Nick Bostrom, one of the first philosophers of the era of artificial intelligence predicts that since artificial intelligence will be smarter than us, it will also be a better scientist, a more accurate engineer, and a more creative artist. The Japanese recently tried this with the advanced large language models
In 2015, Nick Bostrom took the TED stage and coined a saying that has since become associated with his name.
"Machine Intelligence," Bostrom announced, "is the last invention humanity will ever have to create."
Bostrom is known as a particularly provocative philosopher. Among other things, he also invented the simulation theory, according to which we all live, well, in a simulation. He founded and directed the "Future of Humanity Institute" at the University of Oxford, and researched and is still researching the way we should think about advanced artificial intelligence.
One, for example, that will be the last invention that humanity will ever have to create.
But why, actually?
Bostrom claims that once artificial intelligence reaches a sufficiently advanced level, it will be able to perform certain tasks better than humans. And what task is more important than the one that differentiates humans from animals: our ability to develop new inventions?
And so Bostrom predicts that advanced artificial intelligence will come up with our next inventions for us. But she won't do it like humans, excruciatingly slowly and by trial and error. Because she will be smarter than us, she will also be a better scientist, and a more precise engineer, and a more creative artist. She will be able to test her ideas in computer simulations, filter out the failed lines of thought and focus on the successful ones. In the time it takes a person to think of one new idea, she will generate thousands of ideas. While it is necessary for a company to test whether a certain innovative product is really successful, the artificial intelligence will be able to test hundreds of hypothetical products, and will choose to move forward only with the best among them.
And of course, it will also be able to develop the next generation of artificial intelligence, which will develop the generations after them, and so on.
Compared to all this innovation, any human inventor would soon fall by the wayside. We will not be able to compete with the ingenuity of artificial intelligence. We can - maybe - control it and direct it. But we can't compete with her.
And so, Eliba Davostrom, advanced artificial intelligence "is the last invention that humanity will ever have to create."
In the last week, it became closer to reality, thanks to a Japanese company called Sakana, which developed "scientific artificial intelligence".
Science at the push of a button
How do scientists do science? slowly and painfully. Good research begins with understanding the knowledge that exists today, identifying gaps and gaps in knowledge - and formulating theories that explain these gaps. These initial theories are called "hypotheses", and in order to understand which one is correct, you need to conduct experiments. To this end, the scientist needs to plan clever experiments that will disprove the wrong hypotheses, and provide evidence in favor of the correct hypothesis. After each experiment, the results should be collected, processed and analyzed and conclusions should be reached. Then you can move on to the next experiment, until you reach a high level of certainty about the most correct hypothesis.
Sakana's scientific AI does all this, and more.
The computer scientist receives the research subject as input from the user, and access to the information she needs to research the field. It examines the existing knowledge in science and formulates a list of research questions that should be raised. She continues to formulate ideas for experiments, and if it is about computer science - in which the experiments are conducted using code - she is also able to run the experiments herself, collect and analyze the results and reach insights.
Then she does the "more".
It is unpleasant to admit, but a significant portion of the time of every scientist today is taken away in favor of writing academic articles. In these articles, the researcher describes the research he did, the methods he used and the hypotheses he raised and refuted or confirmed. Writing the article can easily require tens of hours of the scientist's time.
The artificial scientist does that too. After performing the experiment, she summarizes the research she performed and the results she received and their meanings in the form of an academic article that can be published or presented at a respected academic conference.
But the scientist's work does not end here either. After writing the article, the scientist must submit the article to "peer review", in which his colleagues are asked to provide a review of the article anonymously. They locate every flaw and every crack in the experiments, and thus the scientist is forced to return many times to the laboratory and perform additional experiments that will strengthen the message in the article.
And yes, artificial intelligence does that too. about. After writing the article, she re-examines it by assuming the persona of a critical scientist. She locates the failures and problems in the article, and decides what grade to give it, and whether corrections are necessary.
And at the end? She can accept the article she herself created - and use it as a basis to improve the original research. Just as a human scientist does.
But there is one big difference. The human scientist has to spend hundreds of hours acquiring knowledge in the field, thinking about new ideas, developing the experiments, processing and analyzing the results, and writing and improving the article. The artificial intelligence can produce any such article in about fifteen minutes, and at a cost of only 15 dollars.
The developers at Sakana concentrated on places where it is easy for the artificial intelligence to achieve initial successes. They didn't send her to do experiments in the lab, but focused on the field of machine learning, so she could do experiments on the computer. She did these experiments impressively, producing fifty papers in 12 hours, and going over them to make sure they were of high enough quality.
At the end of the process, the artificial intelligence produced ten articles of a high level - or at least those that it claims are of a high level. The researchers went through these articles, and found that most of them are indeed of a high level and provide new insights into the research fields. It is fascinating to discover that scientific artificial intelligence has studied the way in which artificial nervous systems work and how they can be improved. That is, she studied the mechanisms that underlie her own action. And the meaning is, of course, that the more artificial intelligence improves in its research capacity - the faster it will be able to improve itself.
Oh, and she lies too.
Not everything is perfect yet
Before I focus on the great significance of artificial intelligence that does science, it is also important to talk about its limitations today. The developers at Sakana have made use of major language models - Claude 3.5, GPT4-O and others - and these are able to provide beautiful results... usually. As many users have discovered firsthand, these AIs can also lie, or provide "hallucinations" or hallucinations.
You can deal with these hallucinations and greatly reduce the chance that they will appear, but they still appear from time to time. This was also the case in some of the articles that the artificial intelligence 'published'. In one of them, for example, she listed the data of the graphics processor she used, but since she did not know what hardware it was - she simply threw the name of a common graphics processor (V100 GPU) into the article. The truth is that a different graphics engine (H100 GPU) was behind the experiment, but readers exposed to the article would not have known that.
It is amusing to note, by the way, that I found no mention of this particular failure in any press coverage of artificial intelligence. To find it, I had to read the… the original scientific paper written by the human researchers about the scientific artificial intelligence, of course.
So we managed to create a scientific AI that lies about its experiments. Already at this stage it will become clear that it is not possible to simply leave the artificial intelligence plugged in - and that it will solve all of humanity's problems with the help of science. This fact does not detract from the importance of the development, but it does make it clear that we must treat every article it produces with suspicion and criticism.
Or perhaps it is more correct to say that when we ask scientific artificial intelligences to think about the experiments of the future, we will have to make it clear to them that they need to examine the articles of their predecessors with hyper-criticality.
Blatant lies are not the only problem In articles written by Sakana's AI. She went ahead and put a positive 'spin' around every result. When she was required to examine the performance of a particular model in a study, for example, she described its results as a "12.8% reduction" (a positive result), but negative results she described as a "3.3% improvement". She also includes in each article the description of all the experiments she conducted and all the intermediate results she received. In other words, the article she writes is less suitable for humans - and more for reading by critical machines.
But that's okay.
I mean, it's wrong. really wrong Scientists who are caught lying in the articles they write are removed from the scientific community. point. They are fired from their senior research positions at universities, and no one is willing to cooperate with them anymore. Their academic careers are cut short with one blow of an ax, and there is no forgiveness.
But artificial intelligence is still in its infancy as a scientist. And Sakana's developers - along with the rest of us - are still learning for themselves how to give her the most precise instructions so that she conducts the most reliable studies. And of course, in a month, two or three, the next model of artificial intelligence will appear - be it GPT5 or some other ground-breaking competitor - and it will already be more accurate, more logical and more reliable.
The important principle has already been proven: yes, artificial intelligence can do research. And even successful research. And even if she still makes mistakes, and even rude ones, she can still be a help against the wise human researcher.
And that will probably be her role in the coming years.
The scientist's best friend
The most likely use of scientific artificial intelligence in the coming years will be to position it as "the scientist's companion". The human scientist will choose to activate the artificial intelligence, choose the areas it will investigate, and define for it the limitations and the boundaries of the analysis. She will produce studies, studies and more studies for him, with an article for each of them, and even rank the most successful articles.
Then his job will be to read them critically.
The human researcher will have to scrutinize every detail in the articles. He will have to examine every point raised by the artificial intelligence and cross-check it with the experiments conducted to make sure that there is no deviation from the truth. He will spend long hours on this - but will see a reward for his labor, since he will be able to publish any such successful article. Thus, the amount of effort required to produce each scientific article - including formulating the hypotheses, performing the experiments and analyzing the results - will be dramatically reduced. It is not an exaggeration to claim that it will decrease by an entire order of magnitude - from hundreds of hours to a few tens.
Given this powerful tool, it is not an exaggeration to claim that the number of articles in computer science, in particular, may skyrocket in the coming years. But this is an almost uninteresting claim. Who cares about articles? The products of the articles - the scientific-technological insights - are the ones that are really important.
This is the first true meaning of scientific artificial intelligence: we are going to see a significant leap in the rate at which we reach new insights, new scientific theories and important technological developments. This surge is going to affect us all, in ways we can't even imagine today. Among other things, it will lead to stronger artificial intelligence being developed faster. And as soon as artificial intelligences can also do experiments in laboratories - and today is not far off - we will be able to bring the same improvement in the pace of research to all fields of science.
Want to stay young forever? Improve the energy harvesting capacity of solar panels to the maximum possible? Figure out how to cause nuclear fusion at room temperature? The scientific artificial intelligence will help us find the way to make all this possible.
And at the same time, it will also democratize science.
Our best friend, period
You must have noticed that the world has recently been filled with color. Everyone is starting to use painterly artificial intelligence - Midjourney, Dali 3, Flux, or any other model - to convey their ideas in a more colorful way to the public. The people who do this do not know how to paint with a brush and canvas, or with an electronic pen and touch screen. But they have learned to tell the artificial intelligence what they want to receive - and it is able to skip most of the conventional creation steps to provide them with what they want.
This process is known as DeSkilling: a decrease in the level of skill required to produce a certain product or service, compared to the level of skill that was required in the past.
The scientist's work will also undergo a similar process of de-scaling in the coming years. It won't happen right away, as the scientific AI is still plagued with problems. But as her abilities improve, and the better we know how to use her, the more she will be able to help anyone carry out research that will help them realize their wishes and dreams.
Have you always wanted blue eyes, but didn't know what to do about it? In a few years you will be able to run your personal scientist on a wealth of existing biological research - and she will come up with ideas, test them in preliminary experiments, and finally also help you synthesize the most suitable material. I still recommend that you test her experiments carefully, before you drop the stuff into your eyes.
Always wanted a love potion? Or maybe a psychoactive substance that especially affects the circuits of desire in the brain and organs that are closer to the earth? The scientific artificial intelligence will help you develop these as well.
Always wanted a chemical warfare agent? Well, you already know what you need to do.
The democratization of science will not arrive tomorrow morning, but on a historical scale, it is definitely around the corner. By the end of the decade, highly intelligent individuals will gain the research power previously reserved only for the most advanced laboratories.
Yes, it is dangerous. And yes, it is promising and exciting and exciting, because every person will be able to do the research that interests them, reach the results and share them with everyone else. We will all be scientists and inventors, each with his own focus and needs. It will be a world of creation and innovation that translates almost immediately into new products. And it may also be a world of magic, because there will be those who will not share the rest of humanity in the research process, but only in its results. As Arthur S. wrote. Clark - "Any technology is advanced enough - it is impossible to distinguish it from magic".
Are we ready for such a world, where the power of the scientist - of Einstein, of Tesla, of Pasteur - is in the hands of every person? of each child?
Ready or not - he is almost here.
More on the subject on the science website
