Head of Claude Code: What happens after coding is solved | Boris Cherny

100% of my code is written by Claude Code.

I have not edited a single line by hand since November.

Every day I ship 10, 20, 30 pull requests.

So like at the moment I have like 5 agents running.

While we're recording this?

Yeah, yeah.

Do you miss writing code?

I have never enjoyed coding as much as I do today because I don't have to deal with all the minutiae.

Productivity per engineer has increased 200%.

There's always this question, should I learn to code?

In a year or two, it's not going to matter.

Coding is largely solved.

I imagine a world where everyone is able to program.

Anyone can just build software anytime.

What's the next big shift to how software is

and Claude is starting to come up with ideas, looking through feedback.

It's looking at bug reports, it's looking at telemetry for bug fixes and things to ship a little more like a coworker or something like that.

A lot of people listening to this are product managers and they're probably sweating.

I think by the end of the year everyone's going to be a product manager and everyone codes.

The title software engineer is going to start to go away.

It's just going to be replaced by builder and it's going to be painful for a lot of people.

Today my guest is Boris Cherny, head of Claude Code at Anthropic.

It is hard to describe the impact that Cloud Code has had on the world.

Around the time this episode comes out will be the 1-year anniversary of Cloud Code.

And in that short time, it has completely transformed the job of a software engineer.

And it is now starting to transform the jobs of many other functions in tech, which we talk about.

Cloud Code itself is also a massive driver of Anthropic's overall growth over the past year.

They just raised a round at over $350 billion, And as Boris mentions, the growth of Cloudcode itself is still accelerating.

Just in the past month, their daily active users has doubled.

Boris is also just a really interesting, thoughtful, deep-thinking human.

And during this conversation, we discover we were born in the same city in Ukraine.

That is so funny.

I had no idea.

A huge thank you to Ben Mann, Jenny Wen, and Mike Krieger for suggesting topics for this conversation.

Don't forget to check out Lenny's Product Pass dot com.

For an incredible set of deals available exclusively to Lenny's newsletter subscribers.

Let's get into it after a short word from our wonderful sponsors.

Today's episode is brought to you by DX, the developer intelligence platform designed by leading researchers.

To thrive in the AI era, organizations need to adapt quickly, but many organization leaders struggle to answer pressing questions like, which tools are working?

How are they being used?

What's actually driving value?

DX provides the data and insights that leaders need to navigate this shift.

With DX, companies like Dropbox, Booking.com, Atlassian, and Intercom get a deep understanding of how AI is providing value to their developers and what impact AI is having on engineering

productivity.

To learn more, visit DX's website at getdx.com/Lenny.

That's getdx.com/Lenny.

Applications break in all kinds of ways: crashes, slowdowns, regressions, and the stuff that you only see once real users show up.

Sentry catches it all.

See what happened, where and why, down to the commit that introduced the error, the developer who shipped it, and the exact line of code, all in one connected view.

I've definitely tried the 5 tabs and Slack thread approach to debugging.

This is better.

Sentry shows you how the request what ran, what slowed down, and what users saw.

Seer, Sentry's AI debugging agent, takes it from there.

It uses all of that Sentry context to tell you the root cause, suggest a fix, and even opens a PR for you.

It also reviews your PRs and flags any breaking changes with fixes ready to go.

Try Sentry and Seer for free at sentry.io/Lenny and use code Lenny for $100 in Sentry credits.

That's s-c-n-t-r-y.io/Lenny.

Boris, thank you so much for being here and welcome to the podcast.

Yeah, thanks for having me on.

I want to start with a spicy question.

About 6 months ago, I don't know if people even remember this, you actually left Anthropic, you joined Cursor, and then 2 weeks later you went back to to Anthropic.

What happened there?

I don't think I've ever heard the actual story.

It's the fastest job change that I've ever had.

I joined Cursor because I'm a big fan of the product and honestly, I met the team and I was just really impressed.

They're an awesome team.

I still think they're awesome and they're just building really cool stuff and kind of, they saw where AI coding was going, I think before a lot of people did.

So the idea of building good product was just very exciting for me.

I think as soon as I got there, what I started to realize is what I really missed about Ant was the mission.

And that's actually what originally drove me to Ant also.

Because before I joined Anthropic, I was working in big tech and then I was, at some point I wanted to work at a lab to just help shape the future of this crazy thing that we're building

in some way.

And the thing that drew me to Anthropic was the mission.

And it was, you know, it's all about safety.

And when you talk to people at Anthropic, just like find someone in the hallway, if you ask them why they're here, the answer is always going to be safety.

And so this kind of like mission-drivenness just really, really resonated with me.

And I just know personally it's something I need in order to be happy.

And that's just the thing that I really missed.

And I found that, you know, whatever the work might be, no matter how exciting, even if it's building a really cool product, it's just not really a substitute for that.

Um, so for me, it was actually, uh, it was pretty obvious that, that I was missing that pretty quick.

Okay.

So let me follow the thread of just coming back to Anthropic and the work you've done there.

This podcast is going to come out around the year anniversary of launching Claude Code.

So I want to spend a little time just reflecting on the impact that you've had.

There's, um, this report that recently came out that I'm sure you saw by Semanalysis that showed that 4% of all GitHub commits are Authored by Cloud Code Now, and they predicted it'll

be a, a fifth of all code commits on GitHub by the end of the year.

The way they put it is, while we blinked, AI consumed all software development.

The day that we're recording this, Spotify just put out this, uh, headline that their best developers haven't written a line of code since December, thanks to AI.

More and more of the most advanced senior engineers, including you, are sharing the fact that you don't write code anymore, that it's all AI generated, and many aren't even looking

at code anymore is how far we've gotten.

In large part, thanks to this little project that you started and that your team has scaled over the past year.

I'm curious just to hear your reflections on, on this past year and the impact that your work has had.

These numbers are just totally crazy, right?

Like 4% of all commits in the world is just way more than I imagined.

And like you said, it still feels like the starting point.

These are also just public commits.

So we actually think if you look at private repositories, it's quite a bit higher than that.

And I think the craziest thing for me isn't even the number that we're at right now, but the pace at which we're growing, because if you look at QuadCode's growth rate kind of across

any metric, it's continuing to accelerate.

So it's not just going up, it's going up faster and faster.

When I first started QuadCode, it was just gonna be a, like, it was just supposed to be a little hack.

Broadly knew at Anthropic that we wanted to ship some kind of coding product.

And for Anthropic, for a long time, we were building the models in this way that kind of fit our mental model of the way that we build safe AGI, where the model starts by being really

good at coding, then it gets really good at tool use, then it gets really good at computer use.

Roughly, this is the trajectory.

And we've been working on this for a long time.

And when you look at the team that I started on, it was called the Anthropic Labs team.

And actually Mike Krieger and Ben Mann, they just kicked this team off again for kind of round 2.

The team built some pretty cool stuff.

So we built QuadCode, we built MCP, we built the desktop app.

So you can kind of see the seeds of this idea.

You know, like it's coding, then it's tool use, then it's computer use.

And the reason this matters for Anthropic is because of safety.

It's kind of, again, just back to that.

AI is getting more and more powerful.

It's getting more and more capable.

The thing that's happened in the last year is that for, at least for engineers, the AI doesn't just write the code.

It's not just a conversation partner, but it actually uses tools.

It acts in the world.

And I think now with Cowork, we're starting to see the transition for non-technical folks also.

For a lot of people that use conversational AI, this might be the first time that they're using a thing that actually acts.

It can actually use your Gmail.

It can use your Slack.

It can do all these things for you and it's quite good at it and it's only going to get better from here.

So I think for Anthropic, for a long time there was this feeling that we wanted to build something, but it wasn't obvious what.

And so when I joined Anthropic, I spent one month kind of hacking and built a bunch of weird prototypes.

Most of them didn't ship and weren't even close to shipping.

It was just kind of understanding the boundaries of what the model can do.

Then I spent a month doing post-training.

So to understand kind of the research side of it.

And I think honestly, that's just for me as an engineer, I find that to do good work, you really have to understand the layer under the layer at which you work.

And with traditional engineering work, if you're working on product, you want to understand the infrastructure, the runtime, the virtual machine, the language, kind of whatever that

is, the system that you're building on.

But yeah, if you're working in AI, you just really have to understand the model to some degree to do good work.

So I took a little detour to do that and then I came back and just started prototyping what eventually became QuadCode.

And the very first version of it, I have like a, there's like a video recording of this somewhere, 'cause I recorded this demo and I posted it.

It was called Quad CLI back then.

And I just kind of showed off how it used a few tools.

And the shocking thing for me was that I gave it a bash tool and it just was able to use that to write code.

To tell me what music I'm listening to when I asked it like, what music am I listening to?

And this is the craziest thing, right?

Because it's like, there's no, I didn't instruct the model to say, you know, use, you know, this tool for this or kind of do whatever.

The model was given this tool and it figured out how to use it to answer this question that I had that I wasn't even sure if it could answer, what music am I listening to?

And so I started prototyping this a little bit more and made a post about it and I announced it internally and it got 2 likes.

That's the, that was the extent of the reaction at the time.

Because I think people internally, you know, like when you think of coding tools, you think of like, you think of IDEs, you think about kind of all these pretty sophisticated environments.

No one thought that this thing could be terminal-based.

That's sort of a weird way to design it.

And that wasn't really the intention.

But, you know, from the start, I built it in a terminal because, you know, for the first couple months, it was just me.

So it was just the easiest way to build.

Uh, and for me, this is actually a pretty important product lesson, right?

It's like you want to under-resource things a little bit at the start.

Then we started thinking about what other form factors we should build, and we actually decided to stick with the terminal for a while.

And the biggest reason was the model is improving so quickly.

We felt that there wasn't really another form factor that could keep up with it.

And honestly, this was just me kind of like struggling with kind of like, what should we build?

You know, like for the last year, quadcode has just been all I think about.

And so just like late at night, this is just something I was thinking about, like, okay, the model's continuing to improve.

What do we do?

How can we possibly keep up?

And the terminal was honestly just the only idea that I had.

And yeah, it ended up catching on.

After I released it pretty quickly, it became a hit at Anthropic and, you know, the daily active users just went vertical.

And really early on, actually, before I launched it, Ben Mann, nudged me to make a DAU chart.

And I was like, you know, it's like kind of early, maybe, you know, should we really do it right now?

And he was like, yeah.

And so the chart just went vertical pretty immediately.

And then in February, we released it externally.

Actually, something that people don't really remember is Cloud Code was not initially a hit when we released it.

It got a bunch of users.

There was a lot of early adopters that got it immediately, but it actually took many months for everyone to really understand what this thing is.

Just again, it's like, it's just so different.

And when I think about it, kind of part of the reason QuadCode works is this idea of latent demand, where we bring the tool to where people are and it makes existing workflows a little

bit easier.

But also because it's in a terminal, it's like a little surprising.

It's a little alien in this way.

So you have to, you have to kind of be open-minded and you have to learn to use it.

And of course now, you know, QuadCode is available, you know, in the iOS and Android Quad app.

It's available in the desktop app.

It's available on the website.

It's available as IDE extensions and Slack and GitHub, you know, all these places where engineers are.

It's a little more familiar, but that wasn't the starting point.

So yeah, I mean, at the beginning it was kind of a surprise that this thing was even useful.

And, you know, as the team grew, as the product grew,

as it started to become more and more useful to people, just people around the world from, you know, small startups to the biggest FAANG companies started using it.

They started giving feedback.

And I think just reflecting back, it's been such a humbling experience because we just, we keep learning from our users.

And just the most exciting thing is like, you know, none of us really know what we're doing.

And we're just trying to figure it out along with everyone else.

And the single best signal for that is just feedback from users.

So that's just been the best.

I've been surprised so many times.

It's incredible how fast something can change in today's world.

You launched this a year ago and it wasn't the first time people could use AI to code, but, uh, in a year, the entire profession of software engineering has dramatically changed.

Like, there's all these predictions.

Oh, AI is going to be written, 100% AI's, uh, our code is going to be written by AI.

Everyone's like, no, that's crazy.

What are you talking about?

Now it's like, oh, of course it's happening exactly as they said.

It's just so things move so fast and change so fast now.

Yeah, it's really fast.

Back at, uh, back at Code with Quad back in May, that was like our first, uh, you know, like developer conference that we did as Anthropic.

I did a short talk and in the Q&A after the talk, people were asking, what are your predictions for the end of the year?

And my prediction back in May of 2025 was by the end of the year, you might not need an IDE to code anymore.

And we're going to start to see engineers not doing this.

And I remember the room audibly gasped.

It was such a crazy prediction.

But I think at Anthropic, this is just the way we think about things as exponentials.

And this is very deep in the DNA.

If you look at our co-founders, 3 of them were the first 3 authors on the scaling laws paper.

So we really just think in exponentials.

And if you kind of look at the exponential of the percent of code that was written by Claude at that point, if you just trace the line, it's pretty obvious we're gonna cross 100% by

the end of the year, even if it just does not match intuition at all.

And so all I did was trace the line and yeah, in November that, you know, that happened for me personally and that's been the case since.

And we're starting to see that for a lot of different customers too.

I thought it was really interesting what you just shared there about kind of the journey.

Is this kind of idea of just playing around and seeing what happens.

This comes up with OpenClaw a lot, just like Peter was playing around and just like a thing happened.

And it feels like that's a central kind of ingredient to a lot of the biggest innovations in AI is people just sitting around trying stuff to pushing the models further than most other

people.

I mean, that's the thing about innovation, right?

Like you can't force it.

There's no roadmap for innovation.

You just have to give people space.

You have to give them, maybe the word is like safety.

So it's like psychological safety that it's okay to fail.

It's okay if 80% of the ideas are bad.

You also have to hold them accountable a bit.

So if the idea is bad, you know, you cut your losses, move on to the next idea instead of investing more.

In the early days of QuadCode, I had no idea that this thing would be useful at all.

Because even in February when we released it, it was writing maybe, I don't know, like 20% of my code, not more.

And even in May, it was writing maybe 30%.

I was still using, you know, cursor for most of my code.

And it only crossed 100% in November.

So it took a while, but even from the earliest day, it just felt like I was onto something and I was just spending like every night, every weekend hacking on this.

And luckily my, you know, my wife was very supportive.

Um, but it, it, it just felt like it was onto something.

It wasn't obvious what, and, and sometimes, you know, you find a thread, you just have to pull on it.

So at this point, 100% of your code is written by Claude Code.

Is that, is that kind of the current state of your coding?

Yeah.

So 100% of my code is written by Claude Code.

Um, I'm a fairly prolific coder.

Um, and this has been the case even when I worked back at Instagram.

I was like one of the top few most productive engineers.

Um, and that's actually, that's still the case, uh, here at Anthropic.

Wow.

Even as head of the team.

Yeah.

Yeah.

Do still do a lot of coding.

Um, and so every, you know, every day I ship like 10, 20, 30 pull requests, something like that.

Every day.

Every day.

Yeah.

Good God.

Uh, 100% written by quad code.

I have not edited a single line by hand.

Since November.

And yeah, that's been it.

I do look at the code.

So I don't think we're kind of at the point where you can be totally hands-off, especially when there's a lot of people, you know, like running the program, you have to make sure that

it's correct, you have to make sure it's safe and so on.

And then we also have Claude doing automatic code review for everything.

So here at Anthropic, Claude reviews 100% of pull requests.

There's still a layer of like human review after it.

But you kind of like, you still do want some of these checkpoints.

Like you still want a human looking at the code, unless it's like pure prototype code that, you know, it's not going to run.

It's not going to run anywhere.

It's just a prototype.

What's kind of the next frontier?

So at this point, 100% of your code is being written by AI.

This is clearly where everyone is going in software engineering.

That felt like a crazy milestone.

Now it's just like, of course, this is the world now.

What's, what's kind of the next big shift to how software is written that either your team's already operating in or you think will head towards?

I think something that's happening right now is Claude is starting to come up with ideas.

So Claude is looking through feedback, it's looking at bug reports, it's looking at

telemetry and things like this, and it's starting to come up with ideas for bug fixes and things to ship.

So it's just starting to get a little more, you know, like a little more like a coworker or something like that.

I think the second thing is we're starting to branch out of coding a little bit.

So I think at this point, it's safe to say that coding is largely solved.

At least for the kinds of programming that I do, it's just a solved problem because Claude can do it.

And so now we're starting to think about, okay, like what's next?

What's beyond this?

There's a lot of things that are kind of adjacent to coding.

And I think this is going to be coming, but also just, you know, general tasks, you know, like I use Cowork every day now to do all sorts of things that are just not related to coding

at all and just to do it automatically.

Like for example, I had to pay a parking ticket the other day.

I just had Cowork do it.

All of my project management for the team, Cowork does all of it.

It's like syncing stuff between spreadsheets and messaging people on Slack and email and all this kind of stuff.

So I think the frontier is something like this.

And I don't think it's coding because I think coding is, you know, it's pretty much solved.

And over the next few months, I think what we're going to see is just across the industry, it's going to become increasingly solved, you know, for every kind of codebase, every tech

stack that people work on.

This idea of helping you come up with what to work on is so interesting.

A lot of people listening to this are product managers and they're probably sweating.

How do you use Claude for this?

Do you just talk to it?

Is there anything clever you've come up with to help you use it to come up with what to build?

Honestly, the simplest thing is like open Claude Code or Cowork and point it at a Slack thread.

Um, you know, like for us, we have this channel that, that's all the internal feedback about Claude Code since we first released it.

Even in like 2024 internally, it's just been this fire hose of feedback.

And it's the best.

And like in the early days, what I would do is anytime that someone sends feedback, I would just go in and I would fix every single thing as fast as I possibly could.

So like within a minute, within 5 minutes or whatever.

And this just really fast feedback cycle, it encourages people to give more and more feedback.

It's just so important because it makes them feel heard.

Because you know, like usually when you use a product, you give feedback, it just goes into a black hole somewhere and then you don't get feedback again.

So if you make people feel heard, then they want to contribute and they want to help make the thing better.

And so now I kind of do the same thing, but Quad honestly does a lot of the work.

So I pointed at the channel and it's like, okay, here's a few things that I can do.

I just put up a couple PRs, want to take a look at them.

And I'm like, yeah.

Have you noticed that it is getting much better at this?

Because this is kind of the holy grail right now.

It's like, cool, building solved.

Code review became kind of the next bottleneck with all these PRs.

Who's going to review them all?

The next big open question is just like, okay, now we need to, now humans are necessary for figuring out what to build, what to prioritize.

And you're saying that that's where Claude code is starting to help you.

Has it gotten a lot better with like, say, Opus 4.6 or what's been the trajectory there?

Yeah, yeah, it's improved a lot.

I think some of it is kind of like training that we do specific to coding.

So, you know, obviously, you know, best coding model in the world and, you know, it's getting better and better.

4.6 is just incredible.

But also actually a lot of the training that we do outside of coding translates pretty well too.

So there is this transfer where you teach the model to do X and it gets better at Y.

Yeah, and the gains have just been insane.

At Anthropic over the last year, since we introduced QuadCode, we probably, I don't know the exact number, we probably 4x the engineering team or something like this, but productivity

per engineer has increased 200%.

In terms of like pull requests.

And like this number is just crazy for anyone that actually works in the space and works on dev productivity.

Because back in a previous life, I was at Meta and, you know, one of my responsibilities was code quality for the company.

So this is like all of our code bases, that was my responsibility, like Facebook, Instagram, WhatsApp, all this stuff.

And a lot of that was about productivity because if you make the code higher quality, then engineers are more productive.

And things that we saw is, you know, in a year, with hundreds of engineers working on it, you would see a gain of like a few percentage points of productivity, something like this.

Um, and so nowadays seeing these gains of just hundreds of percentage points, it's, it's just absolutely insane.

What's also insane is just how normalized this has all been.

Like we hear these numbers, like, of course AI is doing this to us.

It's just, it's so unprecedented.

The amount of change that is happening to software development, to building products, to just this, the world of tech.

It's just like so easy to get used to it, but it's important to recognize this is crazy.

This is something like I have to remind myself once in a while.

There's sort of like a downside of this because the model changes.

So there's actually like, there's many kinds of downsides that we could talk about.

But I think one of them on a personal level is the model changes so often that I sometimes get stuck in this like old way of thinking about it.

And I even find that like new people on the team or even new grads that join, do stuff in a more kind of like AGI-forward way than I do.

So like sometimes, for example, I had this case like a couple months ago where there was a memory leak.

And so like what this is, is, you know, like quad code, the memory usage is going up and at some point it crashes.

This is like a very common kind of engineering problem that, you know, every engineer has debugged 1,000 times.

And traditionally the way that you do it is you take a heap snapshot, you put it into a special debugger, you kind of figure out what's going on, you know, use these special tools to

see what's happening.

And I was doing this and I was kind of like looking through these traces and trying to figure out what was going on.

And the engineer that was newer on the team just had Quadco do it.

And it was like, hey Quad, it seems like there's a leak, can you figure it out?

And so like Quadco did exactly the same thing that I was doing.

It took the heap snapshot, it wrote a little tool for itself so it can kind of like analyze it itself.

It was sort of like a just-in-time program.

And it found the issue and put up a pull request faster than I could.

So it's something where like, for those of us that have been using the model for a long time, you still have to kind of transport yourself to the current moment and not get stuck back

in an old model because it's not Sonnet 3.5 anymore.

The new models are just completely, completely different.

And just this mindset shift is very different.

I hear you have these very specific principles that you've codified for your team that when people join you, You kind of walk them through them.

I believe one of them is, what's better than doing something?

Having Claude do it.

And it feels like that's exactly what you described with this memory leak is just like, you almost forgot that principle of like, okay, let me see if Claude can solve this for me.

There's this, uh, there's this interesting thing that happens also when you, um, when you underfund everything a little bit, uh, because then people are kind of forced to Claudefy.

And this is something that we see.

So, you know, for work where sometimes we just put like one engineer on a project.

And the way that they're able to ship really quickly, because they want to ship quickly, this is like an intrinsic motivation that comes from within.

It's just wanting to do a good job.

If you have a good idea, you just really want to get it out there.

No one has to force you to do that.

That comes from you.

And so if you have Claude, you can just use that to automate a lot of work.

And that's kind of what we see over and over.

So I think that's kind of like one principle is underfunding things a little bit.

I think another principle is just encouraging people to go faster.

So if you can do something today, you should just do it today.

And this is something we, we really, really encourage on the team.

Early on, it was really important because it was just me.

And so our only advantage was speed.

That's the only way that we could ship a product that would compete in this very crowded coding market.

But nowadays it's still very much a principle we have on the team.

And if you want to go faster, a really good way to do that is to just have Claude do more stuff.

Um, so it just very much encourages that.

This idea of underfunding, it's so interesting because in general, there's this feeling like AI is going to allow you to not have as many employees, not have as many engineers.

And so it's not only you can be more productive.

What you're saying is that you will actually do better if you underfund.

It's not just that AI can make you faster.

It's you will get more out of the AI tooling if you have fewer people working on something.

Yeah.

If you, if you hire great engineers, they'll figure out how to do it.

And, uh, especially if you empower them to do it.

This is something I actually talk a lot about with CTOs and all sorts of companies.

My advice generally is don't try to optimize, don't try to cost cut at the beginning.

Start by just giving engineers as many tokens as possible.

And now you're starting to see companies like, at Anthropic, everyone can use a lot of tokens.

We're starting to see this come up as a perk at some companies.

If you join, you get unlimited tokens.

This is a thing I very much encourage because

it makes people free to try these ideas that would've been too crazy.

And then if there's an idea that works, then you can figure out how to scale it.

And that's the point to optimize and to cost-cut, figure out maybe you can do it with Haiku or with Sona instead of Opus or whatever.

But at the beginning, you just want to throw a lot of tokens at it and see if the idea works and give engineers the freedom to do that.

So the advice here is

just be loose with your tokens, with the cost on using these models.

People hearing this may be like, of course he works at Anthropic, you want us to use as many tokens as possible.

But what you're saying here is the most interesting innovative ideas will come out of someone just kind of taking it to the max and seeing what's possible.

Yeah.

And I think the reality is like at small scale, like, you know, you're not going to get like a giant bill or anything like this.

Like if it's an individual engineer experimenting, the token cost is still probably relatively low relative to their salary or, you know, other costs of running the business.

So it's actually not a huge cost.

As the thing scales up, so let's say they build something awesome and then it takes a huge amount of tokens and then the cost becomes pretty big, that's the point at which you want

to optimize it.

But don't do that too early.

Have you seen companies where their token cost is higher than their salary?

Is that a trend you think we're going to find and see?

At Anthropic, we're starting to see some engineers that are spending hundreds of thousands a month in tokens.

So we're starting to see this a little bit.

There's some companies that are, we're starting to see similar things.

Yeah.

Going back to coding, do you miss writing code?

Is this something you're kind of sad about that this is no longer a thing you will do as a software engineer?

It's funny for me, you know, like when I learned engineering, for me it was very practical.

I learned engineering so I could build stuff.

And for me, I was self-taught, you know, like I studied economics in school, but I didn't study CS, but I taught myself engineering kind of early on.

I was programming in like middle school.

And from the very beginning, it was very practical.

So I actually like, I've learned to code so that I can cheat on a math test.

That was like the first thing.

We had these like graphing calculators and the, you know, I just programmed the answer into the— TI-83?

TI-83 Plus.

Yeah, yeah, exactly.

Plus.

Plus, yeah.

So I programmed the answers in and then, The next math test, whatever, the next year, it was just too hard.

I couldn't program all the answers in because I didn't know what the questions were.

And so I had to write a little solver so that it was a program that would just solve these algebra questions or whatever.

And then I figured out you can get a little cable, you can give the program to the rest of the class and then the whole class gets A's.

But then we all got caught and the teacher told us to knock it off.

But from the very beginning, it's always just been very practical for me where programming is a way to build a thing.

It's not the end in itself.

At some point, I personally fell into the rabbit hole of kind of like the beauty of programming.

So like I wrote a book about TypeScript.

I started the, actually at the time it was the world's biggest TypeScript meetup, just because I fell in love with the language itself.

And I kind of got deep into like functional programming and all this stuff.

I think a lot of coders, they get distracted by this.

For me, it was always sort of,

there is a beauty to programming and especially to functional programming.

There's a beauty to type systems.

There's a certain kind of like this like buzz that you get like when you solve like a really complicated math problem.

It's kind of similar when you kind of balance the types or, you know, the program is just like really beautiful, but it's really not the end of it.

I think for me, coding is very much a tool and it's a way to do things.

That said, not everyone feels this way.

So for example, you know, like there's one engineer on the team, Lena, who, you know, was still writing C++ on the weekends by hand, because, you know, for her, she just really enjoys

writing C++ by hand.

And so everyone is different.

And I think even as this field changes, even as everything changes, there's always space to do this.

There's always space to enjoy the art and to kind of do things by hand if you want.

Do you worry about your skills atrophying as an engineer?

Is that something you worry about or is it just like, you know, this is just the way it's going to go?

I think it's just the way that it happens.

I don't worry about it too much personally.

I think for me, like programming is on a continuum and, you know, like way back in the day, you know, like software actually is like relatively new, right?

Like if you look at the way programs are written today, like using software that's running on a virtual machine or something, this has been the way that we've been writing programs

since probably the 1960s.

So, you know, it's been, you know, like 60 years or something like that.

Before that, it was punch cards.

Before that, it was switches.

Before that, it was hardware.

And before that, it was just, you know, like literally pen and paper.

It was like a room full of people that were doing math on paper.

And so, you know, programming has always changed in this way.

In some ways, you still want to understand the layer under the layer because it helps you be a better engineer.

And I think this will be the case maybe for the next year or so.

But I think pretty soon it just won't really matter.

It's just going to be kind of like the the assembly code running under the program or something like this.

At an emotional level, you know, I feel like I've always had to learn new things.

And as a programmer, it's actually not, it doesn't feel that new because there's always new frameworks, there's always new languages.

It's just something that we're quite comfortable with in the field.

But at the same time, I, you know, this isn't true for everyone.

And I think for some people, they're going to feel a greater sense of, I don't know, maybe like loss or nostalgia or atrophy or something like this.

I don't know if you saw this, but Elon was saying that, uh, why isn't the AI just writing binary straight to binary?

Uh, because what's the point of all this, you know, programming abstraction in the end?

Yeah, that's a good question.

I mean, it totally can do that if you wanted to.

Oh man.

So what I'm hearing here is in terms— there's always this question, should I learn to code?

Should people in school learn to code?

Uh, what I heard from you is your take is in like a year or two, you don't really need to.

My take is I think for people that are using quad code, that are using agents to code today, you still have to understand the layer under.

But yeah, in a year or two, it's not going to matter.

I was thinking about what is the right historical analog for this?

Because somehow we have to situate this thing in history and figure out when have we gone through similar transitions?

What's the right mental model for this?

I think the thing that's come closest for me is the printing press.

And so, you know, if you look at Europe in, you know, like in the mid-1400s, literacy was actually very low.

There was sub-1% of the population, it was scribes that, you know, they were the ones that did all the writing.

They were the ones that did all the reading.

They were employed by like lords and kings that often were not literate themselves.

And so, you know, it was their job of this very tiny percent of the population to do this.

And at some point, Gutenberg and the printing press came along.

And there was this crazy stat that in the 50 years after the printing press was built, there was more printed material created than in the 1,000 years before.

And so the volume of printed material just went way up.

The cost went way down.

It went down something like 100x over the next 50 years.

If you look at literacy, it actually took a while because learning to read and write is quite hard.

It takes an education system, it takes free time, it takes not having to work on a farm all day so that you actually have time for education and things like this.

But over the next 200 years, it went up to 70% globally.

So

I think this is the kind of thing that we might see is a similar kind of transition.

And there was actually this interesting historical document where there was an interview with some scribe in the 1400s.

About like, how do you feel about the printing press?

And they were actually very excited because they were like, actually, the thing that I don't like doing is copying between books.

The thing that I do like doing is drawing the art in books and then doing the bookbinding.

And I'm really glad that now my time is freed up.

And it's interesting, like, as an engineer, I sort of felt like a parallel with this.

Like, this is sort of how I feel, where I don't have to do the tedious work anymore of coding.

Because this has always been sort of the detail of it.

It's always been the tedious part of it and kind of like messing with like Git and kind of using all these different tools.

That was not the fun part.

The fun part is figuring out what to build and coming up with this.

It's, uh, it's talking to users.

It's thinking about these big systems.

It's thinking about the future.

It's collaborating with, you know, other people on the team.

And that's what I get to do more of now.

And what's amazing is that the tool you're building allows anybody to do this.

People that have no technical experience can do exactly what you're describing.

Like, I've been doing a bunch of random little projects and any, it's just like, anytime you get stuck, just like, help me figure this out.

And you get unblocked.

Like, I used to, I was an engineer for, earlier in my career for 10 years.

And I just remember spending so much time on like libraries and dependencies and little things and just like, oh my God, what do I do?

And then looking on Stack Overflow and now it's just like, help me figure this out.

And here's a step by step, 1, 2, 3, 4.

Okay, we got this.

Yeah, exactly, exactly.

I was talking to an engineer earlier today.

They're writing some service in Go and it's been like a month already and they built up the service, like it's working quite well.

And then I was like, okay, so like how do you feel writing it?

And he was like, you know, like I still don't really know Go, but,

and I think we're going to start to see more and more of this.

It's like if you know that it works correctly and efficiently, then you don't actually have to know all the details.

Clearly the life of a software engineer is changed dramatically.

It's like a whole new job now as of the past year or two.

What do you think is the next role that will be most impacted by AI within either within tech, like, you know, product managers, designers, or even outside tech?

Just like, what do you think?

Where do you think AI is going next?

I think it's going to be a lot of the roles that are adjacent to engineering.

Um, so yeah, it could be like product managers, it could be design, could be data science.

It is going to expand to pretty much any kind of work that you can do on a computer because the model is just going to get better and better at this.

And you know, like this is the Cowork product is kind of the first way to get at this, but it's just the first one.

And it's the thing that I think brings AI to, agentic AI to people that haven't really used it before.

And people are starting to just to get a sense of it for the first time.

When I think back to engineering a year ago, No one really knew what an agent was.

No one really used it.

But nowadays it's just the way that we do our work.

And then when I look at non-technical work today, so, you know, like, or maybe semi-technical, like product work and, you know, like data science and things like this.

When you look at the kinds of AI that people are using, it's always these like conversational AI.

It's like a chatbot or whatever, but no one really has used an agent before.

And this word agent just gets thrown around all the time and it's just like so misused.

It's like lost all meaning.

But agent actually has a very specific technical meaning, which is it's an AI, it's an LLM that's able to use tools.

So it doesn't just talk, it can actually act and it can interact with your system.

And this means it can use your Google Docs and it can send email, it can run commands on your computer and do all this kind of stuff.

So I think any kind of job where you use computer tools in this way, I think this is going to be next.

This is something we have to kind of figure out as a society.

This is something we have to figure out as an industry.

And I think for me also, this is one of the reasons it feels very important and urgent to do this work at Anthropic, because I think we take this very, very seriously.

And so now, you know, we have economists, we have policy folks, we have social impact folks.

This is something we just want to talk about a lot.

So as a society, we can kind of figure out what to do because it shouldn't be up to us.

So the big question, which you're kind of alluding to, is jobs and job loss and things like that.

There's this concept of Jevons' paradox of just as we can do more, we hire more, and it's not actually as scary as it looks.

What have you experienced so far, I guess, with AI becoming a big part of the engineering job?

Just are you hiring more than if you didn't have AI?

And just thoughts on jobs.

Yeah, I mean, for our team, we're hiring.

So Quadco team is hiring.

If you're interested, just check out the jobs page on Anthropic.

Personally, it's, you know, all this stuff has just made me enjoy my work more.

I have never enjoyed coding as much as I do today because I don't have to deal with all the minutia.

So for me personally, it's been quite exciting.

This is something that we hear from a lot of customers where they love the tool, they love Claude Code because it just makes coding delightful again.

And that's just so fun for them.

But it's hard to know where this thing is gonna go.

And again, I just like, I have to reach for these historical analogues.

And I think the printing press is just such a good one because what happened is this technology that was locked away to a small set of people, like knowing how to read and write, became

accessible to everyone.

It was just inherently democratizing.

Everyone started to be able to do this.

And if that wasn't the case, then something like the Renaissance just could never have happened.

Because a lot of the Renaissance, it was about like knowledge spreading.

It was about like written records that people used to communicate, you know, because there were no phones or anything like this.

There was no internet at the time.

So it's about like, what does this enable next?

And I think that's the very optimistic version of it for me.

And that's the part that I'm really excited about.

It's just unimaginable.

You know, like we couldn't be talking today if the printing press hadn't been invented, like our microphones wouldn't exist.

None of the things around us would exist.

It just wouldn't be possible to coordinate such a large group of people if that wasn't the case.

And so I imagine a world a few years in the future where everyone is able to program.

And what does that unlock?

Anyone can just build software anytime.

And I have no idea.

It's just the same way that in the 1400s, no one could have predicted this.

I think it's the same way.

But I do think in the meantime, it's going to be very disruptive and it's going to be painful for a lot of people.

And again, as a society, this is a conversation that we have to have.

And this is a thing that we have to figure out together.

So for folks hearing this that want to succeed and, you know, make it in this crazy turmoil we're entering, any advice?

Is it, you know, play with AI tools, get really proficient at the latest stuff?

Is there anything else that you recommend to help people stay ahead?

Yeah, I think that's pretty much it.

Experiment with the tools, get to know them, don't be scared of them.

Just, you know, dive in, try them beyond the bleeding edge, beyond the frontier.

Maybe the second piece of advice is try to be a generalist more than you have in the past.

For example, in school, a lot of people that study CS, they learn to code and they don't really learn much else.

Maybe they learn a little bit of systems architecture or something like this.

But some of the most effective engineers that I work with every day and some of the most effective, you know, like product managers and so on, they cross over disciplines.

So on the quad code team, everyone codes, you know, our product manager codes, our engineering manager codes, our designer codes, our finance guy codes, our data scientist codes, like

everyone on the team codes.

And then if I look at particular engineers, people often cross different disciplines.

So some of the strongest engineers are hybrid product and infrastructure engineers, or product engineers with really great design sense, and they're able to do design also.

Or an engineer that has a really good sense of the business and can use that to figure out what to do next, or an engineer that also loves talking to users and can just really channel

what, what users want to figure out what's next.

So I think a lot of the people that will be rewarded the most over the next few years, they won't just be AI native and they don't just know how to use these tools really well, but

also they're curious and they're generalists.

And they cross over multiple disciplines and can think about the broader problem they're solving rather than just the engineering part of it.

Do you find these three separate disciplines still useful as a way to think about the team?

They're, you know, engineering, design, product management.

Do you find like those, even though they are now coding and contributing to thinking about what to build, do you feel like those are three roles that will persist long-term, at least

at this point?

I think in the short term it'll persist, but one thing that we're starting to see is there's maybe a 50% overlap in these roles where a lot of people are actually just doing the same

thing and some people have specialties.

For example, I code a little bit more versus Cat RPM does a little bit more coordination or planning or forecasting or things like this.

Stakeholder alignment.

Stakeholder alignment.

Exactly.

I do think that there is a future where I think by the end of the year, what we're going to start to see is these start to get even murkier where I think in some places the title software

engineer is going to start to go away.

And it's just going to be replaced by builder, or maybe it's just everyone's going to be a product manager and everyone codes or something like this.

Who says hiring has to be fair?

Every founder and hiring manager I've been speaking with these days is feeling the same pressure: hire the best people as fast as possible.

But recruiting is time-consuming, alignment is hard, and competition for great talent keeps getting tighter.

That's why teams like ElevenLabs, Brex, Replit, Deel, and 5,000 other organizations use Metaview, the AI company giving high-performance teams a real unfair advantage in hiring.

They give you a suite of AI agents that behave like recruiting coworkers.

They find candidates for you based on your exact criteria, take interview notes automatically, gather insights across your hiring process, and help you identify the best candidates

in your pipeline.

AI handles the recruiting toil and gives you a real source of truth.

That means hours saved per hire and a team focused on what matters most: winning the right candidates.

Don't let your competitors out-hire you.

Metaview customers close roles 30% faster.

Try Metaview today for free and get an extra month of sourcing at metaview.ai/Lenny.

That's M-E-T-A-V-I-E.W.U.A.I/Lenny.

You talked about how you're enjoying coding more.

I actually did this little informal survey on Twitter.

I don't know if you saw this, where I just asked, I did 3 different polls.

I asked engineers, are you enjoying your job more or less since adopting AI tools?

And then I did a separate one for PMs and one for designers.

And both engineers and PMs, 70% of people said they're enjoying their job more.

And about 10% said they're enjoying their job less.

Designers, interestingly, only 55% said they're enjoying their job more and 20% said they're enjoying their job less.

I thought that was really interesting.

That's super interesting.

I'd love to talk to these people, uh, you know, both in the more bucket and the less bucket, just to understand.

Did you get to follow up with any of them?

They, a few people replied and we're actually doing a follow-up poll that we'll link to in the show notes of going deeper into some of this stuff.

But a lot of, there's like, you know, the factors that make it more fun and less fun.

The designers, they didn't share a lot actually of just like the people that are actually asked just like, why are you enjoying your job less?

I didn't hear a lot.

So I'm curious what's going on there.

Yeah, I'm seeing this a little bit with, at Anthropic, I think everyone is fairly technical.

This is something that we screen for, you know, when people join, we have, there's a lot of technical interviews that people go through even for non-technical functions.

You know, our designers largely code.

So I think for them, this is something that they have enjoyed from what I've seen, because now instead of bugging engineers, they can just like go in and code.

And even some designers that didn't code before have just started to do it.

And for them, it's great because they can unblock themselves.

But I'd be really interested just to hear more people's experiences because I bet it's not uniform like that.

Yeah.

So maybe if you're listening to this, leave a comment if you're finding your jobs less fun.

And you're enjoying your job less because what you're saying and what I'm hearing from most people, 70% of PMs and engineers are loving their job more.

That's like, if you're not in that bucket, you could— something's going on.

Yeah.

Yeah.

We do see that people use also different tools.

So for example, our designers, they use the Quad desktop app a lot more to do their coding.

So you just download the desktop app.

There's a code tab.

Uh, it's right next to Cowork and it's actually the same exact Quad code.

So it's like the same agent and everything.

We've had this for, you know, for many, many months.

Uh, and so you can use this to code in a way that you don't have to open a bunch of terminals, but you still get the power of quad code.

And the biggest thing is you can just run as many, you know, quad sessions in parallel as you want.

We can, you know, we call this multi-quadding.

So this is a, it's a little more native, I think, for folks that are not engineers.

And really this is back to bringing the product to where the people are.

You don't want to make people use a different workflow.

You don't want to make them go out of their way to learn a new thing.

It's whatever people are doing, if you can make that a little bit easier, then that's just going to be a much better product that people enjoy more.

And this is just this principle of latent demand, which I think is just the single most important principle in product.

Can you talk about that actually?

Because I was going to go there, explain what this principle is and, and, and just what happens when you unlock this latent demand.

Latent demand is this idea that if you build a product in a way that can be hacked or can be kind of misused by people in a way it wasn't really designed for to do kind of something

that they want to do,

then this helps you as the product builder learn where to take the product next.

So an example of this is Facebook Marketplace.

So the manager for the team, Fiona, she was actually the founding manager for the Marketplace team, and she talks about this a lot.

Facebook Marketplace is sort of based on the observation back in, this must have been like 2016 or something like this.

That 40% of posts in Facebook groups are buying and selling stuff.

So this is crazy.

It's like people are abusing the Facebook groups product to buy and sell.

And it's not, it's not abuse in kind of like a security sense.

It's abuse in that no one designed the product for this, but they're kind of figuring it out because it's just so useful for this.

And so it was pretty obvious if you build a better product to let people buy and sell, they're going to like it.

And it was just very obvious that Marketplace would be a hit from this.

And so the first thing was buy and sell groups, so kind of special purpose groups.

To let people do that.

And the second product was Marketplace.

Facebook Dating, I think, started in a pretty similar place.

And I think that the observation was, if you look at people looking at, if you look at profile views, so people looking at each other's profiles on Facebook, 60% of profile views are

people that are not friends with each other that are opposite gender.

And so this is this kind of like, you know, like traditional kind of dating setup.

But, you know, people are just like creeping on each other.

So maybe if you can build a product for this, it might work.

And so this idea of latent demand, I think is just so powerful.

And for example, this is also where Cowork came from.

We saw that for the last 6 months or so, a lot of people using Claude Code were not using it to code.

There was someone on Twitter that was using it to grow tomato plants.

There was someone else using it to analyze their genome.

Someone was using it to recover photos from a corrupted hard drive.

It was like wedding photos.

There was someone that was using it for, I think like,

they were using it to analyze an MRI.

So there's just all these different use cases that are not technical at all.

And it was just really obvious, like people are jumping through hoops to use a terminal to do this thing.

Maybe we should just build a product for them.

And we saw this actually pretty early.

Back in maybe May of last year, I remember walking into the office and our data scientist, Brendan,

had a quad code on his computer.

He just had a terminal up.

And I was like, I was shocked.

I was like, Brandon, what are you doing?

Like you figured out how to open the terminal, which is, you know, it's a very engineering product.

Even a lot of engineers don't want to use a terminal.

It's just like a, it's like just like the lowest level way to do your work.

Just really, really kind of in the weeds of the computer.

And so he figured out how to use the terminal.

He downloaded Node.js, he downloaded quad code.

And he was doing SQL analysis in the terminal and it was crazy.

And then the next week all the data scientists were doing the same thing.

So when you see people abusing the product in this way, using it in a way that it wasn't designed in order to do something that is useful for them, it's just such a strong indicator

that you should just build a product and people are going to like that.

Do something that's special purpose for that.

I think now there's also this kind of interesting second dimension to latent demand.

This is sort of the traditional framing is look at what people are doing, make that a little bit easier, empower them.

The modern framing that I've been seeing in the last 6 months is a little bit different, and

it's look at what the model is trying to do and make that a little bit easier.

And so when we first started building QuadCode, I think a lot of the way that people approached designing things with LLMs is they kind of put the model in a box and they were, here's

this application that I want to build, here's the thing that I wanted to do, model, you're going to do this one component of it.

Here's the way that you're going to interact with these tools and APIs and whatever.

And for QuadCode, we inverted that.

We said the product is the model.

We want to expose it.

We want to put the minimal scaffolding around it, give it the minimal set of tools so it can do the things.

It can decide which tools to run.

It can decide in what order to run them in and so on.

And I think a lot of this was just based on kind of latent demand of what the model wanted to do.

And so in research, we call this being on distribution.

You want to see what the model is trying to do in product terms.

Latent demand is just the same exact concept, but applied to a model.

You talked about Cowork, something that I saw you talk about when you launched that initially is you, your team built that in 10 days.

That's insane.

Yeah.

Uh, I think it came out, I think it was like, you know, used by millions of people pretty quickly.

Something like that being built in 10 days.

Uh, anything there, any stories there other than just, it was just, you know, we used Cloud Code to build it and that's it.

Yeah, it's funny.

Uh, Cloud Code, like I said, when we released it, it was not immediately a hit.

It became a hit over time and there was a few inflection points.

So one was, you know, like Opus 4, it just really, really inflected.

And then in November it inflected and it just keeps inflecting.

The growth just keeps getting steeper and steeper and steeper every day.

But, you know, for the first few months it wasn't a hit.

People used it, but a lot of people couldn't figure out how to use it.

They didn't know what it was for.

The model still like wasn't very good.

Cowork, when we released it, it was just immediately a hit, much more so than Claude Code was early on.

I think a lot of the credit honestly just goes to like Felix and Sam and Jenny and the team that built this.

It's just an incredibly strong team.

And again, the place Copilot came from is just this latent demand.

Like we saw people using quad code for these non-technical things and we're trying to figure out what do we do?

And so for a few months the team was exploring, they were trying all sorts of different options.

And in the end someone was just like, okay, what if we just take quad code and put it in the desktop app?

And that's essentially the thing that worked.

And so over 10 days, they just completely used QuadCode to build it.

And, you know, Cowork is actually, there's this very sophisticated security system that's built in and essentially these guardrails to make sure that the model kind of does the right

thing.

It doesn't go off the rails.

So for example, we ship an entire virtual machine with it and QuadCode just wrote all of this code.

So we just had to think about, all right, how do we make this a little bit safer, a little more self-guided for people that are not engineers.

It was fully implemented with quad code, took about 10 days.

We launched it early.

You know, it was still pretty rough and it's still pretty rough around the edges.

But this is kind of the way that we learn, both on the product side and on the safety side, is we have to release things a little bit earlier than we think so that we can get the feedback,

so that we can talk to users, we can understand what people want, and that'll shape where the product goes in the future.

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 like, is a unique reason to start releasing things early.

That'll help you as you exactly described this idea of what is the latent demand in this thing that we didn't really know.

Let's put it out there and see what people do with it.

Yeah.

And Infranthropic is 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.

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 it in a synthetic situation and just say, "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, 4 or 5 months or something before we released it because we weren't really sure.

This is the first agent that, 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's 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.

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.

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 you get, you know, the god can escape and cause damage.

And just finding that balance must be so challenging.

What I'm hearing is there's kind of these 3 layers, and I know there's like— 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 3 layers you work with.

Uh, there's kind of like 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 observe ability tool that can let you peek inside the model's brain and see how it's thinking and where it's heading.

Yeah, you should, uh, you should at some point have Chris Olah on the podcast because, uh, he's just the industry expert on this.

He invented this field of, uh, we call it mechanistic interpretability.

Uh, and the idea is, uh, you know, like at its core, like what is your brain?

Like what are, what is it?

It's like, it's a bunch of neurons that are connected.

And so what you can do is like in a human brain or 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.

Model neurons are not the same as animal neurons, but they behave similarly in a lot of ways.

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 is 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, the way to 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 superposition.

Together it represents this more sophisticated concept.

It's just something we're learning about all the time.