Where the wild Discovery Loops are | Terra Incognita
Skip to content
Home<br>»2026<br>»June<br>»Where the wild Discovery Loops are
Philosophy
Where the wild Discovery Loops are
26/06/202626/06/2026 by Christian S. Perone
Sir Karl Raimund Popper (1902 – 1994).<br>Karl Popper offered an elegant normative account of science as a process of conjectures and refutations : we formulate hypotheses or theories, expose them to criticism, scrutiny and experiment, and retain only what survives our best attempts at falsification. Even though we couldn’t ever know if we reached the truth, we can approximate it and get closer to it, just like Popper noted from Xenophanes: "(…) for all is but a woven web of guesses.". I really like Popper’s views on philosophy of science and also his work on political philosophy in "The Open Society and its Enemies" (1945) and I also wrote about some other aspects of his philosophy in relation to professional ethics here in the past as well, however, Popper’s philosophy helps describing this process of conjectures and refutations, but not the origin of genuinely novel hypotheses, and this is where many other philosophers like Thomas Kuhn (who proposed his "paradigm shift") were closer on the account of how many scientific discoveries actually unfolded.
Thomas S. Kuhn (1922 – 1996).<br>Now, one interesting connection of the scientific method is the modern automated "discovery loop " (I just invented this term because we don’t have a name for that yet, but we are full of loops now don’t we ?) that we are starting to see on many works, where Google DeepMind (GDM) is I think one of the main interested parties and pioneers on it too, like AlphaEvolve, Aletheia, Co-Scientist, etc (we need better open-source ecosystem for these loops to be honest, but that is another topic). When I look at these discovery loops, they are very similar to Popper’s views on the scientific method, but science is much more messy than that and there were many examples of cases from the major scientific discoveries where we can see that Popper’s view on the method and also falsificationism (I cannot really pronounce this word very quickly) worked well as an account for a logic of criticism, but was weaker as an account of the genesis of new hypotheses. Scientific discovery often involves a lot of intuition, analogy, social incentives, ignored anomalies, tacit knowledge, instrument-building, aesthetic judgment, accidental observations, etc, but most importantly, some of the major discoveries required a completely new "framework", or a "paradigm shift" like Kuhn described. It seems to me that the major gap today in this automated "discovery loop" is really on creativity, specifically on the hypothesis generation (obviously not limited to).
Margaret Boden (1936 – 2025).<br>Therefore, the question now shifts to how we should characterize "creativity", Margaret Boden did a very original work on separating three different types of creative work. The first one she called "exploratory creativity " where you push existing rules or forms to their limits while still working within them. Then you have the "combinational creativity " where you can create something new by bringing together ideas, tools, or styles from different areas and lastly the "transformational creativity ", which is the rarest one where you change the rules or removes assumptions, so you basically recreate the entire framework itself instead of playing by its rules. If you have been working with ML for a while, you might have noticed that it is clear that exploratory and combinational creativity is something we have been seeing growing a lot in modern ML systems (I’m saying systems here to avoid the word LLM as today’s systems are really an entire ecosystem working together), and that makes total sense because it is the kind of generalization that we keep seeing on many different domains ranging from (conditioned and unconditioned) diffusion/flow matching models to robotic policies. However, we are yet to see transformational capabilities arising, and this seems to be one of the biggest gaps (besides the slowness of scientific experimentation) that is holding ML systems back into major breakthroughs. It is not without a reason that Kant also tied imagination as a central piece on his schematism, which to me has a lot of connections to generalization and how we go from concepts to particular experiences.
It seems to me that achieving this transformational leap on automated discovery loops would require a paradigm (no pun intended) change on how we propose hypotheses, and it might mean that breaking the rules would require much more than just sampling from models, hence why I think methods incorporating evolutionary computation (EC) algorithms not only to evolve the hypothesis but also to evolve the models generating hypothesis are a very exciting area (same for perturbation methods). Obviously there might be a lot of scientific discoveries that might not need this...