Why Anthropic Hides Its AI R&D Speedup | Nathan Calvin
MTS
November 15, 2026
Nathan Calvin of Encode AI breaks down evidence from METR regarding AI R&D automation at Anthropic and discusses why the company keeps its research-acceleration methods guarded.
Nathan Calvin
I think the main reason why they didn't, which they said, is that they are afraid that sharing too much information about this will accelerate competitors, that if they basically talk about the ways that they are using AI to accelerate AI development in a lot of detail, then their competitors will do that. And I think they kind of say that they're both commercial reasons for that, if they just wanna keep their secret sauce and keep the valuation of the company, but also I think they justify it also with safety reasons saying that we are worried that the race intensifying or people catching up to us will give us less ability to invest in safety.
SPEAKER_2
And we're back.
SPEAKER_3
And we're back with Nathan Calvin in the flesh, who is the Chief Strategy Officer at Encode AI. Nathan, welcome back.
Nathan Calvin
My pleasure to be here in the studio. It's a lovely, lovely studio.
SPEAKER_3
So we were just talking, before we went on, we were talking about many things, but I think the main thing right now is that a lot of people are radically shortening their timelines based on evidence from the METR report about AI R&D automation at Anthropic. And at the same time, Anthropic, who themselves have admitted to increased rates of progress due to AI R&D automation, doesn't seem to be acting like it very much. So you actually had an exchange with Dario Amodei about this on Twitter in a very rare Twitter post by him.
What did you make of it?
Nathan Calvin
Yeah, I mean, my initial reaction was, I was just super, super surprised when he responded. I mean, we were going back and forth about some federal legislation that Anthropic had proposed. And the legislation just for all of the comments that Anthropic has made about how severe they think these risks are and how urgent and important oversight is. Like it was like preempted a bunch of state laws and made it really hard for the government to have additional powers to block the release of a model or block internal deployment. And so anyway, we went back and forth like 5,000 words. I don't know if anyone needs to read all of it, but I do think it was this kind of important question of like, what is the best way to understand what Anthropic is? Like, is it a big tech company trying to make money? Is it an AI safety organization trying to improve AI safety? Is it a company that's really just trying to like increase in the rates to transformative AI with the power and risks that brings?
SPEAKER_3
All of the love?
Nathan Calvin
Yeah, I think that's right. I think that they contain multitudes. Yeah.
SPEAKER_2
So I was briefly asking you, what do you speculate could be the evidence that METR has that was given by Anthropic as to why the rate of AI is increasing?
Nathan Calvin
Yeah. I mean, so they released in the public data this thing that basically shows the epoch capabilities index graphed over a bunch of the recent models, and you can basically see a tick up of the doubling time of capabilities becoming faster around, I think, around mythos. And so there's one explanation that I think Toby Ord has brought up, like, oh, maybe this is just because they did this, like, one really successful pre-train that has brought it, and it's not, honestly, because, like, AI is accelerating development. METR has gone in, and they put in the report that, like, they specifically examined and have evidence that rejected that hypothesis.
So, I mean, I think basically, like, if you view the two main types of ways that AI development can be accelerated, kind of one of it really is, um, for reasons of, of compute, and the other one is, is algorithmic improvements and efficiency. And so, this does seem to indicate to me that, uh, Anthropic is using AI to make its, um, uh, you know, the way in which, you know, it's designing new RL environments, their, their harnesses, like, all sorts of things, like, those, those things are much, much, much more efficient. But this is, this is speculation, and I'm not, I'm not sure. Um, I think that, uh, but, uh, yeah, it, it does seem like if METR is saying that it's not compute, then the other thing that it seems like it has to be is, is some, some kind of algorithmic progress.
SPEAKER_2
Yeah. Do you think that that's present in, like, Open AIs or on some other labs?
Yeah.
Nathan Calvin
I, I would, I don't, again, I, I don't know for sure. Um, I expect that Open AI is kind of closest on some of the, these measures of how aggressively they're, they're automating research. And then I think there's probably a decently large gap between the, the next one. Like I think my rough estimate is, you know, like it's Anthropic the most followed by Open AI, you know, maybe followed by Google or something. But I think Anthropic and Open AI seem to be, you know, they've released more information about it, just like the amount of, you know, agentic compute and bespoke systems and the amount that they've also invested in just like making their coding harnesses really good.
Read the other 16 minutes of this episode here.
$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.
Not what you expected? Full refund within 14 days.
Thousands of transcripts fetched by people building searchable podcast archives
Fetch transcripts with code
The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:
curl -H "x-api-key: pt_demo" \
https://spoken.md/transcripts/1000651996090Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.
From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.
Using your own key:
curl -H "x-api-key: YOUR_KEY" \
https://spoken.md/transcripts/1000802087081