Diego Basch

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.

@amartino Lo que sí es cierto es que por default los modelos escriben muy sencillo, para las masas.

Y yo solo quería saber la opinión de @dbasch sobre esto porque no creo que sea una observación correcta

@allisonology I don't think you can discover any of those things, they are all fleeting. What you need, what you want and who you are changes all the time. Perhaps who you are changes less. But I don't think you discover who you are, it's more of a building process.

Out of the things that one can discover, which is the hardest to do?

You don’t understand IQ. It’s a relative measure. 130 is roughly the highest 2% of the population, always.

Assuming this is not a troll, human alignment has been talked about for ages. Have you not read Nudge by Thaler, for example? We already know how to make people save for retirement, that's the specific example they use in the book. If you want, I can connect this to how governments use it in practice. Or to engagement tactics used by AI companies.

Counterpoint: no, it wont.

It's about to crash because someone on the internet says so? For one, IQ is relative. If the population becomes dumber, 130 means something else but the absolute number of people with 130 IQ tracks the population (it's some fixed percentage). But if you mean that the number of people who would score 130 on a test today will go down, this is just random speculation backed by nothing. You could make a chart showing the opposite and say "the number of smart people in the world is about to grow tremendously."

The total number of smart people in the world has just peaked. And now it's about to crash.

We need to turn this frown upside down.

I've been doing some machine learning work lately, and because it's 2026 I'm using Claude Code. Do you think it has made me 10x faster? No. It has made me more productive for sure, but how much? Read on. Well, ML (and software in general) is not a matter of cranking code as fast as you can. It's nice to be able to tell Claude "code this for me." But how do you know you're coding the right thing? You talk to the stakeholders. You show them what's happening. They take their time. They get back to you. What happens when you train? You use Azure, or GCP, or AWS. But everyone has free credits on one or the other, that run out. And for example when you move to GCP, you run into quota issues. You don't even get access to GPUs. You have to start from scratch, and prove to Google that you deserve access to an A100. Then your training run breaks, or it doesn't work out as planned. You have to go back to the drawing board. Maybe you need better or more data. How do you get it? You need to make deals with those who own it. Perhaps more importantly, if you produce more lines of code you are inevitably producing more bugs. And the apparent productivity of putting code out there faster is debt, and you pay it later when the bugs surface. Unless you're extremely diligent in scrutinizing everything that Claude Code does under your command, which nobody does because it would be faster to just write it yourself. So when I factor all that in, I'd say I'm 10 to 20% more productive than I use to be. And that's only because I believe I'm significantly better than average at knowing when to use Claude, and when to tell it to get out of the way. Conclusion: Software is still for humans, even if it is less *by* humans than it used to. And humans aren't faster. We still change our requirements, we still know it when we see it, we are still overwhelmed by the extra complexity of generated code (more features, more bugs). The code was never the bottleneck, and we still have not unlocked a new era of software productivity. Don't take my word for it, see for yourself.

I think the world has never been more unpredictable. Trying to predict 2036 is like trying to predict 1980 in 1950.

Earlier today I had to split 50k reports into sections. I thought "I know! I'll use regular expressions." But it didn't work very well. So I thought "wait, this is 2026! I'll use an LLM!" But I didn't want to send my confidential data to OpenAI. So I generated some fake data and asked it to make me a few-shot prompt for a smaller LLM. I ended up doing it with Qwen 2.5 3B quantized to four bits, served locally by vLLM. It's only 2.7GB, so I did this on my gaming laptop with an 8GB gpu. I didn't even need the few-shot examples, one example was enough. I expected I would need a significantly better model! This was once a really annoying natural language parsing problem, and now we have the technology to solve it trivially with a toy computer. LLMs are amazing when used properly, and entity extraction is one area where they shine.

@PabloSabbatella Even for work it's unsafe. I don't want it looking into my work email, presentations, corporate databases, nothing at all. I don't want it sending emails or taking any unsupervised actions in my name that have consequences in the outside world.

He says he was using it for work but then it accessed his personal messages and files. That's why he must have separate environments. I think this is an issue for most companies (we are having lots of people asking us about this), that will eventually be solved. In the meanwhile: hire someone to do proper setup in your company, or wait until it's more "tested"

"OpenClaw consultants have since expanded their services to removal, with one 21-year-old consultant saying he was already fielding more requests for deletion than installation." I'll tell you in simple words when you should use OpenClaw: never.

@PabloSabbatella yeah, you could not give it access to anything private. But then why do you even want it? Hence never.

@dbasch Maybe they should first know how to use them in a safe way.

Está haciendo exactamente lo que tiene que hacer. No sabés las cagadas que te podés mandar por no respetar el robots.txt, aunque sea el tuyo propio. Te lo dice uno que escribió crawlers y tiene mil anécdotas. Por ejemplo, te pueden hacer que te DOSees vos mismo (ejercicio para el lector).

@JulianD @claudeai Es su propio sitio … en el mismo Proyecto!

Que frustrante algunas cosas de @claudeai a veces

@DamianCatanzaro Que yo recuerde, nunca hubo healthy life en el ambiente tech de San Francisco. En las oficinas teníamos degustación de vinos, los pibes hacían fiestas con todo tipo de sustancias. Si se proyectó esa imagen, era marketing.

Algo que me da gracia de SF es que pasaron de ✨ healthy life ✨ a necesito drogarme porque estoy siendo poco productivo sin escalas.

Si hasta el presidente de YC está pensando que es algo normal para una persona hacer 16 mil líneas de código por día lamento decir que necesitamos que nos la peguemos contra la pared fuertisimo a ver si vuelven a tener por lo menos la mitad de sentido común que había hace unos años. Después leí el segundo tweet donde dice que básicamente se está drogando para mantenerse despierto y hacer más código, realmente no sé si es verdad o bait lo que publica (creería que si siendo el presidente de YC), lo que si se es que si la gente de SF sigue por esta línea van a quedar todos esquizofrenicos de por vida.

@gptcrosa 16K LOC por día no significa nada. Es como los que posteaban diciendo "Cloné Reddit en 142 líneas de código (import reddit-clone era una de ellas). Decime concretamente qué estás haciendo más rápido porque generar LOC es fácil.

La diferencia entre quienes están en San Francisco y quienes no, es que, ante algo que suena descabellado, los primeros reaccionan con curiosidad en lugar de juzgar o descartarlo

Yeah, the Greeks were famous for their uplifting stories that led to the greatness of the Roman Empire. Unfortunately this horrendous Jesus character became popular, and led to the fall of the empire and the Middle Ages.

@juanicaffa eso es TaskRabbit (ahora parte de IKEA). Les fue muy bien, yo los usé varias veces en San Francisco.

Una idea obvia y maravillosa: una app o marketplace de plomeros, pintores, albañiles, carpinteros… muchos probaron pero por algún motivo el proyecto fracasa, porque es tan difícil de hacer funcionar?

In a few decades there will be headlines like this for AGI doom prophets.