Has not been the year Skynet became self-aware (that, according to the movie Terminator, should have taken place at 2:14 on 29 August 1997), but almost: 2020, the year just passed, a terrible year for humanity, struggling with coronavirus, contagion, lockdown and quarantines and near-total freezes, it’s been a great year for AI. Perhaps the best year since the development of these began incredible machines, capable of learning, to act even without the need for the intervention of those who created them e maybe, in the future, even to understand our emotions.
To certify it is theAI Index Report of the University of Stanford, now in its fourth edition (this is it): leafing through it we discover, for example, that the quantity of scientific publications dedicated to the topic has grown by almost 35% between 2019 and 2020, over one and a half times more than what had already grown between 2018 and 2019, and also that if in the United States most of these publications are financed by private companies, in China and the European Union most of the funds come from governments. On this last point it is interesting to note that also in this field the US-China dualism already seen in other technological sectors, but also how now over 50 countries have set up (or are setting up) a national strategy for the development and control of AI.
The search for jobs related to artificial intelligence
Economic interests and effects on the world of work
This obviously has repercussions on employment: the demand for workers specialized in artificial intelligence is growing more or less all over the world, so much so that in Singapore it has come close to 2.5% of all personnel searches (graphic above); it fell slightly only in the United States, but according to the Stanford researchers this decline could also be linked to the achievement of a sort of maturity of the market, at least temporary. In detail, the countries that between 2016 and 2020 hired more professional figures of this type, in addition to Singapore, were Brazil, India, Canada from the South Africa.
More: in 2020 almost 68 billion dollars were invested in the development of artificial intelligences, a jump of 40% compared to the previous year and a impressive growth compared to 2015, when the economic commitment of governments and private companies did not reach 13 billion. But what is all this money invested in? What are AI’s enormous computational, processing, and problem-solving capabilities being used for? Apart from that for create Tom Cruise clones virtually indistinguishable from the original, of course. Especially in the development and discovery of new medicines, for the fight against cancer or the coronavirus (è aAlso thanks to artificial intelligences that the vaccines were developed in 6-8 months instead of in 5-10 years): of total investments, this sector alone has collected almost 14 billion, a jump of four and a half times compared to 2019.
Because with AI, like told to VentureBeat by Jack Clark, co-director of the team that compiled the Stanford report, the point is that “we went from ‘hey, this doesn’t work’ to ‘now it works, how can we exploit it?’ in a much shorter time than we expected ”. In short: you can earn it, and now everyone wants to earn it.
Machine learning, language processing and other applications
What Clark said is (obviously) right: the steps forward are not only large, but also faster than expected. Much faster. An excellent example is that of Gpt-3, a natural language processing software developed by OpenAI and funded a year ago by Microsoft with 1 billion dollars: it is based on 175 billion parameters and is able not only to recognize the language (which Siri or Google Assistant on our phones can also do), but also to create it. He can write, and he can do it quite well, even posing as a lawyer (tweet below).
twitter: Gpt-3 and the first law of robotics
How GPT3 works. A visual thread.
A trained language model generates text.
We can optionally pass it some text as input, which influences its output.
The output is generated from what the model “learned” during its training period where it scanned vast amounts of text.
1/n pic.twitter.com/imM66oyTIC
– Jay Alammar (@JayAlammar) July 21, 2020
twitter: Gpt-3 tries to write like a lawyer
GPT-3 performance on “write this like an attorney” is insane. It even includes relevant statutes if you mention a jurisdiction. This will put a lot of lawyers out of work.
Only the first 2 are prompts… pic.twitter.com/u7eOyuSd9b
— Francis Jervis (@f_j_j_) July 15, 2020
Not only that: using the same concepts learned to create or complete texts, the OpenAI it would also be able to “repair” damaged or partial images in some way sensing the missing part and reconstructing it (an example can be seen here). And if this last detail suggests science-fiction police investigations, with the agents and their artificial “colleague” sifting through the grainy footage of some surveillance camera, the applications that those of Microsoft have in mind are much more practical: Gpt-3 capabilities could be used to improve search capabilities within the Azure cloud computing service and also within Teams, as well as to summarize long text documents or to converse with gamers during intense gaming sessions on the Xbox .
Some AIs can (still) be misled (Granny Smith is an apple variety)
Despite the enormous progress and the enormous investments made to allow them, however, these software are still not perfect: at the beginning of March, the same OpenAI researchers discovered that another of their image recognition program can be easily fooled opponent techniques similar to those we talked about here at the end of 2019. In practice, if you show him an object, he is able to more or less always recognize that object; but if you show him the same object with something written in front (an apple with the word “pizza”, image above), in most cases he is wrong. Because it is still a machine. At least for now.