An autonomous research system called ASI4AI has succeeded in developing new architectures for artificial intelligence without human intervention – is this the harbinger of a global revolution or an inflated illusion?
There are sentences that mark the end of an era and the beginning of a new one. The following sentence, from a study published in recent days, is one of them –
"We have demonstrated that large-scale research advances can be achieved using computing resources, rather than human skill."
And so, with these words, the era of human research begins to come to an end – and the era of artificial intelligence research begins.
So what exactly happened in this study?
the background
Contrary to popular myth, humans are bad at technological inventions. Yes, we invented the spear, fire, the wheel, tea and beer, and a host of other technological advancements. And yet, it took us tens of thousands of years to do so. The only reason we credit ourselves so much for these advancements is that we are better at them than any other creature on earth.
But we're still doing it very, very slowly.
Why? There are several reasons, but one of the most important is that the process of invention requires human labor. Human experts need to come up with new ideas based on the most advanced research in their field. They then need to design prototypes of new technologies, deploy them, test them, figure out why they don't work (as is usually the case), and go back to the drawing board to start the whole process over again.
If this whole process sounds tedious, agonizing, and especially long to you – you are absolutely right. But as Churchill said – “Democracy is the worst system of government that exists, except for all the other systems of government that have been tried.”
The same principle applies to current methods of scientific and technological development. They are bad, they are slow, and they are clumsy – but we have not found an alternative to them. And despite the fact that they crawl forward at the pace of a lame-limbed turtle, we have still managed to double the human lifespan, reach the moon, and produce baby videos that speak in the voice of Udi Kagan.
So yes, we have made progress and we continue to make progress. But surely the pace of progress can be accelerated! What if, for example, we replaced human intelligence in the process with… artificial intelligence?
That's exactly what the researchers did in the new study. And they found, not surprisingly, that AI can perform the entire research process faster than humans—and bring new, more efficient technologies to the table than existing ones. And if all that wasn't enough, they used AI to develop even more advanced AIs.
About architectures
The most important artificial intelligence engines today run on an infrastructure of "artificial neural networks." These are computer structures that simulate the operating principle of real neural networks like the brain. And just as the brain has different regions, each of which specializes in performing different operations, so too can artificial neural networks be 'pruned' and shaped. It is possible to build different 'architectures' of such networks, which will be more successful at completing words, for example, or at recognizing images, or at any other task.
The problem is that developing any new architecture like this requires a lot of human effort. Human researchers have to come up with ideas for developing new architectures, based on the most advanced research. After receiving approval from their managers (or their academic supervisors), the researchers also have to deploy these architectures, train them on data, and test their performance. If the idea works – congratulations, they have a breakthrough on their hands. And if it doesn’t work? Then it’s back to the drawing board. And again, more work, more effort, more approval requests, more experiments, and so on.
Sound familiar? This is exactly the process of developing new technologies, but with an emphasis on developing more advanced artificial intelligence.
The new study automates this entire process, using three modules – that is, three different parts that work together. The first is the “researcher,” who comes up with creative ideas for new architectures, drawing on existing research from the scientific literature. Next comes the “engineer,” who builds initial, limited prototypes of the proposed architectures to test whether the “researcher”’s ideas are successful. And finally comes the “analyst,” who analyzes the results of each run and derives new insights into what worked better and less well.
But that's just the first step.
In the second stage, the most successful architectures are selected from those tested, and the "researcher" is called into action again. He now needs to improve these architectures even further. To do this, he uses the insights gathered by the "analyst" and compares them with existing research knowledge. He suggests new ideas, which the "engineer" will test and the "analyst" will analyze. New insights again, and voila – the entire system has become smarter, because these insights are stored in its main memory. In the next run, it will be able to develop even more successful architectures, and so on and so forth, ad infinitum.
Or at least, until the computing power at its disposal runs out. Because all these runs require quite a bit of it.
But what is the result?
The results
The researchers called their research system ASI4AI. This stands for Artificial Superintelligence for AI Research. This name includes a very pretentious claim that the system is at the level of "superintelligence." That is, it reaches a superhuman level in research.
But it is not entirely clear that the results support this assertion.
The entire system consumed about 20,000 computing hours to perform 1,773 experiments developing new AI architectures, all by itself. Of these, the system identified 106 new AI architectures. When these were tested against existing architectures invented by humans, some of them achieved better results. The improvements were not large – perhaps one or two percent in total. But they were certainly there. In some parameters, by the way, there was actually a decrease in capabilities.
What does that mean? At least right now, it's not clear whether it's a "superintelligence." If so, it's not doing its job very well. But the beauty is that it doesn't have to be a "superintelligence." It doesn't have to significantly outperform human researchers. It just needs to do work that's comparable in quality to what human researchers do to change the world.
Because human researchers cost a lot, and artificial intelligence is very cheap.
Research enterprises
It's common to complain about the cost of running artificial intelligence, the high energy it consumes, and the data center infrastructure needed to run it. All true, but how much does it cost to produce a "human" to conduct research?
A quick calculation shows that in the United States, raising a child through high school—including food, health care, clothing, and all other necessities—costs about $300,000. Add to that the costs of a bachelor’s degree and a doctorate, and you easily reach $XNUMX million. The government subsidizes about half of that amount. From that point on, the young doctor can begin conducting research and returning value to the country.
But can that doctor compete with artificial intelligence?
According to the authors of the study, they have uncovered a new law of “science at scale.” They claim that developing a new AI architecture requires 2,000 hours of work by human researchers. The AI they have put together, however, is capable of discovering almost twenty times as many new architectures in the same amount of time. And if you add hours of work—that is, hours of computing—then the system’s performance improves dramatically. With an investment of 7,000 computing hours, their system developed 106 innovative architectures.
If the researchers' claims are correct, and if they can be applied to other fields of science, then it would not be more profitable for governments to provide human researchers. What would be more profitable for them? Supporting computing infrastructures and systems that will conduct research better than human researchers.
Without despair (and researchers) at all
Does this mean that we won't need human researchers at all? Of course not. Anyone who thinks this way falls into the "either-or" fallacy that Daniel Burroughs defined in his book "Flash Foresight." We like to think about the future in dramatic terms: either there will be e-books, or paper books. Or there will be newspapers, or there will be news websites. Or motorized ships, or rowboats. Or carriages, or automobiles.
Or – or. Either one, or the other.
Reality shows that the future is moving more in the direction of “both-and.” People read both paper books, and e-books, and audiobooks, and comics on computers. Newspapers continue to exist both in their printed versions and in the digital world. And ships powered by steam, coal, oil, and electricity appear alongside rowboats on rivers and Venetian gondolas. And in London, you can still find a few carriages in well-defined areas.
There is room for everyone. Both this and that.
This is the argument I like to use when thinking about the future, and there is a lot of truth in it. But it is also impossible to ignore the nuances in it: Yes, there is room for everyone, but how much exactly? True, there is room for both cars and horse-drawn carriages, but it is clear to everyone that cars have already taken over the world, while the era of horses is a thing of the past. And rowboats are reserved for tourists who want a one-time exotic experience. Those who want efficiency – get on a motor yacht.
So there will be room for everyone, even in the world of research, but it is very likely that the "mix" of researchers will change. There will still be human researchers, of course, but they will be divided into two types.
The first human researchers will be more "managers" than actual researchers. They will manage the AIs that will do the actual research. They will also manage the labs, the infrastructure, and the organizations that will run the AIs. We will see them at first mostly in computer science, because experiments in this field require very little lab work.
The second type of human researcher is more commonly seen in other fields, such as biology, chemistry, physics, and more. These will be the "research workers." Artificial intelligence will suggest things to research, guide them in each experiment, and provide them with emotional support when they are disappointed with the results and when the grant money runs out. But they will be the ones who will work in the lab itself. They will be the fingers, hands, and feet of artificial intelligence. They will hold the pipettes and clean the biological hoods at the end of the day. And we will need many of them, at least in the coming years, because as artificial intelligence makes research easier and more convenient, there will be more labs in academia and industry.
And then the labs will also become robotic. As I've reviewed here before, hundreds of millions of dollars are currently being invested in automating research labs. And even if they can't be fully automated, there will still only be a need for a lab manager, who will also perform the few tasks that will still require human hands.
It is fair to argue that all of this will happen in the future, but our present is currently shaped by the expectations of governments and capitalists about what is to come. The analysis I have described – of autonomous research producing products at a speed that exceeds that of any human researcher – is currently leading the thinking of forward-looking governments, such as China and the United States. These expectations are also responsible for the new Cold War between them, for computing power and the development capabilities of more sophisticated artificial intelligence. The powers understand very well that whoever can use more powerful artificial intelligence will gain a huge advantage over others in every field that technology can integrate into and advance. There is no field that technology cannot advance.
The pleasant revelation
And now that I’ve (hopefully) excited you about the potential for the future, let’s address the elephant in the room: the current study isn’t particularly impressive. It was published on arXiv, a repository of research that hasn’t yet undergone proper peer review. The experts who have reviewed it agree that the lead researcher behind the study is a renowned figure in the field of artificial intelligence. On the other hand, it’s also clear that the authors use particularly pompous language, and that the paper is still in its rough draft stage.
And yet, this is yet another study that joins many others and reveals a simple fact: Artificial intelligence can dramatically accelerate technological research and development processes. Countries that are not ready to adopt artificial intelligence will be left behind. Companies that do not know how to integrate artificial intelligence into their work processes will slowly die while their competitors run ahead. And people who do not know how to activate artificial intelligence and manage it for the benefit of their work will have to understand what sets them apart – what they can bring added value to – compared to their colleagues who do use it.
It will be interesting. Good luck to us!
More of the topic in Hayadan:
