fallpeak

On OpenRouter there are 11 third-party inference providers hosting DeepSeek V4 Pro right now, of which 8 are US-based and 7 of those have zero data-retention policies (which I mention to rule out any claims of "oh they're making up the money by logging all your data"). This is a 1.6T-A49B model, so a bit bigger than Sonnet (~2/3rds the size) and a bit smaller than Opus (~3x as large). These third parties are almost perfectly interchangeable via OpenRouter as a marketplace, so they have no incentive to offer any sort of "growth pricing" below costs, and they serve it at $3.48/Mtok out.

Kimi K2.6 is 1T-A32B with a slightly less computationally efficient architecture, and is served at around $3.50/Mtok out by 9 US ZDR providers.

Unless you think that either the generally accepted size estimates for Anthropic/OpenAI models are wildly off or those companies are a lot worse at serving models efficiently, Anthropic and OpenAI are probably making around 5-8x margins on their API costs.

The cost of training new models is of course a major factor not counted here. Depending on how you want to think about that this may or may not make them net profitable. I remember one of those CEOs gave an interview a while back where they described it as a series of independent investments, where each model they train is net-positive in revenue by EOL just from its own inference, but I don't know whether that's still true or not.

Regardless, the point is that if they stopped training new models today, both Anthropic and OpenAI are making incredibly generous profits on their API inference.

You are incorrect. There, now we've both made unsupported assertions. Care to provide any evidence for your position?

For what it's worth, when I provide a Pangram link it's because I can already tell something is AI and I'm attempting to provide objective third-party confirmation so the conversation doesn't just degrade into me asserting that I have superior taste to you.

This phenomenon (and/or the fear of being falsely accused) is why I find myself being an annoying unpaid evangelist for Pangram every few days.

Right now a lot of people are developing their own heuristics, some much more well-calibrated than others, for sniffing out AI generated text but it's kind of an unfalsifiable accusation since you're really just saying "this sounds like slop, I can tell from the words and from seeing lots of slop in my time".

Linking to an AI writing detector serves as a somewhat objective second opinion, but half the time if I say "Pangram confirmed it" I'll see at least one person pulling the old "yeah but GPTZero is crap so that means AI writing detection can't work in principle" maneuver, which strikes me a lot like someone who tried ChatGPT in 2023, decided AI was all hype, and never re-evaluated that stance.

It would be really helpful for the discourse if we could get to a point where people generally accept one or more _actually-good_ LLM writing detectors as a reliable tool such that a "this is AI" judgement from them is accepted as reasonable proof of guilt, solely so that we can then extend the benefit of the doubt to anybody else.

This is slightly more tasteful slop than average (I'm thinking probably Claude rather than ChatGPT?), but it's still 100% AI written: https://www.pangram.com/history/c55ab69b-e0a9-49a0-8056-2fcd...

AFAICT they're basically saying "hey lawmakers we're not going to rock the boat or take a stand so in exchange pretty please don't lump us into the app store category"

What does that have to do with anything? These days any piece of text may or may not be AI generated (my money would be heavily on "no" for the post you asked about), but either way it isn't blatant slop so we can't tell.

It feels like you're trying for a lazy gotcha, but the actual point here is something like "AI models often generate writing with specific noticeable characteristics that make it obviously AI output, and TFA is an instance of such writing, and this should be called out when possible"

Have courage and trust your own instincts. Unless one is extremely disagreeable it's very tempting to hedge and avoid outright saying "this is AI" just in case you're wrong, but if you're literate and regularly exposed to AI outputs your instincts are likely quite accurate.

In this particular case the linked article is definitely AI generated.

The Kobo Clara doesn't fit in my pocket while the X4 does. An e-reader that I actually have on-hand beats one that I leave in a drawer at home so it was an instant purchase for me the moment I saw one.

Tokens _are_ as cheap as keystrokes. A single keypress by a full-time SWE averages out to $0.005-$0.02 (depending on typing speed and TC). The relationship is obscured because the keystrokes are usually part of a fixed-price subscription plan but they absolutely have a cost. Prior to AI this was in fact a large reason everyone pontificated about concise programming languages and elegantly factoring problems and DRY and...

The only correct way to ask an AI "why did you do that?" is in the sense of a blameless postmortem. You're the person responsible for giving the LLM appropriate context and instructions and guardrails, so the only reason you should ever ask a question like that is when you're genuinely trying to figure out how to improve those for next time. Every time I see people posting this sort of "apology" from an LLM it makes me cringe, feels only half a step away from outright AI psychosis.

That seems like unnecessary pedantry, only one step short of the common "everything is chemicals" retort. Egg yolks are an ingredient that contributes to the flavor and nutrient profile of the food while also emulsifying. In this context "emulsifiers" is presumably meant to refer to other ingredients added solely to modify the texture.

TL;DR: Devs didn't know what they were doing and turned off autovacuum and eventually it broke, then the author decided to have an AI slop out an article about the incident which may or may not have actually occurred.

It doesn't know that because it wasn't trained on any tasks that required it to develop that understanding. There's no fundamental reason an LLM couldn't learn "what it knows" in parallel with the things it knows, given a suitable reward function during training.

Did you write this or is it AI? Not hating, but it pings several of my "AI writing" heuristics and I'd like to improve my model if possible.

edit: Never mind, given your comment history this is definitely LLM output.

The unreadability of that example has approximately nothing to do with code formatting, which is generally understood to refer to modifying the textual representation of the code while leaving the actual logic more or less unchanged. Can you propose some alternative whitespace or indentation scheme which would make that example significantly more readable?

Now do the comparison if one endpoint is realtime user facing traffic and the other is batch processing which can easily eat up all available capacity and drive up latency.

If visa shoppers are overwhelming the normal processing of applicants who actually live in a particular country, it seems entirely appropriate to say "no, sorry, this location isn't for you" to the people who don't live there.

People still make games for old consoles occasionally as hobby projects, and those are usually released freely as ROM files. I'm not familiar with Japanese law, but in most countries that would constitute a fairly solid proof that there are legal uses to which an emulator can be applied and thus that emulation itself isn't inherently illegal.

"The basic principles of the Python ecosystem" are a dumpster fire to anyone who isn't already desensitized to the situation. Just like 'uv' as a whole, this seems like a meaningful step towards making Python a little less terrible to actually use, and should be applauded.

Sure. But, and I will die on this hill, it's not vibe coding if you're looking at and understanding the code.

I mean, by construction you're only ever going to see the examples where people checked them in and published that. It doesn't mean that other people aren't getting more use out of local instructions customized to their particular work.

Interesting, I'm surprised the results vary so much for a query with such an objectively correct answer. I tested it on my desktop in private browsing mode because I wanted to see whether Kagi was much better, but figured I should report the null result when Google actually did fine.

In my opinion an AGENTS.md file isn't an artifact at all (in the sense of being a checked-in part of the codebase), it's part of the prompt. You should gitignore them and use them to give the LLM a brief overview of the things that matter to your work and the requests you are making.

You're in luck! If you search for "canadian visa application" the first two results are "Visit Canada" and "How to apply for a visitor visa" on https://www.canada.ca/

I don't know if GPT-5 is an exception and is overcooked on XML specifically, but in general Markdown and XML seem to work about equally well for LLM inputs, the important part is just that they like hierarchical structured formats. The example on that page could probably be replaced with:

  ## Code Editing Rules

  ### Guiding Principles

  - Every component should be modular and reusable
  ...

  ### Frontend Stack Defaults

  - Styling: TailwindCSS

Without any meaningful change in effectiveness.

> If you pay in crypto, you have to report every conversion from fiat to monero

That's not what your link says, and as far as I'm aware it's not true. Buying crypto and then using some of it to buy goods and services has no tax reporting requirement, those only start when you're either selling crypto or receiving it as payment. Which is the same situation as the tax reporting for any other currency or valuable item you could deal in.

While this is really neat work, it's not entirely accurate to describe this as a base model or unaligned. Instruct training does two things, broadly speaking: it teaches the model about this "assistant" character and how the assistant tends to respond, and it gives the model a strong prior that all prompts are part of just such a user/assistant conversation. GPT-OSS is notable both because the latter effect is incredibly strong (leading many to suspect that its training was very heavy on synthetic data) and because the assistant character it learned is especially sanctimonious.

This finetune seems to work by removing that default assumption that every prompt is a user/assistant chat, but the model still knows everything it was taught about the assistant persona, so inputs that remind it of a user/assistant chat will still tend to elicit the same responses as before.

This is only a problem when implemented by entities who have no interest in actually solving the problem. In the case of apps, it has been obvious for years that you shouldn't outright tell the app whether a permission was granted (because even aside from outright malice, developers will take the lazy option to error out instead of making their app handle permission denials robustly), every capability needs to have at least one "sandbox" implementation: lie about GPS location, throw away the data they stored after 10 minutes, give them a valid but empty (or fictitious) contacts list, etc.

What an auspicious coincidence, to see this the same day I finally got around to setting up a proper pseudonym.