ElFitz

> 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.

> Shitty will be the new normal.

I’ve heard the same from the best devs, and some who thought themselves to be the best, I’ve known long before LLMs were ever a thing.

I’m sure others heard the same when JavaScript and Python became near ubiquitous. When PHP emerged. When C supplanted Fortran and COBOL. When these two took over from Assembly. When punch cards went the way of the dodo.

There’s always someone for whom shitty is becoming the new normal. If that makes it a rule, what do we make of that rule?

> I understand the impulse to provide a UI to manage codebases, etc. […] 'I don't care what's happening, just ask Claude to do it'.

Reading the first part, I was going to say they don’t even care about whether or not there’s a codebase. It doesn’t matter; it could be all gremlins and hamsters in wheels for all they care, and for all they should care. All that matters is the functionality, the value it gives them.

We’re even getting disposable code now. Entire single-use ephemeral web apps, built on the go to enable, visualise, or simplify a specific thing, then thrown away.

Will it all lead to some trouble? Definitely. So did computers, and so did the internet.

Weird times. Fun times.

Sounds like "observability, for sleep".

It’s funny how many things can boil down to "rich distributed traces" and events / logs.

> You used Python or TypeScript because[…]. because Rust, Go, C++, and many more would give you 10–100x the performance, but you paid for it: […] a build system that fought you.

I would argue I spent more time fighting the TypeScript build system than Rust’s.

But up until recently I only used either just often enough to never remember what magic configuration needed to go in my tsconfig.json and package.json to get TypeScript to work.

The state does. And once a death certificate have been issued and filed, proving you are in fact still around is whole can of worms of its own.

eg https://www.theguardian.com/lifeandstyle/2021/jul/03/they-sa...

Edit: and a few others https://news.ycombinator.com/item?id=48060857

The Romanian man Guardian article has a great and horrible quote:

> The court told him he was too late, and would have to remain officially deceased.

Last time (after years of doing it manually every once in a while) I just gave codex an ephemeral restricted Cloudflare API Token / key / whatever, the screenshot, and it set up all the records on its own.

> I never conflated anything. I said it's a neglible amount of space for current hardware, which I still believe.

>

> If anything, the fact that I think the amount of space is acceptable for the amount of ram a modern laptop has exaggerates the point.

How does the memory usage of your browser tabs relate to the amount of disk space taken up by the downloaded models?

> That's the approach they take for most of their features.

And there’s a reason for that. Same one the EU forced companies to make consent to marketing and consumer data sharing opt-in and not opt-out: we have a bias.

> Which seems to be the motivation of having these local models embedded in the browser's available resources

Sooo… we are basically agreeing on the fact it could prove to be useful? In the long run, if and when they decide to pick a model that isn’t half brain dead (apparently it’s based on Gemma 3).

Offering something like a local Gemma 4 (though apparently not what we get here) to web apps via a browser API could change UX quite drastically. Possibly for the better. We had a project where it could have been nice.

> Including a 4gb is a negligible amount of space for current hardware and Chrome is not known as the browser to run on resource constrained devices.

4gb definitely isn’t a negligible amount of space on most people’s devices.

The quite successful it would seem MacBook Neo has 256GB of storage in its base configuration.

A MacBook Air and a basic sub $1000 Dell laptop starts at 512GB.

> To put 4gb in context, I currently have 2 tabs open that nearly take up 4gb.

You are conflating disk and memory.

> The fact Chrome also has a way to disable this makes it kind of a nothingburger in my opinion.

There’s a reason they picked an opt-out model for this, and not an opt-in approach.

But I also see the point in it. We recently did a hackathon, and I considered relying on Gemma 4 for privacy considerations. The local model could interpret the user’s natural language request and derive less privacy revealing requests to form based on that.

But then, a web app that shows people a loading screen while it downloads a 4GB model probably wouldn’t be a best-selling UX.

> Zed is years behind VSCode in terms of polish

One could think that. But VSCode is the one that occasionally failed to simply render text.

No idea what happened these handful of times, but the UI was just completely screwed up, as if it were one of these "scratch to reveal" games, but with the file’s content (and unresponsive, obviously).

These people worked on applying ML to code and towards producing code using ML before the transformers paper even came out (2017).

https://web.archive.org/web/20170629103718/https://blog.sour...

Oh! Yes, stupid assumption of mine. Thanks for catching it.

Edit: As a kind someone pointed out, I made the wrong assumption that the first response was entirely Anthropic’s, and not the author.

~~I mean, the worse part is the gif at the end of the message.~~

~~What are they even trying to do? What are they trying to convey? It just feels like being given the finger and getting my face rubbed in it on top of that.~~

Hadn’t thought of that. How does it work with web-based apps and simulators / emulators? MCPs + screenshots?

That’s pretty much it. I’d say minimal supervision, and I can’t say whether or not it is doable or desirable, but maybe I’m not bullish enough. Or too much.

There’s a joke about dark factories, by the way:

> The Factory of the Future Will Have Only Two Employees, a Man and a Dog.

> - The human feeds the dog.

>- The dog makes sure no one touches the equipment.

Except one of the major other wrappers was pi, through OpenClaw. With countless hundreds of thousands of instances running every hour on that heartbeat.

I have no idea what the share of OpenClaw instances running on pi was, or third-party wrappers in general, but it was obviously large enough that Anthropic decided they had to put an end to it.

Conversely, from the latest developments, it would seem they are perfectly fine with people running OpenClaw with Claude models through Claude Code’s programmatic interface using subscriptions.

But in the end, this, my take, your take, is all conjecture. We are both on the outside looking in.

Only the people who work at Anthropic know.

That’s what we see.

It may be (but I wouldn’t know) that some of other changes not covered here reduced costs on their side without impacting users, improving the viability of their subscription model. Or maybe even improved things for users.

I’d really appreciate more transparency on this, and not just when things fail.

But I’ve learned my lesson. I’ve been weening off Claude for a few weeks, cancelled my subscription three weeks ago, let it expire yesterday, and moved to both another provider and a third-party open source harness.

Edit: the article prompting this entire thread.

Evidently, all these things you just dismissed matter, else all the changes I quoted from the original post wouldn’t have affected anyone, or half as many people, or half as much. Anthropic wouldn’t have had any complaints to investigate, the article promoting this entire thread wouldn’t exist, and we wouldn’t be having this very conversation.

Defaults matter. A large share of people never change them (status quo bias, psychological inertia). Having control over them (and usage quotas) means Anthropic can control and fine-tune what this fixed subscription costs them.

And evidently (re, the original article), they tried to do so.

By imposing the use of their harness, they control the system prompt:

> On April 16, we added a system prompt instruction to reduce verbosity. In combination with other prompt changes, it hurt coding quality, and was reverted on April 20. This impacted Sonnet 4.6, Opus 4.6, and Opus 4.7

They can pick the default reasoning effort:

> On March 4, we changed Claude Code's default reasoning effort from high to medium to reduce the very long latency—enough to make the UI appear frozen—some users were seeing in high mode

They can decide what to keep and what to throw out (beyond simple token caching):

> On March 26, we shipped a change to clear Claude's older thinking from sessions that had been idle for over an hour, to reduce latency when users resumed those sessions. A bug caused this to keep happening every turn for the rest of the session instead of just once, which made Claude seem forgetful and repetitive. We fixed it on April 10. This affected Sonnet 4.6 and Opus 4.6

It literally is all in the post.

I don't worry about anything though. It's not my product. I don't work for Anthropic, so I really couldn't care less about anyone else's degraded (or not) experience.

Because then they lose vertical integration and the extra ability it grants to tune settings to reduce costs / token use / response time for subscription users.

Or improve performance and efficiency, if we’re generous and give them the benefit of the doubt.

It makes sense, in a way. It means the subscription deal is something along the lines of fixed / predictable price in exchange for Anthropic controlling usage patterns, scheduling, throttling (quotas consumptions), defaults, and effective workload shape (system prompt, caching) in whatever way best optimises the system for them (or us if, again, we’re feeling generous) / makes the deal sustainable for them.

It’s a trade-off

Now we know why Anthropic banned the use of subscriptions with other agent harnesses: they partially rely on the Claude Code cli to control token usage through various settings.

And it also tells us why we shouldn’t use their harness anyway: they constantly fiddle with it in ways that can seriously impact outcomes without even a warning.

Hmm.

Maybe there’s a way to play around with this idea in pi. I’ll dig into it.

Yes, it’s not there yet. But nothing unsolvable. First thing that comes to mind would be generating smaller portion at the same resolution, then expand through tiling (although one might need to use another service & model for this), like we used to do with Stable Diffusion years ago.

Another option would be generating these large images, splitting them into grids, and using inpainting on each "tile" to improve the details. Basically the reverse of the first one.

Both significantly increase costs, but for the second one having what Images 2.0 can produce as an input could help significantly improve the overall coherence.

Damn. There’s a fun game app to make here ^^

And yet running the Claude Code cli with `-p` in ephemeral VMs gets me the "Third-party apps now draw from extra usage, not plan limits. We've added a credit to your organization to get you started. Ask your workspace admin to claim it and keep going." error.

One day you're experimenting just fine. The next, everything breaks.

And I'd gladly use their web containerized agents instead (it would pretty much be the same thing), but we happen to do Apple stuff. So unless we want to dive into relying on ever-changing unreliable toolchains that break every time Apple farts, we're stuck with macOS.

I wouldn’t agree. Even at national scale, these projects cost resources. And the resources of all agents (org, countries) are constrained.

While we could reason in "performance / watt" and "performance / people", "performance / whatever other resource involved", and "performance / opportunity cost of allocating these resources to this use case and not another", "performance / whatever unit of stable-ish currency" is a convenient and often "good enough" approximation that somewhat encapsulates them all.

A simplification, like any model, but still useful.

That was quite the unexpected anticlimactic ending. I’m sure Terry Pratchett would be proud.

> A dollar is an entirely fictional unit and trillions of it can be manufactured at no cost

It’s still a useful proxy for resources allocation and viability.