aix1

> Anthropic's revenue is up 50% in the past two months

What's the source for such recent revenue numbers?

And 50% over what -- two months ago, same months last year etc?

For a few more sudoku techniques, see https://masteringsudoku.com/sudoku-solving-techniques/

Oh wow, I had no idea so had to Google it:

https://en.wikipedia.org/wiki/Bombings_in_Sweden

> They may want to rethink their reward structure a bit to avoid paying out to people like me.

Maybe, but who is Waymo to decide whether your trip was gratuitous? Perhaps you had brief business in Embarcadero and chose the particular mix of modalities for one reason or another (timing; or had to carry something heavy one way but not the other etc).

From a recent FT article

"Investors are questioning whether Anthropic can sustain its extraordinary growth as [...] the potential for AI to destroy humanity cloud[s] its blockbuster IPO."

What a time to be alive.

What makes you conclude that?

I just finished listening to A Man for All Markets (narrated by Thorp himself!)

Bitcoin derivatives arbitrage doesn't sound like something I'd personally be comfortable getting into, but I'd be curious to hear more about the types of things you're doing (to the extent you're happy to share).

Could someone explain the appeal of flow typing?

I can see how it can be useful to start with a broad type, e.g. a union, and narrow it down in a block. However, I don't quite get the opposite direction shown in their example (first an int, then a string, then a union).

Did you do a pixel diff against the original? The output could look plausible at a glance but contain hard-to-spot errors or inaccuracies.

I see, we're talking about different things.

My thought experiment was along the lines of "Let's say I'm Anthropic and I want to significantly improve my frontier model's performance on, say, theoretical physics research. How do I build a fully autonomous process capable of constructing an eval that's somewhat outside the current capability in some useful direction (decided by the autonomous process itself)?"

Would love to hear folks' ideas. :)

Would love to learn more about some techniques that "everybody" uses to do this well. So far, everything I've seen that meaningfully advances the frontier has been high-touch (involving human experts in one way or another).

Are you saying the models are already autonomously constructing next-gen evals for themselves? (Which is what the GP is asking.)

Also firmware updates (on the non-locked variants), no?

What other potential uses are there? Uploading books without WiFi is the only one that comes to mind.

One thing that jumped out at me is a slower-than-I'd-expect rate of adoption of Opus 5. If you're running Opus 4.8 and choosing not to migrate to 5, what's behind that decision?

These are GPU rental price futures, so you won't be investing in the hardware.

The pitch is that, if you know you'll need a certain amount of compute at a certain point in the future, you can lock in the rental price today.

https://www.cmegroup.com/media-room/press-releases/2026/8/11...

What have you been observing? Genuinely curious; I use Fable as my main model and haven't noticed any regression.

Yes they do. They have to file an S-1 with the SEC, which will be made public about a month before the IPO.

The S-1 has to include, among other things, three years of audited financial statements, plus interim statements (unaudited). It will cover both revenue and expenses, the latter breaking out things like cost of revenue, R&D, sales and marketing etc.

Based on the (unofficial but reported) IPO target date of late Sep to early Oct, the S-1 will have to be made public in a few weeks from now.

Here is one hypothesis: https://theconversation.com/problematic-paper-screener-trawl...

<quote> Have you ever heard of the Joined Together States? Or bosom peril? Kidney disappointment? Fake neural organizations? Lactose bigotry? These nonsensical, and sometimes amusing, word sequences are among thousands of “tortured phrases” that sleuths have found littered throughout reputable scientific journals.

They typically result from using paraphrasing tools to evade plagiarism-detection software when stealing someone else’s text. The phrases above are real examples of bungled synonyms for the United States, breast cancer, kidney failure, artificial neural networks, and lactose intolerance, respectively. </quote>

Amazing, thanks for sharing.

It still knows how to speak English. When I tell it to explain something in plain language, it generally does a very good job. The weird thing is that those instructions don't persist: it lapses back into Claude-speak pretty much every turn no matter how hard I try to instruct it not to.

(In my case "it"=Fable; I assume Opus is similar.)

With all due respect, did you read my comment beyond the first paragraph? It addresses both points, TPU economics/pivot to sales + internal shortages making it hard to train models, to the extent they can be addressed based on public sources.

There are other factors at play, but they're more recent/second-order.

The "let's make money by selling/renting out TPUs" faction has won and the "let's make money by training and selling a frontier model" faction has lost.

And it's arguably not crazy, at least if SemiAnalysis's estimates are to be believed:

  * 20% of all TPU shipments from Q3 2026 through Q4 2027 are sold to SPVs serving Anthropic ($150B of contracted revenue); vs
  * ~$12B ARR for Gemini.

https://newsletter.semianalysis.com/p/gemini-is-cooked-but-g...

Because they compete for the same scarce resource, the result is a resource crunch for the group that's lost: https://www.latimes.com/business/story/2026-05-18/inside-ai-...

No, relabelling a Pro model as Flash would make no economic sense (the Pro series is larger than Flash and more expensive to serve).

Google did not sell those TPUs because they had excess capacity. Google sold them because it can make lots more money that way then using them for Gemini inference (the shift to sales is also indicative of some power dynamics within the company).

Good point, thanks.

Having thought about this a little more, it's clear that server-side storage is not compatible with Zero Data Retention (ZDR). However, in non-ZDR settings, it seems likely that the providers are capturing all that data anyway?

> a per user key would have solved this issue for sure

It would have helped with PII leakage, but not with plain-text trace extraction attacks, right?

I really don't understand why server-side storage of the trace isn't a viable approach here, with only a unique key flowing to the client and back. Does it have something to do with how backend load-balancing works?

One detail I missed is that Koray's new title is "SVP of Google DeepMind"[1], i.e. identical to his old title.

So, basically, the third change is even more of a no-op than I thought.

It seems pretty clear that GDM will no longer have a CEO.

[1] https://blog.google/company-news/inside-google/message-ceo/n...

So there are two events here:

Event 1 - Jeff & co leaving.

Event 2 - Demis stepping down. DeepMind doesn't need a "chair" and "Alphabet's chief scientist" is a bullshit title that Jeff invented for himself when he moved from Google Research to GDM, to make it look like we wasn't abandoning the former for the latter.

The third change is basically ratifying the status quo, since Koray has been de factor running GDM for a while now (and directly reporting to Sundar in addition to reporting to Demis, who in turn also reported to Sundar - quite some triangle there).

I actually think there's a low probability of that.

The four have so much influence within the company that they could have trivially set this up as part of Alphabet, if they wanted to. They are also ridiculously wealthy.