«Generative AI and its possibilities. (2/5). The technical foundations of generative AI are simple but very powerful. Human validation is essential for critical decisions. And for businesses, its use should only occur after data enrichment.
«It's astounding. Frankly, it's astounding!» Who hasn't had that thought when trying ChatGPT. And incidentally, the term «bluffing» applies well to this Generative AI: it's both very impressive but also gives a slightly «inflated» impression of what can be done with it. This isn't about condemning this technology, but rather understanding what can be expected from it.
Generative AI has become considerably significant in the press in recent months but isn't actually that new. French researcher Yann LeCun, Head of AI at Meta, even stated some time ago that:
«In terms of the underlying techniques, ChatGPT isn't particularly innovative. There’s nothing revolutionary there, even if that’s how it's perceived by the public. It’s just that, you know, it’s well-packaged, it’s nicely done.» [1]
Indeed, text (or image) generation is an evolution (a culmination?) of the work carried out successively on deep neural networks. The processing of Time Series had initiated efforts to take context into account to make a prediction. And in fact, technically, the basis of Generative AI is to suggest a word (a token from the preceding sequence.
In a caricatured way, it boils down to asking you for the word that follows in the sentence «Merry Christmas and Happy XXXX». No doubt, you have automatically replaced the «XXXX» with «year» because that is usually the word that follows. So, obviously, for GPT (Generative Pre-Trained Transformers) type models A lot, a lot, a lot of examples had to be given to arrive at the result we all know.. However, from a technical perspective, the objective is «simply» to find the word that is statistically most probable in a sentence being constructed, as the machine has no real understanding of it. Furthermore, in order to vary the responses, DeepLearning models offer a «temperature» parameter to adjust the probabilities of generation. Microsoft on its Azure OpenAI platform describes it as a parameter «to use that controls the apparent creativity of the generated completions» [2].
This leads us to ponder the role these new tools can play, and almost philosophically, our own way of working and being.
Since work on these topics has existed, the term Artificial Intelligence has been almost a false friend. We should perhaps talk more about «Artificial Experience».» The mathematical principles governing its existence assume that numerous cases are subjected to its functioning. In the world of work, even more than elsewhere, it is the fact of having encountered numerous cases that allows a senior to react perhaps more quickly and better than a beginner: «This is what needs to be done, because this is how, in this type of case, one gets out of it best.».
How many of us, after several years in a firm, have developed true «libraries» of PowerPoint slides that we use and reuse until we're sick of them? Who hasn't ended up with «ready-made answers» to objections from a client or a colleague. It's the «ready-made thinking» that one can sometimes denounce.
However, within your company, are there roles or tasks that require, first and foremost, a good understanding of the company, its processes, and its documentation? One might naturally think of call centres, for which answers are often scripted or require in-depth knowledge of contracts and their interpretation. FAQs, or even digital assistants if they are prevalent on websites. are good candidates for benefiting from generative AI. The adaptation of a Large Language Model (LLM) to your company's context can then be done through supplementary «training» from an existing document base.
Conversely, and fortunately, not all jobs meet these criteria for automation. For example, if you are looking for an idea to stand out from the competition... well, by construction, you shouldn't do things as usual, like everyone else. At best, one might to draw on generative AI for an idea on what not to do !
Similarly, we mentioned at the beginning of the article the «statistical» functioning of artificial intelligence. A robot like ChatGPT will mechanically give an answer to any question it is asked. Even a term that is unlikely for a human being... has a non-zero probability and is therefore acceptable for the machine. These are the famous «hallucinations» that major companies are working against with massive human validation. One of the most striking examples is that of a lawyer who questioned a robot about case law. The machine provided him with exactly what he hoped to find... until checks proved it had all been made up.[1] «Facts are stubborn things,» as we often say. And this is what is necessary for critical decisions. Would we accept being convicted based on «habitual behaviour»?» Of course not! So, should we make important decisions for our company without seriously substantiated documentation? No, that too. Here again, the machine can be considered an aid, but it is important to ensure the explainability and therefore the acceptability of a proposal.
Thus, generative AI joins with a fanfare the tools that these techniques make available to us. The fruit of training on massive examples, it remains that the answers provided remain the application of statistical calculations. So certainly this is an excellent way to «tap into» a company's knowledge base and thus assist roles that rely on it, but we should not lose sight of the non-creative aspect, often difficult to justify, of the responses produced.
Depending on the level of task automation, this type of solution can, however, prove to be a valuable ally once a base model has been enriched with your own data.
[3] https://www.capital.fr/economie-politique/cet-avocat-fait-plusieurs-erreurs-dans-une-affaire-apres-avoir-utilise-chatgpt-1469903
