A new study by researchers from Tel Aviv University and Stanford University shows that AI can develop general predictive skills using only self-generated training data, challenging a fundamental assumption about how artificial intelligence learns.
NEWS
What are you interested in?
Research
Nov 5th, 2026
AI Learns to Predict the World Without Ever Seeing It
- Environment
Almost every artificial intelligence system we know begins in the same way: it is exposed to vast amounts of information from the real world - books, websites, images, and recordings and trained to identify patterns within it. A new study involving researchers from Tel Aviv University and Stanford University raises another possibility: Can a system develop general learning capabilities even when all the data it trains on is generated entirely from scratch within the system itself? The researchers have shown that it can, marking a significant step toward a goal, long considered the holy grail of AI training.
The study was conducted under the supervision of Dr. Yoav Levine of the Blavatnik School of Computer Science at Tel Aviv University, as part of the postdoctoral research of Dr. Kfir Dolev, in collaboration with researchers from Stanford University. The study is available as a preprint on arXiv.
Two AI Models Learning Together
In the new study, the researchers presented a training method in which two models start from scratch and learn together. The first generates short computer programs that produce sequences of symbols. The second is trained to predict the next symbol in each sequence. The model generating the exercises receives feedback on how much they contribute to learning and gradually becomes better at selecting challenges suited to the learning model’s developing capabilities.
The decisive test came after training: Would what the model had learned from these artificial sequences also help it predict real-world data it had never encountered during training? The researchers evaluated its performance on text, images, sounds, music, code, and DNA sequences. They found that as more computational resources were devoted to the training process, the model’s predictive capability improved consistently across many domains. The model was also able to infer rules from examples presented to it during testing, without any additional training.
What Happens When AI Creates Its Own Lessons?
One of the most intriguing findings concerned the learning material the system generated for itself. During training, familiar mathematical patterns appeared in this material, including the Fibonacci sequence. In other words, the system learned to find ways to generate structured sequences from which the second model could benefit. In another experiment, the researchers found that pretraining with the new method helped the model learn faster when it was subsequently exposed to real-world data.
What Could This Mean for the Future of AI?
The study offers an initial answer to fundamental questions in artificial intelligence research: Does intelligence have a “universal” component that does not depend on any external data? Does part of a model’s learning capacity stem from general patterns that can be acquired without prior exposure to real-world information? The findings indicate that this is indeed possible. However, the researchers emphasize that the method cannot teach a system facts about reality without exposing it to that reality. Rather, it may eventually help AI systems learn from real-world data more efficiently.
Dr. Yoav Levine concludes: “For years, the prevailing assumption was that for an AI model to learn anything useful, it first had to be fed enormous amounts of real-world data. In this study, we disproved this fundamental assumption and showed that a model can acquire general predictive skills even when it is trained only on data generated within the learning process itself. This is still a proof of concept involving small models, but the findings raise fascinating questions for future research: What universal mathematical patterns will emerge as we strengthen AI using these methods? How far can we take AI without exposing it to information from the outside world? And can we prepare models to learn more effectively from reality through experiences they have generated for themselves?”
Research
Oct 26th, 2026
A 99-Million-Year-Old Lizard Tail Reveals an Ancient Survival Trick
A tiny tail preserved in amber provides the clearest direct fossil evidence yet of a defense mechanism still used by lizards today.
- Environment
Imagine a lizard escaping danger by leaving its tail behind, still wriggling as the animal darts away. It's a familiar survival trick among lizards today, but how far back does this ability go? A new Tel Aviv University study offers a rare glimpse into its ancient origins through a piece of amber just 18 millimeters long, discovered in northern Myanmar. Inside it, researchers identified a lizard tail that detached from its body approximately 99 million years ago. According to the team, the fossil provides the best-preserved direct evidence yet of caudal autotomy, the ability of a lizard to deliberately shed its tail as a defense mechanism.
The study was conducted by Dolev Fabrikant, a PhD candidate at Tel Aviv University's New Environmental School, and Dr. Leonidas-Romanos Davranoglou of TAU's School of Zoology and Steinhardt Museum of Natural History, in collaboration with researchers from the National and Kapodistrian University of Athens and Israel's Volcani Institute. The findings were published in Palaeontologia Electronica.
An ancient escape preserved in amber
How Can Scientists Tell If a Lizard Shed Its Tail?
The researchers explain that earlier reptile fossils have provided evidence of tail-shedding through natural fracture planes in tail vertebrae and regenerated tails. These features suggest that an animal had the ability to shed its tail, but fracture planes alone cannot always establish whether it actually did so voluntarily.
In the newly discovered fossil, the detached end of the tail contains exceptionally well-preserved features that allowed researchers to examine exactly how the separation occurred.
A Tiny Fossil with a Remarkable Story
For years, the significance of the amber went unnoticed. What initially appeared to be a tiny twig or a fragment of an insect caught Fabrikant's attention when he examined the specimen more closely. “I realized that what seemed to be a twig was covered in tiny scales,” he recalls. “It was a lizard’s tail.” When Fabrikant shared his discovery with his supervisor, the reaction was immediate.
“I nearly jumped out of my chair,” says Dr. Davranoglou. “I instantly realized that this tiny fossil could open an extraordinary window into the past, helping us understand the history of one of the most widespread survival strategies among reptiles. Until now, we could only speculate on whether ancient lizards could shed their tails. We now have direct evidence that one could.”
An ancient escape preserved in amber
Looking Inside 99-Million-Year-Old Amber
Because the amber is cloudy and the fossil is so small, a conventional microscope could reveal little of its finer details. The researchers therefore used micro-CT scanning at the Volcani Institute. “Micro-CT takes X-ray images from many angles and combines them into a three-dimensional reconstruction. It allowed us to examine structures hidden by the amber without needing to cut this one-of-a-kind piece open,” explains Dr. Ahmed Nasser, Head of the institute's Inter-Institutional Analytical Unit.
At the point where the tail detached, the scans revealed a distinctive ring of tissue resembling the structure found in tails shed by living lizards. The separation was also relatively clean and symmetrical, without the damage patterns typically associated with accidental injury.
What We Still Don't Know About This Ancient Lizard
Despite the fossil's remarkable preservation, the tail alone cannot reveal exactly which group of lizards the animal belonged to. “Although we can securely identify this specimen as a lizard, we lack the skull and other skeletal features that would allow a more secure identification. Possible affinities include the groups containing today’s wall lizards or, alternatively, slow worms and monitor lizards,” says Apostolos Christopoulos, a PhD candidate at the National and Kapodistrian University of Athens.
Research team (Left to right): Dr Ahmed Nasser, Dolev Fabrikant & Dr Leonidas-Romanos Davranoglou
The researchers also cannot determine what caused the lizard to shed its tail. It may have become trapped in sticky resin and detached its tail to escape. Alternatively, it may have shed its tail during a predator attack, with the discarded tail subsequently landing in resin. The fossil cannot tell us whether the lizard survived or later regrew its tail. What it does reveal is that this remarkable defense mechanism was already present in lizards approximately 99 million years ago.
Preserved in a tiny piece of ancient resin, small enough to rest on a fingertip, is evidence of a survival strategy that lizards continue to use in gardens and on sunlit walls today.
Research
Oct 25th, 2026
Can We Really Predict the Future?
A Tel Aviv University study analyzing more than one million forecasts reveals when prediction markets get the odds right, and where they tend to go wrong.
- Psychology
Can crowds of people trading on what they believe will happen collectively predict the future? A new Tel Aviv University study of more than one million forecasts suggests that they often come remarkably close. But when it comes to certain events, from international conflicts to elections, the odds may tell a different story.
A research team from Tel Aviv University's School of Psychological Sciences, led by Prof. Michael Gilead and including Justin Dickerson, Hilla Shinitzky, Maayan Peer, and Mattan S. Ben-Shachar, analyzed 1,014,122 forecasts across four major prediction platforms: Polymarket, Kalshi, Manifold, and Metaculus. Covering 12 domains over a decade, the study, published in the scientific journal Proceedings of the National Academy of Sciences (PNAS), examines how well collective predictions perform across topics ranging from elections and geopolitics to technology and security.
Overview: What Did More Than One Million Forecasts Reveal?
- Crowds can estimate probabilities surprisingly well. Across the full dataset, predicted probabilities closely matched real-world outcomes. Events given a 70% chance of happening occurred approximately 70% of the time.
- Some events are consistently overestimated. Forecasts involving military conflict, international relations, and technology tended to assign higher probabilities to events than the outcomes justified.
- Election forecasts show a different pattern. Prediction markets tended to underestimate the likelihood of election-related events across certain probability ranges, particularly when predicted chances were low or moderate.
- Predictions become more accurate as events approach. However, this improvement mainly reflects better identification of which events will happen, rather than better calibration of probabilities. Forecasts were already well calibrated after just 25% of their allotted forecasting period.
When the Crowd Gets the Odds Right
Prediction markets are increasingly used to estimate the likelihood of political, economic, and technological events. Rather than examining a handful of high-profile predictions, the researchers looked at more than a million forecasts to understand how these systems perform on a much larger scale.
Their findings reveal an important distinction: while prediction markets are generally good at estimating probabilities, their reliability varies depending on the type of event being predicted.
What Does This Mean for the Future of Forecasting?
The findings do not suggest that prediction markets can predict the future with certainty. Instead, they show that collective forecasts can provide valuable information about what might happen, provided we understand their limitations.
By identifying the areas in which prediction markets systematically overestimate or underestimate the likelihood of events, the study offers a clearer picture of how to interpret these forecasts, especially when the outcomes could have significant political, economic, or technological consequences.
Research
Oct 13th, 2026
Can Ketamine Relieve Pain Without Altering Consciousness?
A new TAU study reveals that ketamine's pain-relieving effects and its influence on consciousness involve separate brain mechanisms, opening possibilities for future treatments with fewer psychological side effects.
- Medicine
A new brain imaging study in humans, led by Tel Aviv University's Institute for Psychedelic Research in collaboration with TAU-affiliated Tel Aviv Medical Center (Ichilov), has found that ketamine can relieve pain independently of the sense of dissociation it produces. Dissociation is a feeling of detachment from oneself, one's body, or one's surroundings. The findings suggest that ketamine's effects on pain and consciousness involve different brain mechanisms, potentially paving the way for treatments that retain its therapeutic benefits while reducing unwanted psychological side effects.
The study was led by Dr. Noam Goldway, under the supervision of Prof. Talma Hendler and Dr. Haggai Sharon of the Institute for Psychedelic Research at TAU's Sagol School of Neuroscience, in collaboration with Tel Aviv Medical Center (Ichilov). The findings were published in Neuropsychopharmacology, a journal in the Nature portfolio.
Is Altering Consciousness Necessary for Pain Relief?
In recent years, mind-altering substances have moved from the fringes of popular culture to the forefront of medical research. MDMA is being studied as a potential treatment for post-traumatic stress disorder (PTSD), while psilocybin, the active substance found in magic mushrooms, is being investigated primarily for depression. Ketamine has also attracted considerable interest because of its effectiveness in treating pain and depression, and more recently, its potential for treating PTSD. But alongside this promise, a fundamental question remains: Do these drugs need to alter our state of consciousness to produce their therapeutic effects?
To investigate this question, the researchers focused on ketamine, a drug known both for its ability to relieve pain and for inducing dissociation, a temporary feeling of detachment from oneself, one's body, or one's surroundings. For years, physicians and researchers have assumed that these effects were connected: the more detached a person feels, the less intense or threatening their pain may seem.
The study included 37 healthy volunteers, both women and men. Each participant received a ketamine infusion and a placebo infusion in two separate sessions. During the experiment, participants underwent functional magnetic resonance imaging (fMRI) while exposed to controlled pain stimuli. They were also asked to report the intensity of their pain and the degree of dissociation they experienced.
Two Effects, Two Different Brain Mechanisms
According to Dr. Noam Goldway, although ketamine reduced pain and increased dissociation, the researchers found no connection between the intensity of these two effects. Participants who experienced stronger dissociation did not necessarily report greater pain relief, and those who experienced less dissociation could still benefit from the drug's pain-relieving effects. In other words, the findings suggest that the feeling of detachment is not what causes ketamine to reduce pain.
The brain scans supported this conclusion. Pain relief was associated with reduced activity in regions involved in processing pain, including the insula and the anterior cingulate cortex. Dissociation, meanwhile, was linked to changes in a different brain network known as the default mode network, which plays a role in processing our sense of self. These findings suggest that ketamine's effects on pain and consciousness are not simply two unrelated experiences, but involve distinct brain mechanisms influenced by the same drug.
Prof. Talma Hendler, Head of TAU's Institute for Psychedelic Research, explains:
“In recent years, there has been a lively scientific and clinical debate over whether the subjective experience produced by psychoactive drugs is an integral part of their therapeutic effect. Our study shows that, at least in the case of ketamine and pain relief, the answer is apparently no. We were able to show that the drug's effect on pain is separate from its effect on consciousness - both behaviorally and in terms of brain activity. It is important to note that this cannot be generalized to other psychedelic drugs, such as magic mushrooms. In this case, the burden of proof still rests with us, the scientists.”
Toward Pain Treatments with Fewer Side Effects
According to Dr. Haggai Sharon, Deputy Head and Medical Director of the Institute and a specialist in pain medicine, the study provides further evidence of just how complex the brain's experience of pain is, and the many ways in which the brain can regulate and change it. This understanding could also have important clinical implications. If researchers can find a way to activate ketamine's pain-relieving mechanism without triggering dissociation, it may become possible to develop treatments that maintain the drug's effectiveness while reducing feelings of detachment and other subjective side effects.
