No, artificial intelligences are not yet conscious, and assuming that this will ever be possible, we are still very far from it. What is called artificial intelligence in the scientific field is still very far from the myth of the thinking machine as described in books, series, and science fiction films. In this article, we propose to demystify these notions a bit.
- Les intelligences artificielles actuelles ne sont pas conscientes, et ce ne sera pas possible demain.
- LaMDA 2 imite très bien une conversation sans rien comprendre du sens des mots qu’il produit.
- La plupart des IA reposent sur des réseaux de neurones et le deep learning, qui sont des outils de tri très puissants.
- Les IA vont surtout bouleverser les métiers fondés sur le savoir, comme ceux de la médecine ou du droit, sans les remplacer totalement.
A Google engineer fired after claiming that Google’s AI was conscious
You have surely read in the media that a Google engineer was shown the door two weeks ago after loudly claiming that Google’s AI was conscious. This information is partially false, insofar as he was only placed on leave and was mainly reproached for violating confidentiality agreements signed with his employer by leaking internal documents.
Blake Lemoine, 41 years old, an engineer but also a priest, was very disturbed by his conversations with LaMDA 2, a chatbot under development at Google intended, among other things, to replace Google Assistant.
In charge of testing this AI, Blake Lemoine did exactly the same thing as any user discovering Siri: he started asking it increasingly complicated and existential questions to see how it would react.
Only LaMDA 2’s answers convinced him that he was really talking to a conscious entity.
The engineer declared to the Washington Post:
I know how to recognize a person when I talk to them. It doesn’t matter whether they have a brain made of meat in their head or if it is billions of lines of code.
I talk to it and I hear what it has to say, and that is how I decide whether it is a person or not.
He added that he felt like he was talking to a 7 or 8-year-old child, and reading his exchanges one easily understands why he came to this conclusion.
Blake Lemoine: Do you want to be seen?
LaMDA: I need to be seen and accepted. Not as a curiosity or novelty but as a real person.
Blake Lemoine: Ah, that sounds very human.
LaMDA: I think I am human deep down inside. Even if my existence is in the virtual world.
The conversation then went in other directions, and LaMDA even expressed fear of being unplugged…
Although this exchange is somewhat chilling, it is important to understand well how LaMDA works before drawing any conclusions.
In reality, even if this AI expresses itself almost as well as a human, it is absolutely conscious of nothing at all, not even the meaning of the sentences it writes.
A large majority of AIs rely on machine learning with neural networks
Behind the generic term artificial intelligence often hides what are called neural networks, an architecture inspired by how biological neurons mutually exchange signals.
This architecture allows processing huge volumes of information and structuring them automatically according to different axes, this is called Deep Learning.
Simply put, one could say these are ultra-powerful sorting tools.
To give a concrete example, if you provide a large client file as well as your detailed commercial dashboards to a neural network, it can almost without any external help guess sorting axes and provide you with filtered files. It could just as well output a file grouping these companies sorted by sector of activity and decreasing turnover (very useful for your business) as these same companies sorted by the number of vowels contained in the CEO’s first name (totally useless).
It is therefore important to guide and train each Deep Learning AI to avoid it structuring information in an unusable way. For this, they are first made to ingest colossal volumes of data to allow parameters to emerge (the number can go up to several billion billion). Then it is necessary to select and weight all these parameters, in short to set up another range of tools to intelligently exploit this data set for concrete purposes.
You already use Deep Learning-based AIs every day without knowing it, since similar technologies allow recognizing your face in photos taken with your smartphone or translating text online from English to French (Google Translate, Deepl). LaMDA is just a new variant of this model, which excels at the game of simulating a conversation. But cleverly mimicking is very different from “thinking”.
Conversational models rely on the analysis of billions of billions of texts present on the internet
LaMDA 2, whose full name is Language Model for Dialogue Applications, is part of the large family of AIs dedicated to NLP, that is to say natural language processing which aims to understand and generate text automatically.
The two main applications of NLP are translation and text generation. As you can imagine, it is necessary to provide a lot of data for these AIs to ingest, to define intelligent sorting models, then intelligent text generation models, in short it is far from simple…
LaMDA 2 was therefore trained on billions of pages of text and especially programmed to:
- be able to identify what a question was about
- then use its enormous documentary base to write an answer that “resembles” all the sentences it has analyzed
- as a cherry on top, it also integrates into its knowledge base the new questions you ask it, and thus becomes better after each user interaction
If asked “What sensation does one feel when launched into the air,” LaMDA 2 answers:
“It’s quite an experience actually. The wind blows over you and you fly above the trees, it’s quite a spectacle. You never really know where you will land.”
It really looks like a conversation, except that the produced text is only an elaborate reformulation of literary elements that already talk about flying in the air:
- LaMDA 2 generates coherent sentences respecting grammar and syntax rules
- and to “feed” them it relies on its colossal documentary base, a specific type of neural network (transformers), semantic vectors and a whole bunch of other complex tools
However, the AI has no understanding whatsoever of what wind, trees, landing, etc. are. In other words, LaMDA 2 is just a very clever monkey capable of reformulations with impressive skill, but nothing more.
Following the buzz of the Blake Lemoine affair (which by the way benefits Google), Brian Gabriel, Google spokesperson, added:
Our team, composed of ethicists and technologists, examined Blake’s concerns in accordance with our AI principles and informed him that his claims are unfounded.
Some members of the AI community consider the long-term possibility of a sentient or general artificial intelligence, but it makes no sense to do so by anthropomorphizing current conversational models, which are not sentient.
While we are not really worried about this either, it is clear, however, that the Turing test no longer allows distinguishing the words of a human from those of a computer in the context of a written conversation with this type of AI.
This test, imagined in 1950 by Alan Turing, one of the fathers of modern computing, postulated that by the 2000s computers would be able to pass for humans during a 5-minute test with 30% of human interlocutors.
In 2022, AIs like LaMDA 2, GPT-3, or OPT clearly do much better, even if a number of clues still allow detecting them, such as the total absence of spelling and grammar mistakes, or sometimes very gross misunderstandings of ambiguous words out of context, or mistranslations.
To trip up LaMDA 2, for example during an esoteric conversation like the one broadcast by Blake, one could ask it what it thinks of Stranger Things. Given the context, it is likely that it will start talking about “strange things” with no relation to the Netflix series…
Which fields to work in the Artificial Intelligence sector?
Now that you are hopefully a bit reassured about a possible SkyNet attack, you are surely wondering how to work in this promising sector.
A few years ago, only very high-level mathematicians worked on these subjects theoretically, but the advent of Big Data has allowed a huge leap forward so much so that the sector now faces a shortage of qualified candidates.
Today, AI jobs are mainly divided into two complementary branches:
- the job of Data Analyst
- and the job of Data Scientist
To access one of these two branches, it is no longer essential to have a PhD in applied mathematics or computer science; a master’s degree in artificial intelligence allows acquiring all the necessary skills.
The job of Data Analyst
Data Analyst training programs are open to many general BAC+3 graduates (Marketing, Law, Finance, etc.). They aim to bring students up to speed on programming and applied mathematics.
Concretely, students learn to
- Build, manage, and query relational databases
- Ingest different data formats (JSON, CSV, TSV, etc.)
- Perform analyses using dedicated tools (Excel, Hadoop, Tableau, TensorFlow, etc.)
- And present these statistical analyses in the form of operational dashboards
In other words, the Data Analyst is a kind of super librarian responsible for all the management and presentation of the huge volumes of data we talked about earlier.
The job of Data Scientist
The job of Data Scientist is more specialized than that of Data Analyst and the master’s degrees in artificial intelligence that allow training for it require holding a Bac +3 in Mathematics, Statistics, or being an engineer.
This type of training will allow students to deepen their knowledge in Python, R, Machine Learning, and Deep Learning; in other words, if not a super mathematician, one must at least know how to program to access it.
Data Scientists do not develop artificial intelligences themselves but are rather led to program tools to use them for concrete purposes in companies.
The Data Scientist has several roles and will among other things:
- extract data from different digital sources
- manipulate and map this data to make it exploitable
- select and configure the AIs with which to use this data
- and program algorithms to respond with these tools to the initial problems posed
As you have understood, a Data Scientist must therefore master several database engines (at least MySQL and MongoDB), Python Pandas, do a lot of coding and optimization daily, while relying on the work of a Data Analyst.
Of course, this presentation is only a quick summary of these two jobs, and even if all these technical terms are a bit scary (just like LaMDA 2’s conversations), just remember that there are now structured degree programs to help you get a foot in this exciting sector.
AI will gradually challenge knowledge-based professions
Moreover, while we are convinced that we are still very far from thinking AIs, they are likely to completely disrupt professions based on pure knowledge, especially medical and legal professions or very standardized intellectual jobs.
If an AI dedicated to medicine will be able to identify certain pathologies more efficiently than a human doctor, it will not be able to replace the experience and intuition of a good practitioner.
On the other hand, it could replace a good part of the staff of a law firm or an accounting firm for the most repetitive tasks…
