erwald

The internal service is Microsoft Copilot, so this is probably just a really really dumb decision.

It's defined in the linked article?

Why do you think I used AI to write this? (I didn't.)

Yes, for that newsletter, where I write in my day job as a policy researcher, we use em dashes. For my personal writing (example: https://www.lesswrong.com/users/erich_grunewald), I use double dashes.

Haha, I do actually use double dash, as I write in org-mode. I then convert that org-mode text to Markdown and copy and paste the Markdown into the Substack editor, which preserves the double dash. In case it needs to be said, the article is 100% human-written.

To get the sure-to-be-correct answer to the question they were tasked with answering?

> Furthermore, export controls, for which Dario Amodei openly advocates, are already prompting China to build its own chips. China is, of course, behind the West, but not as much as you may think. Instead of adopting a Western technology stack and relying on our technology to build its AI models, China will increasingly rely on its own chips, reducing Western leverage.

China has been working hard to indigenize AI chip making since well before the October 2022 export controls. I don't think the controls had zero effect, but overall I think the effect is often very overstated, and anyway it's irreversible (there's no way China will stop its self-sufficiency efforts even if the AI chip ban is lifted; the CCP will always ensure there's a large enough domestic market carved out for every Huawei Ascend produced).

Cf. https://newsletter.decisiontreeresearch.com/p/how-much-did-u...

> Anthropic already possesses the compliance, legal, and evaluation infrastructure needed to satisfy such a regime. A university lab, startup, or community does not, even if they somehow manage to train a sufficiently capable model.

All the actual proposals that Dario Amodei has endorsed with regard to safety testing and so on have exempted smaller models. I don't think university labs or startups would ever be hit by any such regulation.

Bans on Chinese open weight models being used in the US and bans on AI chips and semiconductor manufacturing equipment being exported to China are two extremely different things, and it's not inconsistent in any way to oppose one and endorse another.

Lots of people paying for a product they use is more or less the opposite of an MLM

Right, and the comment I replied to was about revenue, not profit. (That said, while I don't think Anthropic is already profitable, it reportedly expects its first operating profit later this year.)

Those seem like reasonable questions about future margins and moats, but I was making the narrower point that revenue is in fact growing quickly, contra the comment above mine

I wasn't saying anything about costs, only that revenue is growing quickly

No, I wasn't claiming that revenue scales with headcount, though it probably does to some extent. The point is that these companies' revenue is large and growing quickly, which is what the comment above mine denied.

It's relevant to the claim I was replying to, which was that revenue isn't growing, not to the question of whether the spending will pay off

Maybe, but that's a different claim. You wrote that the improvements are "not translating to a dramatic increase in revenue", but going from about $1B to about $30B run rate in 16 months seems like a pretty dramatic increase to me!

The point is that revenue is large and growing quickly, which is what the comment above mine denied

That's a point about margins, not revenue

True, but the comment I was replying to was about revenue, not costs

"Anthropic and OpenAI generate a lot of revenue with relatively few employees – an estimated $9M and $5.5M in revenue per employee (RPE), respectively. If either company were to go public, it would have a higher RPE than any public tech company on Forbes’ Global 2000 list." https://epoch.ai/data-insights/revenue-per-employee-ai-compa...

Master craftsmen paid apprentices almost next to nothing, and they were often contractually guaranteed to stick around for many years, so the teaching was a kind of wage and also a cost that could be recuperated later on. (The apprentice even often had to pay the craftsman to take them on.) None of those things are true for junior software engineers, who are paid to contribute and can leave at any moment. Also, yes apprentices often had to do chores. It is just not analogous at all.

No, 2027 was never their median forecast; it was their modal forecast. See https://blog.aifutures.org/p/clarifying-how-our-ai-timelines...

It's more like it moved from 2028-2032 to 2030-2035 (depending on the author).

I think electricity doesn't matter that much (yet) because China is bottlenecked on chips. I think the incentives/directives to build on Huawei also doesn't matter that much yet because it's still such a small percentage of compute relative to NVIDIA, even for Chinese AI companies. (But this too could matter more from 2027-2030 and on.)

Yeah, to be clear I'm pretty excited about confidential computing and startups building on it, like Tinfoil, for some use cases. I just wanted to point out it's far from adequate for some important threat models (e.g., securing model weights for data centers located abroad, I think). (It's also not super widely adopted in AI yet, but that seems to be changing, at least for inference workloads.)

Where did you get this information? I think it's wrong -- I'm pretty sure they used the Export Administration Regulations (EAR) under the Export Control Reform Act (ECRA), which is under Commerce, not ITAR which is under the State Department. See for example https://harvardlawreview.org/blog/2026/06/is-access-to-fable...

Confidential computing is not secure against a potential attacker who has physical access to the hardware. The CC security guarantees explicitly assume the attacker has no physical access.