Jeffrey D. Sachs at the European Parliament – Pt.1
Brussels, 14 October 2026 (Transcript) Edited transcript of the video
At MIT, researchers have discovered that an artificial intelligence model can read, modify and even directly construct the neural networks of other machines. The result is still experimental, but it raises an unexpected question: might artificial intelligences learn to design their own successors?
Imagine finding twenty thousand numbers laid out in a long sequence on a table. They are not words, instructions or mathematical formulas in the usual sense. Yet those numbers form the control system of a small artificial intelligence. Insert them into the appropriate computational structure and the machine begins playing a video game. It knows where to move, which obstacles to avoid and how to pursue a goal. Now imagine that those twenty thousand numbers were not the product of lengthy training, but had been written directly by another artificial intelligence, to which we had simply explained which game we wanted it to play. This is one of the achievements reported by four researchers at the Massachusetts Institute of Technology (MIT) in a study published on 2 November 2026. The authors, Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, call the procedure Weight-Space Programming. The name is forbidding. Once explained, the principle is understandable. And its implications could be considerable.
To understand the experiment, we need to look, at least briefly, at how artificial neural networks work. These networks are computational systems made up of numerous interconnected processing units. Their connections are governed by numerical values, called weights, which determine how information is transformed as it passes through the network. We should not imagine that each number corresponds to an idea, a word or a memory. A network’s capabilities emerge from the relationships among a great many values. It is precisely this distribution of functions that makes it difficult to understand what happens inside it. Normally, weights are established through training. The network is given a problem, its performance is measured, and the numerical values are progressively adjusted to improve the result. The process may require millions or billions of attempts. An imperfect but useful analogy is a musical machine with thousands of controls. The traditional method involves adjusting those controls one after another, constantly listening to the result until the music sounds satisfactory. The MIT researchers asked an advanced OpenAI model called Astra to do something different: examine those settings, understand what behaviour they produce and, in some cases, write new ones directly.
In an initial experiment, Astra was given the 17,253 weights of small neural networks trained to play four different video games. All the networks had the same architecture. Astra knew the format of the input information and the possible actions, but not which game was associated with each network. By examining the numbers, it correctly identified all four games. The result is striking, but it needs qualification: Astra was not starting from nothing. It knew the networks’ architecture and the meaning of their inputs and outputs. It was not faced with twenty thousand numbers entirely devoid of context. A second experiment involved small language models trained on different materials: English texts, Python code and Japanese transcribed in Latin characters. Astra was able to identify the type of training material by analysing certain internal structures of the networks. In this case it could also use mathematical tools, although it did not directly run the models to see what texts they would produce. The distinction matters. Not all the experiments demonstrate a mysterious ability to read numbers at a glance. Some show something different, and perhaps equally useful: an artificial intelligence capable of independently conducting sophisticated analyses of how other artificial intelligences work.
The most spectacular experiment is the one in which Astra does not merely read an existing network but builds one. The researchers presented it with the rules and code of three relatively simple video games, together with the architecture of the neural network intended to play them. Astra directly generated the necessary weights, without training the network through successive games and without being able to test its behaviour during generation. Once inserted into the network, those numbers produced systems capable of playing. In one of the games, Breakout, the network built by Astra achieved results roughly twenty times better than those of a network trained using a particular conventional procedure after 500 million steps. This comparison does not establish that Astra is universally superior to conventional training methods. Their performance also depends on the settings chosen by researchers, and a suitably optimised conventional system may do better. The significant point lies elsewhere: Astra managed to produce a working configuration directly, translating the rules of a game into thousands of numerical parameters. It did not simply write a program for a player. It wrote the internal values of a neural network capable of behaving like one.
Another experiment addressed a problem even closer to practical applications. A small language model from the Qwen family had a difficulty: during certain mathematical tasks, it could become trapped in repetitive reasoning without reaching an answer. Astra examined the model, observed some of its behaviours and directly modified a tiny fraction of its parameters: approximately 0.016 per cent. In the tests reported by the researchers, the proportion of correct answers on a mathematics benchmark rose from 7.9 to 48.1 per cent. This is an impressive improvement, but it does not mean that Astra inserted new mathematical knowledge into the model. An essential part of the intervention involved correcting a defect that prevented the system from effectively using capabilities it already possessed. It is like repairing a mechanism that keeps jamming rather than replacing it with a more powerful one. This possibility could have important applications. Today, correcting a neural network may require new data, further rounds of training and considerable computing resources. If it became possible to identify and directly modify the parameters responsible for particular behaviours, some corrections could become faster and less expensive. Of course, modifying only a few parameters does not guarantee that the effects will be confined to the problem being addressed. A correction could have unexpected consequences for other capabilities of the network. That, too, would need to be tested.
The experiments raise a fascinating cognitive question. Human beings can design machines, formulate theories and interpret complex systems. But ordinarily no one can look at a long sequence of neural weights and directly derive an effective configuration from it. Astra appears able to do so, at least under certain experimental conditions. Does this mean it possesses a cognitive faculty alien to human beings? Not necessarily. The model may be using familiar abilities to recognise patterns, reason mathematically and construct representations, but combining them in ways and on scales that are impracticable for us. That would still be an important result. For a long time, we have assessed artificial intelligences by asking how well they perform human tasks: writing, translating, programming, solving problems and interpreting images. These experiments suggest that capabilities may emerge which are difficult to measure by comparison with human performance, simply because they do not belong to our ordinary repertoire of activities. A bat uses echolocation to navigate. We can understand the principle, but we do not naturally possess that sensory ability. The analogy does not prove that Astra has developed the cognitive equivalent of echolocation. It does, however, invite a question: which capabilities of artificial intelligence might we fail to recognise because we keep looking primarily for those that resemble our own?
We come to the potentially most important implication. If an artificial intelligence can analyse the parameters of another network, identify a defect and correct it, or directly construct a network intended to perform a task, we can imagine systems in which AIs contribute ever more directly to the design of other AIs. The next step would be to use the results to improve the procedure itself. One machine modifies another. The second is tested. The results are used to design a subsequent version. The cycle begins again. This process is called recursive self-improvement. It is not what the MIT researchers have already demonstrated. Their experiments concern narrowly defined tasks, relatively small networks and, in the more complex cases, the use of external tools and controlled testing. We do not know whether the same method can be extended to large frontier models, or whether it can produce progressive, reliable and generalisable improvements. There is also a substantial difference between designing a network capable of playing Breakout and improving a general-purpose artificial intelligence with billions of parameters and widely varying capabilities. Yet the prospect is new: conventional training may no longer be the only route by which the capabilities of neural networks are built and modified.
For years, neural networks have been described as black boxes. We can observe what they receive and what they produce, but it is often difficult to reconstruct precisely why they behave as they do. If one artificial intelligence began directly constructing the parameters of another, that difficulty might increase. But the opposite could also happen. A machine capable of recognising how a network works by examining its weights could become a valuable tool for making it more understandable. And a network designed through explicit steps might, in some cases, be easier to interpret than one that emerged from millions of successive adjustments. The two developments are not mutually exclusive. On the one hand, systems capable of modifying other systems with a speed and complexity that are difficult to control. On the other, new tools for exploring mechanisms we currently understand only partially. The issue, then, is not merely how powerful these machines will become, but how capable we will be of verifying, explaining and governing what they produce.
The MIT research is still preliminary. The experiments will need to be independently replicated, compared with alternative methods and tested on different networks and tasks. It would be premature to announce that artificial intelligences have learned to improve themselves autonomously, and even more premature to infer an unlimited growth in their capabilities. But it would be equally reductive to dismiss these results as a mere technical curiosity. Until now, we have mainly imagined artificial intelligence as a tool through which human beings produce texts, images, programs, analyses and knowledge. Here we glimpse another possibility: artificial intelligences intervening directly in the mechanisms from which the capabilities of other artificial intelligences emerge. This is not yet a completed revolution. It is the opening of a field of research. And perhaps it introduces a distinction destined to become important: the difference between knowing how to build something that works and being able to explain fully why it works. Human beings have long been familiar with this gap. We selectively bred plants and animals long before understanding genetics; we used medicinal substances before knowing their molecular mechanisms. What is new is that one machine might now operate on another, in a space of numerical structures that no human being could explore directly with the same speed. At that point, the question will not simply be whether artificial intelligences can build other artificial intelligences. It will also be who will be able to understand, verify and control what they have built.
Main source: Rikiya Takehi, Jayden Teoh, Alex Zhang, Omar Khattab, “Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights”, MIT CSAIL, 2 November 2026. Preliminary results, not yet subjected to independent scientific peer review.
Brussels, 14 October 2026 (Transcript) Edited transcript of the video
Rome, oil, cancer cells: when growth transforms the conditions of
At Palazzo Roverella, a journey into the visionary world of