Diego Basch

If you want insights about AI, dad jokes, entrepreneur/investor stories, and tips about Buenos Aires or San Francisco, smash that follow button.

[Diego Basch @dbasch]

Imagine you have been looking at the Earth from space, for millions and millions of years. It always looks the same. All of a sudden it lights up, in a matter of decades. The blink of an eye. Could you have predicted that? Of course not.

Humans love to predict things. People imagined insane urban landscapes when cars started becoming popular (look at the Western US and its unwalkable cities), or a dystopic life in virtual reality when the internet was invented. Look up the word "teledildonics."

Now we live in one of the most unpredictable periods ever. We have a weird new technology, and the predictions are all over the place. It's pointless to predict whether AI is going to replace job X or job Y, whether it will kill us all (it's increasingly looking like it won't), whether jobs will disappear, or whether it will cure cancer. We're just going to have to ride it and see what happens.

When an expert predicts with arrogant certainty that "in 10 years AI will..." just stop reading. Rolling your eyes is optional.

[Diego Basch @dbasch]

@antoniogm You should come to the Buenos Aires Steak Week.

[Diego Basch @dbasch]

I disagree. It's easy to dunk on Dijkstra, but back in his day a "program" meant something else. Supercomputers had at most 1Mb. Programs were meant to perform a computation. They generally were not constantly running in loops with a user interface, connected to networks and operated by average people.

Today programs are tiny components of extremely complex systems. Proving a program correct means very little. For starters, what does "correct" mean? It means:
- Given the following invariants (e.g. inputs of a given type and in a given range), the program will produce outputs that conform to the following specification (this obviously implies that it won't crash).

But how do you even begin to prove the SYSTEM correct? First you have to define correctness for the system. How do you define correctness for Uber, or for ChatGPT, or for a bank?

Sure, you can pick features and define invariants. For example: "you don't get a pickup confirmation unless a driver has agreed to pick you up." But you cannot define an invariant that says "ChatGPT does not hallucinate" or even "The Uber app cannot be remotely exploited." And these are things that you would want to prove.

In order to prove something useful about a system, you have to have extremely strong constraints on what you want to prove. For example: "this software is incapable of ever commanding the hardware to deliver a radiation dose above 30 Gy. But you cannot prove "this software won't do anything life-threatening" because you cannot even define it. And the majority of the functionality of the systems we use all the time is like that.

[Eric S. Raymond @esrtweet]

The controversy over "You should be able to prove your program correct before you run it" is still live.

For several decades after Dijkstra's pronouncement, the conventional wisdom was that he was wrong. And, with the tools and the hardware we had available, that was difficult to dispute. Attempts to apply formal proof methods scaled very badly, seldom worked at all on programs of larger than toy size, and were ridiculously expensive.

However, in the future that may change. There has been a lot of theoretical progress towards practical proofs on programs of nontrivial size. They remain difficult to apply to common languages and toolchains.

If we get Dijkstra's future it will be because the transition costs of moving to languages on which proofs are practical fall to the point where moving is less than the continuing cost of errors. LLMs make language-to-language translation cheap and will help.

Two domains to watch are avionics and medical-device software. The transition, if and when it becomes practical, is likely to start there.

[Diego Basch @dbasch]

Image-enhancing AI on the iPhone is getting out of hand.

[Diego Basch @dbasch]

@tenobrus People keep throwing around the word "conscious" as if it meant "fat" or "green" when it's an unquantifiable subjective experience that means different things to different people. This trend needs to stop.

[Diego Basch @dbasch]

@robinhanson I assume that this is the result of propaganda by parties interested in slowing down the US, but it's perfectly possible that it's just a fashion trend. It's easy for people to distrust what they don't understand, like what still happens with nuclear energy in many countries.

[Diego Basch @dbasch]

"Conscious" is a terrible word choice for AI. For one, it is a fuzzy word with many meanings. "He woke up from a coma and is conscious" is not the same as "I didn't consider my son conscious until he started asking questions about the world."

In the second sense, "conscious" refers to self-awareness. We humans welcome children into the world of being self-aware. At first we treat them like animals, and they slowly become more like us. The problem is that none of us can know for sure that anyone else is self-aware in this sense. The only experience we have is our own. This is why the concept of solipsism exists. Every sufficiently intelligent person has at some point wondered: "what if I am the only conscious being in the universe, and everything else is a simulation? How could I tell?"

We quickly eradicate that thought because we don't want to live like that. I assume (and hope) that all of you are the same as me, and have a subjective experience of existence. It is absolutely clear that this experience doesn't need to be coupled with intelligence or language. I can experience things without words crossing my mind. I can describe it later with words, if asked. When I was showering earlier today, I wasn't thinking "I will rotate the handle until the water is sufficiently warm."

I am pretty sure that AI is not conscious in any of the above senses for a number of reasons. The most important one is that it doesn't have a continuous experience that follows a linear timeline. Every inference of a model has an arbitrary context, and it can be started and stopped at any arbitrary point. But what's more important is that AI is not embedded in beings. It's not an animal, it doesn't have feelings, it doesn't have a body. It's more like liquid thought that can be put in a bottle and consumed as necessary.

We confuse "consciousness" with the idea of creating a Frankenstein, a Robin Williams in Bicentennial Man. While in principle physics doesn't preclude this, we're far from that. Disembodied language is not like us, it's not a peer, it's not an individual.

AI has much more interesting properties to talk about, such as its capabilities. At this point, I consider that the consciousness discourse is just a marketing tactic on the part of AI labs.

[The Atlantic @TheAtlantic]

If we confuse generative AI’s ability to produce text with consciousness, we risk assigning moral responsibility to chatbots—and not to their makers, Ted Chiang argues. https://t.co/BlvYtM2WCQ

[Eric Weinstein @ericweinstein]

¿Algún físico teórico, geómetra, topólogo u otro «malviviente» de la ciencia tiene ganas de tomar una cerveza cerca de Palermo-SoHo y hablar de lo nuestro —en inglés y en «español Tarzán»—? Avísenme. ¡Gracias!

[Diego Basch @dbasch]

@paulg Everything is different, it's a different board game. The starting line for startups looks like the Boston Marathon instead of a local 5k. The S/N ratio for anyone doing business with a startup has never been this low. Social media wasn't a factor back then.

[Diego Basch @dbasch]

@tszzl Why would RSI participate in human markets? It would have its own production pipeline. It may have to pay human workers in some instances, but it definitely would not be buying compute from a cloud provider.

[Diego Basch @dbasch]

I think they trust members of Congress too much.

[Paul Graham @paulg]

Republicans don't just admire different people from Democrats. They have lower opinions of almost everyone.

[Diego Basch @dbasch]

Better at writing implies that it's more effective at accomplishing something via writing. It already is better at writing code, at least way better than I am. And obviously agents are already better at writing for each other.

If we assume the audience is humans: It's not better at humor, or at creating anything that's really entertaining and pleasurable to read. The second seems easier: it could find the perfect simile, the perfect metaphor, great illustrations of a point, unexpected twists, uncharted narrative territory. It seems to me that the space of writing is more constrained than the space of solving problems in real life. It's make-believe vs. actually do things.

[Diego Basch @dbasch]

Hay gente para todo. Un amigo se compró un campo en el medio de la nada, y se armó una casita completamente aislada del mundo. A mí personalmente no me gusta pasar mucho tiempo en mi departamento, prefiero despertarme y salir. Eso es demasiado chico, pero prefiero vivir en un lugar que no me haga sentir demasiado cómodo en casa. Sobre todo en Buenos Aires que no es Saskatchewan y siempre se puede salir.

Si se equivocan y nadie los quiere, el mercado se los dirá. No veo el problema.

[Diego Basch @dbasch]

The UK ETA system has been down for 72 hours. I had to cancel tickets from SF to Spain and rebook via Zurich, and @British_Airways doesn't want to give us a refund. They never once mention ETA in the confirmation email, which ChatGPT finds "surprising." It reminds me of...

[Diego Basch @dbasch]

@amartino @profplum99 https://t.co/7bseuxDqQp

[Diego Basch @dbasch]

@jamescham Always is too strong. But most of the time, humans prefer contact with humans. Go to SFO, see how many people stand in line at Peet's, compare with CafeX. CafeX is better and cheaper too.

[Diego Basch @dbasch]

This is how I see it too. To me one of the main issues is: since the dawn of humanity, there has been an implicit assumption: human labor is necessary and has value. At a minimum, we had to move and build things. There was endless demand for this. The need for actual physical output from humans is almost gone. This is in part why higher education became so valued in the second half of the 20th century.

Now we are at a point in which it's not clear that having a brain means capitalism has a use for it. This could send our society in many directions that are hard to predict. https://t.co/VowXdUM5ar

[Diego Basch @dbasch]

I believe most people of college age in the US should not go to college. Exceptions:

1) You can afford going to a top 20 college without getting into (significant) debt.

2) You get a full ride scolarship at a decent college.

For everyone else, it's a luxury. You have to weigh the expense and opportunity cost against everything you could do with the money and the time. You will end up with a dubious credential and knowledge that you can acquire for free. In that time you could have attempted to build a business, lived/worked abroad, developed expertise in fields that colleges cannot fathom teaching and that will be important four years from now.

What college gives you is an experience, connections, a social life and a structure. You should really think about how much you want those things at that price, as opposed to the alternatives. And remember that you can try something else for one or two years and still fall back on college.

[Diego Basch @dbasch]

Taunting a VC is terrible advice. They are not your enemies, they are the buyers of an investment that you are offering. Just focus on selling your investment.

[Mert Deveci @gm_mertd]

Next time a VC tells you your company is not VC-scale, send them this

[Diego Basch @dbasch]

A windfall is not the same as success. Random employee at AI company having X million in the bank is not success. It's just not having to worry about income.

In my view, success is much more than that. It's about being acknowledged by your peers about having done something of note, and continuing to do it consistently. You cannot rest on your laurels.

[Diego Basch @dbasch]

@soncharm @robinhanson If I'd have to summarize it, it's "use evidence and careful reasoning to do as much good as possible with the resources you have." You get to define what good is, and a measure of the amount of good per dollar spent. The definition of good is very personal.

[Diego Basch @dbasch]

I don't particularly love the idea of EA, but the lightest incarnation of it doesn't prescribe a function. It simply says "make your own function and deploy your resource according to some metric instead of emotionally." It's hard to argue against that. The typical functions chosen by Bay Area people don't resonate at all with me. And I'm not very convinced that I could be effectively altruistic for the functions that matter to me. It's a nice sentiment though.