Lenny's Podcast: Product | Career | Growth
00:54:05 · Lenny Rachitsky
[00:54:05] S1: Yeah. I think that point is so interesting, and it's so unique. There's always been this idea, release early, learn from users, get feedback, iterate. The fact that it's hard to even know what the AI is capable of, and how people will try to use it is a unique reason to start releasing things early. So, that'll help you as you exactly describe this idea of what is a latent demand in this thing that we didn't really know? Let's put it out there and see what people do with it.
00:54:30 · Boris Cherny
[00:54:30] S0: Yeah. And for Anthropic as a safety lab, the other dimension of that is safety, because when you think about model safety there's a bunch of different ways to study it. The lowest level is alignment, and mechanistic interpretability. So, this is when we train the model we want to make sure that it's safe. We, at this point, have pretty sophisticated technology to understand what's happening in the neurons to trace it. And so, for example, if there's a neuron related to deception, we're starting to get to the point where we can monitor it and understand that it's activating. And so, this is alignment, this is mechanistic interpretability, it's the lowest layer. The second layer is evals, and this is, essentially, a laboratory setting, the model is in a Petri dish, and you study it. And you put in the synthetic situation and just say, "Okay. Model, what do you do?" And, "Are you doing the right thing? Is it aligned? Is it safe?" And then the third layer is seeing how the model behaves in the wild. And as the model gets more sophisticated, this becomes so important, because it might look very good on these first two layers, but not great on the third one. We released Claude Code really early, because we wanted to study safety. And we actually used it within Anthropic for I think four or five months, or something before we released it, because we weren't really sure. Like, this is the first big agent that I think folks had released at that point. It was definitely the first coding agent that became broadly used. And so, we weren't sure if it was safe. And so, we actually had to study it internally for a long time before we felt good about that. And even since there's a lot that we've learned about alignment. There's a lot that we've learned about safety, that we've been able to put back into the model, back into the product. And for Cowork, it's pretty similar. The model is in this new setting. It's doing these tasks that are not engineering tasks. It's an agent that's acting on your behalf. It looks good on alignment, it looks good on evals, we tried it internally, it looks good. We tried it with a few customers, it looks good. Now we have to make sure it's safe in the real world. And so, that's why we release a little early. That's why we call it a research preview. But, yeah. It's constantly improving.
00:56:26 · Boris Cherny, Lenny Rachitsky
And this is really the only way to make sure that over the long-term the model is aligned, and it's doing the right things. [00:56:33] S1: It's such a wild space that you work in where there's this insane competition and pace. At the same time, there's this fear that if the God can escape and cause damage, and just finding that balance must be so challenging. What I'm hearing is there's these three layers, and I know there's ... This could be a whole podcast conversation is how you all think about the safety piece, but just what I'm hearing is there's these three layers you work with. There's observing the model thinking and operating. There's tests, evals that tell you this is doing bad things. And then releasing it early. I haven't actually heard a ton about that first piece. That is so cool. So, you guys can ... There's an observability tool that can let you peak inside the model's brain and see how it's thinking and where it's heading.
00:57:16 · Boris Cherny
[00:57:16] S0: Yeah. You should, at some point, have Chris Olah on the podcast, because he's just the industry expert on this. He invented this field of we call it mechanistic interpretability. And the idea is at its core what is your brand? What is it? It's a bunch of neurons that are connected. And so, what you can do is in a human brain or in an animal brain, you can study it at this mechanistic level to understand what the neurons are doing. It turns out surprisingly a lot of this does translate to models also. So, model neurons are not the same as animal neurons, but they behave similarly in a lot of ways. And so, we've been able to learn just a ton about the way these neurons work, about this layer or this neuron maps to this concept, how particular concepts are encoded, how the model does planning, how it thinks ahead. A long time ago we weren't sure if the model was just predicting the next token, or is doing something a little bit deeper. Now I think there's actually quite strong evidence that it is doing something a little bit deeper. And then the structures that let it do this are pretty sophisticated now where as the models get bigger it's not just a single neuron that corresponds to a concept. A single neuron might correspond to a dozen concepts, and if it's activated together with other neurons this is called super position. And together it represents this more sophisticated concept. And it's just something we're learning about all the time. And for Anthropic as we think about the way this space evolves doing this in a way that is safe and good for the world is just this is the reason that we exist, and this is the reason that everyone is at Anthropic. Everyone that is here, this is the reason why they're here. So, a lot of this work we actually open source. We publish it a lot. And we publish very freely to talk about this just so we can inspire other labs that are working on similar things to do it in a way that's safe. And this is something we've been doing for Claude Code also. We call this, "The race to the top" internally. And so, for Claude Code, for example, we released an open source sandbox. And this is a sandbox they can run the agent in, and just make sure that there's certain boundaries and they can't access, like, everything on your system.
00:59:20 · Boris Cherny, Lenny Rachitsky
And we made that open source, and it actually works with any agent, not just Claude Code, because we wanted to make it really easy for others to do the same thing. So, this is just the same principle of race to the top. We want to make sure this thing goes well, and this is the lever that we have. [00:59:37] S1: Incredible. Okay. I definitely want to spend more time on that. I will follow up with this suggestion. Something else that I've been noticing in the field across engineers and product managers, others that work with agents is there's this anxiety people feel when their agents aren't working. There's a sense that [inaudible 00:59:57] has a question [inaudible 00:59:58] answer, or it's, like, blocked on something, or, "There's all this productivity I'm losing. I need to wake up and get it going again." Is that something you feel? Is that something your team feels? Do you feel like this is a problem we need to track and think about?
01:00:11 · Boris Cherny, Lenny Rachitsky
[01:00:11] S0: I always have a bunch of agents running. So, at the moment I have five agents running. And, at any moment, I wake up and I start a bunch of agents. Like, the first thing I did when I woke up was like, "Oh, man. I really want to check this thing." So, I opened up my phone Claude iOS app code tab." Like, agent, do blah, blah, blah. Because I wrote some code yesterday and I was like, "Wait. Did I do this right?" I was double guessing something, and it was correct. But now it's just so easy to do this. So, I don't know. There is this little bit of anxiety maybe. I, personally, haven't really felt it just because I have agents running all the time. And I'm also just not locked into terminal anymore. Maybe a third of my code now is in the terminal, but also a third is using the desktop app. And then a third is the iOS app, which is just so surprising, because I did not think that this would be the way that I could even in 2026. [01:01:03] S1: I love that you describe it as coding still, which is just talking to Claude Code to code for you, essentially. And it's interesting that this is now coding. Coding now is describing what you want, not writing actual code. [01:01:16] S0: I wonder if the people that used to code using punch cards, or whatever, if you showed them software what they would have said. [01:01:22] S1: Isn't that great? Yeah.