Okay, so first he talks a bit about biologic systems...
I will not deny, it is cute to see a Lisper discovering subsymbolic methods, but he is romanticizing it something fierce. While biological evolution can, under the right circumstances, result in a level of robustness, it inherently tends to produce horrible and unmaintainable spaghetti systems with only incidental organization.
He also brushes aside security issues by referring to bacteria, which misses the point. Bacteria are not trying to kill you at all; they do not constitute a determined attacker. Ultimately the immune system provides robustness, not security.
If you really wanted to kill a human, you could easily make a killer virus or bacteria to do so, or you could kill in a thousand other way.
Then he loses the biology and spends the rest of the talk doing little programming tricks. Automatic derivation, deduction systems, truth maintenance systems. Okay, fine, that's neat, but it looks like very classical Lisp thinking. Just playing around with symbolic stuff.
How will any of that help users do what they want to do? What practical system could you compose out of those pieces? If you have a barrel-load of picoprocessors inside your concrete, it is presumably not to distributedly compute square roots, fibs and the height of the building by barometer.
But I lie, because a very relevant application of intelligent concrete could be to calculate load and detect early signs of structural weakness... This would probably involve distributed, robust, iterative calculation of the kind he at one point discusses, but, you do not need a mythical language of the future to do that. You would figure out the formulas, and then you would write it in C, or assembly, or whatever format lies close to the nature of the hardware.
But, he points out, he is not advocating that specific approach, he's advocating any approach which has a chance of freeing us from the current state of affairs.
I am led to think he is asking for the Silver Bullet.
There is essential complexity in programming, argued Fred Brooks, details that you cannot escape because they matter. Do not sweat the details, says the Sussman, but he does not know what details we should not sweat. When we throw away our blindfold of architecture, what will we see? He does not tell.
This video was by no means a horrible watch, but I feel that he failed to fully address the topic of the talk.