overrun11

This sentiment aged extremely poorly with Anthropic's S1 now released. Turns out the "vibes" guys were right! I expect the goalpost will shift to claims that the S1 is fabricated in a larger than Enron fraud involving both the auditors and Anthropic.

Most Americans directly own stocks and a college graduate is even more likely to. This isn't the 1860's so a lot of these critiques of capitalism are anachronistic. The reality is "shareholders" are fairly ordinary people and not a tiny and mysterious group of elites.

> AI represents the ultimate dream of the capitalist: the elimination of the need of human labor entirely

Decreasing human toil for the same level of production should be the dream of _everyone_. If it's only capitalists in favor then that's a massive indictment of the non-capitalists.

This reminds me of the famous Bastiat quote: "If, then, the utility of any branch of industry is to be estimated not by the amount of satisfactions it is fitted to procure us with a determinate amount of labour, but, on the contrary, by the amount of labour which it exacts in order to yield us a determinate amount of satisfactions, what we ought evidently to desire is, that each acre of land should yield less corn, and each grain of corn less nourishment…"

The misunderstanding that labor and not production is the basis of prosperity leads to some pretty silly conclusions.

They aren't profitable on a GAAP basis and no one claims this. This obsession over profits is misguided. These are hyper growth companies growing at a scale never seen before. It is both deliberate and uncontroversial to invest in growth rather than slowing down to produce profits.

Yes that is exactly what is happening. OpenAI and Anthropic are the fastest growing companies by revenue ever and their gross profit margins are healthy.

Arguably nothing even has to change with training for this to be sustainable. Dario has claimed that Anthropic is profitable on a per training run basis. They aren't profitable because they choose to keep investing in increasingly large training runs.

> hypothetically, if a construction company sold houses for more than it cost to build them, that company could be considered profitable.

Construction companies capitalize and depreciate over many years so they can answer "yes" they are profitable even when they are very cashflow negative. This is exactly Dario's point: model training costs are treated as expenses but in practice are much closer to construction costs. Model training effectively produces an asset, the model weights, which will generate revenue for many years into the future.

> Zitron argues that this isn't going to happen because training is actually something that companies need to do to retain customers at all.

This is exactly why Dario's point about each training run being profitable is so important. It suggest that this is not true. Customers are happy to use old models long enough to fully pay off their costs.

> there's a long history of companies using creative accounting

Zitron seems to know very little about accounting evidenced by him using terms like "gross margin" wrong in this article. He's pattern matching against his limited exposure to company financials to find superficial similarities between the AI labs and famous frauds. Find me a company that doesn't report non-GAAP measures. Google search claims 96% of SP 500 companies do it. Are they all frauds too? Sometimes non-GAAP adjustments are eye roll inducing but they are tolerated because they can be genuinely useful to get a fuller picture of the business.

Google can't do any of these things

Gross margins and cost of revenue are well defined accounting terms that apply to any type of business.

> Does it include:

> Inference used for training? Modern training pipelines aren't just gradient descent, there's a ton of inference used in them too.

No because this is training and not inference. Just like how R&D costs for a drug aren't part of COGS either.

> Gradient descent itself?

No

> The CPUs and disks storing and managing the datasets?

Yes

> The web servers?

Yes

> The people paid to swap out failed components at the dc?

Yes to the extent they are swapping for inference and not training. If the same employees do both then the accountants will estimate what percent of their time is dedicated to each and adjust their cost accordingly.

The article is about compute cost though. By "lose money on inference" I mean the assertion that inference has negative gross margins which a lot of people truly believe. This is important because it's common to reason from this that LLM's are uneconomical and a ticking time bomb where prices will have to be jacked up several orders of magnitude just to cover the compute used for the tokens.

A huge number of people are convinced that OpenAI and Anthropic are selling inference tokens at a loss despite the fact that there's no evidence this is true and a lot of evidence that it isn't. It's just become a meme uncritically regurgitated.

This sloppy Forbes article has polluted the epistemic environment because now theres a source to point to as "evidence."

So yes this post author's estimation isn't perfect but it is far more rigorous than the original Forbes article which doesn't appear to even understand the difference between Anthropic's API costs and its compute costs.

> If they are, they should be compensated for the constant risk they bear.

I suspect that they are. US tech workers likely make dramatically more than the country you are from with better worker protections.

To the extent that this is even true it appears to be caused by three things: stock option compensation accounting, R&D deductions and bonus depreciation.

Stock option compensation rules have been a boon because Meta stock has risen 6x in three years. It's unlikely to do that again. My understanding is that this is symmetrical so if the stock trends down we will see an inflated effective tax rate for Meta.

Recent R&D rule changes allowing software engineering salaries for R&D to be written off seem reasonable and were quite popular on Hacker News. Previously these expenses were amortized over five years so this just pulls it forward. Subsequents years will see depressed expenses.

Bonus depreciation is once again just pulling forward legitimate expenses earlier than before. At worst they are just delaying giving the government its taxes and the corporations gain a few points of interest in between.

All of the tax rules used here are open to debate but none seem obviously wrong or nefarious. This is why people like Reich choose to keep things vague. Corporations brazenly stealing from your pocket is much more interesting than the mundane reality.