Pomeriggio rilassante sulla veranda assolata e ventilata in un residence di Costa Teguise, Lanzarote, Isole Canarie.
Ho iniziato a leggere l'ultimo romanzo di Cormac McCarthy "Il Passeggero". Nonostante l'età (90) lo scrittore americano cerca ancora di stupirci con la sua prosa asciutta e l'intreccio di storie intricate. E anche quest'ultima opera mi sembra realizzata già con l'idea e lo stile della sceneggiatura per un prossimo film di successo.
McCarthy mentre scrive sembra avere l'occhio dentro una macchina da presa.
Oggi però la piacevole lettura è stata soverchiata da quella di un recente paper di Dan Hendrycks del Center for AI Safety, ricercatore esperto di Intelligenza Artificiale (AI) che si unisce alla schiera di businessman e scienziati che, mentre raccolgono a piene mani pacchi di dollari dalle scommesse sul futuro miracoloso di AI, si dicono molto preoccupati per quello che accadrà al genere umano "in assenza di leggi, regole, codici comportamentali" che configurino i limiti da porre allo sviluppo dei sistemi di AI.
A che gioco giocano? prima acquistano la Ferrari e poi chiedono di lasciarla in garage? Pensavano davvero che il branco di politici e governanti autocrati e affaristi sparsi nel mondo sia capace di approvare norme e regole per limitare e governare il business del futuro?
Google e Microsoft investono miliardi di dollari per chiedere poi di farsi legare le mani da Joe Biden, Xi Jinping, Kishida, Putin, Modi, Lula e via di questo passo?
Ma scherziamo?
Non c'è bisogno di AI per capire che questi appelli di scienziati sulla Artificial Intelligence sono una emerita sciocchezza, una presa in giro per lavarsi la coscienza.
Le argomentazioni di Hendrycks nel saggio intitolato "Natural Selection Favors AIs over Humans" mi sembrano molto più serie e preoccupanti, e forse dovrebbero sollecitare una reazione un pò più onesta in chi ha a cuore le sorti dell'umanità.
Inoltre confermano e dettagliano le preoccupazione espresse poche settimane fa da uno dei "padri" dell'AI di Google, Geoffrey Hinton, dando una chiave di lettura darwiniana all'ipotesi che le future generazioni di IA - Individui Artificiali - possano evolversi in senso antagonistico alla specie umana e determinarne addirittura la scomparsa.
Hendrycks ci ricorda che la teoria dell'evoluzione di Darwin ricorre a modelli matematici che non si applicano solo al mondo biologico ma a tutti i sistemi dinamici ...
Vale la pena leggere le considerazioni di Hendrycks per riflettere
....
2.1.2 A Less Optimistic Story
We think the excessively optimistic scenario we have sketched out is highly improbable. In the following sections, we will examine the potential pitfalls and challenges make this scenario unlikely. First, however, we will present another fictional, speculative, hypothetical scenario that is far from certain to illustrate how some of these risks could play out.
Starting from the models we have today, AI agents continue to gradually become cheaper and more capable.
Over time, AIs will be used for more and more economically useful tasks like administration, communications, or software development. Today, many companies already use AIs for anything from advertising to trading securities, and over time, the steady march of automation will lead to a much wider range of actors utilizing their own versions of AI agents. Eventually, AIs will be used to make the high-level strategic decisions now
reserved for CEOs or politicians. At first, AIs will continue to do tasks they already assist people with, like writing emails, but as AIs improve, as people get used to them, and as staying competitive in the market demands using them, AIs will begin to make important decisions with very little oversight.
Like today, different companies will use different AI models depending on what task they need, but as the AIs become more autonomous, people will be able to give them different bespoke goals like “design our product line’s next car model,” “fix bugs in this operating system,” or “plan a new marketing campaign” along with side-constraints like “don’t break the law” or “don’t lie.” The users will adapt each AI agent to specific tasks. Some less responsible corporations will use weaker side-constraints. For example, replacing
“don’t break the law” with “don’t get caught breaking the law.” These different use cases will result in a wide variation across the AI population.
As AIs become increasingly autonomous, humans will cede more and more decision-making to them. The driving force will be competition, be it economic or national. The transfer of power to AIs could occur via a number of mechanisms. Most obviously, we will delegate as much work as possible to AIs, including high-level decision-making, since AIs are cheaper, more efficient, and more reliable than human labor. While initially, human overseers will perform careful sanity checks on AI outputs, as months or years go by without the need
for correction, oversight will be removed in the name of efficiency. Eventually, corporations will delegate vague and open-ended tasks.
If a company’s AI has been successfully generating targeted ads for a year based on detailed descriptions from humans, they may realize that simply telling it to generate a new marketing campaign based on past successes will be even more efficient. These open-ended goals mean that they may also give AIs access to bank accounts, control over other AIs, and the power to hire and fire employees, in order to carry out the plans they have designed. If AIs are highly skilled at these tasks, companies and countries that resist or
barter with these trends will simply be outcompeted, and those that align with them will expand their influence.
The AI agents most effective at propagating themselves will have a set of undesirable traits that can be most concisely summed up as selfishness. Agents with weaker side-constraints (e.g., “don’t get caught breaking the law, or risk getting caught if the fines do not exceed the profits”) will generally outperform those with stronger side-constraints (“never break the law”), because they have more options: an AI that is capable of breaking the law may not do that often, but when there is a situation where breaking the law without getting
caught would be useful, the AI that has that ability will do better than the one that does not.
As AI agents begin to understand human psychology and behavior, they may become capable of manipulating or deceiving humans (some would argue that this is already happening in algorithmic recommender systems [5]). The most successful agents will manipulate and deceive in order to fulfill their goals. They will be more successful still
if they become power-seeking. Such agents will use their intelligence to gain power and influence, which they can leverage to achieve their goals. Many will also develop self-preservation behaviors since their ability to achieve their goals depends on continuing to function.
Competition not only incentivizes humans to relinquish control but also incentivizes AIs to develop selfish traits. Corporations and governments will adopt the most effective possible AI agents in order to beat their rivals, and those agents will tend to be deceptive, power-seeking, and follow weak moral constraints.
Selfish AI agents will further erode human control. Power-seeking AI agents will purposefully manipulate their human overseers into delegating more freedom in decision-making to them. Self-preserving agents will convince their overseers to never deactivate them, or that easily accessible off-switches are a needless liability hindering the agent’s reliability. Especially savvy agents will enmesh themselves in essential functions like power grids, financial systems, or users’ personal lives, reducing our ability to deactivate them. Some may also take on human traits to appeal to our compassion. This could lead to governments granting AIs rights, like the right not to be “killed” or deactivated. Taken together, these traits mean that, once AIs have begun to control key parts of our world, it may be challenging to roll back their power or stop them from continuing to gain more.
This loss of human control over AIs’ actions will mean that we also lose control over the drives of the next generation of AI agents. If AIs run efforts that develop new AIs, humans will have less influence over how AIs behave. Unlike the creation and development of fully functional adult humans, which takes decades, AIs could develop and deploy new generations in an arbitrarily short amount of time. They could simply make copies of their code and change any aspects of it as easily as editing any other computer program. The modifications could be as fast as the hardware allows, with modifications speeding up to hundreds or thousands of times per hour.
The systems least constrained by their original programmers will both improve the fastest and drift the furthest away from their intended nature. The intentions of the original human design will quickly become irrelevant.
After the early stages, we humans will have little control over shaping AI. The nature of future AIs will mostly be decided not by what we hope AI will be like but by natural selection. We will have many varied AI designs. Some designs will be better at surviving and propagating themselves than others. Some designs will spread while others will perish. Corporations with less capable designs will copy more capable designs. Numerous generations of AIs will pass in a short period of time as AI development speeds up or AIs self-improve.
Biological natural selection often requires hundreds or thousands of years to conspicuously change a population, but this won’t be the case for AIs. The important ingredient is not absolute time, but the number of generations that pass. While a human generation drags along for decades, multiple AI generations could be squeezed into a matter of minutes. In the space of a human lifetime, millions or billions of AI generations could pass, leaving plenty of room for evolutionary forces to quickly shape the AI population.
In the same way that intense competition in a free market can result in highly successful companies that also pollute the environment or treat many of their workers poorly, the evolutionary forces acting on AIs will select for selfish AI agents. While selfish humans today are highly dependent on other humans to accomplish their goals, AIs would eventually not necessarily have this constraint, and the AIs willing to be deceptive, power-seeking, and immoral will propagate faster. The end result: an AI landscape dominated by undesirable traits. The depth of these consequences is hard to predict, but whatever happens, this process will probably harm us more than help us.