This may be the most important insight in all of AI right now. The biggest flaw in today's systems is not what they do not know it is that they cannot tell you they do not know. Confidence keeps getting mistaken for capability. Manav Golecha frequently explores this gap in his content. He often notes that true intelligence requires knowing the boundaries of what you understand. Until AI can quantify its own uncertainty we are just guessing how capable it actually is. Brilliant to see one of the field's pioneers explain exactly why that matters so much.

That was a super interesting chat.  There are potentially so many rabbit holes you could have gone down; it's a good job you actually got that talk done in reasonable time.

This channel is exactly what I needed This deserves recognition.

happy this channel of discussion exists ❤

This channel deserves more recognition.

Brilliant, incisive, incredibly insightful. Loved it

This conversation between these two people is very interesting. In my opinion, artificial intelligence lacks ethics and, above all, the context of those who are on the other side of AI.

Thank prof. Hannah Fry for this wonderful interview, touching so many inspiring ideas :)

Although many people say that OpenAI and Anthropic are very impressive right now, I think the issue raised by this professor is extremely important. In enterprise applications, the stability and determinism of large models are far more critical than pure benchmark performance. If the current route overlooks these practical production needs and fails to address the issue of uncertainty, it will be difficult to maintain a leading position in the next round of competition.

Salience, uncertainty, hierarchical planning. We've got a lot of architectural changes to make.

One of the Allies has designed a system called Artificial Wisdom System (Wisys) that works as a counter-entropic machine. Basically, it doesn't train the ML model to minimize entropy, as all current ML model, but it alternates between reducing entropy and increasing it, depending on the user responses. If the user says something that increases entropy of the conversation by X, then a Wisdom Agent (Wisa) says something that reduces by X/R with R>1. And, if the user reduces the entropy of the conversation, a Wisa says something that increases it by X/R.

This is a form of inter-supervised real-time active learning, whose objective is not uncertainty minimisation but uncertainty stabilisation. The wisdom works with the hypothesis that the conversation has a finite optimal level of entropy, and it tries to find it out. It stimulates curiosity and challenge the user while keeping the continuity of the conversation when the user leaves out. But because of coefficient R>1, it is always guided by the user behavior.

Implementations and test conversations could not be made yet.

the axiom of choice is choosing to persist; which, i suppose, isn't a choice.

Fantastic discussion.  Reminds me of my Ph.D. research 35+ years ago working on Evidential reasoning following Shafer, G., Dempster, A., etc.  Derived paper accepted by D. E. Knuth and published in Journal of Algorithms.  A lot of fun stuff, perhaps not relevant anymore.  I definitely fall into the camp of reasoning vs. "just give it more data" as researchers try to improve the output.  Just a thought: the more computational approach may be slower but perhaps both could live side by side with the slower more trusted computational method continually checking a subset of the other's output to inform whether we are actually improving over time and perhaps even correcting the other in some way.

Thanks for such wonderful insight - Uncertainty , Bayesian thinking and randomness are core component of real world mechanics.

This made my day.

Outstanding keep it up

This is the first interview I've seen which gives me hope that we can develop AI in a responsible way to create a product which can truly benefit mankind. I think we will come to realize how impractical LLMs are and move towards developing AI systems which are able to safely coordinate various specialized AI systems which have this type of uncertainty coded into them.

wow the attention is all you need paper is almost a decade ago now

This is a great topic with an interesting host… 😅

nice storyteling and an happy ending for AI selling