The speed at which artificial intelligence (AI) is evolving is already insane. The expected arrival towards the end of 2023 of GPT-4, the next generation of the system that brings the ChatGPT and Dall-E interfaces to life, will further accelerate the trend.
People who are already out of breath from the advances in automated natural language understanding and generation in recent months are encouraged to catch their breath, take a look at a dictionary and jot down the definition of “exponential.” Because that’s pretty much what we’re experiencing these days: an improvement that seems to be ever faster in the computerized systems available to just about everyone.
Where will these advances stop? It seems hard to predict. On the other hand, we must already think of the perverse effects that these systems will have on the daily lives of many people if they are not supervised very quickly.
artificial general intelligence
Because the balloon must immediately be prevented from inflating to a certain level of bursting. Remember that, despite the most recent progress, the AI interfaces that we have seen emerge in recent months are not perfect. They come a little closer to this other concept dear to computer research, that of an artificial general intelligence, or “AGI” according to its English acronym, but they are not yet related.
Such a system, according to the accepted definition, would take the form of software that could be asked about almost any type of problem and which would know how to find the right solution. Such an AI could in principle perform all the intellectual tasks that humans are capable of performing.
Obviously, to achieve this, such an AI will have to be able to distinguish between what is true, and not just probable, and what is false, even if it is probable. Ideally, his creations, his works, should be authentic, and not just a collage of styles or inspirations obviously taken from others. This is more apparent on the side of the Dall-E image generator, which is better at imitating the pencil or brush strokes of existing artists than it is at developing its own visual signature.
That said, the track is already blurred: on certain Internet forums where visual artists meet, we have started to ban visitors whose creative style looks too much like that of an AI. We could see the same thing happening in schools later this winter, when the first big papers are marked, and teachers may be worried about reading papers that suddenly seem more researched than usual.
In other words, even imperfect, the AI currently accessible via the Internet is already making its presence felt everywhere. As usual, hardly anyone was ready to see a seemingly mundane website radically transform their habits. Especially not governments.
It would be nice if a framework to limit the abuses and failings of tomorrow’s AI could see the light of day… today. And not in five years. Europe could adopt its own framework next summer.
It won’t be too soon, as the next Dall-E and ChatGPT upgrade of this world is expected later this year.
GPT-4
Going live in early December 2022, ChatGPT only needed a week to shake up the entire internet. This web-based software is based on a machine learning model called GPT-3. It was, at that time, the most recent version of the “Generative Pre-trained Transformer” model produced and constantly developed by researchers at OpenAI, the company behind this system.
In its understanding and generation of text or language, the GPT-3 system processes the information it receives from 175 billion different criteria. It is enormous ! This is also a serious problem with this technology: it is extremely energy-intensive.
Yet that’s nothing compared to what will be GPT-4, its next iteration. According to what OpenAI executives said lately, GPT-4 could process information according to a number of criteria that are rumored to be between 1 and 100 trillion (or between 1000 and 100 trillion criteria).
This system will be much more powerful; he will almost read our thoughts!
He could better understand our intentions behind the questions we ask him, without them being explicit. He could also respond in a way that was even more natural, or at least able to better imitate human conversation.
OpenAI knows the limitations of its technology and may include safeguards in GPT-4 that will limit the AI’s ability to invent falsehoods. This could take the form of reinforcement learning algorithms taking more into account the reaction of users.
Because, yes, interaction with humans allows the machine to improve. Taken upside down, this means that each time we have fun with Dall-E or ChatGPT, we contribute to the acceleration of the development of a machine that will undoubtedly replace us one day in many daily tasks.
And it should happen in 2023…