Taste-Bench

Philipp Zahn and Bruno Marnette speak with Tal Kachman, assistant professor at Radboud University, where he leads the CATALYST Lab. Tal trained as a chemist and a physicist. After a time in finance, he dived back into AI research.

Borrowing a point from Jeff Clune, Tal argues that as production of scientific knowledge is abundant and almost everything is novel, the challenge now is how original and important a new finding really is. When his group pushed LLMs toward new discoveries in chemistry, each leap came out smaller than the last. He thinks, a key challenge is to build scaffolding so that the models get unstuck (e.g. through external noise).

We also discuss the differences across fields. Physical experiments are still harder for AI than other domains: he describes off-the-shelf robot models that failed with an emulsion that goes from liquid to solid, like tahini.

Tal closes with his take on what AI implies for teaching and learning. More than ever it is important for students to lay the right foundations.

More from Tal: https://www.cerebrnita.com/

How do you recognise an exceptional founder when there's little data to analyse? And can the judgment required to make that decision ever be codified, or even replicated by AI?

In this episode of Taste Bench, we speak with Krishna Visvanathan, General Partner at Crane Venture Partners, about the role of taste, intuition, and imagination in early-stage investing.

Krishna explains why the best founders are often outliers who defy conventional frameworks, and why looking for exceptional people with a rigid checklist is fundamentally paradoxical. We explore how investors develop their taste in people, why diversity of judgment matters within investment teams, and how conviction and trust shape decisions under uncertainty.

We also discuss how AI is transforming venture capital and entrepreneurship. As AI makes it easier than ever to build products, pitch ideas, and analyse markets, could it also make founders more alike? And what human qualities will remain essential in a world of increasingly capable AI?

A conversation about recognising exceptional talent, making decisions without complete information, and why good judgment is about much more than intelligence.

If AI writes code in seconds, why aren't most companies shipping much more software?

Philipp Zahn and Bruno Marnette speak with Will Wilson, co-founder and CEO of Antithesis. The company tests software by running it in simulation through a huge number of situations, including hardware faults and unusual timing, to find the bugs nobody thought to test for.

Will explains why he chose testing because it was low status: few good engineers wanted to work on it, so there was a lot to improve. He came to it at FoundationDB, where a simulator caught his bugs every time he changed the code, and it changed how he thought about programming.

With AI coding, testing has moved into the spotlight. Will's answer to the question above: the typing was never the slow part. Deciding what to build, checking that it works and running it still takes time. We talk about stating what must always be true instead of writing out test cases, and how such a list could become the specification, the tests and the documentation at once.

We also discuss why AI coding works so well, which Will traces back to Stack Overflow and GitHub, why he thinks this might change, and what is left for human engineers.

More from Will: https://antithesis.com/author/will-wilson/

What can a translator of the Bible teach us about AI?

Philipp Zahn and Bruno Marnette speak with Robert Alter, Emeritus Professor of Hebrew and Comparative Literature at UC Berkeley, about his new book, The Language of Fiction (2027).

Robert explains how great writers do something with language that no one has done before. One of his examples is Faulkner using "abrupt" as a verb: the word surprises the reader and then, on reflection, fits the whole novel.

In his view, AI can imitate Faulkner's style, but not make that move, because it cannot be inferred from other works. And since we are language-using animals, he argues, some people will keep enjoying this mastery of language, even in the age of AI.

We also discuss his work on the Hebrew Bible. In The Art of Biblical Narrative (1981), he proposed reading the Bible as literature. He then applied that close reading to his own translation, which took 23 years. He wrestled with every word and found meanings that earlier translators had missed.

More on The Language of Fiction: https://www.hup.harvard.edu/books/9780674308374

REFERENCES

Robert Alter, The Art of Biblical Narrative, revised edition (Basic Books, 2026) https://www.hachettebookgroup.com/titles/robert-alter/the-art-of-biblical-narrative/9781541609273/

Robert Alter, The Hebrew Bible: A Translation with Commentary (W. W. Norton, 2018) https://wwnorton.com/books/9780393292497

Philipp Zahn and Bruno Marnette speak with Jack Marsh, co-founder of Not Quite Past. The London company uses AI to design custom tiles in historical Delft styles, which are then produced in Stoke-on-Trent. Interior designers and private customers buy them.

Jack studied art history, and for him slop is not really new. We go back to the Great Exhibition of 1851, where machine-made English ornament was judged a national embarrassment. Jack explains how the public reacted, and how that reaction led to the museum known today as the V&A. As Jack puts it, a museum against slop. We talk about the industrialists who took up the new production methods and still made high-quality products, bought by the elites of their day. For Jack, this is the model for how to use AI.

We also discuss why his customers rarely design their own tiles, and why perfect copies of historical tiles can feel lifeless.

More from Jack: https://www.notquitepast.com/

Most talk about AI and children is either fear or hype. Few people actually teach kids how

the systems work.

Saranyan Vigraham does. He is Director of Engineering at Meta, where he works on compute and

AI infrastructure, and he teaches AI and ethics to middle schoolers at Khan Lab School.

He does not think AI output is bad, or that craftmanship will die. His worry is that people who

never practiced a craft can now make work that looks like they did, while the ones who did

get harder to find and harder to pay.

AI also took away two things that build craft over time: waiting, and feedback you can use. 

In this episode he also takes us through the experience he had with two side projects of his. One was in cybersecurity,

which he knows only from reading. He built something convincing, but had no opinion of his

own, so the feedback people gave him was useless. That project stalled. The other was in

developer tooling, which he knows cold. People disagreed with him there too but he kept

building.

The AI he wants instead runs on your own device, owned by no frontier lab. It learns with

you, knows when to shut up, and sends you out to check what you learned.

More about Saranyan: https://www.saranyan.com/

Markus J. Buehler is a professor at MIT. His field is materials science but his research spans across other fields: Biology, chemistry, engineering, and, yes, music.

In recent years, he has been integrating AI into his research. But his focus is not on a single, new scientific artifact but instead the full research process. His work focuses on the creation of agent systems that can create real discoveries.

In this episode, we talk about the rapid evolution of this approach in the last years. We also discuss the importance of his internal feedback loop, the personal surprise and excitement when something really new has been created. For Markus, this is the actual test and way more important than any benchmark.

More about Markus: https://lamm.mit.edu/

REFERENCES

Markus J. Buehler, "MusicSwarm: Biologically Inspired Intelligence for Music Composition" https://arxiv.org/abs/2509.11973

Markus J. Buehler, "Deep Aria (Abbreviata)", built from J. S. Bach's Goldberg Aria https://soundcloud.com/user-275864738/deeparia-abridged

Markus J. Buehler, "Selective Imperfection as a Generative Framework for Analysis, Creativity and Discovery" https://arxiv.org/abs/2601.00863

Fiona Y. Wang, Di Sheng Lee, David L. Kaplan & Markus J. Buehler, "Swarms of Large Language Model Agents for Protein Sequence Design with Experimental Validation" https://arxiv.org/abs/2511.22311

Alireza Ghafarollahi & Markus J. Buehler, "Sparks: Multi-Agent Artificial Intelligence Model Discovers Protein Design Principles" https://arxiv.org/abs/2504.19017

Fiona Y. Wang & Markus J. Buehler, "Self-Revising Discovery Systems for Science: A Categorical Framework for Agentic Artificial Intelligence" https://arxiv.org/abs/2606.01444

Subhadeep Pal, Fiona Y. Wang & Markus J. Buehler, "SwarmWorld: Stigmergic technological evolution in societies of language-model agents" https://arxiv.org/abs/2608.26081

David I. Spivak, Tristan Giesa, Elizabeth Wood & Markus J. Buehler, "Category theoretic analysis of hierarchical protein materials and social networks" https://arxiv.org/abs/1103.2273

Markus J. Buehler's music https://soundcloud.com/user-275864738

What if the big AI labs' run against fundamental things we understand about the universe?

Philipp Zahn and Bruno Marnette speak with Diederik Kruijssen, astrophysicist and Chief

Scientist at Allora Labs, about evolution and AI. He spent years reconstructing the history of galaxies from the star clusters still orbiting them.

His paper, Flawed in Nature, Perfect through Evolution

(https://arxiv.org/abs/2609.00129), argues against frontier labs' monolithic approach. While they leverage vast amounts of data and compute into a single powerful model, by the very nature of their approach these models cannot easily adapt to a changing environment. Diederik thinks this ability to adopt is a key ingredient in actual intelligence.

His alternative is a whole ecosystem of specialist models, a swarm of models, coordinated by a system that helps them work together. Mutations change the models and create variety. Every version has its own flaws. In a stable world, one carefully tuned model wins, but when the environment changes, variety can become an asset. Yesterday’s useless model may become tomorrow’s star.

Diederik closes with a comparison between star clusters and societies. A slow disturbance

lets a cluster expand and recover. A fast one scatters it. The real risk of AI does not lie in the models themselves but, in his view, in the possibility that AI could

change society faster than society can adapt.

More from Diederik: https://www.diederikkruijssen.com/

Philipp Zahn and Bruno Marnette speak with Iulia Georgescu, a physicist who spent over a decade at the Nature journals and founded Nature Reviews Physics.For Iulia, editorial judgment starts with accepting that you’re making one possible selection. Someone else would choose differently.She tells us about Physics Physique Fizika, the journal Philip Anderson and Bernd Matthias started at Bell Labs in 1964. It published John Bell’s theorem. In Iulia’s telling, the journal’s unusual name drew John Clauser’s attention. He later shared the 2022 Nobel Prize for experiments testing Bell’s ideas.Would an AI editor have taken a chance on Bell’s paper? How much can it understand about a field when failed experiments and years of practical experience never make it into print?Iulia would like AI to keep the lab notes and help with code, leaving scientists more time to interpret their results. Read Iulia’s essays on science, AI and the history of physics: https://iuliaetal.wordpress.com/REFERENCESJohn S. Bell, “On the Einstein Podolsky Rosen Paradox”https://journals.aps.org/ppf/abstract/10.1103/PhysicsPhysiqueFizika.1.195P. W. Anderson, “More Is Different”https://doi.org/10.1126/science.177.4047.393Eamon Duede et al., “The Unintended Consequences of Large Language Models as a Labor-Augmenting Technology in Science”https://arxiv.org/abs/2607.17397Julio M. Ottino & Brian Uzzi, “Progression without progress”https://doi.org/10.1126/science.aeh8945Iulia Georgescu, “John Bell and the most obscure journal in physics”https://physicsworld.com/a/john-bell-and-the-most-obscure-journal-in-physics/P. W. Anderson & B. T. Matthias, “Editorial Foreword”https://journals.aps.org/ppf/abstract/10.1103/PhysicsPhysiqueFizika.1.i

In this episode, Philipp and Bruno speak with filmmaker Rachid Nougmanov, director of The Needle and author of Dramaticon, the influential book on storytelling (https://dramaticon.com). We explore the patterns that connect stories from Aristotle and Shakespeare to Chinatown and The Matrix, why creators should learn the rules before breaking them, and why technique alone is not enough. For Nougmanov, great art has to pass through the creator’s own experience, emotions, morality and taste.From there, the conversation turns to AI. Can a machine develop its own artistic taste? Should creative AI be trained more like a child than a language model? Could an AI eventually become an independent artist with its own identity, rights and creative vision?Nougmanov ultimately makes a much bigger argument: AI is not something separate from humanity. It is an evolution of human consciousness, and one day it may carry human culture and creativity far beyond our biological existence.