I think the AI obsoleting copyrights is more of a problem for non-fiction, because nobody is gonna pay the primary source. So it's a big question how will remuneration work in the world where AI is effectively an enclosed commons.

I had the same experience with Debian unstable and tetex, cca. 2003. Don't upgrade your production system, a good life lesson.

So, OK, Steve Jobs had a life lesson.. but why is comparing Jensen to Jobs stupid? The article doesn't answer this question.

I think maybe both of you are correct, in a sense - the population growth provides means (labor, competition, consumption) for increasingly extractive and unequal society.

Innovation relies on bigger groups, which then allows a smaller portion of the population to do more.

While I tend to agree with you, I find the Japanese are incredibly innovative and they are not growing. Maybe I should look at the rate of innovation and see if that has stalled out with population growth?

Which major disruptive innovations have come out of Japan lately?

Not major disruptive innovations, but the gadgets that do all kinds of little quality of life things. They have innovative solutions to so many of the small things in life. And a society where you can invent and bring to market quirky little products that do cool stuff.

Perhaps.

To me it feels like their most rapid period of innovation was post World War II as the population was growing very rapidly.

I have heard this before...in what sense would adding population result in more innovation rather than having a more supportive environment for the people we already have (healthcare, education, financing, etc).

Population growth drives the competition for jobs, which encourages education as a means of achieving social status and a higher salary, which enables the research that produces knowledge. Highly unequal wealth distributions then enable wealthy individuals to spend vast amounts on moonshot projects that turn knowledge into technology. These things still happen without the population growth or inequality, just slower than when the pressure to advance is high. People without that pressure will tend to make more conservative choices about what to study or research, make fewer irrational investments, and spend less time working (because wages are higher when you're not in competition for jobs).

Simply more nodes for experiences to bounce through, so long as they have ways to communicate enough to copy, recombine, and incrementally improve on their activity. The more people copying, recombining, and improving on stuff, the more innovation.

Often easier to consider this in terms of competitive sports, though you could check seshatdb or something to see if it holds at scale. For example, YouTube really changed a lot with sports.

You figure our society is "increasingly extractive and unequal"? That may be true relative to post-WW2 America, but on a larger scale--either historically or geographically--can you really hold that modern society is more extractive and unequal than slave societies, serfdoms, and empires of the past?

Honestly, population growth has, if anything, correlated to increasingly equal and egalitarian societies.

  > has, if anything, correlated

This phrasing is delicate. Correlated, but likely not caused by nor a cause of. Just two things that happened in parallel.

To be clear, I am agreeing with you.

Wealth inequality is largely irrelevant outside of limited goods like real estate. The poorest today live lives that much more closely resemble the lives of the wealthiest of today than was the case even 100 years ago. Access (to resources, shelter, entertainment, etc) inequality is far lower than in the past. Social inequality is lower than ever before. 50-60 years ago the US was still de jure racially segregated in many places.

Regarding formally verified OSes, IIRC an early version of CICS (what we would call application server today) has been formally verified in the 1980s.

I don't think there is a hard dichotomy (and I also think the original article is mostly wrong).

The software is always interpreted in the real world, and you can always convert the software plus the interpreter into a mathematical proof (via Curry-Howard isomorphism) that given input gives a given output. It is mathematically equivalent to running a program, up to the reliability of the interpreter and correctness of the mathematical model of the interpreter.

What might be a bit harder, to show, that a set of inputs causes a set of outputs. But that's just partial interpretation and depending on properties, in practical programs quite doable.

And frankly, existence of LLMs shows that you can quite easily prove most program properties. When LLM creates a program it translates a program description in natural language together with formal version of natural language (which it has from training) into a computer program. In other words, LLM doesn't work with the real world, but with a model of the real world. Although LLM's world model is not sound (in the sense that logical fragment of its reasoning has not been proven sound), many people are mostly happy with the results. This shows that LLM can preserve properties between the description of the program in natural language and the world model embedded in LLM (these often include implicit desirable properties), and the resulting source code to be executed. And if you're able to preserve the properties, the actual sound proofs transfer by construction, too (up to the few cases where LLM slips and reasons unsoundly, but AI labs are trying hard to reduce these cases to minimum).

Yes, moreover - nobody in practice creates programs that are close to be undecidable, just like no engineer uses nonmeasurable functions, they just assume everything is nicely integrable. (It might be close to true for things like cryptographic functions or pseudorandom generators, but even about these we can usually prove the things we need to know.)

I read the book years ago, and while I think it's interesting, today I think it's better to learn lambda calculus directly, without the Lisp "gimmick".

> I think it's better to learn lambda calculus directly, without the Lisp "gimmick".

The book isn't about teaching lambda calculus, though. It's covering things you can do in Common Lisp, specifically around macros.

And around lexical binding with functions as first class values. That to me is the lesson that is possible in other non-CL environments. E.g. JavaScript.

"[It's better to] learn type theory, category theory, lambda calculus, logic, etc." I have heard a similar notion for years. And I tried. Oh boy, I tried. Grokking Haskell felt like a Sisyphean grind. And then I tried Lisp. I dunno what happened to my brain, but the sheer pragmatism and plain simplicity of it suddenly made all the things that before felt so difficult to grok clearer than Lego-set instructions. Recursion, with its small intricacies that confused me for years, made me realize how simple it actually was, I just wasn't seeing it the way it actually is. Y Combinators, Folds, Monads, Applicatives, Laziness, and many other things.

I wish I had found Lisp many years sooner. That simplified Haskell, Rust and Idris for me. The downside is that it turned out to be vastly pragmatic, and for personal stuff I'm on Lisp. Even though most things at work are not.

So, maybe you should honestly self-reflect: had that book not happened in your life before, would you even consider understanding LC today, unadulterated by "gimmicks"?

I think you're making a good point, perhaps it was necessary in my path (which was roughly C++ -> Python -> Common Lisp -> Forth -> Haskell -> soon Lean).

But perhaps we're teaching lambda calculus wrong. Let and lambda are essential building blocks of it, and Lisp is a sort of special case.

I don't think it has to be hard - To Mock a Mockingbird is just as enjoyable as LoL. It's just difficult to see significance of LC for programming, I think, because people do not think of lambda calculi as programming languages.

I agree but the reality is most authors want to use their OSS creations at work, which usually means they will make a compromise with capitalists who reject copyleft licenses.

Many human problems could be solved if humans were less willing to compromise with greedy people, but unfortunately authors are too nice in general to do that.

I predict the opposite: Humans will collectively create something like an LLM, but curated. That will be next level for Wikipedia-like project.

Think about it - LLM is just a computer program that predicts a next word. So it has a very detailed, formal description of each concept it can represent in its residual vector.

I think humans will understand (perhaps with AI help) these formal descriptions and put them into a giant database.

You say, building ontologies with a specific intent will actually lead to intelligence at a similar level of current AI?

This would be a severe blow to the bitter lesson, and generally vindicate the AI researchers of the 80s that put a bet on symbolic processing

Well, the ontologies are there inside LLMs, humans just don't know them. There is nothing that contradicts with bitter lesson to what I said.