ElFitz

> Many website will request my personal physical address for trivial matters like billing or delivery.

Some will even require it for no actual reason at all.

Do I need to give my living address when I buy a sandwich? Then why would I need to when buying an online service?

Similarly, fast foods nearly all have these automated kiosques. They don’t need any info. So why do they require an email address when ordering to the table through the app, while in the restaurant?

They don’t need them. They just demand them because they can and everyone online is used to giving them without a second thought.

I can’t wait for personal data to become digital radioactive waste.

In "Language Models are Unsupervised Multitask Learners"[0]. Not sure whether it’s "the" GPT-2 paper.

3.7 Translation

> Performance on this task was surprising to us, since we deliberately removed non-English webpages from WebText as a filtering step. In order to con- firm this, we ran a byte-level language detector2 on WebText which detected only 10MB of data in the French language […]

[0]: https://cdn.openai.com/better-language-models/language_model...

> LLMs can only reconstitute things in its training data

Such as a 4D raytracing engine in Metal? Or integrating APIs for features first released months after their knowledge cut-off date?

LLMs have shown an ability to transfer "knowledge" and capabilities across domains, languages, and use-cases outside their training data.

Case in point: GPT-2 "learning" to translate English to French and vice versa despite non-English examples having been voluntarily (and almost entirely) removed from the dataset.

I still remember learning about habeas corpus. And loved Terry Pratchett’s take on it.

Some people even had some fun de-minifying JS and disassembling binaries. Successfully.

First one was Stable Diffusion. Especially the image to image, and the first gos people had at making videos with it.

Second one was trying to bootstrap what would come to be called a "harness", back in 2023, initially serving as the go between between api calls and file edits, feeding back the logs and gradually stepping back as step by step the llm bootstrapped the cli.

And finally, using Claude or codex to do ops work. Diagnosing issues on my machine, provisioning servers and VMs via ssh, debugging them, all on its own.

Agreed. Some have even trained neural networks made of actual biological neurons to play doom[0]. Brain cells! Doing smart things!

I still wouldn’t argue that this brain in a Petri dish is, in any way, conscious. Despite it sharing the exact same substrate as everyone around me.

[0]: https://www.theguardian.com/games/2026/mar/16/petri-dish-bra...

How are decompiling code or making a design system inspired by another one even remotely illegal?

The first thing that came to mind reading it was "backtracking". Which I guess is the recursive pathfinder piece?

That is fascinating. Thanks!

I’d gladly take a few of these self-contained rack-clusters off their hands when they do.

I’d even get a house with a garage or something just for that.

> Placement is probably the big issue that needs to be solved now.

Would some sort of constraint-solving algorithm help with that? Something like (but not necessarily) Cassowary[0]? Maybe I'm misunderstanding what is meant by placement though; I don't have much domain knowledge in PCBs / electronics.

[0]: https://news.ycombinator.com/item?id=43362528

And are exposed as local files stored in documented, stable, and sane paths.

Their sync is good too, in my limited experience.

Isn’t that the usual EU startup playbook once they give up on the B2C or world-scale SaaS markets? Refocusing on large (European) enterprise B2B and government contracts?

It always felt to me this (enterprise B2B) was where European startups went to die.

> tiny <2-3B -- could run in a browser even, mac neo

Or a phone. I’m running Gemma 4 E2B in one of my apps on my 14 pro (which may or may not be killing my display through overheating. It might just be a coincidence).

I never considered the implications and impacts on datacenters' architecture and organisation. It’s fascinating.

Over here our CTO replaced Intercom with an internal equivalent that costs less than $20 / month to run, haiku and sonnet support agent costs included. In less than a few weeks, in his spare time.

I recently switched from GitHub Actions to Buildkite + self-hosted runners.

Setting it all up would have been tediously annoying eight months ago (Buildkite requires setting up GitHub webhooks for each repo).

Last week I just had codex set up everything, ephemeral vm runners and all, using a couple of low-spec refurb mac minis, Buildkite’s API, a short-lived API token, and migrate my repositories one by one.

So far so good, it’ll pay for itself within two to three months, and following today’s outage I suggested at work that we experiment with the same set up.

They’re considering it.

Another aspect is that selling to all of Europe as a B2C business is hard. Until recently you ended up having to register for VAT all over the place, god only knows how many different specificities, bureaus, and rules, with most payment solutions not helping you in any way, and most accountants (in my experience) being at best unable to help you in any significant fashion, at worst being very confident in their ability to help you.

It is done, but by very few. And the EU has made progress on uniformizing and simplifying it, but it seems to have done more progress on the B2B side than on B2C.

While the US is a much bigger consumer market than any single EU country, with significant differences in disposable income and spending power. 18% to 20% of full-time workers in the US make over $100.000. That's nearly half the entire population of France. A third that of Germany.

And even if there are differences and administrative hurdles when selling across US states borders, that road has mostly (or seems to have, from here) been paved.

> Why TF haven't enterprise made it more painful for these events?

They mostly have. By mostly refraining from dealing with startups and companies they deem either “too young” or "too small" to be reliable partners. And, when they do, imposing long sales cycles.

And thus the enterprise well is poisoned for most startups.

I looked at Punkt.

They keep saying "If you don’t pay for the product, you are the product". Okay, all fine and well.

But what will my phone still actually be able to do if / when I stop my subscription? Not a single clear answer besides "[…] gradual feature deactivation, and ultimately reverting to a device running AOSP".

Doesn’t really inspire confidence.

The risk is people who incidentally get into software engineering through using agents probably won’t be aware of any of that. And probably wouldn’t care even if they did. And for most software they’ll create, single user, perhaps even single use, it’ll be fine.

And for those who might care about these things, they’ll probably just be facing constant pressure to deliver more, faster, perhaps with less.

One of the issue about literacy in algorithms, data modeling, efficiency, development patterns, systems designs… is that people either aren’t aware of them (and in the best case scenario reinvent the wheel), don’t care about them, or feel they aren’t given the time to invest in learning them (or worse, might be penalized for it).

A punchcard with extra steps?

Enlighten us.

Instead we’ll all run LLMs that generate UI from scratch, at run time, using up gigabytes of RAM and who knows how much power on every request and response.

Progress!

(I’m half kidding)

> Something is lost along the way.

Definitely.

My point is more that it seems to have been the way of the world for the past few decades. I’m arguing it’s nothing new, basically a trope at this point.

And once that is said, if we care about it, what do we do about it? Besides just repeating it.

There will always be dataset and training quirks, and the provider’s own biases and focus, granting one model an edge over the others in some specific domain.

I am afraid we just aren’t going to be able to understand each other.

No, current models and tools are more than capable of dissecting entire codebases.

And even if they somehow weren’t, we’d just do what we used to do with documents to turn them into "chewable bites": chunking, extracting, summarising,…

I thought so too.

But 1) people use other models with that same harness. 2) I moved on from Claude Code and all the features I cared for up and running in less than a couple days. Without even looking for available plugins or extensions.

> The only genuine moat the frontier labs have is their product take-up

And even then, their is no stickiness. For most use cases there isn’t much value in one frontier model over the other.

Just have to look at the people flocking from one to the other for whatever reason.