"Today I have the pleasure of speaking  with Dario Amodei, the CEO of Anthropic,   and I'm really excited about this one. Dario, thank you so much for coming   on the podcast. Thanks for having me.  Fi..."

"distribution. What's not clear at all is why does  it scale so smoothly with parameters? Why does it   scale so smoothly with the amount of data? You can think up some explanations of why   it's linea..."

"Does that imply that the circuit or  the process for doing addition was pre   existing and it just got increased in salience? I don't know if there's this circuit that's weak   and getting stronger. I..."

"we reach human level intelligence, looking  back on it, what would be your explanation?   What do you think is likely to be the  case if that turns out to be the outcome?  I would distinguish some pro..."

"that way for a number of reasons. But if you  told me — Yes, you trained your 2024 model.   It was much bigger and it just wasn't any better,  and you tried every architecture and didn't work,   that'..."

"basically saying something similar to this. This view that I have formed gradually from   2014 to 2017. My first experience with  it was my first experience with AI.   I saw some of the early stuff ar..."

"There are zillions of things people do with  machine learning. But I'm like, weird, this   seems to be true in the speech recognition field. It was just before OpenAI started that I met Ilya,   who yo..."

"as a counterexample, but I just thought,  well, it's hard to get data for robotics,   but if we look within the data that  we have, we see the same patterns.  I think people were very focused  on solv..."

"for generations to come. What explains  this discrepancy between super impressive   performance in these benchmarks and the things  you could describe versus general intelligence?  That was one area w..."

"Like, write a sonnet in the style of Cormac  McCarthy. I'm not very creative, so I couldn't   do that but that's a pretty high level human  skill. And even the model is starting to get   good at stuff..."

"an intelligence explosion or something? This kind of stuff is really hard to know   so I'll give that caveat. You can kind of predict  the basic scaling laws and then this more granular   stuff, which..."

"Let me ask about those thresholds. If Claude was  an employee at Anthropic, what salary would it be   worth? Is it meaningfully speeding up AI progress? It feels to me like an intern in most areas,   ..."

"the details, it's going to be weird and different  than we expect. That in all the detailed models,   we're thinking about the wrong things  or we're right about one thing,   and then are wrong about ..."

"I think it gets a little murky after that and  all of those thresholds may happen at various   times after that. But in terms of the base  technical capability of — it kind of sounds   like a reasonab..."

"workflow? How do you actually interact with it? It's very different to say, here's a chat bot that   looks like it's doing this task or helping the  human to do some task as it is to say, okay, this  ..."

"and it's going to happen fast. How do those different exponentials   which we've been talking about net out? One was the scaling laws themselves are   power laws with decaying marginal loss parameter ..."

"Again, I'm not making a normative statement  here. This is what should happen. I'm not   even saying this necessarily will happen because  there's important safety and government questions   here whic..."

"Shouldn't we be expecting that kind of stuff? I'm not sure. These words. Discovery. Creativity.   One of the lessons I've learned is that in the big  blob of compute, these ideas often end up being   ..."

"On that point. Last week in your Senate testimony,  you said that these models are two to three years   away from potentially enabling large scale bio  terrorism attacks. Can you make that more concre..."

"and they're not explicit knowledge. They're  more like, I have to do this lab protocol,   and what if I get it wrong? Oh, if this happens,  then my temperature was too low. If that happened,   I neede..."

"I'm excited about, but I believe it's happening. Somebody might say, listen, you were a co-author   on this post that OpenAI released about GPT-2  where they said, we're not going to release the   wei..."

"wouldn't say it's 100%. It could be 50-50. Okay, let's talk about cybersecurity,   which in addition to bio risk is another thing  Anthropic has been emphasizing. How have you   avoided the cloud micr..."

"competitors architecture’s leaking  is narrowly helpful to Anthropic,   it's not good for anyone in the long run.  Security around this stuff is really important.  Could you, with your current  securi..."

"Okay, let's talk about alignment and let's  talk about mechanistic interpretability,   which is the branch you guys specialize in. While  you're answering this question, you might want to   explain wh..."

"really understand what's going on inside the  models at the level of individual circuits.  Eventually when it's solved, what does the  solution look like? What is the case where   if you’re Claude 4, ..."

"more like an X-ray of the model than modification  of the model. It's more like an assessment than   an intervention. Somehow we need to get into  a dynamic where we have an extended test set,   an ex..."

"ability together in a way that actually works.  And not in a stupid way, there's lots of stupid   ways to do this where you fool yourself. I still don't feel like I understand the   intuition for why ..."

"I suspect that it may roughly work to think of  the model as if it's trained in the normal way,   just getting to above human level. It may be a  reasonable assumption, you should check, that   the in..."

"on one hand, I've actually always been a fan  of studying these circuits at the lowest level   of detail that we possibly can. And the reason for  that is that's kind of how you build up knowledge.   ..."

"thing. Others are starting to do  mechanistic interpretability now,   and I'm very glad that they are. A part  of our theory of change is paradoxically   to make other organizations more like us. I'm ..."

"things go wrong with them. I just feel like  you discover ten new ideas and ten new ways   that things are going to go wrong by trying these  in practice. I think that empirical learning is   just not..."

"The second option is you just find a way. You  just accept the trade offs. And the trade offs are   more positive than they appear because of a  phenomenon that I've called Race to the Top.   I could ..."

"problems as you mentioned but in the long scheme  of things, say 30 years down the line, which do   you think will be considered a bigger problem? I think it's going to be much less than 30 years.   I..."

"have the superhuman models, what does that look  like? Who are the right people? Who is actually   controlling the model five years from now? My view is that these things are powerful enough   that I ..."

"is a much narrower thing. This is something  that makes decisions for Anthropic. This is   basically a body. It was described in a recent  Vox article. We'll be saying more about it later   this year...."

"cure all the diseases, solve all the fraud –  things all humans would say, 'I'm down for that.'   But now it's 2030. You've solved all the real  problems that everybody can agree on. What happens   ne..."

"On the opposite end of things going well  or good actors having control of AI.   We might want to touch on China  as a potential actor in the space.  First of all, being at Baidu and seeing  progress ..."

"How do you think China thinks about AGI? Are  they thinking about safety and misuse or not?  I don't really have a sense. One concern I  would have are people saying things like,   China isn't going t..."

"I think it will defend against most attacks and  against a state level actor who's less determined.   But there's a lot more we need to do, and some of  it may require new research on how to do securi..."

"What is the timescale on which you think alignment  is solvable? If these models are getting to human   level in some things in two to three years,  what is the point at which they're aligned?  This i..."

"statistical systems and you can ask them a  million things and they can say a million things   and reply. And you might not have thought of a  millionth and one thing that does something crazy.   Or w..."

"it wants to turn them into pumpkins, just some  weird shit. Because the models are so powerful,   they're like these giants that are standing in  a landscape and if they start to move their arms   aro..."

"I do think that over the next two to three years  we're going to start eating that probability mass   of ways things can go wrong. It's like in the  core safety views paper, there's a probability   ma..."

"But the kind of thing that would really be  like, oh man, we can't solve this is like,   we see it happening inside the X-ray. I think  right now there's way too many assumptions,   there's way too mu..."

"that isn't actively optimizing against us. Let's talk about the specific methods other   than mechanistic interpretability that you  guys are researching. When we talk about   RLHF or Constitution AI,..."

"instead of telling you it can make the bioweapons. It's not that much of a concern with today's   passive models. If we were to fine tune a  model, we would do it privately and we work   with the expe..."

"more careful about what it is you're doing. On Constitution AI, who decides what the   constitution for the next generation of  models or a potentially superhuman model   is? How is that actually writ..."

"lessons from how societies work and how  politics works, that strikes me as fanciful.  Even after we've mitigated the safety issues,  any good future, even if it has all these   security issues that w..."

"kind of awareness as well as discovering stuff. It  was when I read that book that when I wrote this   big blob of compute doc and I only showed it to a  few people and there were other docs that I sh..."

"Regarding cybersecurity, what should  we make of the fact that there's a   whole bunch of tech companies which have  ordinary tech company security policy   and it's not obvious that they've been hack..."

"about cybersecurity is, it's not something you can  trumpet. A good dynamic with safety research is,   you can get companies into a dynamic and  I think we have, where you can get them   to compete to..."

"things are scaling up, we're anyway heading to a  world where the networks of data centers cost as   much as aircraft carriers. They're already going  to be pretty unusual objects but in addition to  ..."

"but there's reports that the human brain, from  the time it is born to the time a human being   is 20 years old, is on the order of 10^14  Flops to simulate all those interactions.  We don't have to g..."

"or maybe we'll understand why the discrepancy  is present, but at the end of the day, I don't   think it matters, right? If we keep scaling the  way we are. I think what's more relevant at this   poin..."

"happened in a far enough distant past, then it's  like the compute doesn't flow. The spice doesn't   flow. The blob has to be unencumbered. It's not  going to work if you artificially close things   o..."

"it's being blocked until you need to free it up. I love the gradients changing that to spice.   On that point, though, do you think that another  thing on the scale of a transformer is coming down   t..."

"And then the other possibilities you  mentioned. RL, you can see it as...  We kind of already do RL with RLHF. Is this  alignment? Is this capabilities? I always think   in terms of the two snakes, th..."

"of the founders and the employees at Anthropic  are physicists. We talked in the beginning about   the scaling laws and how the power laws  from physics are something you see here,   but what are the ..."

"Are you concerned that there's a lot of people  who would have been doing physics or something,   they would’ve gone into finance  instead and since Anthropic exists,   they have now been recruited to..."

"Do you think that Claude has conscious  experience? How likely do you think that is?  This is another of these questions that just  seems very unsettled and uncertain. One thing   I'll tell you is I u..."

"What would change if you found out that they  are conscious? Are you worried that you're   pushing the negative gradient to suffering? Conscious, again, is one of these words that   I suspect will not..."

"We talked about this initially, but I want to  get more specific. We talked initially about   now that you're seeing these capabilities ramp  up within the human spectrum, you think that the   human s..."

"In terms of watching what the models can do, how  has it changed my view of human intelligence? I   wish I had something more intelligent to say  on that. One thing that's been surprising   is I thoug..."

"other surprising and interesting thing is, many  years from now, it'll be one of those things that   you’ll wonder why it wasn't obvious to you?  If you're seeing these smooth scaling curves,   why we..."