Oh no!
11.03.2026 09:29
👍 0
🔁 0
💬 1
📌 0
Oh no!
Outsourcing does work. What doesn't work is one-shot outsourcing where you tell somebody (person or organisation) what to do and leave them to it. One-shot LLM solutions don't work either, but iterative development of solutions does. The challenges of scale are universal.
I don’t think LLMs allow us to program in natural language. I disagree that what’s happening now is just another rung on the abstraction ladder. Instead, we’re directing a programmer in natural language, which is what we’ve always done.
Indeed. It’s very much being able to define unambiguously what success looks like. That’s one reason it’s a force multiplier. If you can already think clearly and can define the objective, it’s fantastic. If you’re muddled or unclear, then slop-in, slop-out.
I don't think anything gets upgraded in space. It becomes obsolete in space with five years and disposed of in the upper atmosphere.
I've just spend three days modelling head-rejection from space-based data-centres. There are challenges for sure, but I think they're surmountable. The chips themselves are basically disposable, and if a few die or get taken out by micrometeorites, is that a problem? You just design for resilience.
An all too common misunderstanding. Runaway feedback and stabilising feedback would be better terms to use with people who've never learned about feedback. They describe behaviour not mechanism or polarity.
40 years of experience. I don’t deny that there are enormous unanswered questions about the pipeline of programmers and developer education. What we do can be taught, but is will take a decade or more to figure out how. System qualities and constraints are more important than ever.
We've departed the syntax era of software development (how) and entered the specification era (what). I written hardly any syntax for six months now. I've written more specification in English than ever before.
And yes, I've built very non-trivial apps from scratch and performed deep surgery on apps that were coded in the before times. I've gone out of my way to push these tools hard. They're a force multiplier, so if you know how to build software with discipline they're incredible.
Rather than writing about on social media, Allen, use your time to build something "non-trivial" with these incredible tools. Use the skills I know you have, thoughtful specs, iterative development, TDD or at least test-first. You can get a _very_ long way; probably all the way for 80% of apps.
Interesting to come back to C++ after over 15 years. While I've been gone they've finally admitted that operating systems exist and provide enough useful OS integration out of the box to make it a half-reasonable platform.
This would be a very good week for anyone who has not already abandoned the Mercator projection for maps of the world to do so
Finally, finally getting off my butt and getting off all possible US goods and services. Some will just take longer than others! www.goeuropean.org
Thanks for the tip!
First US tech company/service dropped from my life, long overdue: Backblaze. Lets go for Norwegian company Jottcloud instead.
10 minutes work to change my rclone backup script over to use Jotta instead for backups, and hey - a waaay nicer UX in their web console, to boot 🙂🪄
The best use of American AI today would be to help Europe build its own sovereign technologies: operating systems, social media, communications, semiconductors, defence, and yes, more AI. While the subsidies are flowing, let them underwrite a future of genuine technological independence.
I think we're lucky to have grown up with much of the abstraction ladder, having a pretty good idea in principle (though not necessarily in detail) how computers work from the gate level up to what we use every day via many intermediate layers.
I worked in C++ from 1995 to 2010, then in Python from 2010 to 2025. I've been building complex native application in C++ again recently, almost entirely mediated by AI. It's hard to leave a 100x performance gain on the table when the dev- experience of interpreted versus native code is eliminated.
Yes, but precious little of what goes on in the computing industry is science, with or without AI. LLMs (and their weights) are hugely complex quasi-natural artefacts which are amenable to scientific study.
Well, at least it finally has chance to become a science.
One Earth radius of cycling in 2025 achieved. For the other geo-pedants out there, a distance slightly greater than that from the summit of Mount Chimborazo in Ecuador, which at 6384.4 km is the furthest point from the centre of the Earth. 🌍 🚴♂️
"what's really new about it?" 🤯
It's not another rung on the program representation abstraction ladder. It's a sideways hop into agent management, and likely on to a different and new abstraction ladder.
This misses what’s happening: a categorical shift from specifying program representations in ever higher-level languages, to delegation. I’m no longer expressing programs; I’m directing an agent that expresses and manipulates programs. The directions themselves are ephemeral.
Rent a mini excavator.
Neither is it what I said just now.
Fair, but I would argue that little to no financial benefit accrues for the vast majority of book authors. For those authors for whom that is the case, are LLMs a loss? It’s not zero sum, is it?
So far, yes.
Extending work by Harvard economist Stefanie Stantcheva and others, my analysis of polling by More in Common finds that in the US, UK, France and Germany zero-sum beliefs on the left (eg people only get rich by making others poor) and the right (eg immigrants succeed at the expense of the native-born) are related expressions of the same underlying worldview. Namely that there is only so much to go around and we must therefore use restrictions, exactions and preferential treatment to redress the balance between winners and losers.
On zero-sum thinking. www.ft.com/content/30a4...