felipeerias

Asimov’s laws of robotics were a literary device. The plot of his robot stories usually revolved around someone finding a way to get the robot to inadvertently break the laws.

You need to provide agentic tools with enough context about the project so they can find their way around. In Claude Code this is typically done via a CLAUDE.md document at the root of the codebase.

I work on Chromium and my experience improved immensely by using a detailed context document (~3,000 words) with all sorts of relevant information, from the software architecture and folder organisation to the C++ coding style.

(The first draft of that document was created by Claude itself from the project documentation.)

Perhaps you should consider adding “be more Dutch” to the system prompt.

(I’m serious, these things are so weird that it would probably work.)

If pensions depend exclusively on what each person has contributed during their working years, then lowering the retirement age might theoretically help young people by opening up more senior positions for them.

However, in many developed countries today’s pensions are mainly paid for by today’s workers, often through a combination of social contributions and taxes.

In Spain, for example, the social security has a large structural deficit that gets covered by large taxes on salaries, taking money from other social services, and sovereign debt.

In other words, young workers are forced to support retirees through means that make it much harder for them to build a future of their own.

This system is kept in place because pensioners are a huge 1-issue voting block that no political party wants to antagonise. I believe that this is somehow downstream of cultural changes that have made all of society more individualistic and self-centred, incapable to work together for a common future.

These regulations may end up creating a trap for European companies.

Essentially, the goal is to establish a series of thresholds that result in significantly more complex and onerous compliance requirements, for example when a model is trained past a certain scale.

Burgeoning EU companies would be reluctant to cross any one of those thresholds and have to deal with sharply increased regulatory risks.

On the other hand, large corporations in the US or China are currently benefiting from a Darwinian ecosystem at home that allows them to evolve their frontier models at breakneck speed.

Those non-EU companies will then be able to enter the EU market with far more polished AI-based products and far deeper pockets to face any regulations.

That it is very likely not going to work as advertised, and might even backfire.

The EU AI regulation establishes complex rules and requirements for models trained above 10^25 FLOPS. Mistral is currently the only European company operating at that scale, and they are also asking for a pause before these rules go into effect.

The experience with other industries like cars (specially EV) shows that the ability of EU regulators to shape global and home markets is a lot more limited than they like to think.

Translations are a very subjective matter because the emotional punch of a story is far greater when it is speaking to you in your mother tongue.

Shakespeare is perhaps impossible to translate properly to Spanish, just like Don Quijote to English, and yet we keep doing it because even the small glimpse afforded by the translation gives you an idea of the greatness behind it.

Funnily, I’ve always found the Spanish translation of the Lord of the Rings significantly more readable than the original, perhaps because Tolkien went out of his way to write in an old form of English that is a bit too distant for me. Or maybe it is because I read the story in my youth and re-reading it is a way to recapture some of the wonder that I felt then.

I've watched it with my kid several times. The characters are great, the visual effects hold on, and overall it is still a tremendous adventure movie.

The missing bit is culture: the concepts, expectations, practices, attitudes… that are evolved over time by a human group and which each one of us has picked up throughout our lifetimes, both implicitly and explicitly.

LLMs are great at predicting and navigating human culture, at least the subset that can be captured in their training sets.

The ways in which we interact with other people are culturally mediated. LLMs are not people, but they can simulate that culturally-mediated communication well enough that we find it easy to anthropomorphise them.

Presenting LLMs with a dramatic scenario is a typical way to test their alignment.

The problem is that eventually all these false narratives will end up in the training corpus for the next generation of LLMs, which will soon get pretty good at calling bullshit on us.

Incidentally, in that same training corpus there are also lots of stories where bad guys mislead and take advantage of capable but naive protagonists…

It can be enlightening to look at well-known Biblical stories from a different point of view. For example, you might be familiar with the Book of Job: the story of a good man who loses everything—his wealth, children, and health—despite doing nothing wrong.

At the same time, a koan is a paradoxical question or story used in Zen Buddhism to break logical thinking and provoke a deeper understanding beyond words and concepts.

Even though both are separated by an immense span of time and distance, the story of Job functions surprisingly well like a koan because it presents an apparently impossible paradox: if God is just, why do bad things happen to a good person?

This question has no logical answer, and when God finally speaks to Job, He doesn't explain the suffering but instead shows Job the overwhelming mystery of creation. Job's response isn't intellectual understanding but a transformative experience.

Just like a koan, Job's story seeks to break through logical thinking to produce spiritual enlightenment.

If someone asked you about the usages of a particular element in a codebase, you would probably give a more accurate answer if you were able to use a code search tool rather than reading every source file from top to bottom.

For that kind of tasks (and there are many of those!), I don't see why you would expect something fundamentally different in the case of LLMs.

People are having three different conversations at the same time:

– the concrete causes of this specific blackout; – how the existing grid is not prepared to deal with the current energy mix; – the energy policy of the past decades, from the nuclear moratorium in the 80s to the large subsidies for renewable generation of the past couple decades.

A person's strong opinion on any one of these issues will inevitably influence their opinion on the others.

If we did that, would we be really replicating what nature does, or would we be just simulating it?

Human intelligence and consciousness are embodied. They are emerging features of complex biological systems that evolved over thousands and millions of years. The desirable intelligent behaviours that we seek to replicate are exhibited by those same biological systems only after decades of growth and training.

We can only hope to simulate these processes, not replicate them exactly. And the problem with such a simulation is that we have no idea if the stuff that we are necessarily leaving out is actually essential to the outcome that we seek.

The article hints at that when it describes vibe coding as “fancy UX”, but fails to connect the dots.

Essentially, we now have a system that can turn a simple problem description into an interactive tool for that problem. Even if it still does so very imperfectly, it’s easy see already the beginnings of a powerful and empowering new paradigm.

The problem with trying to reason everything from first principles is that most things didn’t actually came about that way.

Both our biology and other complex human affairs like societies and cultures evolved organically over long periods of time, responding to their environments and their competitors, building bit by bit, sometimes with an explicit goal but often without one.

One can learn a lot from unicellular organisms, but won’t probably be able to reason from them all the way to an elephant. At best, if we are lucky, we can reason back from the elephant.

It is hard to convey just how important it is to be able to provide additional context, ask follow-up questions, and reason about the text.

I live in Japan. Almost every day I find myself asking things like “what does X mean in this specific setting?” or “how do I tell Y to that specific person via this specific medium?”.

Much of this can be further automated via custom instructions, so that e.g. the LMM knows that text in a particular language should be automatically translated and explained.

To prevent an attack where someone steals your username and password, triggers the 2-factor notification, and waits for you to accept it. This can be automated and repeated until you eventually click the wrong button for one reason or another.

By requesting a short-lived code, attackers now need to communicate with you at the same time of the attack and somehow convince you to give them that code. Much harder.

Planning is indeed a very underrated use case.

One of my most productive uses of LLMs was when designing a pipeline from server-side data to the user-facing UI that displays it.

I was able to define the JSON structure and content, the parsing, the internal representation, and the UI that the user sees, simultaneously. It was very powerful to tweak something at either end and see that change propagate forwards and backwards. I was able to hone in on a good solution much faster that it would have been the case otherwise.

The emergence of AI tools and closed platforms will reduce the importance of advertising on the open Web, and eventually of the open Web itself.

One possible way this plays out is that attention and investment move towards proprietary ecosystems, with large AI companies being able to secure exclusive access to closed information sources while everyone else is reduced to getting what they can from a dwindling open Web.

Another possibility might be that new standards allow interoperability between AI agents and open content providers, including microtransactions between them, and creating a new marketplace for information.

https://stratechery.com/2025/the-agentic-web-and-original-si...

Devs have a hard time seeing vibe coding as UI, but that is what it effectively is for the average user. Describe the problem, get an interactive tool to handle it.

The key of the teaching comes after that metaphor, when the disciples ask how one can be saved and Jesus answers that “with man this is impossible, but with God all things are possible.”

These words eventually became the foundation of Christian charity in an Ancient world that until then despised and rejected the weak and infirm.

Many of our modern day institutions trace their origin to that Christian charity, to those few enigmatic words.

(Peter Brown's “Through the Eye of a Needle” is a great book to know more about this process.)

Nothing is more valuable for an AI company than exclusive access to large amounts of human-generated data.

Both Claude 4 Sonnet and Opus fail this one, even with extended thinking enabled, and even with a follow-up request to double-check their answers:

“What is heavier, 20 pounds of lead or 20 feathers?”

I have been trying to use Anki for years. Every time it is the same story: I keep it up for a few months until I miss a couple days, then due cards accumulate far beyond what can be reasonably managed, and I end up spending more time trying to fix the app than actually learning anything.

Recently I had to define a remote data structure, specify the API to request it, implement parsing and storage, and show it to the user.

Claude was able to handle all of these tasks simultaneously, so I could see how small changes at either end would impact the intermediate layers. I iterated on many ideas until I settled on the best overall solution for my use case.

Being able to iterate like that through several layers of complexity was eye-opening. It made me more productive while giving me a better understanding of how the different components fit together.

If we survive far into the future, we will learn a lot more about the structure and evolution of the Universe. It might be that the questions that our scientists can ask now will turn out to be trivial or meaningless to our descendants. Perhaps the Universe is far stranger than we can imagine.

That is because it is thinking out loud. Producing tokens is how it thinks.

One sci-fi example could be based on the replicators from Star Trek, who are able to synthesize any meals on demand.

It is not hard to imagine a "cooking robot" as a black box that — given the appropriate ingredients — would cook any dish for you. Press a button, say what you want, and out it comes.

Internally, the machine would need to perform lots of tasks that we usually associate with intelligence, from managing ingredients and planning cooking steps, to fine-grained perception and manipulation of the food as it is cooking. But it would not be conscious in any real way. Order comes in, dish comes out.

Would we use "intelligent" to describe such a machine? Or "magic"?