Post

They already settled. authorsguild.org/news/court-g...

Well, it turns out there can be only one best thing. Claude Code >>> Codex, basically. Anthropic really won the agentic development race. I don't even think it's that close anymore.

Probably asked Claude

None of these measures include capital expenses directly.

Seems like what happened is they cornered one large market where the tool will be (or maybe already is) ubiquitous. Apart from bespoke in-house stuff (probably built on Claude anyway) is anyone coding with other brands?

Yahoo! and Ask Jeeves were earlier, no?

Us elder millennials remember our fallen search engines like AltaVista, Lycos, Excite, and Ask Jeeves.

I'm just making this up, but -- the money they had to pay Apple with is the difference, that they had that from making the best product and having lots of customers and revenue Anthropic making money and doing a galactic IPO will give them so much power to make these sorts of moves too?

This only works if it's free

There is certainly an element of time this right so the balance sheet is at its best though. Plenty of reasons to think it wouldn't and won't look so good 6 months from now, including their own statements and rushed behavior.

Dario coming in strong on the side of Claude

Claude probably has the answer, but Dario would need three safety reports and a 40-page essay to approve it. 😂

I think the big disconnect over everything is that the one industry that LLMs absolutely *are* going to devastate is software development.

battery storage? no he doesn't, CATL is way bigger than Tesla

Yeah, “just make them do coding first and foremost” turned out to be correct bsky.app/profile/dmnd...

"let's make the AI that is best at using computers which also makes it the best at improving AI" *should* have been an obvious play. i mean, if you gave me *one* infinity stone, i'd use it to try to get the others, not immediately figure out how to short-term monetize it!

addendum: this piece on truly open models is good and I think relevant to AI companies' long term profitability www.argmin.net/p/the-world-...

FWIW, this also feels a lot like "the truism that B2B is a better bet to be profitable than anything else continues to hold".

The two best use cases for AI are - fold some proteins for me - here's this discrete unit of code that's small enough that I'm pretty sure you won't screw it up; get that done while I work on this other stuff Anything else is somewhere between pointless and actively harmful

the steel man case for that nonsense is that it got CAPITAL sufficiently excited to throw around the insane gobs of money in enough directions that one (1) AI company was able to land on a legitimate business case which, being boring B2B disrupting major players, would never have landed on its own

Anthropic doesn’t even _have_ audio or image generators.

@garymarcus.bsky.social points at the same. mu.social/profile/jess...

Inference is not that capital intensive and at scale probably not hard to make it profitable. Training is what is expensive.

Yeah, if you connect this to the layoff news at MSFT (1/4 of the workforce), it seems clear that Claude Code and the like are well worth it.

Yes that’s very helpful thank you

I didn't phrase that like a real question but I was genuinely asking, thank you.

Consumption vs subscription

Mobsters, scammers, and malicious hackers also know how to make money, but we generally try to discourage those business models in a civilized society. Also doesn't address longer-term prospects (like companies switching to smaller internal chatbots for cost saving/data protection purposes).

Crime sometimes eventually pays out pretty good.

Someone call a preemptive welfare check on Ed Zitron

Okay I'm legitimately surprised to see this.

Question is, can Anthropic sustain this?? bsky.app/profile/rasu...

Great, that means this waking nightmare will be the blueprint for future IPO's...

I bet I could also make money if I was able to steal from others without consequences.

This is not GAAP accounting but yes the Bros could make a lot of money selling today’s AI tools. Their next goal is much bigger than that—to drop a general purpose agent into any office job and take that job, and in the medium term to do the same to non-office jobs through integration with robotics.

it's not really opaque. the vast vast majority of this revenue is enterprises paying per token at API rates

Yeah I mean, IPOs are quick money…and I don’t think the company will go tits up anytime soon (unless there’s some accounting shenanigans going on) but reality comes for everyone.

this isn't a leak. this is the public filing

Seems like they are profitable now. Training models is expensive, but serving those models to users is quite lucrative. They just needed enough users to balance the equation. finance.yahoo.com/markets/arti...

I do not want to live in a world where people who wholesale stole their product get to be successful.

They just asked Claude

The same way PT Barnum did.

That's an interesting real-world example of how model choice affects perceived product quality. I'd be curious whether the improvement holds across different business tasks. Comparing the same prompts, source material and tool permissions would make a useful practical benchmark.

yep, I've heard similar things. like any tech product there is a learning curve, and these are pretty weird ones

Furiously writing in my business school notes "make the product...that is better..."

all those same-looking cowork ppts for the qbr

Am I wrong in thinking that one reason people will keep using US models is that if something bad happens you can't sue a Chinese lab?

Interesting. I wonder how big companies split on this. At the one I work for, they push us to use free models but allow for limited Claude model use.

Claude is the only AI that any of my clients have wanted for their spreadsheet lords to use. Just zero fucking interest into ChatGPT, Gemini, etc. Copilot lmao.

Maybe. Do you remember their social network?

Gemini is frustrating in that it should work better with built in connectors to everything but it really just doesn’t.

i think they probably will but consumer sentiment is such a strong force that "but claude is the one I like" might set in fast

IMO, people have been sleeping on the non coding abilities of LLMs since Gemini Pro 3 back in late 2025. The frontier labs had done a terrible job setting up the non-coding use cases, outside of trivial gimmicks, and couldn't articulate what you would do beyond "make an X". Things are different now

This is my experience as well; at NOAA where I work we get Google AI, but I pay for Claude for myself and have found it is way better at helping me do actual scientific programming

Corporate finance guy here and I second this. I’ve seen Claude make my job 100 times easier. Copilot in particular is what’s available at a lot of companies (because they’re already in the Microsoft ecosphere due to Excel) and its native capabilities so far seems like a joke to me.