.162

λ3 ≈ 4.838

As the polynomial used to find the roots was -λ(λ^2 - 14λ + 50), I asked “don’t you think 0 could be an obvious answer, and then you could use quadratic forms to solve the second degree polynomial”.

In a very polite way, ChatGPT admitted its mistake, saying that 0 was an obvious answer and started to calculate the discriminant (b2-4ac) = (-14)2-4*50=-4

To conclude that there was no real roots to this second degree polynomial as the discriminant was negative… Not too bad… but I realized that then it would probably mean that the characteristic polynomial was the wrong one.

To push it forward I did answer “yes but what about complex roots?”.

ChatGPT did calculate two complex roots and concluded that the 3 eigenvalues were 0, 7-i, 7+i

Hopefully for the student, I am a lazy boy, but I don’t trust LLM when it comes to do even simple maths. Eventually the true polynomial was (λ-2)*( λ-3)*( λ-7) with obvious roots then.

Lessons that can be learnt from this little experience:

1.??????Use LLM for what they are good for, i.e. handling a conversation without forgetting they can hallucinate. But don’t expect them to be good at anything else and do not mistake them with AGI despite they might sound like it as they always provide a structured-perceived answer (but that can be totally “invented”). Here ChatGPT was good at explaining how we could find eigenvalue, i.e. by solving the roots of the characteristic polynomial (OK, any student who has read her/his maths lessons knows that too). LLM are made to provide a Natural Language based interface to a "system"... not to be the system itself and provide the functions for which the "system" has been designed. If you think about Iron Man's lab assistant Jarvik : LLM will only handle the natural voice and polite interface... not the complex scientific problem solving capabilities underneath.

2.??????To best leverage the LLM and AI in general, you have to understand they way they work internally, know a little bit of their “architecture”, to understand their weaknesses and strengths. Here, despite setting a context asking for reliable questions… it did not prevent ChatGPT from hallucinating… because it cannot provide reliable answers to maths problems that are “deterministic” while it is all based on probabilistic approach of the next good action/letter/word/sentence… without understanding the rules to which matrixes obey. AI is somehow like a MooC… everyone can attend a MooC on “Astrophysics” or “Quantum Physics”, but only those who have the right academic equipment / prerequisites can leverage it to increase their knowledge.

3.??????Use the right tool for the right issue. LLM are not good at doing maths, complex reasoning, hardcore computational tasks … just simply because they are not meant for that. You have something that can do this… it is called a brain. And ok, if you’re that lazy, you can also use specific software such as “Mathematica”, or even simply some online “matrix calculators”. Those specific tools/Software will provide you with the appropriate good answers… at a fraction of the energy needed by ChatGPT to provide an hallucinated wrong answer.

?As Openai (cognitive) services are now available as part of Azure Cognitive services, you can easily leverage the power of LLM (knowing their strengths and weaknesses) to build AI-infused solutions that will enforce your data security (your data are stored in your tenant in which cognitive services are run) and provide you with all the trustworthiness of Azure when it comes to data privacy and cybersecurity. Just a simple example: If you combine “Speech to text” capabilities, with “chatGPT”, and something like “Vall-E” (voice generator), you can build a NLP solution based on voice interface. You can then set a specific context and specific rules to ground this “voicebot” to a particular context… hum being a French guy… Why not a “E-sommelier/E-Wine Waiter”...?Tell me what would you suggest to best accompany my "Boeuf Bourguignon" dish for lunch ?”.

Let's give it a try... Industry Solutions Digital Advisors can assist you in defining new products and services that could benefit from AI for disruptive differentiation while Industry Solutions Delivery teams can help you building your AI-infused applications to provide an enhanced UX for your partners and customers... to build your own co-pilot.