Generative AI is on a roll. And with good reason!
ChatGPT is the fastest-growing application (> 100 million users) in the history of technology. Ahead of TikTok, Instagram and the rest. Its astonishing ability to respond to (almost) all requests (“prompts”) in a polished style, implements a characteristic that was until now unique to humans: communication in everyday language, in other words, natural language.
Then there are the multimodal applications of generative AI, i.e. where the product is not (just) text:
- Dall-E, Midjourney and Jasper Art are capable of generating elaborate images from textual instructions;
- MusicLM ((Google) can now transform text into musical composition;
- Avatars Zae-In, Hermes, Lisa and Fedha present the news of the day, simulating ‘real’ TV or Internet presenters. They are even capable of answering interviews, as Zae-In recently did for the Guardian.
The spectacular nature of generative AI, immediately accessible to the general public, has overshadowed other uses of AI.
And yet, the equation AI = chatGPT or more broadly AI = LLM (Large Language Model) is wrong, by a long way.
To stay within the realm of automatic language processing (NLP), other technologies are opening up very different possibilities, in particular conversational AI.
Generative AI
Generative artificial intelligence is a solution capable of generating text, images or other products in response to prompts expressed in natural language.
The architectures used are generative adversarial networks (GANs) and pre-trained generative transformers (GPTs). GPTs are more widespread, using large language models (LLMs).
LLM: big is beautiful
From a mathematical point of view, LLMs are functions whose input and output are lists of numbers; these numbers represent words, images, sound…
From a computational point of view, LLMs are deep neural networks, whose billions of parameters (up to a thousand billion (10^12)) are trained from a considerable quantity of unlabeled text, which can also exceed a thousand billion elements.
They really are “large“.
A “magical” (or at least counter-intuitive) feature of LLMs is the sudden emergence of their competence. Below a certain level of training, they suck. In any case, no better than a random draw. The dotted red line in the following diagram is quite revealing. And suddenly they’re very good at the task in hand! A bit like learning to swim or ride a bike: you quickly go from before to after.
What's the point of generative AI?
Generative AIs can perform a wide range of tasks, most of which involve exploiting knowledge bases. For example, chatGPT’s knowledge base is quite simply the Internet.
More and more business applications and e-commerce sites are using LLMs (mainly those from openAI for the moment), as assistants or co-pilots for users.
Some examples of information retrieval:
- Finding the right document in an unstructured file library;
- Finding the right legal text or jurisprudence to apply in a given situation;
- Answering general questions (on pre-trained knowledge) or specific questions (in the case of customized training);
- Assistance with medical diagnoses…
Some examples of text generation:
- Produce summaries from documents too long or complex to be read in full;
Writing texts and articles, e.g. for the press, corporate communications, etc.
Answering assignments and exams on behalf of (overly) rushed students…
Other uses :
- Automatic text translation ;
Software development;
Customer support ;
Recruitment support…
Security, confidentiality and bias
Over and above the usual precautions applied to cloud or on-premise services, it’s worth noting the opacity of the LLMs’ pre-training conditions and corpus. This makes it impossible to seriously assess the level of bias in the answers provided.
It’s best to be aware of this, especially for applications aimed at the public (e.g. recruitment criteria), but also at employees (appraisals, analysis of annual interviews, etc.).
Generative AI hallucinations
LLMs have an original faculty: they hallucinate.
An LLM doesn’t know that he doesn’t know. So, it will produce the most probable answer based on its learning data, even if this “probability” has no basis in fact. Besides, it’s almost impossible for an LLM to provide his sources.
Sometimes this is amusing, other times more worrying. It’s possible, with a set of rules, to anticipate classic chicken-and-egg situations. But the range of queries that an LLM can’t solve is infinite. So it’s up to users to sort them out.
How do I get to Mulhouse by canal from Dijon?
To reach Mulhouse from Dijon via the canals, follow these steps:
1.Depart from Dijon, located on the Saône.
2.Sail southeast on the Saône.
3.Join the Rhône.
4.Sail southeast on the Rhône to the entrance to the Canal du Rhône au Rhin.
5.Access the Canal du Rhône au Rhin from the Rhône.
6.Follow the Canal du Rhône au Rhin south to Mulhouse.
Mulhouse south of Lyon…! This is obviously a credible formulation, but a very bad idea in reality (by the way, Dijon isn’t on the Saône either).
What LLMs are available?
With ChatGPT having had the media success and impact we’re all familiar with, the field is abuzz with activity, and a host of LLMs have appeared in recent months. Some have been in the pipeline for a long time, others are more recent. It’s impossible to keep a truly up-to-date list, but we can point to a few trends:
Proprietary LLMs: Open AI (GPT3 and its successors), Google (LaMDA);
Open access, pre-trained LLMs: Meta (LLaMA);
Hybrid LLMs, often based on a GPT model with specific training (OpenMind, Bloomberg, Hugging Face…).
- Les LLM propriétaires : Open AI (GPT3 et ses suivants), Google (LaMDA) ;
- Les LLM pré-entraînés en accès libre : Meta (LLaMA) ;
- Les LLM hybrides repartant souvent d’un modèle GPT avec un entraînement spécifique (OpenMind, Bloomberg, Hugging Face…)
Generative AI or LLM: Powerful, but expensive
Maryam Ashoori, product director of watsonx.ai (IBM) recently published an article on the cost of generative AI, depending on its use.
In a nutshell, the costs can be broken down as follows:
- The cost of calling (inferring) an LLM and generating a response; this depends on the number of tokens (word elements) in the prompt and the response: typically a few € cents ;
- The cost of training an LLM to generate customized responses from a pre-trained model; for a 48-hour ‘fine tuning’ session, count €1,000;
- The cost of pre-training a new LLM from scratch: at least €1 million (5-month training), much more for very large GPT-type models;
- The cost of hosting, deploying and maintaining a model behind an API, supporting inference and training: around €20k/month.
Naturally, these costs depend on the implementation model: access to a commercial service via an API, fine-tuning of a pre-trained model or cold pre-training.
For his part, Nicolas Oulianov (QuantumBlack, AI by McKinsey), highlights another phenomenon: the almost quadratic increase in the number of tokens (and therefore the price) when using an LLM in conversation mode.
In conclusion, widespread use of an LLM may require a significant budget, and a thorough cost/benefit analysis is needed to make the right decisions, both for the technology (choice of LLM and production method) and for the uses we want to promote.
Did you enjoy this article?
Do you now have a clear vision of what generative AI is?
Want to understand how it differs from conversational AI?
We recommend you read our second article “Conversational AI or generative AI… Ready to rumble”.
Want to understand how our conversational AI platform optimizes your users’ productivity and engagement by effectively complementing your enterprise applications?