cloudbonsai

Your point is fair.

Another big component of the $80B loss was $3B in stock-based compensation. This line item is very expense-like (“instead of paying your drivers in cash, give them shares of your taxi company”), and Anthropic also adjusts it away when reporting their adjusted net income. Without this adjustment, their non-GAAP operating income was negative in 2026 1H.

Overall, I feel like a man in the parable of The Elephant in the Dark while writing this. Not only is my sight limited, but this elephant (Anthropic) is dancing wildly!

This is correct. Out of the $80.7B loss, $72.4B was due to changes in the fair value of their convertible notes (and similar financial arrangements).

Going a bit meta, I'm curious why we don't observe a faster convergence of opinions here. I assume that HN users are well-versed in AI tech, and the current generation of LLMs will happily break down the relevant information in Form S-1. Nevertheless, we see a chaotic storm of opinions, with wildly varying S/N ratios.

Is it a limit of current LLMs, or is the bottleneck perhaps human nature?

I want to express a very different perspective here as a counterpoint: I'm really grateful for the Signal team's work.

Today, I discussed my father's health condition with my family. Signal worked smoothly (even for my aging mother!), and it's free and secure.

Over the past five years, I have donated $500 to the foundation in total. I hope they continue to maintain the app into the far future.

I'm curious about your perspective on the impact of LLMs on physics.

My impression is that (theoretical?) physics is increasingly becoming a branch of abstract math. It has somewhat reached a practical limit for experimental verification -- perhaps the last landmark result was the confirmation of the Higgs boson 15 years ago?

So I see no reason LLMs cannot crank out a lot of theoretical physics papers, maybe combining forgotten/obscure theoretical branches explored in the past. After all, the most advanced theoretical physics is about internal consistency just as pure math is.

It seems that Meta assigns a unique hostname to each shared conversation. For example:

> fiscal-review-b4a5ff.cf.metaaiusercontent.com

So when someone opens the link, the DNS resolution exposes the unique hostname. The original bug reporter found dozens of these URLs in passive DNS logs.

I'm curious why such a large gap exists between our numbers. My hypotheses are:

A. The average utilization rate of inference GPU cards is not that high (i.e. it peaks during the daytime, but is idle at night). So the time window for earning $38/h is fairly limited.

B. They are renting costly GPU capacity from 3rdparty hyperscalers because they can't meet the peak demand with their own hardware.

C. Free-tier users are eating the gross margin.

D. Subscription users are eating the gross margin.

Another observable point is that DeepSeek's company-wide gross margin was reportedly 44.6%. It's intriguing what the full picture looks like here.

Your estimation feels too high. It implies >1000% ROI on a GPU card (= $300k gross annual profit on a $30k card).

As a reference point, Z.ai is operating at 26% gross margin as of 1H26. So assuming ~90% gross margin is likely a bit off.

Zebrafish have retinas very similar to mammals. For whatever reasons, they are able to regenerate their retinas when injured, while we mammals can't.

Scientists hypothesized that the presence of a certain protein is the reason why we can't, and they experimented with mice to see if blocking the protein would allow them to regenerate their retinas. It did.

What’s more interesting to me is the fact that these two kernels (con-leche and con-ron) were actually implemented by Claude (with “human supervision”).

Programming as a hand-crafted art is really diminishing fast.

They created an implementation of a Lean kernel in Lean (con-leche) and verified it as consistent.

Since there can be bugs in its compiler and runtime, they translated the code into Rust (con-ron), and verified the resulting kernel with another Rust-based verification tool.

There is also an ongoing verification project of Lean4, which is written in Lean4 itself. (My mental model is Lean4 := Raku and Lean := Perl)

So we have two verified kernels now, and another is awaiting release.

Yeah. I experimented with a random Japanese news article. Scanning it with my eyeballs, I found:

    Correct .. 80 chars
    Missing .. 12 chars
    Wrong ....  8 chars

So I guess the accuracy is roughly 80%. Maybe useful, but not super reliable.

Reading your comment, I’m curious if AI models will eventually “fork” mathematics. (Sorry for beating the same cranky analogy)

Like, not expecting human experts to review and merge their results, but developing math branches on their own. I guess it’s at least doable for applied math. For example, if AI agents find a fast method of prime factorization, they can demonstrate it by cracking RSA encryption. No human review required.

I’m less sure how it plays out for pure math, though. My intuition is fairly limited here.

Do you think it’s possible to operate math research in a fully autonomous way (i.e. with no human expert involvement)?

As a software engineer, my mental model so far is that OpenAI made a gigantic (but potentially sound) pull request to the math community, and the mathematicians just did not have enough bandwidth to review and merge it.

If my model is correct, I guess it makes a lot of sense to fund mathematicians, because their lack of resource is the bottleneck here.