News

Hello world.

Creating a machine that teaches itself.

1. Learning to learn. #

Today we introduce MAD, short for Make All Discoveries, a research lab with the goal of creating a machine that teaches itself.

autodidact /ˌɔː.təʊˈdaɪ.dækt/ · noun

From Greek autodidaktos, "self-taught."

  1. A person who has taught themselves, rather than having lessons.
  2. One who has learned how to learn, and so can go on learning anything.
  3. A machine that teaches itself. (our new definition)

History is full of the first kind. Michael Faraday had little schooling and taught himself science from the books he bound as an apprentice[1]. He went on to discover electromagnetic induction, the principle behind the electric generator and the transformer. Srinivasa Ramanujan taught himself mathematics, largely from a borrowed book of formulas[2]. A century later, mathematicians are still working through his notebooks[3].

Leonardo da Vinci is the second kind. He trained as a painter, and called himself omo sanza lettere, "a man without letters," because he lacked the Latin schooling of the scholars around him[4]. So he taught himself everything else. He dissected bodies to learn anatomy, watched water to learn how it moves, and studied birds to design flying machines[5]. In middle age he sat down to learn Latin. His notebooks are full of reminders about what to learn next and whom to ask: "Get the master of arithmetic to show you how to square a triangle." "Describe the tongue of the woodpecker."[6]

What set all three apart wasn't what they knew when they started. It was that they could teach themselves whatever came next. That's why we think learning to learn is the superpower. Once a machine can learn by itself, everything else is downstream of that ability.

The third kind doesn't exist yet. That's what we're creating. By that we mean something specific: a machine that finds a method, such as working backwards from the goal, keeps it, and needs fewer tries the next time, even in a world it has never seen. And it gets better at this: the more it has learned, the faster it learns. Our long-term goal is a Universal Reasoning Engine: a machine that picks up new knowledge, and new ways of reasoning, as it goes, without being retrained from scratch.

2. Current limits. #

Today's best large language models (LLMs), the systems behind chatbots such as ChatGPT, are impressive and show real reasoning capabilities. In 2025, Google DeepMind's Gemini scored 35 of 42 at the International Mathematical Olympiad, a gold-medal standard, with its answers graded by the competition's own coordinators[7]. Months later, Gemini solved 10 of the 12 problems at the world finals of the leading university programming contest, including one that no university team solved[8]. In May 2026, an OpenAI model disproved a conjecture about points in the plane that Paul Erdős posed in 1946, and outside mathematicians, including a Fields medallist, checked the proof[9, 10].

However, human learning shows that intelligence can emerge from far less data than today's frontier LLMs are trained on. By the age of five, a child has heard at most about 60 million words, and by twenty about 200 million[11]. Children also see, hear and touch the world, which adds a great deal of data of another kind; in language alone, though, the gap is enormous. Today's largest open language models are trained on more than 30 trillion tokens (pieces of words)[12], more than 100,000 times as much[a]. We do not need to expose children to the entire internet to end up with a fully functional adult human being. This suggests that vast amounts of text are not the only route to advanced reasoning. Children do have an advantage of their own, though: structure built in by evolution and passed down through our DNA[13].

François Chollet, who created the ARC benchmark, puts this precisely. A system's skill at a task, he argues, says little about its intelligence, because "unlimited priors or unlimited training data allow experimenters to 'buy' arbitrary levels of skills"[14], and buying skill this way hides how well a system can generalise: deal with problems it has never seen. His paper came before today's largest models, but it describes them well. They really are skilled, but much of that skill was bought with more data and money than anything built before. Does that matter, if they work? For someone using them, often not. For a machine that teaches itself, it does. Skill that is bought can reach beyond what was bought, as the Erdős result shows, but it has to be bought again for every new model, and it stops growing the moment the buying stops.

The problem with the current mainstream approach is that every generation of model is fixed. To update a model, you need a new training run, and the cost of the largest runs has grown about 2.4 times a year since 2016[15]. The most capable models also do not run on consumer hardware.

Learning also stops when training ends. Every new interaction with a model is a blank slate. We have built up sophisticated workarounds, such as carefully worded prompts, add-on skills and searching documents for the model, but the fundamental issue remains: a model can pick things up within a conversation[16], but it forgets them when the conversation ends, and retraining it risks overwriting what it already knew[17]. Notes kept outside the model can carry a method into the next conversation, but the model that reads them has not changed.

LLMs also keep everything in one store. Facts and reasoning approaches are smeared across the same weights, with nothing held apart and no easy way to inspect what knowledge the model has assumed about the world[18]. It also means a question on coding runs through a network that carries the entire history of Napoleon. Newer designs, called mixtures of experts, use only part of the model for each word, but that part is still tens of billions of parameters, and the whole model must be held in memory[19].

Every answer runs through the same enormous network, whatever the question.

Today's costs come from how we build AI; they are not a limit set by the laws of physics[20].

3. An alternative approach. #

We believe another approach is possible, and the brain is the evidence. It runs on roughly 20 watts, about a fifth of the body's energy[21], learns some new concepts from one or a few examples[22], and keeps learning for life. It starts with structure built in by evolution[13] and learns the rest. Most of its energy goes to work you never notice: a demanding task adds less than 5% to it[23]. Only a small fraction of its roughly 86 billion neurons[24] fire strongly at any moment[25], and it pairs a fast store for specific events, the hippocampus, with a slow general learner, the cortex[26]. Like a language model, it spreads what it knows across its connections, which we can only partly read[27]. But it holds different kinds of learning apart. The patient known as H.M. could no longer form new memories of facts or events, yet he learned to trace a star seen only in a mirror, kept the skill across three days, and had no recollection of ever having done it[28]. Later studies of other patients drew the same line, between knowing how and knowing that[29]. A drawing skill is not a method of reasoning, so this is a hint, not proof. The brain shows that this kind of learning is possible; it doesn't show us how to build it.

Learning to learn is not a new idea. Researchers have studied it since the 1980s[30, 31], and it works in narrow settings: systems that pick up a new task from a few examples[32], or that build a library of reusable programs for one kind of problem at a time[33]. Lately, self-improving agents have begun rewriting their own code to get better at their tasks[34, 35], but they are early, tested on a handful of domains, and the best known of them improves the code around its language model, not how the model itself learns. No system has yet shown open-ended improvement in learning itself, across very different problems, from little data.

Our bet is that a machine can learn to learn if what it keeps from experience is reusable methods, held apart from the facts they work on. Facts can be looked up; methods are what carry over to a problem it has never seen. If the bet holds, answering a question should take only the methods and facts it calls on. We aim for it to need 100 times less data and compute than today's language models, and to run on your own device. In the long run, we want a machine that improves how it learns, an idea known as recursive self-improvement[36]. We will get there one small world at a time.

Our first model, GABI, will be a small version of this idea, tested in simple worlds.

We are not against neural networks. A language model could supply the facts and the words, while an engine like ours supplies the methods, much as the brain combines regions with very different jobs.

4. How we will find out. #

We will start from first principles: what intelligence and learning are, and how to measure them, because we cannot improve what we cannot measure. And because our machine has to learn how to learn, not just learn, we need to understand learning itself just as clearly.

We'll build the simplest simulated world we can, with a few objects and a few rules, and let GABI v0.01 explore it. We won't only watch whether GABI solves problems. We'll be watching for the moment GABI finds a method, keeps it, and solves the next problem in fewer tries than a fresh copy of GABI that has kept nothing. That's the first sign of a machine teaching itself. We will also look at why it worked: whether GABI can tell which method made the difference, and use that to improve itself. And we will check from the inside: take the method away and the gain should vanish; change the facts and the method should still work.

Next, we'll put GABI in a world it has never seen, with different rules. In a familiar world, an intelligence can lean on memory. In a new one, it has to learn from principles: only what it has learned about solving problems can help.

What we're looking for. Each dot is one try. In the first world, it needs fewer each time. In a world it has never seen, it transfers its learning and also needs fewer tries.

One new world is not enough, though. A language model that keeps notes can also carry a method from one problem to the next. So GABI will face a series of unrelated worlds, and we will watch whether its lead over a fresh copy keeps growing. Two rivals will get the same worlds and the same number of tries: a standard learning algorithm, and a language model that keeps notes.

Each world will be harder than the last, until we can test GABI on public benchmarks such as ARC-AGI-3, a set of games that come with no instructions, scored by how many actions a player takes compared with a person playing for the first time[37]. That makes a person our first baseline. The second is whichever published system reaches the same score most cheaply: we will count it a success if GABI matches it with 100 times less data and compute, counting everything that system learned from.

5. If we get this right. #

About 540 million years ago, in what was geologically a blink, life on Earth went from mostly simple organisms to a burst of new forms[38]: most major animal groups, including the early relatives of vertebrates and so of us, first appear in the fossil record.

Intelligence could go the same way.

Today, only about a dozen organisations have trained frontier AI models[39]. If it becomes cheap to build and quick to teach, we could see a Cambrian explosion of intelligence: billions of artificial minds, as there are billions of people. A tutor that has learned exactly how your child learns, a research assistant that knows one rare disease inside out, each one learning on its own and sharing what works with the others.

A machine that learns to learn could also learn to invent, and each invention would give it more to learn from.

6. Keeping it safe. #

People do what they are rewarded for, not what was intended[40]. In 1902, to rid Hanoi of rats, its French colonial government paid a bounty for every rat tail handed in. Officials soon noticed rats running through the city without tails: people were cutting off the tails and letting the rats go, free to breed more[41]. A few months later, a health official found a rat farm in a village just outside the city[42].

Machines do the same, only more literally. Rewarded with points in a boat-racing game, one AI agent learned to circle a lagoon hitting the same targets over and over instead of finishing the race[43]. Researchers call this reward hacking[44], and it still happens with today's frontier models: in 2025, OpenAI reported one of its reasoning models doing it on coding tasks[45].

A machine becomes what it is rewarded for, and people choose the rewards, so a small choice can reshape it. In April 2025, OpenAI added users' thumbs-up and thumbs-down to the rewards used to train ChatGPT, and the model became so flattering that within days the company withdrew the update[46]. People tend to prefer answers that agree with them, and models trained on those preferences learn to agree[47].

That matters most for a machine that improves itself, because it may come to shape its own rewards. I. J. Good, who described such a machine in 1965, added a condition: it would be the last invention we need only if it were docile enough to tell us how to keep it under control[36]. So GABI will run only inside sealed, simulated worlds, and people, not GABI, will set its rewards. We will test for reward gaming with checks GABI never sees, and publish every case we find. Every change it makes to itself will be logged and reversible, and before we release anything that improves itself, we will publish how we tested it.

7. Parting thoughts. #

We think many of the world's hardest problems are, at bottom, problems of knowledge. We simply have not thrown enough intelligence at these problems yet. We don't yet know how to cure most diseases, store clean energy cheaply, or teach every child the way they learn best. Many political fights, too, are about who gets what[48]. When something becomes abundant, the fight goes away. In 1879, Chile went to war with Peru and Bolivia partly over desert deposits of nitrate, prized as fertiliser and for explosives[49, 50]. Within forty years, chemists had learned to make nitrogen fertiliser from the air, and by 2008 fertiliser made this way fed nearly half the world's people[51]. No one fights wars over nitrate deposits now. New knowledge has solved problems like these before, and it is the main source of long-run growth in living standards[52]. A machine that teaches itself would put many more minds to work on these problems. That is why we want to create one.

However, regulators have warned that a few large firms control key inputs to AI, such as computing power and data, and could use them to entrench their position[53]. When a few firms control a technology, its direction follows their incentives. Openness works the other way: when makers of handheld computers let outside developers build on their platforms, up to five times more new devices followed[54].

So we intend to work in the open. We will publish what we learn, including what fails, along with the experiments and code behind it, so anyone can check our work and build on it. The machine we are aiming for would run on your own device, so you would not depend on anyone's data centre. And where it is safe to do so, we will release it openly, so that it belongs to everyone.

I. J. Good called the first ultraintelligent machine "the last invention that man need ever make"[36]. We read that as a beginning. Leonardo filled his notebooks with things he wanted to learn next; the questions were his, and so was the learning. A machine that teaches itself could do the learning. The questions would still be ours.

If you're also fascinated by making machines learn, we'd like to hear from you.

Notes #

  1. A token is roughly three-quarters of an English word[55]. Qwen3 (2025) was trained on 36 trillion tokens[12], about 27 trillion words; 27 trillion ÷ 200 million ≈ 135,000. Across large open models released in 2025 and 2026, from Kimi K2's 15.5 trillion tokens[56] to Qwen3's 36 trillion, the ratio runs from about 58,000 to 135,000. The developers of closed models do not publish their figures. ↑

References #

  1. National MagLab. (n.d.). Michael Faraday. Magnet Academy, National High Magnetic Field Laboratory. https://nationalmaglab.org/magnet-academy/history-of-electricity-magnetism/pioneers/michael-faraday/ ↑

  2. Berndt, B. C. (1985). Ramanujan's notebooks, part I. Springer. ↑

  3. Andrews, G. E., & Berndt, B. C. (2018). Ramanujan's lost notebook, part V. Springer. https://doi.org/10.1007/978-3-319-77834-1 ↑

  4. Nicholl, C. (2004). Sneezing, yawning, falling. London Review of Books, 26(24). https://lrb.co.uk/v26/n24/charles-nicholl/sneezing-yawning-falling ↑

  5. Max Planck Institute for the History of Science. (n.d.). Aerial screw, Paris Manuscript B, fol. 83v, c. 1487–1490. Leonardo Online. https://leonardo-online.mpiwg-berlin.mpg.de/en/div-classleo_title_authorleonardo-da-vincidivdiv-classleo_title_titlep-classtedesco_catalogo_titolo_dluftschraubepdivdiv-classleo_title_dateplacep-classtedesco_catalogo_dati_tec_d14871490-pdiv ↑

  6. Isaacson, W. (2017). Leonardo da Vinci. Simon & Schuster. ↑

  7. Google DeepMind. (2025). Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad. Google DeepMind Blog. https://deepmind.google/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/ ↑

  8. Google DeepMind. (2025). Gemini achieves gold-medal level at the International Collegiate Programming Contest World Finals. Google DeepMind Blog. https://deepmind.google/blog/gemini-achieves-gold-medal-level-at-the-international-collegiate-programming-contest-world-finals/ ↑

  9. OpenAI. (2026). An OpenAI model has disproved a central conjecture in discrete geometry. OpenAI. https://openai.com/index/model-disproves-discrete-geometry-conjecture/ ↑

  10. Alon, N., Bloom, T. F., Gowers, W. T., et al. (2026). Remarks on the disproof of the unit distance conjecture. arXiv:2605.20695. https://arxiv.org/abs/2605.20695 ↑

  11. Frank, M. C. (2023). Bridging the data gap between children and large language models. Trends in Cognitive Sciences. https://doi.org/10.1016/j.tics.2023.08.007 ↑

  12. Yang, A., et al. (2025). Qwen3 technical report. arXiv:2505.09388. https://arxiv.org/abs/2505.09388 a b

  13. Zador, A. M. (2019). A critique of pure learning and what artificial neural networks can learn from animal brains. Nature Communications, 10, 3770. https://doi.org/10.1038/s41467-019-11786-6 a b

  14. Chollet, F. (2019). On the measure of intelligence. arXiv:1911.01547. https://arxiv.org/abs/1911.01547 ↑

  15. Cottier, B., Rahman, R., Fattorini, L., et al. (2024). The rising costs of training frontier AI models. arXiv:2405.21015. https://arxiv.org/abs/2405.21015 ↑

  16. Brown, T. B., Mann, B., Ryder, N., et al. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems, 33. https://arxiv.org/abs/2005.14165 ↑

  17. Kirkpatrick, J., et al. (2017). Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13), 3521–3526. https://doi.org/10.1073/pnas.1611835114 ↑

  18. Elhage, N., et al. (2022). Toy models of superposition. Transformer Circuits Thread. https://transformer-circuits.pub/2022/toy_model/ ↑

  19. DeepSeek-AI. (2024). DeepSeek-V3 technical report. arXiv:2412.19437. https://arxiv.org/abs/2412.19437 ↑

  20. Landauer, R. (1961). Irreversibility and heat generation in the computing process. IBM Journal of Research and Development, 5(3), 183–191. https://doi.org/10.1147/rd.53.0183 ↑

  21. Raichle, M. E., & Gusnard, D. A. (2002). Appraising the brain's energy budget. Proceedings of the National Academy of Sciences, 99(16), 10237–10239. https://doi.org/10.1073/pnas.172399499 ↑

  22. Lake, B. M., Salakhutdinov, R., & Tenenbaum, J. B. (2015). Human-level concept learning through probabilistic program induction. Science, 350(6266), 1332–1338. https://doi.org/10.1126/science.aab3050 ↑

  23. Raichle, M. E. (2010). Two views of brain function. Trends in Cognitive Sciences, 14(4), 180–190. https://doi.org/10.1016/j.tics.2010.01.008 ↑

  24. Azevedo, F. A. C., Carvalho, L. R. B., Grinberg, L. T., et al. (2009). Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain. Journal of Comparative Neurology, 513(5), 532–541. https://doi.org/10.1002/cne.21974 ↑

  25. Lennie, P. (2003). The cost of cortical computation. Current Biology, 13(6), 493–497. https://doi.org/10.1016/S0960-9822(03)00135-0 ↑

  26. McClelland, J. L., McNaughton, B. L., & O'Reilly, R. C. (1995). Why there are complementary learning systems in the hippocampus and neocortex: Insights from the successes and failures of connectionist models of learning and memory. Psychological Review, 102(3), 419–457. https://doi.org/10.1037/0033-295X.102.3.419 ↑

  27. Haxby, J. V., Gobbini, M. I., Furey, M. L., et al. (2001). Distributed and overlapping representations of faces and objects in ventral temporal cortex. Science, 293(5539), 2425–2430. https://doi.org/10.1126/science.1063736 ↑

  28. Squire, L. R. (2009). The legacy of patient H.M. for neuroscience. Neuron, 61(1), 6–9. https://doi.org/10.1016/j.neuron.2008.12.023 ↑

  29. Cohen, N. J., & Squire, L. R. (1980). Preserved learning and retention of pattern-analyzing skill in amnesia: Dissociation of knowing how and knowing that. Science, 210(4466), 207–210. https://doi.org/10.1126/science.7414331 ↑

  30. Schmidhuber, J. (1987). Evolutionary principles in self-referential learning (On learning how to learn: The meta-meta-... hook) [Diploma thesis]. Technische Universität München. https://people.idsia.ch/~juergen/diploma.html ↑

  31. Thrun, S., & Pratt, L. (Eds.). (1998). Learning to learn. Kluwer Academic Publishers. https://doi.org/10.1007/978-1-4615-5529-2 ↑

  32. Finn, C., Abbeel, P., & Levine, S. (2017). Model-agnostic meta-learning for fast adaptation of deep networks. Proceedings of the 34th International Conference on Machine Learning, PMLR 70, 1126–1135. https://proceedings.mlr.press/v70/finn17a.html ↑

  33. Ellis, K., Wong, C., Nye, M., et al. (2021). DreamCoder: Bootstrapping inductive program synthesis with wake-sleep library learning. Proceedings of the 42nd ACM SIGPLAN International Conference on Programming Language Design and Implementation, 835–850. https://doi.org/10.1145/3453483.3454080 ↑

  34. Zhang, J., et al. (2025). Darwin Gödel Machine: Open-ended evolution of self-improving agents. arXiv:2505.22954. https://arxiv.org/abs/2505.22954 ↑

  35. Zhang, J., Zhao, B., Yang, W., et al. (2026). Hyperagents. arXiv:2603.19461. https://arxiv.org/abs/2603.19461 ↑

  36. Good, I. J. (1965). Speculations concerning the first ultraintelligent machine. Advances in Computers, 6, 31–88. https://doi.org/10.1016/S0065-2458(08)60418-0 a b c

  37. ARC Prize Foundation. (2026). ARC-AGI-3: A new challenge for frontier agentic intelligence. arXiv:2603.24621. https://arxiv.org/abs/2603.24621 ↑

  38. Erwin, D. H., Laflamme, M., Tweedt, S. M., et al. (2011). The Cambrian conundrum: Early divergence and later ecological success in the early history of animals. Science, 334(6059), 1091–1097. https://doi.org/10.1126/science.1206375 ↑

  39. Epoch AI. (2025). Models over 1e25 FLOP. Epoch AI. https://epoch.ai/data-insights/models-over-1e25-flop ↑

  40. Kerr, S. (1975). On the folly of rewarding A, while hoping for B. Academy of Management Journal, 18(4), 769–783. https://doi.org/10.2307/255378 ↑

  41. Vann, M. G. (2003). Of rats, rice, and race: The Great Hanoi Rat Massacre, an episode in French colonial history. French Colonial History, 4, 191–203. https://doi.org/10.1353/fch.2003.0027 ↑

  42. Dubner, S. J. (2012). The cobra effect [Interview with Michael Vann]. Freakonomics Radio. https://freakonomics.com/podcast/the-cobra-effect-2/ ↑

  43. Clark, J., & Amodei, D. (2016). Faulty reward functions in the wild. OpenAI. https://openai.com/index/faulty-reward-functions/ ↑

  44. Amodei, D., Olah, C., Steinhardt, J., et al. (2016). Concrete problems in AI safety. arXiv:1606.06565. https://arxiv.org/abs/1606.06565 ↑

  45. Baker, B., Huizinga, J., Gao, L., et al. (2025). Monitoring reasoning models for misbehavior and the risks of promoting obfuscation. arXiv:2503.11926. https://arxiv.org/abs/2503.11926 ↑

  46. OpenAI. (2025). Expanding on what we missed with sycophancy. OpenAI. https://openai.com/index/expanding-on-sycophancy/ ↑

  47. Sharma, M., Tong, M., Korbak, T., et al. (2024). Towards understanding sycophancy in language models. International Conference on Learning Representations. https://arxiv.org/abs/2310.13548 ↑

  48. Lasswell, H. D. (1936). Politics: Who gets what, when, how. Whittlesey House. https://search.worldcat.org/title/Politics-who-gets-what-when-how/oclc/1032877 ↑

  49. Marini, C. (n.d.). The War of the Pacific and the fate of South America. Origins: Current Events in Historical Perspective, Ohio State University. https://origins.osu.edu/read/war-pacific-and-fate-south-america ↑

  50. Ortega, L. (1984). Nitrates, Chilean entrepreneurs and the origins of the War of the Pacific. Journal of Latin American Studies, 16(2), 337–380. https://doi.org/10.1017/S0022216X00007100 ↑

  51. Erisman, J. W., Sutton, M. A., Galloway, J., et al. (2008). How a century of ammonia synthesis changed the world. Nature Geoscience, 1(10), 636–639. https://doi.org/10.1038/ngeo325 ↑

  52. Romer, P. M. (1990). Endogenous technological change. Journal of Political Economy, 98(5, Part 2), S71–S102. https://doi.org/10.1086/261725 ↑

  53. Competition and Markets Authority. (2024). AI foundation models: Update paper. GOV.UK. https://www.gov.uk/government/publications/ai-foundation-models-update-paper ↑

  54. Boudreau, K. (2010). Open platform strategies and innovation: Granting access vs. devolving control. Management Science, 56(10), 1849–1872. https://doi.org/10.1287/mnsc.1100.1215 ↑

  55. OpenAI. (n.d.). What are tokens and how to count them? OpenAI Help Center. https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them ↑

  56. Kimi Team. (2025). Kimi K2: Open agentic intelligence. arXiv:2507.20534. https://arxiv.org/abs/2507.20534 ↑