dofm

They probably don't. But some words acquire their full sense meaning from historical and-non-programming context.

Words that are or were fairly common in tech circles — "bailiwick", "gatekeeper", "daemon", "boundary", "infer", "method".

It's true to say that most of these words have some specific meanings in tech, but you don't necessarily want the LLM to have only those narrow meanings when you are talking to it. A broader grasp of the word means the LLM won't go down a specific pathway before understanding you by analogy. That broader grasp comes from non-programming understanding.

They do, to the extent that general intelligence training is what gives you the ability to talk in broad natural language to a coding agent.

Words only have their meanings through general context.

It’s not, of course; you’re going to act on your own requirements as you should.

But I think it becomes a little more relevant when making a price-performance comparison such as the article makes between “open” models and “closed” models.

Because the reader is invited by the term “price-performance” to think about efficiency, when in reality the results are being skewed by the fact that the user is not the only one paying for the service: investors are chipping in fairly directly for every token, at inference time.

Users are going to continue to do their own thing for their own reasons, and they should; that’s the market at work. But this is not an efficient market where meaningful comparisons can be made; all of the major closed providers are loss-leading.

It seems way, way easier to switch LLMs than it is to even change from one iPhone to another iPhone!

Granted Anthropic (and Google, maybe less OpenAI actually) do what they can to make it similar. But switching phone brands would take me a day or two of planning and I don’t even use the Apple mail app.

> A banana left out in the walkway still requires the user to go for a walk.

This is some Eric-Cantona-level gnomic mysticism. I approve.

> They do not have Claude. I use Pi and other harnesses, but it's not the same.

This is going to sound tautological but of course it's not the same, because it's different?

Participants in markets make their own choices from the available options, for their own reasons.

Another way to frame my point is this: not the same for whom? You are but one market participant, and it's at least wobbly thinking for any of us to assume that our set of needs is the archetype.

There's any number of harnesses and people have any number of preferences. Sure, overchoice of imperfect substitutes tends to entrench better-resourced, better-resourced (or better-marketed) alternatives, but it's not a slam dunk. Things come out of left field all the time.

(I mean from my own perspective, I like Pi — more than OpenCode. And the "betterness" of Claude or Codex over Pi, or their models over even Muse Glimmer or Qwen, is completely irrelevant to where I am now in terms of the needs I have. So I don't use either of them and I remain unfazed. When my needs change I will revisit, but the theoretical superiority of these products for the majority of users just isn't a factor in my decision, just like I don't care that any tool I use might not be the best for most people. You might think me insane but I'm no less a rational market participant than you.)

No this is Ned Kaspersky, from Moscow (Idaho)

Here I am saying “well, shit”, over and over again, but so quickly that not even dogs can hear it.

This is very much aligned with Steve Eisman’s take: that it’s a terrible overvaluation but for Anthropic it is now or never.

OpenAI are openly suggesting 5%, but my instinct is that this is because they think Trump is foolish enough to demand double that.

And even then, the path for investors to ever make their money back is really essentially non-existent. There’s so much “total addressable market” fantasy here.

Everything we know about Anthropic’s unit economics is still riddled with EBBT. Earnings before bad things.

I assume some of this is the escalating cost of training, indeed, and that being faster means abandoning efficiency.

But you’re right. I have often wondered why it is that such a small amount of extra compute (300MW) from SpaceX was entirely enough for them to uncap so many customers.

I think the (in-house) data centre burdens of ever-more complex models may be considerable, but I think the data centre businesses expecting to serve inference tasks are probably heading into a Chinese real estate bubble.

Despite its simplicity, price-performance is not the only metric; this obsession is a typical consequence of extreme overextension of the Friedman doctrine, and it is why so much software and hardware has been qualitatively shitty even when it matters that it’s not. That something can be measured doesn’t automatically make it preferable as a decision-making basis over intangibles.

It’s pretty obvious that local AI has a higher TCO if you look only at money. But there are other things to be considered.

Three strikes, in particular, is inherently unjust. There's no serious argument that it is not.

Silly fake faux-old-testament solutions like this don't actually work, as is demonstrated by the fact that most crime has fallen faster since they were abandoned.

It’s not, particularly, without unjust laws like three strikes.

For one thing, it’s a revolving cast of a few dozen people at any one time (people do go to prison, get rehabilitated, move on or get moved on). For another, they are often vulnerable or exploited: a lot of shoplifting is done to order, to supply the black market.

The UK does need to take shoplifting more seriously, but it’s important to understand that it is a consequence of both the design of retail (self-checkout and fewer staff) and of the cost of living.

Most of it is Harp or Scavino, as far as I have read.

Unless it is running on a government issued iPhone, it is not the Individual #1 you are alluding to.

He very famously does not know how to use computers and can only basically manage tweeting from his phone, which he is mostly not, actually, doing himself.

Now, there are two or three terminally online people around him. I would be more quick to point the finger at his foul-mouthed spokesman but it’s admittedly because I hate that guy and he seems rather stupid. But perhaps it could be the devoted portable printer lady, or the overpromoted unblinking former golf caddy guy.

All of them are around to help with the, er, modern day presidential communications.

> and an astronomical crime level (160 per 1000 people? So I have a 16% chance of being a victim??)

Not even close, unless you own a shop (in which case, in some areas, life can undeniably be exasperating and miserable). The vast majority of high street crimes are shoplifting or possession (or drinking in alcohol control areas) and an astonishingly large number are attributed to a tiny, tiny number of people.

There are high crime areas in most towns in the UK where you will feel pretty safe, because the crime is simply not directed at you.

When it does spill over into personal violence, it is the aforementioned handful of people with drug problems. It’s worth noting that one incident involving three people is likely to result in a dozen recorded crimes. And a British town of sixty thousand people will have so few of these prolific individuals that they themselves commute to neighbouring towns to steal.

Worth remembering also if you are an American that in the UK we have not decriminalised marijuana, so there are drug possession and small-scale dealing crimes in there too.

I am committed to learning about it; it would be nuts not to.

Not committed to using it.

Until recently my mental health post-burnout has been so absolutely poor that I have been learning all of this just to try to concentrate on anything so I don’t feel further out of the loop. Most days I still feel like I should quit IT.

In the limited ways I am actually using it in a productive rather than self-educational sense, I tend to be using it to combat burnout: helping get past blank page problems, rough ideation and keeping track of ideas, doing slightly more cohesive research than just ending up with dozens of open tabs and some unclear notes, generating helper processes (like content migrations) that I might otherwise not be able to make myself write, and exploring using it to do the unthrilling work like automated test generation.

I don’t plan on becoming someone who sits back and specifies loops, and that doesn’t fit with the work I am doing. I have far too many disjointed things I have to work on.

But it is definitely a more varied rubber duck, and that has some potential. I feel a little less stuck, sometimes.

1) none

2) Local LLMs on an M1 Max (Unsloth, LM Studio, occasionally direct llama.cpp). Muse Glimmer (because it is good at what it is optimised for: interactive coding sessions). Qwen 3.8 27B occasionally and a couple of other things for fun (an LFM MoE, experimenting with Nemotron Bolt 74B-9B, enjoying the really super Gemma 4 12B QAT when working on creative photography research or non-code planning).

No openrouter (yet) but that will definitely happen.

3) Currently focusing my attentions on Pi via Paseo (and trying Pendant). Pi is just much better in smaller context sizes, and it is the right fit if you are keen to learn from first principles.

I tried OpenCode first. I couldn’t quite get on with it and what ultimately made me switch was trying to understand its terribly messy source code.

4, 5) I am mostly using AI as a coding sidekick so I don’t really have much need for orchestration yet. I did some experiments with Osaurus (really quite impressive) and I actually think that is almost enough by itself, except that I want a little more control over the sandbox container.

I feel that my usage pattern is still settling, because I am not keen to surrender all of my coding to the machine. But I am trying more purely “directed” coding to learn the process, and finally beginning to get a feel for how I will work with that.

Yes. It's that thing — the fact that a language helps you read what a single function does really easily, but in so doing makes the entire codebase quite dense, is not optimal for us, but is good for them.

And the things that we do to manage this, that are symbolic or based on visceral real-world analogy, are good for us and bad for them.

Go is kind of in the middle, in the way that many vaguely Wirth-like and CSP-like languages possibly are, I guess.

I hope they never have to call each other about anything. It could get confusing if neither knows who is the base case.

So we have a Recursive and a Ricursive, both making LLMs?

This is mostly why I have come back to Go.

I am using local LLMs and Go has a bunch of strengths where smaller contexts are concerned; less complicated flow of control being one of them.

I… think this is good?

I'm just getting the neuron-band back together where Golang is concerned. I have done some in the past for little system tasks, but I am switching away from interpreted languages entirely for web stuff, and initially to Go, as I switch away from hand-coding to at least LLM-supported coding.

So I am a bit out of date on the pain points (I never really felt pain points when I last used it, because the jobs I was doing had very small, quite Go-friendly shapes).

Oh, for sure.

But you can't trust any LLM, or any output from any LLM. You need to protect against stuff like this regardless. They could all do it, and quite accidentally. Even an LLM you trained yourself could do this to you, based on the training data and based on RAG.

But if you want to avoid the specific case we are talking about where your inference provider sees everything you do, and ignores your preferences, then running someone's LLM on your own inference is the only way forward.

Why do you need to train a model at all?

You surely don't need frontier models for any job that isn't going to replace a human. If the human is still in the loop, just use a model the human can work with. It can be quite small. You could even be running Qwen 3.8 27B for each developer and it would be fine, and you are in control of your own destiny and can make more progressive judgements about things.

I don't really understand how the AI world has caused so much black-and-white thinking to come about. You are not under mortal threat if you don't use the latest models. This is just FUD. You still have people with brains and you might get a better result over the longer term if you use AI to assist them in their choices rather than make them sit there and hit return.

Thank you for your reply — I think it sounds like you're in a good place here.

I still definitely wish this wasn't something we had to care about still, but for a while at least I suspect it is.

I am a small/open-weights LLM proponent, really, so at some point I will test this a bit more, because I tend to use LLMs close to bare-metal where prompts and skills are concerned, due to limited context. (In general I think post-training of smaller coding-oriented models like Muse Glimmer and Qwen has made long system prompts and extensive skill files far less significantly important)

Muse Glimmer seems to have some knowledge of Toit without searching the web.

e.g. I asked it: "Could you write an embedded webserver that blinks the LED the number of times specified in a GET query string parameter?"

And what it wrote without access to documentation appears to be plausible (from what its human overlord can see from the documentation at least).

Even I (as a sort of grudging cynic) now find myself thinking in these terms:

Is this a language that an LLM can write well?

Especially a small LLM, given the scripting and friendly IoT focus, where it might be useful for a smaller model in a home assistant to be able to write in it.

A packageable scripting language with fewer footguns that has LLM-friendly features feels like it probably has a place in the world.

But my own experience so far suggests that smaller LLMs in particular do better with simpler languages with obvious flow of control and less cleverness like metaprogramming.