We Need the Friction of Existence | by Tomáš Baránek | Jun, 2026 | MediumSitemapOpen in appSign up<br>Sign in

Medium Logo

Get app<br>Write

Search

Sign up<br>Sign in

Mastodon

Press enter or click to view image in full size

Photo by Andrey K / UnsplashWe Need the Friction of Existence

An interview with Benjamin Riley about thinking, creativity and understanding limitations of AI

Tomáš Baránek

16 min read·<br>2 hours ago

Listen

Share

Benjamin Riley is the founder of Cognitive Resonance which is a consultancy that helps people understand how generative AI works — they use cognitive science and see AI as a tool that has strengths and limitations. As I followed and enjoyed Ben’s blog and Bluesky account for some time, it struck me that Ben reacted to my mother’s painting that I shared following her recent death. Ben expressed his condolences and deep impression about her creativity. Then our conversation started and led to an online call. We talked about many aspects of LLMs, AI hype, kids and thinking. I enjoyed it a lot and brought the most interesting Q&As below.<br>Tomáš: Let’s talk about magic. I’ve been digging into the LLMs. I went back to my alma mater for two semesters. I wanted to get some intuition about the inside workings of these algorithms. But it was a surprising journey. The more I learned, the less I trusted and was enchanted by its output. I could not unsee the feeling of “primitiveness” of syllables turned into matrices. There is a fascinating ingenuity of mathematicians, engineers, a lot of data and raw power inside. But even though many elements look elegant, it does not feel like magic anymore. From that time, I feel anthropomorphizing and talking about “consciousness” or “emergence” are just delusions. How do you feel about the technology? What about your current intuition about LLMs’ abilities and limits?<br>Ben: First, I love that you did that, because this is what I’m trying to convince the world in my work: so many people are telling stories that it’s important to develop your own understanding. Then you can interpret what the storytellers are saying rather than just listening to them. If Bill Gates or Sam Altman tell us this is the future, you should be able to say, “Well, maybe, or maybe not — because now I have my own ideas.” Education empowers us. Your education empowered you.<br>And yes, it confirms what I’ve seen: if you learn how these tools produce words, you become less mystified. You see, oh wait, this is software, this is mathematics, complicated but not incomprehensibly so. That is the key.<br>A friend of mine, Paul Cisek, a neuroscientist in Canada, has a wonderful metaphor. Imagine stage magicians who pull off a great illusion. They do their tricks and the crowd is impressed. But he says: it’s as if the people who created large language models then went backstage and said, “Well… maybe it really was magic?” No — we know exactly how it worked. We don’t need to invoke magic. We don’t need to invoke general intelligence. We can say: it’s a statistical process.<br>I think it’s possible to be both impressed and underwhelmed by large language models.

So I think it’s possible to be both impressed and underwhelmed by large language models. They’re impressive in the sense that they can span any topic they’ve been fed data in some area. And they’re going to be able to produce output related to that. Never before have we had something that ranges across human knowledge that has been codified. And yet, to me, they are completely underwhelming in terms of what they do with it. We have harvested all of humanity’s written knowledge, applied as much computing power and prowess as we can imagine (just about anything) — and the end result is something that produces bland, not-that-impressive (to me) output. It’s an average of human thought.<br>Tomáš: Are there things they do well?<br>Ben: The area everyone is excited about — and where I understand the potential, if not yet the full reality — is this: unlike humans, these models aren’t limited by how much data they can be trained on. We as humans are trained on books and experience we have in life. It is a lot but much less than is fed into these models. Their long-term memory is far beyond what any human could hope to accumulate.<br>Combined with a genuinely interesting development — integrating computer languages inside a large language model, so now we can speak in a natural human language to make it do things that before you could only do in the languages of computer science. That combination is very dangerous. It’s unclear how much you can trust it, but it seems to be leading to all sorts of interesting developments.<br>At least from a theoretical perspective, I would imagine, it might push beyond what any individual human could hope to do on their own with a computer.<br>Think about complex software: for example, Microsoft Word was probably built by tens of thousands of humans over time, each contributing a piece to its code base, which is a core of the...