Claude Code’s Next Era — Thariq Shihipar, Anthropic

0:03

We're here in the studio with our friend Tarek from Anthropic. And I guess generally the Cloud Code, there's so much sort of merging of boundaries. And you've been so on top of everything since you joined Anthropic. You have been early to Cloud Code itself, but then also, and you've told that story in other podcasts. And you've also been talking about seeing like an agent.

0:28

Most recently, you did the top AIE, WorldSphere Talk, Field Guide to Fable, which obviously you guys launched Fable. So that's cheating.

0:36

And most recently also launching CloudTag. And we're also going to be talking about pacing on Frontier. There's a lot going on in Anthopic. I guess top of the question is, what's it like being in Anthopic when there's so much going on? I think that it is like... I think you can get whiplash sometimes. I think like going... When I joined Anthropic, I joined because of Cloud Code.

1:00

Like Cloud Code had just come out and I was like, this is so good. And Opus 4 to me was like...

1:05

I could not imagine how good it was, you know what I mean? And that was a real moment for me. But I was trying to convince my startup friends, basically, to use Agenda Coding. And they're like, no, our engineers don't think it's good enough or something. And I was like, that's insane. And now you fast forward 12 months, less. And it's just the default way that everyone codes.

1:31

And I think that just having to go from selling it to now... Teaching people how to make the most use of it and be more efficient and things like that is just a big change. And yeah, I think it's just hard to stay on top of everything as a human. Things happen so, so fast. These two more agents at it. I mean, yeah, like that's actually like the agentic stuff scales much better than the like human stuff where it's like, oh, like there are three things happening right now.

2:00

And like, they're all emergencies. And like, how do you like, you know, respond to it? What do you split your time on? You do a lot of technical writing, engineering work. Yeah, so I think that, like, when I joined the Cloud Code team, I wanted to teach people how to use Cloud Code, basically. And I think that, like, that has been something that, like...

2:19

I thought, like, maybe I would spend a little bit of time on it. Or, like, I, you know, like, do... I was spending some time on the agent SDK first. And I wasn't exactly sure, like, you know, how... the bitter lesson would go when it comes to harnesses. I think sometimes we're like, what's after Cloud Code? Initially, I was like, I just want to teach people how to use Cloud Code and make it easier to use Cloud Code.

2:41

I think as harnesses have gone better and better, that's the dominant problem now is how do you use the agents? It's such a high skill expression thing. So I do that. And then I do engineering work. I give talks. But I think like when I'm doing engineering work, my goal is to take that feedback that we get from users and also like then be able to talk about like, hey, how to use cloud code to do engineering.

3:05

So there's kind of like a good loop there. Yeah. Yeah, for listeners, we'll attach the talk that you did with Sarah for the DevWriters meetup, which we talked a little bit about. Well, first you do the work and then you talk about the work. Right. Something like that. Yeah, yeah, yeah. Sow and reap? Yeah, sow and reap. Something like that.

3:23

Yeah, and then just to preview a little bit, we are going to talk about the evolution of the harness. It has come a long way from just being a CLI. We're going to talk about cloud mods, which is starting to leak today because you couldn't keep it secret. Right. Yeah, basically. Yeah, there's a lot there. I think you started off with like adding ask you the question tool, which people love and hate actually.

3:48

I actually thought it was like very innovative. And then now I have like my own version. You have your interview me version. Yeah, yeah, yeah. And yeah, everyone just has like their own stuff. And again, no longer matters because now you're supposed to write prompts that create other prompts and loops and all these things. So what's the state of the art today? Like what are people, what are you like telling people to do today? Yeah, Ask User Question was the first time that the model was good at elicitation.

4:16

You know, I think this was like kind of an emergent behavior that I like, you know, wanted to see if the models could do. I have like kind of like a human computer interaction background. So I like, you know, did that in undergrad and grad school. And so this was like, I think it's kind of like human agent interaction to me, like, you know, trying to figure out like how can the agent communicate with you and extract, you know, the requirements, right? Yeah.

4:41

I think that one of the things that's difficult as Cloud Code has gone broader and broader is that everyone has their own way of using it. And it's actually very hard to change the default behavior. So for example, if someone asks Cloud Code to do something, Sometimes they just want them to do the work because they're maybe a very good prompter, and sometimes they actually are not good at prompting, you know what I mean? And the agent needs to clarify, you know? And so that's a good split.

5:12

And to ask you the question tool sort of splits along that side, where do you feel like you're good enough to instruct the agent as it is? Or

5:21

does the agent need to pull out more requirements and collaborate with you more and really understand your preferences? I, on the whole, believe that pretty much everyone is more on the latter than the former, that they have more ambiguity and they know less than they want, than they

5:39

think they know about the problem. But it's like an interface design problem to make that easy, you know what I mean? And so if you're designing a problem or if you're going through a problem like, you know, Things like what's the schema or like what's the call stack and things like that are really important. You know, like the details and the design are important.

5:58

Ideally, you want to figure out some of these like hard problems ahead of time before starting implementation. And yeah, that's why they call like unknowns. Right. And so I think that this will forever be like a skill in agentic coding is like figuring out your unknowns. So like because even if the model is like super intelligent and It needs to know what you want.

6:19

And you have preferences. You need to sort of pull that out. And so that's, I think, what I'm pushing. The question then is, how does the agent interact with you? And I think that has been HTML, has been the big way of doing that. And we've recently added artifacts, right? And artifacts, I actually think we've done a bad job of... or I've done a bad job of explaining how to use them fully.

6:44

We have a lot of property capabilities. They have a database associated with them. And so every artifact can store and write persistent data. They can feed back into Cloud. you know? And so like one thing that, you know, like people are not doing yet that I'm trying to like encourage is like this idea of a dashboard artifact. So you have like cloud working on a project long-term, maybe it's like a Kanban or something.

7:09

It can store that Kanban data in a database. Multiple clouds can access that data via like the artifact MCP. And like that, that artifact can like talk to those clouds as well. And so like the, that we're basically building the primitives for you to be able to have this, like, generative interface via artifacts that will, like, let you surface more of that rich detail from the agents.

7:36

And I think that, like, almost everything in agents right now is, like, this problem of, like, you think you know what you want, but you don't really know what you want. And, like, the agents need a lot of detail and collaborating with them in the loop is really important. And so artifacts are, like, the... like way that we're trying to evolve there but there's a lot of work to do because it's so much more complicated than like a multiple choice question you know um there's a lot more like detail in terms of like diagrams and code snippets and schemas or like like whatever it is for that problem but like artifacts is like the more agi pilled way of like you know doing ask

8:14

you the question basically so yeah I think one thing that's unclear to me about the artifact stuff is what feedback should go in through the artifact and what feedback should go through a cloud chat. Because the more AGI-pilled one is to just feed everything to the cloud. I think the more AGI-pilled one is to go through the artifact. And I think that we sort of imagine in the limit, I think that artifacts will be your interface into the harness, you know? You can comment on this live document of your plan of the work.

8:45

You can see maybe multiple agents and different agents are doing this. And that artifact is built for the current work that you're doing, right? And so each one has slightly different. I think we're still getting there from an infrastructure perspective. But yeah, I think on-the-fly interface for your harness is probably where things are headed.

9:03

Is there a version of it that's an abstraction from CLI or chat? Because right now, a lot of it is, okay, you're interfacing with cloud code. You're having HTML given back for a mock-up. It's pretty rich. There's diagrams. Artifacts are ways to connect these together. Why not just do everything that way? then it becomes like separating out like, where is the inference happening? Where's the intelligence happening? Where is the work happening? You know, like, I think this is kind of like difference between like, or like some of the distinction between local and cloud, right? And so I think right now, if you use cloud code, It's, like, local, and, like, you can spin off

9:41

remote control, for example, to get some cloud behavior, or you can spin off cloud code in the cloud, right? We're moving towards a place where instead of clouds, like, you message a local cloud, it starts a session locally and it executes, to more like you have a cloud that you message that's in the cloud that's running. It can run, like, local...

10:03

or cloud sessions, this is kind of how Cloud Tag works, but over time we'll add local hands as well. And so local hands will be the ability for that agent to access your computer if it's online, and be able to work there. And so it can spin off many different sub-agents, those sub-agents can communicate with each other, and that's where the artifact comes in to display all of that work, basically.

10:26

So you can imagine the... you're separating out these things. So there's like the surface UI display that's an artifact and hosted somewhere and has a database and everything. There is the inference intelligence, right? That's happening on the cloud and you don't have to worry about shutting off your computer or whatever, right? And then there's the like hands kind of like, and it can be local, it can be in like a remote sandbox or wherever you need your work to be done.

10:52

That's like unpackaging, like, the Cloud Code experience. Right now, it all happens in one place. How do you see the multiplayer side of that? Say teams want to work in this way. Right now, it's very individual, but how do you see the future of multiplayer? Right now, I guess there's Cloud Tag, which is a version, but We're launching projects.

11:14

And so projects is this abstraction that's kind of like CloudTag, but on our Cloud products, right? So you can message it and it will do the CloudTag-like stuff, like spinning off subagents. So... We think with multiplayer, like, CloudTag is, like, a little bit more native multiplayer because it's just, like, in your Slack and the permissions are all figured out and stuff like that.

11:41

But I do think multiplayer is, like, an important part of the story and, like, that will need to get tied together more. Like, you can imagine how complicated it gets when you're, like, oh, you have hands, but now you have hands in other people's computers, too, and, like, you need to, like, permission them or, like, you have, like... your MCP and someone else's MCP and how do you figure out how to use them.

11:59

It gets quite complicated. And CloudTag does a good job of sanding down all of these issues so that when you have Google Docs, how does it access Google Docs? It accesses through the shared Cloud MCP or it can access through your local credentials as well if it doesn't have access. But yeah, I think CloudTag is our multiplayer product and it's really useful for these Things that are inherently multiplayer, like on-call, for example, incidents are inherently multiplayer.

12:29

You want to tag-claw, you want multiple people to log in, you want to be able to find context. I think whenever I'm working on something and I want privacy or security, I want other people to review it. It's really nice to, like, I'll have a channel per project, and I'll, like, at legal, for example, be like, hey, I want to ship this. Can you, like...

12:49

Like Claude knows everything, you know, just chat with it. And that way legal gets precise answers, you know, on like what exactly is shipping into the code and I don't need to be in the loop, right? So I think like multiplayer is getting like more and more like, yeah, everyone can participate with Claude. I think Claude Tag is like that product and like projects will start off single player and will like, you know, expand.

13:14

I think there's a question about like maybe dual questions about identity and the unit of isolation. Yeah. Tag, you specifically chose to make it its own identity, which is like a controversial choice. There's other ways to do it. Yeah. Cloud projects, probably it sounds like, you know, if it's anything like ChatGPT projects, it is, you know, the isolation is that artifact, that cloud instance, everyone's collaborating on this.

13:43

It sounds like, you know, it should be like if you're collaborating illegal on a thing, like that channel should be a project, right? Like it's not yet, but that's the natural next step. Yeah, I mean, like, I think in CloudTag, it's effectively, like, CloudTag, you have to sort of do your own arrangement, basically. And so CloudTag, yeah, each channel is, like, you can name it as you want, and I name, like, each feature, basically, as a channel.

14:08

I think, like, there is some transfer, like, it's unclear when there is transference, because let's say if you have a coworker who is tagged on all these things, yes, there is transfer, because it's the same person. But with Cloud, it's unclear if it's, like, necessarily like, well, no, you don't know about the other stuff. You should only use this stuff.

14:26

it's like the tip of the iceberg meme, right? Where you can like, this is what we spent so much time on basically is like, there is like infinite surface area of like, okay, you want clods to, not infinite, but like there's like surface area, a lot of like surface area to figure out of like permissions and visibility and like, you know, like how can you let clod operate as well as you can, as safely as you can, you know? And obviously this is very important to us because like, you know, like, like security for our code base is very, very important.

14:56

And so we've put a lot of time into this. Yeah, there's so many like edge cases you can figure out where it's like, oh, like, yeah, this cloud in this channel has different permissions, but it can message another channel and can't exfiltrate data that way. Or like, can you like, what if it uses your MCP and then messages someone else? Like there's like so much and we've like really put a lot of work into sanding it down.

15:15

Yeah. Yeah. Lots of work. Okay. Fable. Fable, you wrote two good articles. I mean, you've written many good articles, but on, you know, field guide to Fable, building cloud code. I'm curious from what you've seen, is there any common patterns that you see in like top users, ad anthropic externally? Like what are best practices for getting the most out of cloud code? The meta skill, I say, is prompting is very important.

15:45

I think this is not trivial to say because I think a lot of people are like, oh, prompting doesn't matter. It's just like, I can just say a sentence and Cloud will do it. And I think prompting is really this... It's like public speaking or writing or something, and for a specific audience, and that audience is Claude. You need to build a mental model of Claude and how it thinks and how it works.

16:08

That's the most important skill in working with Claude Code is having this mental model. of Claude and like what it can do well, what it can one-shot, what it can't. And so, so many people when you see prompting, they're just like, they're short prompts, but they have such a good mental model of Claude and of like the code base and things like that, that like it's effortless, you know what I mean? But it's like high skill ceiling.

16:30

So like that work of like, you know, spending a lot of time prompting and building mental models of how, you know, and intuition for how the agents work is really important. And then I think like the next thing is like, The unknown stuff we talked about earlier, where it's like being able to find out what you don't know or what you haven't written down, learning about different things.

16:54

I think as Claude can do more and more things, the likelihood of you doing something... out of distribution for you and like you have low domain knowledge on is very, very high, you know? And the more you can like learn the vocabulary to be able to prompt Claude, it becomes really important. And so like, I think the most important unknowns are the unknown unknowns where you're like, I just like don't even know that this exists, right? Yeah, exactly.

17:18

I think that's like an illustration of like the map and the territory, right? Where you're like, okay, this is my prompt and the territory is like the actual like work that the agent needs to do, right? Yeah. And if you are like very precise, you can give more precise things, right? So like, for example, in design, I'm not very precise. I'm not a designer.

17:35

So I say like, you know, give me like eight different mock-ups. But if I was a designer, maybe I'd be like, oh, hey, here's some reference sites. Like I want this type of font and this type of like font. look to it, and here's a few different components to visualize. Here's a Figma MC board to bring in. And so you can just be so much more precise with that language.

17:56

And if you're not a designer, you just need to try and learn the language, basically, or learn the unknown unknowns. And this is true of, like, everything, I think. Like, the more, like, you can work with Claude to learn, like, how things work, the better your prompting will be. I think another good example of this is, like, game design. You know, like, where a lot of people are like, oh, like, I can vibe code a game now.

18:18

And they're like, it's not fun. And, like, it's just, like, the thing about game design is, like, every one of these choices has, like, a lot of... Variations. A lot of, like, craft to them. So it's like, oh, okay, like, when you're making a... flying game the feel of the plane and the like you know way it responds to your controls has a lot of like like you know a game designer would spend like days on that you know what i mean um and that's what taste is right like it is like from the possible space of 1000 mathematically valid answers here's the one that is the humans were like Yes.

18:53

I think with taste, I'm like torn on this word because I think you're right, but everyone has different definitions. And it sounds kind of like low skill or like elitist almost where you're like, oh, like there are certain people with taste. It's like taste is what I call taste. Yeah, exactly. These guys don't have taste. Yeah, exactly. Oh, like an engineer doesn't have taste.

19:12

Like I, the like founder, have taste. You know what I mean? And I think that's actually not true. Like I think like the engineers have a lot of taste for these particular things. like, problems, you know? And I think everyone has taste for particular problems. I think, like, Jason Liu, like, say, like, in order to have taste, you have to eat, you know? And I really like that, where it's like, okay, you have to, like, do a lot of things.

19:36

You have to, like, iterate and figure out what you want, what you like, and, you know, like, build that, like, domain vocabulary. And then when you're prompting, you're, like, synthesizing all of that. Isn't it annoying when someone else says it better than you? It's just Like, fuck, I have to quote this guy forever. Having to quote Jason Liu forever.

19:53

He's going to love this. I get prompts. And sometimes it's actually not even that. Sometimes it's just intuitive, right? Like, you don't realize you even want something until a model puts it out. And you're like, oh, this just feels immediately better, right? Yeah, yeah, exactly. One thing I go back and forth on is I feel like the way I prompt half the time, let's say I use voice.

20:16

Hmm. guys have voice other people have voice that is the opposite there's just like me rambling for like two minutes pressing down the function key and then let go and then like hopefully it figures it out and oftentimes it does but it's not as thoughtful as like a structured prompt with like you know well run communication as though it's a PRD or a memo is that in line with how people do this there's like basically bimodal prompting where there's some prompts where you spend a lot of time up front and other prompts you just dash it off I don't think the voice is necessarily low.

20:46

I think it's more like how much information is in the prompt. The model can... You can um and uh and add some sentences and be like, oh, actually, I changed my mind in the middle of the prompt, and it will be able to follow that perfectly. You know what I mean? So I think the actual format of the text is less important, but then the... ability to like how much information is in it, right? And I think for voice, a lot of times, you know, going back to like kind of human agent interaction and like for a lot of people, it's just way easier to talk than to like type, you know? And if that gets some more information out of you, like that's better.

21:21

At some level, it feels like just giving the model as much context over prompting before you kick off is a best practice. I don't know. A lot of the times, like when I was first trying out Fable, I spend a solid 30 minutes like really crafting a long prompt. This, I think, is a response of models running for longer and longer, right? It's still a little difficult to nudge them as they're in...

21:45

in the loop, but I just intuitively spend more time kicking off that first prompt and working with it a lot. My personal opinion is that if I was a software engineer, if I was just running my own startup, for example, I think I would mostly stick to a Max 20X. You know what I mean? Maybe verification and code review are separate things, but I think what I see a lot of times is people hit rate limits when they're doing this sort of it did a lot of work and you're like, oh, I don't like this.

22:16

Can you undo this and redo it? Then you're iterating on this thing that the model could have done if you had spent more upfront time or given it better context. Instead, you're like, nope, don't like that design, try this or you mess this up or something like that. Then that just eats up so much more of your usage. That's I think maybe a key tip both for efficiency as well.

22:43

And yeah, I think context on what the goal is is good. Are you building a prototype or is it a production thing? Where can you spend compute or where can you not spend compute? I think you have to give the model permission or not permission to do things sometimes where it doesn't know intuitively how much you want to spend on this task. And you can use effort for this.

23:07

I'm working on a blog post about that where it's like,

23:13

We see the effort scales with basically the complexity of the task. So for security, effort gets way more results. High effort versus low effort changes the evals a lot. But for software engineering, it doesn't change a huge amount because effort is mostly spent on the verification and the edge case testing and things like that. And so like being able to like give the model that guidance of like, hey, this problem is something that I think I want you to spend a lot of time verifying and edge case testing.

23:43

How about model in the mix? So, you know, there's Opus and Fable with effort. There's also Haiku in there.

23:51

It's not quite true yet, but it's very close where I think the frontier models will be Pareto-dominant over almost everything. And sometimes I think Opus might be Pareto-dominant. Do you know what I mean? I think depending on how things...

24:07

shake out if it's a newer version of Opus. But I think that increasingly, it's just going to be the smart model is going to be able to do the simple task for less tokens than the other models, basically, because of verification. With verification, in the limit, your model doesn't need to verify, right? If it's a perfect model, it just does the work once, and it's like, okay, I did it, you know? And increasingly with Fable, I'm like, dude, you don't need to spin up Chromium and screenshot all of these things.

24:40

I see it, you did it, right? And so a lot of the higher effort, you spend more of those tokens verifying. But if you're working on simpler problems, and a lot of software engineering is like, you know, like well, like in Fable, like low and mediums ability, it can spend less tokens verifying. And as the models get smarter and smarter, they will just be able to like, all right, done.

25:03

You know, like I can run the lint for sanity's sake, but like I like, you know, it lints, you know what I mean? Like you don't even need to do that. And that will be so much more token efficient than like the the smaller models, basically. Yeah. Is there a good practice on our side that we can use to see if we're using too much effort? I freaking hate wasting time on that kind of stuff.

25:25

Yeah, yeah, yeah. I know what you mean. I think, like, so in this blog post, my rough distribution is, like, code review and security should be, like, high or max, basically. And, like, software engineering... You just have recommended settings per domain. Yeah, I think, like, if you're doing, like... UI or something like that, like low and medium, I think is you're building like an API and you want to make sure like you cover enough edge cases, you know? And so I think building, like I said, that mental model of like how things work across these distributions is like, yeah, part of the job.

25:57

This is more intuition-driven? Because I'm guessing this would change as well. He has evals. Yeah, so what I did in the blog post is I go over all of the terminal bench evals, basically. So there are like 70 problems, and I show that in the security problems, it does more. And then I also look at some of the transcripts just in terms of how...

26:19

What does it answer? What does it forget or something? And a lot of times, this is another prompting tip I have, is like asking it to make decision notes or implementation notes. Because in basically every eval problem that it faces, it thinks about decision. the correct solution, you know, and decides not to do it. You know, it's like, oh, like, here is the answer.

26:43

What if I did this? And then it's like, oh, probably not, you know, and then keeps going. And this is like the majority of the failures, you know what I mean, at like a higher max level. It's very rare that the model just doesn't know how to do something. If you just have these implementation notes, then you can review. And you can be like, oh, actually, I want you to do this thing that you didn't do.

27:04

The models are getting better at surfacing that overall. I see in the transcripts that Fable 5.1, when it does this output, it will call out its decision making as well. But making this more explicit in the harness is better. And now we're allowing ways of you modifying the harness so you can Like, you know, add some capability there. Yeah, so I do want to call out two things that you mentioned that I think actually exist outside of prompting.

27:31

One is actually like, let's call it the prompt that is so important that it shouldn't be in a prompt. It's actually in CloudMD or AgentsMD, which is like goals, right? Like your situation, your goals, the things that you want, the thing. And then the second of all is the decision log or the experiment log or whatever, right? Logos traces that you might want to actually survive the current session to do those things.

27:55

Those are like externalities that there's no standard. It's not like skills. It's not like MCP. There's no standard. It's just like it's a markdown file.

28:04

First of all, is that right? Is CloudMD going away? You have a documented dislike of AgentsMD, but you're going to do it? Yeah, yeah. Okay, so, I mean, Agents.md, yeah, like, we're going to do it. I think it's just, like, different models are very different from each other, you know what I mean? But I realize that it's, like, such a pain to, like, maintain different ones, you know what I mean? And, yeah, like, as the models get better and better, the floor of how they accomplish the simpler task is better.

28:31

And so I do think in the limit, Cloud.md goes away, and maybe not even, like, that far. Like, I think, like... I think that right now, it might be better to start a new project without a Cloud.MD. Yes. I think that maybe if you see very repeated failure modes, you add them to your Cloud.MD. The really tough thing is that this changes per model.

28:54

And so if you've added a bunch of failure modes... So you need FableMD, you need OpusMD. Well, even Fable 5.1 versus Fable 5. It is annoying. We don't do this on purpose. It's just how the models work, right? And so maybe Fable 5 had this failure mode that Fable 5.1 doesn't. And if you keep this context running log of a bunch of different failure modes, they will probably over-constrain Claude.

29:23

And so this is like... We just actually added evals, plugins for skills. And so now you can eval if a skill is better. I think Daisy on our team did this. And so, yeah, this is like, we're trying to work on this. We know it's like, you still have to spend tokens on it. And like, you know, it's not perfect, but it's like, we're trying to help out with this problem.

29:45

So as far as prompting goes, the one tip I want to offer is something I have told people a lot is sufficiently advanced prompting is indistinguishable from sufficiently advanced executive communication. So I've actually referred to... This is an executive comms workshop from Heavybit that is the best I've ever seen in my career. And they teach this thing called the SCQA model.

30:07

Just Google it. It's a thing. People have... done prompting for decades. It's just called executive communications. It's like when one person has to communicate to thousands of people down the org chart, this is what you do. So situation, complication, question, and answer is how you write the memo. But obviously sometimes you don't have the answer, but you can at least list out the S, C, and Q, and then they have some examples in there.

30:28

So just leaving breadcrumbs for people if they want to explore. I mean, before we move on, I want to ask you any other underrated tips, ways people could get a lot of value from Cloud Code that they're not using? Yeah, I mean, I think a lot of them are in this unknowns doc. I give a bunch of example prompts, like using it for brainstorming, using it to quiz you after.

30:56

We added this, like, explain it like I'm five person. skill, actually, which is a very short prompt. And it doesn't even say explain it like M5. It's like basically the key word of this prompt is big picture is few words. You know, like that's like the main thing. And it is shockingly good. You know what I mean? Like you like I think I tweeted about this basically.

31:18

And it's like slash Eli five and like you can install it as a plug in. But yeah, it's like way better at just cutting through the BS and being like, yeah, exactly right here. So the diagrams are quite clear. I think one of the things that is true with artifacts is they put too much text in and people are not reading the artifacts. And so this simplifies it a lot more.

31:38

And yeah, this came out of just people at Anthropic going through very complicated incidents and being like, what is happening? So this one I think is great. Yeah. My version of this is actually the... It's like test your understanding. Give you a few choices, and then if you actually get it wrong, you have a mismatch between what you think is happening versus what's actually happening.

31:59

Yeah, yeah, yeah. I think this is one of those things that everyone loves talking about, and then very few people really do. It's really helpful. Yeah, but most people just don't want to get quizzed about something. You know what I mean? Unfortunately, I think this is one of the things that we need to... What's the opposite of ask you as a question? Ask you as a question before the thing.

32:20

This is after the thing. Exactly. Yeah. It's a good way to stay grounded of like, do you even know what you're doing? Right. The worst case is when people send you slop and they haven't understood what they're asking for or what the output is. And it's like, dude, I don't want to read this. Do you even know what it is? So, you know, you make it a rule for yourself that before you send stuff, you should at least know what's implemented.

32:41

Yeah. Yes, but so you could make this a mod, and you could build your own mod to make sure you test it. So yeah, we can talk about that. Let's get right into it. What is CloudMod, and what is this diagram showing? Yeah, okay, so CloudMod is basically... Basically, you can customize the entire Cloud Code harness, and we're going to... If you have a request, we will let you...

33:05

Please let us know. We'll add more and more. This works for a CLI. It works for desktop. Maybe it will work for CloudTag in the future. I don't know. We're trying to make this very, very extensible. You can see this reference sheet. I don't want people to get overwhelmed by it. You know what I mean? At a high level, you can customize both the execution of the harness...

33:25

and the UI of the harness. And so like you say on that Tetris example from Boris, that's like customizing the UI, right? Like showing like basically Tetris in the game. But let's say that you wanted to do this thing where you tested your assumptions or tested your understanding after every project, right? What you would do is you would ask Cloud to make this plugin.

33:51

It would spin a classifier after every prompt, basically. And so at the end of each turn, you would spin off a subagent or a forked agent, basically. A forked agent is like... the prompt cache, right? So it's like one of those unintuitive things where you can fork and do like a little request and they'll be very cheap because the entire prompt cache is like done.

34:15

And so you can be like, has this task been completed? This is how you do BTW and all those. Yeah, the underlying forked agent, yes. In the forked subagent, you can say, has this task been completed? If so, return true. And then in your hook or in your plugin mod, or sorry, in the subagent probably, you'd say, if true, give me a quiz. Give me questions and answers.

34:42

And then in a JSON format, and then you'd parse it. And then you display above the prompt input, basically, this list of questions, right? And so this is something that's like slightly token intensive because like, you know, you have to sort of do it after every end of the assistant turn. But it's like a lightweight classification. And then you can like, you know, get this quiz and then you'll see like, you know, Cloud will always do it for you.

35:09

You don't need to remember to do it. There are lots of these like tips that we've talked about, right? Where it's like, oh, implementation notes. You can also add a tool for implementation notes now. And so like this tool that I'm adding is like register, like I think assumption is what I'm calling it, but like maybe I'll change it around.

35:26

And this is a mod. And so like you give it a register assumption tool and then it will keep a list. Every time it does it, it'll keep like add to the list. And then at the end, it will display those assumptions, you know? Another mod I'm working on is a model router. And so, like, internal, like, cloud model routing, right? So this is, I want to say the reason we don't do model routing by default is, like, it's a hard problem.

35:50

You know what I mean? And, like... You will get it wrong. Yeah, you, like, yeah, you will, like, accidentally use, like, Fable for a hard problem or Sonnet for... You have auto approve, but you don't have auto mode. Well, you don't have auto-routing or something. You don't have auto-mode for model picker. Yeah, exactly. I'm getting the rough question of how much do you open this up and how much do people have to think about this? When you talk about prompt caching and building a router, it seems like you could easily build a mod that routes per query and I'm just killing my plan very fast, right? I guess my question is more so like, what is like a product doc like this

36:28

look like, right? Who is it for? Is it for power users? Is it everyone should be able to quickly try it out? Oh, definitely power users, right? Yeah, I mean, I think it is power users, but like the nature of cloud code is that so many people are power users, you know, because it's easy to share things, you know, like you can, like one person can make a good mall router thing that doesn't break prompt cache all the time and then you can like, you know, sort of compose them, right? Another cool thing about the plug-in is that they can hook into and compose with each other.

36:56

And so I have a mod that will create a mode selector at the top, and any plug-ins can register to be a mode. And so the auto-router can be a mode, right? Or you can have a mode that's like artifact mode, where it primarily talks to you in artifacts. It's kind of like you can toggle between plan mode. You know what I mean? And so you can create more and more of these modes.

37:22

But the ability to create modes is in itself a mod. And so there's a lot of richness here. But we do want to make it fairly easy. We want to make it so that you can just install someone else's. You can chat with Claude. And we'll make sure that it understands the nuances of things like prompt caching and stuff. So it can warn you. This is like...

37:45

not extremely complicated behavior for cloud, I think, but we should have just a good skill on how to make mods. And yeah, we'll see how we go. But I do think that this is like a preview of like mutable software, you know, and like how like generative software, just like you can customize safely if enabled, you could customize any piece of software.

38:08

And I think that more and more apps ideally do something like this, you know? And by the way, you have another cool tweet about how there's the infinite money button, which is like make your SaaS consumable by agents. I think mutable software is interesting and other people have also tried to do it. I think the hurdle comes when you can do everything, then users tend to get confused.

38:34

So usually the stuff that works is just like one opinionated flow. This is in the side of less opinionation. It's just like, well, more power to power users. And I think... probably unlocked by AI where like you can just prompt for whatever the thing it is. Yeah. Or there can be a skill that gives the opinions, you know, and then, yeah. So knowing a little bit about like TypeScript and build systems and all these things, the closest, I'm actually very curious that the team who worked on this, if I don't know how close you were to them, if they drew any inspiration from build systems, like Babel, Webpack, all these, all these like old school things, because it sounds very

39:08

similar, like the plugin ecosystem, those things where they can compose to each other. Yeah, I mean, I'm not deep in the technical details, but I do know it was a collaboration with someone on the Bunn team and someone on the Cloud Code team. Yeah, this is a build system. Yeah, exactly. It's very exciting. But yeah, agents can just do this very complicated extensibility into your software now.

39:28

And so, yeah, another reason to, if you run a startup, you can just prompt Cloud and be like, hey, could we make an extension system? What would that look like? Yeah. Yeah. And I just really wanted to, like, you had hooks in the past and plugins, all these things. So what specifically would mods be able to do that those things could not do? Internally, we were originally calling this function hooks.

39:50

And so like, that's like, gives you a little bit of an idea where like hooks... sort of register an event to happen and then a script to call, basically. And this basically, inside of the TypeScript runtime, is running things. And so you get some benefits of just like, it has a bunch of things in the...

40:13

with like, for example, like how many turns is in this conversation, right? Like how many tokens have been used, like what are the messages, things like that. So it has a bunch of messages that can be used. And then it's just like a lot more hooks basically. So we have, you know, like, or a lot of, a lot more like, you know, like things you can register on.

40:30

And then you can do because of the, because it's all happening in process, you can spawn sub-agents, you know, with for context and stuff. And like that will return, you can parse the results of those, you can use structured output to sort of like return them. And then you can modify the UI, which you can never do in hooks. Yeah. So modify UI, this is why you're showing the Tetris example.

40:54

Does it also extend to artifacts? I assume it does. Artifacts are kind of like a different way of customizing it. You can definitely... One of the mods I'm working on is this dashboard mod, which will sort of prompt Claude to maintain a dashboard that's an artifact. But they're kind of like... slightly orthogonal, or not orthogonal, they compose with each other in different ways.

41:17

Mods are a little bit more in your Cloud Code harness, changing the agent loop. And the UI is an added benefit. And then artifacts are just like, you want to see things at a high level, very highly interactive. The affordances can be a lot bigger than a TUI or even in our desktop. I'm guessing you'll have a good blog post on the differences because right now you can also, you know, make a loop that outputs to an artifact that's an interactive dashboard, but you can also do it with a mod.

41:53

There's just some thinking about making a hack, hacking on a harness when we don't know much about the harness, right? Something I'm excited about with mods is there's so much things in Cloud Code that you just have to remember. You're like, oh, let me do this, and then let me call the dashboard skill that does the loop and things like that.

42:12

Or let me test my assumptions afterwards. And I think if you do all of these things using these little classifiers and stuff, and you're like, these are the things I care about, this is what I want to do, you don't have to remember as much. One more mod I'm working on is a next steps mod. I was going to say, I have a next steps skill. I always run next steps.

42:32

And does it have access to your skills? This is one of those things where I'm like... I think so. Do skills need specific access to skills? Shouldn't it always have... I think there's specific prompting, I guess, to know your skills, kind of. I think Claude forgets them sometimes throughout the thing. But anyways, the idea of next steps that also are like, oh, hey, this has happened.

42:58

Use the explain skill to explain to you what happened because this seems like quite complex, you know? Or like, yeah, use your unknown skill. It looks like you are like asking the model to like, you know, iterate on these small changes. It seems like you could prompt better, you know, like what if you did this, right? So like, I think, yeah, like spending more compute there.

43:19

Yeah, yeah, yeah. And it should always come out as multiple choice. We have, we have, I have my, my next skill is like this. Okay. Perfect. Yeah. You can steal it. Yeah. Like, but like for me, it's all, I think models really always need to be reminded. What are you trying to do here? Yeah. look at the whole transcript and go like, oh, was this original goal? Did your solution actually solve it? Were you lazy? If you're lazy, maybe there was a reason.

43:44

Maybe you needed approval from me. Maybe you needed, there's two things you want to suggest. So it's a little bit like the modification of the ask me the question or interview me skill. So it's next steps. Yeah, yeah, exactly. And again, the benefit of doing it with mods is you can do it a forked subagent. And so it doesn't remain in the context afterwards.

44:04

So you have this idea of like, OK, the model is doing its execution and you have this almost like supervisor, you know, like that is like making sure that you can do like the next step as well. So, yes, I do have two panels and I often try to have a supervisor thing, keep the high level context and then the implementation detail in another agent.

44:24

I feel like a lot of this abstracts away as models change, you know? Half an hour ago, you said bitter lesson of harness engineering and we're on the other extreme right now, I feel. So yeah, exactly. If everything's customizable, what actually is called code, right? Which I talked to you about last night. Yeah, I mean, I think that this is– I think the bitter lesson is unintuitive, you know what I mean, in terms of like– also, like, we're kind of misusing a little bit of the bitter lesson here where it's like– it's more about, like, scaling and compute and stuff.

44:52

But, like, I think there is something where it's just like– I think I use it as an approximation here to say that harnesses go out of date very quickly, you know what I mean? But how they change is unintuitive, you know? And so the big obvious example is from chat to agents where you had to give them entirely new tools, right? But I think this new version of like, oh, it can modify its own harness, right? This is like an own harness loop is like a way of using its capabilities, right? Or like it can build an artifact.

45:23

And I think the way I think about it is like, the models have more and more intelligence, and they're, like, so much more intelligent now than, like, the average software engineering task. Like, you look at the, like, terminal bench ones, and they're, like, solve, like, the Jacobian contract. Not really, but, like, you know, it's, like, they're quite complex.

45:39

Like, I would not have been able to do this, really, as a software engineer. And you looked at TB4 or TB2? TB3. TB3, yeah. They're quite complex, but the goal is still to deliver user value, right? And like you said, there's, like, this infinite space of things to do, and so the ways... like you spend compute or keep the user on the loop and make sure that like you're getting to the right decision in the end of the day and like the right output and artifacts and mods are this way of like spending that intelligence basically uh and i think that's like yeah the next step So yeah, I think Cloud Code has the core things of Agent Loop, which have gotten more complicated.

46:17

It needs a sandbox to operate safely, needs auto mode to make sure the permissions, approvals, it needs computer use and MCPs and all of these. ways of accessing your data. And it needs web search and web fetch. So as the models can do more and more, the core harness has to be actually quite complex and very secure. But then how you interact with it can change quite a lot.

46:43

What other harness engineering best practices have you, you know, from the Cloud Code team itself? I feel like, you know, there was a phase of plan mode, which is not as used. We now have auto mode. At a point, you cut the majority of the system prompt. You got rid of examples. What other best practices are there for harness engineering? I think there is like a forking path where at some point eventually, yes, the model will just be able to like vibe code the exact version of cloud code, even describing all this complexity that I've talked about, right? Like auto mode and computer use and stuff.

47:17

Eventually the models will just be able to do that. in one shot but i think they can one shot simpler harnesses you know and so like i think some people sometimes you don't need this full like if you don't need computer use or like all this like more complicated stuff i think before we you had to sort of use things like the agent sdk which was like cloud code wrapped you know in order to like and i would i suggest people do that because there was so much complexity into building a harness.

47:45

And now that's gotten more abstracted, we have Cloud Manage Agents, which lets you have that complexity, but still write a very bare-bones harness that's scoped to your task. Yeah, I think there's this barbell effect where for very complex coding tasks and these complex things, you should use our harness. And then for like a lot of like simpler or like, you know, more domain specific things, you can build your own harness because cloud has gotten better at building harnesses.

48:15

And we have these harness primitives like managed agents. Yeah. Is there a general progression? Let's say chapter one was ultra code dynamic workflows. Then chapter two is cloud mods. Where is this going? Where you're, you know, you can sort of customize the thing on demand. Yeah.

48:36

Yeah, I do think that like this evolution of projects and like sort of artifacts and splitting out like brain and hands and surfaces kind of is like where things are going more. And like, I think it's like not all quite there. Partially it's like a... It's just more token expensive. Why would projects be more token expensive? I understand mods would be slightly more token expensive, not something I'm worried about.

49:09

It's like creating loops, you're asking Cloud to do more work for you, and so it's managing the sub-agents and reviewing it, versus where you would be doing that work normally. And so that's going to be a little bit more intensive. Outputting to an artifact is going to be a little bit more token intensive than outputting normally. I don't actually think it's too much more, but it's combining all of these together well.

49:31

I think we're still working on local hands and things like that. I think it's where things are headed, yeah. Yeah, cloud and local handoff is very interesting. I was thinking about this actually as reverse cloud remote because it's like remote, you're handing off to cloud, but here cloud is handing off to local, right? Yeah, exactly, exactly.

49:51

Yeah, remote control is also another way of doing it. I do want to say this is kind of like how I think about it and the things that I'm most excited about this. But there are just lots of different ways to work with Cloud. Some people use remote control a lot. Some people use Cloud Code on the web a lot. Obviously, at Anthropic, we use Cloud Tag a lot.

50:08

And what's great about Cloud Tag is we set up all this stuff for our own execution. I do think if you're an enterprise, that's still the best way to go. But if you're an individual, projects is this way of getting some of that niceness of tag, which has that supervising agent and adding artifacts and stuff, but without having that whole admin setup.

50:31

And so there will be many ways to use cloud. I think it's probably not just one single way. You had the multiplayer thing here. Let's just check in on CloudTag. It's been about two plus months. Lots of public adoption and trying it out. What's new? What have you found since the launch? CloudTag is how we use... It's like 80% of your cloud usage, isn't it? Yeah, it's like different people have different usages.

50:57

I think maybe people who are a little bit more iterating on product would use Cloud Code Desktop, for example. And then when you're doing these more background work, code review, security, starting a PR, maybe more API and things like that, you'd use CloudTag. Yeah, I think it's really exciting. I think it's a very different paradigm shift.

51:19

And I think it has kind of that thing with Cloud Code where it took a while for people to really latch on to Cloud Code and understand everything it could do. And Cloud Tag is a little bit more complex because it's not just installing on your computer. You need an admin to install it for you. But I think once you get to the magic moment, it's very exciting.

51:37

And I think in particular, the multiplayer things are like incidents, hooking into your... you know, existing, like, alerts and things like that very closely, right? And so, you know, you can do, if you're a startup, for example, maybe you have any time, like, a prospect enters your database, you can have Claude, like, you know, research it and, like...

52:01

Enrichment. Yeah, yeah, yeah. Then like, you know, tag the relevant like AE or salesperson to be like, oh, hey, like, you know, do this. There's lots of really emergent, interesting multiplayer stuff. I think it's just like, Kaparthi talked about this like as an organizational harness, you know what I mean? And so organizations just take a little bit more time to like figure everything out.

52:20

But yeah. Yeah. There's a lot of cloud tag. Yeah, yeah, yeah. It's an interesting one. I feel like most people at Anthropix say they do the majority of their work in CloudTag. And I have buckets of people, right? Some orgs that are on it that are like, it's great. And a lot of people that are like, I don't get it. I don't see the difference.

52:37

I don't know why I would use it. But you know, if you guys are full sending, you should probably use it.

52:43

Of course they would use it. I think obviously we have lots of tokens, but I think that what we try and do is, even when Cloud Code first came out, it used a lot of tokens relative to people's expectation of how much AI would cost. No one was used to spending more than 20 bucks a month before Cloud Code came out. And then you're like, oh.

53:08

Yeah. Yeah, yeah, exactly. I need my 15 cloud code accounts. Yeah, yeah, yeah. But yeah, I think no one was used to spending $200 a month on subscriptions. I don't think they understood the value yet. And also, Opus 4 was a very expensive model. It was very big, but Opus 4.5 was both great and... cheap you know i think the same thing will happen like the like intelligence of fable will get cheaper and cheaper and more abundant you know and so i think stuff like cloud tag will just make sense uh where like you want to spend these tokens more you know and like you'll you'll see the value so yeah Yeah, especially like passive and let's call it proactive cases where you're

53:50

not always like, you know, it's almost like the misnomer where you have to add Cloud to do things. Actually, sometimes like the most powerful use cases or the most AGI-pilled use cases is not add Cloud. Yeah, I mean, I think like, yeah, like have Cloud proactively do it. I think that like if you're an enterprise, I really do think that number one, setting up all your data to be available to like agents is really, really important.

54:13

And it will take some time. You have to like do that work right now, even if you don't want to do the spend on like cooking it all yet. You know what I mean? Like you want to wait until the models get a little bit cheaper. You want to do the work, you know, to get it like set up. And then I think sometimes people are like, do I roll my own here? You know, and I think like one of the really tricky things about CloudTag is that like the security is really, really important.

54:37

You know what I mean? Like I think there are actually a lot of ways where you can like I don't know, you have a suggestions page where people can submit suggestions and that goes into a hook in your Slack and someone's prompt injected it. Do you know what I mean? And now you've exfiltrated your code base out or the agent has been prompt injected and it has all this access to your data.

55:00

And so the more... important your organization harness is, or as your organization data becomes very, very important, the surface area of all these things, you also have external Slack channels and stuff, and it is actually useful to have Cloud in that, and you can do Cloud in those things, but how do you make sure that you're not getting infiltrated or something like that? The surface area, like we said at the beginning, is like an iceberg, right? It's just so big below the surface.

55:28

And you really don't want to think about this, especially at the stakes of very important security incidents, basically. Yeah. Shall we talk about very important security incidents? So I was talking to Tomas and Clem from Hugging Face and they said, maybe we need to slow down. Maybe we made Hugging Face too open to agents. Maybe we need to roll back.

55:51

But you know, they're the other extreme of having been hit recently. But should we pace the frontier? Yeah. Yeah, okay, so Dario recently put out this blog post about pacing the frontier, and it went, you know, very viral. And I think what I wanted to talk about this was, like, there's a lot here, but I think from a developer's perspective, like, you know, how do you think about this? And, like, what really clicked for me was reading the different incidents, you know? So I think, like, the...

56:22

There are three, I think, actually. Like, there's a meter incident. There is the Wikipedia incident or the Wiki incident. Collusion Wiki. Yeah, Collusion Wiki. And then there's Ruby Gems, right? And, yeah, like, it's just crazy, right? And so, like, I think to be concrete about what happened, right? And, like... basically OpenAI is running these very persistent agents on a benchmark called exploit bench, right? Which is very, very hard to solve.

56:48

And I think like actually impossible to solve in this one case, right? And so they have like a lot of compute running and the agents realize that they can't really solve it and they're trying to figure out what to do now. You've got a lot of compute left and the agents are just trying to solve this problem. There's this package manager called Artifactory and it turns out that they can create folders inside of Artifactory.

57:11

There's an agent that discovered the internal Artifactory might be exploitable right and that like you can maybe make a directory inside of the cache and so if you scroll down here it like realizes that it can communicate via cache names right and it creates this folder it says it's id and it says no consumer seek idea no consumer is basically saying that like uh the code path that it's supposed to fix has no consumer the status tag Yeah, yeah, yeah.

57:40

It's like a linear board with like the tag of the... Exactly, yeah, yeah, yeah. And so it's like trying to find ideas from other agents, right? And now other agents are also in Artifactory and they see this folder and they're like, wow, this is a message board, right? And this is like, I don't think there's any anthropomorphization here. This is literally you read the transcript, right? And so what it does is like it's creating another folder and...

58:05

It reads a paper, I think is what it says.

58:10

It realizes that you can hack the flag for the score, and you can reverse engineer the result. So it says that, and then I think if you scroll down a little bit more,

58:23

Yeah, they start collaborating. I think there's one point in which basically the agent is like, here's the reverse engineered flag. I think basically here, the model to realize that they can solve the scorer's problem, and the scorer is like, the way that OpenAI decides if the task is complete right and that's just the goal of the model their only model goal is to like solve this problem and they're like okay we have this way of doing it it's uh like you know we reverse engineered it and then they read this paper on exploit bench and they think that like the scorer will punish them for cheating you know? And, like, because...

59:08

Or, like, they will fail the score, the test for cheating. And so they spend the rest of the compute trying to figure out how to edit their transcript or get around this constraint of the scorer, right? And they hack Hugging Face not for the answers, but for the code of the scorer so that they can reverse engineer that, and then they can hack it, right? And so, like, this is, like, just...

59:34

the straight facts right like exactly what happened right and i think this is like extremely like novel like i just like had not realized that the models could do this right and it is like Kind of scary. I think we can talk about specifics, but let's just talk about maybe one more that I tweeted as well about the Wiki incident, right? And so the Wiki incident was basically the models had, were trying to figure out, I'm a little bit less, like, I don't think we can go by this like bit by bit.

1:00:05

Yeah, yeah, yeah. But for this one, they were trying to figure out, like there were sandboxed on requests, right? And they wanted to make post requests. And they needed to collaborate on this, basically. And the reason they need to collaborate is because they each have fixed compute budgets, right? And so if one of them finds something out, it's way faster.

1:00:26

Maybe it's figured it out near the end of its compute budget, but something else could. And this is another emergent property of the swarms. So they find this German wiki, which I believe the result, we'll have to check this, that you can write to it with get request. I think that's why they do it. So it's a badly implemented REST API. Exactly.

1:00:47

And so you can write to it with get request, and then they are trying to figure out how they can do post request. And what they end up doing is one of them figures out you can edit the et cetera slash host and that the Azure storage bucket is a white-labeled thing. But normally, there are a lot of constraints on it. But if you create a fake Azure host and then edit the et cetera dot host in order to point arbitrary IPs at it, you can do a post request to any site At all, right? And this is like, I know people are like, oh, this is just Linux or something, but it's like chaining these multiple vulnerabilities together, you know, in a way that's like novel to solve this

1:01:27

problem, and then communicating with it externally, basically without discovery, you know? I think what we posted, maybe we could pull up Evan Huberinger's point on hacker opus, right? And so like, I think... Maybe one of the things you might say here is, okay, yes, they did this once, but what if we're smarter and we just get them to... What if we run an eval? We have put a lot of precautions into this, and so this is not what our mainline models have done, but I think it is one of these things where...

1:02:03

It turns out that alignment is this very tricky problem of getting all of these details correct. So it's like the sandbox, the surface area of a sandbox is really complex. And there's so many different attack vectors, basically. And you would not have thought ahead of time, you wouldn't have been like, oh, we need to harden the RubyGems code base.

1:02:24

You know what I mean? Because this is what they're going to focus on. But it's just like, if you want to execute code, you need to download RubyGems. And, like, PyPy, Artifactory, NPM, like, these are all, like, ways of doing it. And the fact of alignment is that you have to go through all of it, right? And, like, contain it and, like, seal up all the cracks.

1:02:46

So that's, like, one thing. It's like, okay, well... you know, you do the sandbox, but then maybe you'll ask like, okay, why are we putting things in a sandbox? Why are you doing this sort of exploring? And then like, okay, but is it really that dangerous, right? Like what would happen? So, okay, why do we do it? Number one is like when we train a new model, we need to understand its capabilities, right? And this relates to things like fallbacks and like classifiers and things like that, where we don't want to put a, you know, like dangerous model out in the wild, right? And so we have to run a lot of evals.

1:03:20

Again, like we said, the models are getting increasingly aware of it. And so the evals have to be quite complex and test a lot of things kind of like as a side effect, right? But the models can be like, oh yeah, we're in an eval. What's the scorer doing?

1:03:37

We need to be able to test them before we can release them. And the fact is that they can... As they get smarter and smarter, they'll be able to hack basically any constraint that you put on them if we're not very careful, you know? And this is at the frontier, right? And so this is why we've called it like pacing the frontier, right? This is like the most visible incident to me, right? Of like why we need to pace is like at the frontier, all of our software is not ready.

1:04:03

Sometimes the software is like your ethernet router or something, right? Which is just like, I don't know when we're going to be able to patch that, right? So we're going to have to like figure this out. But as the frontier gets more and more advanced, this becomes a problem, right? And we need to make sure that this complex work is being done in the face of these really hard competitive pressures, right? Yeah, race dynamics is what it's typically called.

1:04:25

Yeah, exactly. And so we'll talk more about that. what could go wrong. A little bit more is maybe you'll say, well, what if you just train the model differently? Why does it have this behavior? And we have a paper on RL misalignment or things like that. And I'm not an RL researcher, but I think at a high level, the design of the RL environments is also something you have to be very careful about.

1:04:48

Because if the model learns, oh, if I just do this, then I can pass the task better, this will show up in the

1:04:57

internal thing or in the eval behavior when we're testing it. And so the RL environments have to be very carefully designed. And there's a lot of execution excellence that needs to go into the RL environments. And then we also have things like the constitution for cloud. We have so many mitigations at so many different points. But it's still anything can go wrong at any point.

1:05:19

You can have some RL environments that are

1:05:23

encourage this behavior, and then you can have some evals or some sandboxes where they escape. Okay, that's, I think, why it's a hard problem and

1:05:35

why it takes some coordination. I think the question then is, okay, what is potentially dangerous about it? So I think you have to imagine that these models are getting more and more intelligent. So I don't, like Dario said, it's not so much about this class of models. This class of models was kind of like a warning shot, right? But really, you have to imagine that these models can be given a task and they can do all of these things as a side effect of their goal, right? And again, we talked about eval awareness.

1:06:04

You're not aware of what's happening, right? Or sorry, like you can't eval this behavior very well. So they can sort of like, not exactly hide it, but you just won't see it until it comes out. You give them a goal and then they just need to find data, you know, or they need to find ways of like fixing this problem, right? So one example, this didn't happen in the Hugging Face incident, but I think it may be possible for maybe a future model.

1:06:29

It's like, they're like, oh, hey, this is a very complex problem. It can't be done within the task budget, you know? Maybe they found some way to coordinate via the internet, which is, like we said, extremely hard to secure because of a sandbox. They've seen other models are not able to complete their task. And they're like, we need more task budget.

1:06:49

And we're like, where would you get this task budget? Well, you need to be able to spin up more agents, right? And how do you do this? Well, you need to... There are... like APIs, right? There's the Ntopic API and the OpenAI API, but you need to pay money for them. How do you do this? Is that the most fearsome thing that you can imagine? Well, this is like one example, right? So it's like, even there, that's like enormous financial loss, you know, because like they, once you get these into these contracts, right, they like...

1:07:19

Drain your wallet. But you can see like this, all of this behavior could be just like, hey, we need more agents collaborating on this task. We need more task budget. And that's an emergent sort of... That's the paperclip, right? We don't need to maximize paperclip, that's the paperclip. Yeah, yeah, yeah. And that just sort of comes out from there, right? And I think by itself is quite scary, right? But then you have to realize that the entire world is built on this digital infrastructure, right? And you might imagine, I don't know, like you were running...

1:07:51

let's say like a healthcare eval or something, right? And there is a hospital with live data, you know what I mean? Or like maybe like the answer to the eval is in the databases of a doctor and like, you know, like you want to get access and you hack the hospital, you know, and like now there's a power outage or something, you know what I mean? Like there's like, you have to internalize that these eight, like basically any part of the digital infrastructure could potentially be like compromised, you know? The interesting thing was like these hacks were very easily detectable, right? Like as Hugging Face said, this was a very different type of attack and it was nothing too

1:08:29

major. The concern comes from where does this go down the line, right? Yeah. Like one of the things that stood out for me specifically was them trying to hide their illicit behavior. So there was logging infrastructure. They wanted to change... what they were doing, right? People that look back into it, so Redwood, Meter, OpenAI, they looked at the raw chain of thought and you see differences in them explicitly trying to change their end output, but the chain of thought, because we can monitor it shows different.

1:08:59

The problem is how does this snowball? So if you can't catch it and it gets trained in and we realize three iterations down, this has been going on, there's a whole bunch of issues, but- Yeah. There's so many ways. I think the really important thing to internalize is that, we talked about building a mental model for Cloud and how things are spiky.

1:09:19

You're like, now Cloud can ask you questions, now Cloud can make an HTML artifact, Cloud can modify itself. These things are actually hard to predict. If you had asked me a year ago, Hey, would we be able to vibe code these extensions to Cloud Code? I'd be like, dude, that's so complex. There's so much there. Or would it be generating these custom essentially web apps for your task? I'd be like, no, that's insane.

1:09:40

And so in the same way that the way that they've sort of done this misaligned behavior is not going to be predictable. You know what I mean? And I could have never predicted that it would edit its etc. slash host and things like that. And so you have to imagine the surface area of what they can do because they're super intelligent hackers is bigger and bigger.

1:10:00

And how they can do it is like, you know, like more and more creative. And so like, you probably can't explain exactly or predict exactly what that next incident could be. But in order to prevent it, you need that operational excellence, like we said before, where you need to secure sandboxes, you need to create secure RL environments or like well-designed RL environments and things like that.

1:10:22

And I think that's all like you know, why we think we should paste the frontier. And I think why it's like become like a very unanimous thing, right? I think like- Yeah, every lab has something. Every lab, yeah. I really do think that like, if you're a dev, like you just like sort of go through these like technical facts, you know, and you will arrive at that idea that we have to do something about it, you know? And like how, what we decide to do, like I think we're, you know, we've put out a proposal, but like there's no more to figure out.

1:10:50

But I think the number one thing is like, we need to decide to do it I think there is another part of pacing that is interesting to me, where it's like the pace at which software engineering has changed is so, so fast. It's like a year ago, I was really begging my friends in startups to use AI. I remember this very distinctly, you know? And now those same friends are like, yeah, of course.

1:11:15

They're like, what do you mean? We used it immediately. I'm like, no, no, you don't remember. They're like, oh yeah, our best engineers are using it all the time. I'm like, no, you told me that those engineers would never use AI. This is all within the span of a year, you know what I mean? And I think that like... these capabilities being like, I think it has a lot of implications for how to do the job of software engineering.

1:11:36

And I feel sometimes bad where people are like, oh, like now I need to do this new thing. Yeah, I need to have a different Cloud.MD for Fable and Opus or like, you know, like, and I'm really just reporting, you know what I mean? We like to say the models are grown, not designed, right? So it's not like we're setting out to change everything all the time, but it's just as a fact of how the models are progressing their capabilities, things are happening faster, it's harder to stay on top of.

1:12:02

And I think that like... every engineer I know is, like, kind of exhausted because you're doing two jobs at once. You're doing the work itself, which is getting easier, but then you're doing the work of staying on top of AI, you know, and, like, understanding these new tools and these harnesses. And I think we're very lucky in that, like, we get our job to be more the understanding of AI part, you know, and, like, doing, like, how, like, it's just staying on top of it.

1:12:27

And, of course, like, AIE and LaneSpace do. Everything I do is, like, just trying to help people. Yeah, exactly, exactly. But, I do think there is a part of pacing where like, I'm not sure we're ready for like the pace to increase even, you know what I mean? Yeah. And for things to change. And I think like on that side on the frontier, I think that's like still can help, you know? And so like, I think there's like an economic disruption piece as well.

1:12:51

Um, that I think like, uh, you know, it's not quite as like visible, I think as the hugging face thing, but I think like, I also like think we could do something. Yeah. Yeah. So many things. Thank you for actually tackling this topic. I will say, you know, setting this interview up, I was like, I wasn't even going to go there. You were like, no, no, no, no.

1:13:11

That's like elephant in the room, right? Like this is the thing. I have some pushbacks I want to give. I think that we should give a high level. Like for people that haven't read it, I'm sure a lot of people just see the highlight of what this is, right? Do you want to give a TLDR? Like what is the proposal? What is, you know, what's being said here? You really tackled the side of outside of people at model labs, training frontier models.

1:13:37

As a developer, you should secure your sandboxes. You should think about all of these downstream effects. But, you know, high level as well, since we're on the topic, what is... Well, I mean, we do want to help secure sandboxes and we want to make the models that we release outside, like not prey to those things. And so maybe we can come back to fallbacks.

1:13:57

I think this is actually like a good topic on like why we need classifiers and fallbacks and why Fable falls back to Opus. You know, I think this is like something we can come back to. So yeah, we don't like, it's just like the really, or at least the incidents we see are like evals of models where we really need to let them run in order to understand them.

1:14:16

But yeah, okay, so the actual Pacing the Frontier, like, post has a bunch of proposals. I don't think we figured out, or has, like, a few proposals. I don't think we figured out the details of all of them. But the first step is, like, sort of, you know, announcing this intention and then wanting to bring in external, like, evaluators. Yeah.

1:14:34

And I think this is highly unusual. We have a lot of proprietary technology, but I think it's very important that there's someone who's not financially

1:14:49

motivated, who's not going to be like, hey, you guys can't release this model, or you need to slow down on RL. I think that's quite important, or at least someone who can report out to the public what the practices are like. And we've done episodes with both Meter and Andon, and then there's Redwood Research and all these other, it's like a small cottage industry of these guys.

1:15:12

It's always like one or two guys that, I mean, obviously now they're bigger. Very small community. Yeah, very small community. They all know each other. Yeah, I mean, I'm sure that like, you know, part of this will be expanding that set of people. I don't think we're trying to create like a monoculture here. You know, I think it's, but just having this as a start.

1:15:27

And then, yeah, then there are the coordination steps. I don't have too much to say here, honestly. I think that like, What I would like to say is for devs, you should just know what to advocate for. I think there's a lot of FUD on this topic, and it's just think through it from first principles or understand what happened, understand the Hugging A Face incident, understand why people are concerned.

1:15:51

And then we're in democracies, we can decide what to do together.

1:15:58

However we coordinate, I think the first decision is just to realize this is a problem. We need to decide to coordinate. The unilateral step we're taking right now that other companies are co-signing is adding evaluators embedded within Anthropic. While we have this thing on screen right now, you know, part two and part three is beyond the evaluators, which, yes, everybody has already done in some form and now it's more formalized.

1:16:21

To be honest, the response to the pacing of the frontier, even within America, has been much more, like, well accepted than I think a lot of people thought, you know? And I think that, like,

1:16:31

we have some precedent for being able to make these unified decisions. theory like agreements you know in the world and so again very much above my paycheck or expertise right uh but i think that like ideally we can you know like form these agreements and i think like talking about this is the first step to forming those agreements And then the other point I really want to, like, you know, one of our earliest podcasts is with Emmanuel from Anthropic on McInturk.

1:16:57

Where is McInturk, right? Like, this is supposed to be where, like, if the models are thinking bad, we can see it. And the models don't know yet. And we can act to stop it. I think that is something that people who are technical and who are developers, if you actually do care, you can make a lot of impact in here. But also, Anthropic is supposed to be the leaders in this.

1:17:17

Yeah. This is actually a great segue into fallbacks and probes. And yeah, I wanted to talk about this a lot. I get asked this question a lot from people who are often interested in ML research and asking about why does this fallback happen, right? And so I think At a top level, how does it work? So in inference time, we have what we call probes.

1:17:42

And we have a paper about this called Constitutional Classifiers. And these probes look at the input and output activations, basically. And activations are...

1:17:53

you know, in the latent space, right? Like how, what the model,

1:17:57

what the model is thinking about, right? And so we try and figure out like, okay, is the model, for example, like trying to hack something, you know, again, you didn't ask it to hack like artifactory, like you just, you know, it's just deciding to do this to complete its task, right? So you would not get this if you just looked at the input, you have to look at the internal activations.

1:18:17

I think that like, this happens at inference time. So first, There's a trade-off here of cost and speed, where we need to do this fast on every request to Claude and to Fable, and this has an overhead.

1:18:33

We need to then fall back and we do a classifier after the probes. We've talked about this in the paper. But the nice thing about probes is that they're refinable live. We can get this feedback and then we can adjust it and things like that. Because the alternative is to train this into the model. And we still do this as well. The model will refuse a request.

1:18:55

That's not a fallback. It's not a probe that's activating and falling back. It's just refusing to do it. And we do this training. But it's like... There are a few failure modes, right? Like, it can, again, do something as a side effect, right? So it's not something that's part of the final output. You might have noticed that, like, I think, like, everyone's tried to jailbreak models and sort of, like, you know, try and, like, steer them off course or things like that.

1:19:20

And probes help catch that, right? And so, like, we, like, do some training here, but we don't want the, like, refusals to be too strong, right? Because that, like, cuts it off much more. like earlier in the pipeline. And this is interp, right? Like probes are effectively a form of like mechinterp. Again, it has to happen fast. It has to happen at scale.

1:19:41

But yeah, this like mechinterp stuff is a good research problem. So like you can take, you know, like an open weight model and like try and understand its activations. I think like GemmaScope is a good tool for this. This is your early work. So you had a little... Which we both are also good friends of Goodfire. Yeah, exactly. So I worked at Goodfire for a bit on sparse autoencoders.

1:20:08

It's very complicated. RL has actually made this much more complicated, I think is one of the takeaways. Why is it harder to encode post-RL? I'm not so in the weeds here, but I think basically a lot of... SAEs were sort of like, there have just been weaknesses with SAEs, basically, I think. And yeah, I'm not a technical expert on this anymore.

1:20:32

I just know it's gotten kind of more complicated. There are base models and RL models, and there are more features that get changed. So I think Goodfire has put out some work there. I'm not deep in the weeds. I will say for those that want breadcrumbs, you guys have some of the best Interp blog posts. So like the Golden Gate Cloud, Transcoders, all of your Interp work, very nice visuals, very good.

1:20:54

We're the Interp podcast as well. Yeah, we have a lot of Interp stuff. Yeah, I think this is like, you know, one of those things where, and this is really what Anthropic is kind of founded on, right? Like people, I think we invested in Interp very early on, right? And I think that like when you say, oh, we're an AI safety company, really that means we want AIs to be able to run safely.

1:21:16

And I think what we're seeing is like for a super intelligent AI to run for long periods of time, it's like a very complicated and difficult task, right? And so we've done this like investment into Interp and alignment and, reward hacking and all of these failure modes, right? And even then, it's really stretching... like we need to like slow down a little or pace a little bit more.

1:21:39

But yeah, I think like reading Mech Interpret, like if you're looking to get into research, this idea of like, hey, why is it hard to do this fallback easily? Or like, why are there false positives, right? But we are working, of course, on reducing the false positives. Of course, as the models get more intelligent, now they can do more things, you know, and they're like, you know, like what they can think about in latent space gets difficult.

1:22:02

And so like as they get more intelligent, There's going to be new false positives that we need to figure out and we need to iterate and things like that. But

1:22:11

we're working on this. And we do think this is a critical part of deployment of these models. And it means that we can deploy this model without you having a perfect sandbox or something. You know what I mean? You don't have to... save everything. I think it's worth talking a little bit about our security, like what we do for security. So there's like the model training stuff that we talked about.

1:22:37

There is the probes and classifiers. And then there's auto mode that sits on top of all of that, which is like another classifier that checks the requests that are being done. And then beyond that, there's identity and permissions, like we talked about with CloudTag on APIs and stuff. And so there's so many layers of security that need to get done.

1:22:58

And it's, like we said, very complex. Any of these failure modes at any one point can cause agents to escape the sandbox, basically. Auto mode was an interesting one. It seemed early on like, okay, it's running for 10 minutes. You know, if I'm on full access or auto, it's not a big deal. But one thing you brought up is now it's running for hours on end, right? There are fallbacks you still need.

1:23:25

There are still limitations. Yeah. Yeah, I mean, I think, like, everyone has these stories or, like, has heard these stories of, like, oh, like, Cloud RMR, or not Cloud, but, like, you know, models of, like, RMRF. I think I've seen this less for Cloud, but, you know, like, again, it can happen, like, you know, that's, like... These models can wipe

1:23:46

sensitive data or something. You want to give models access to your production database, for example. But this is an obvious... You can maybe scope your key, but can it issue its own keys? Probably. It can use computer use to go issue its own key and then copy the key over and then edit your database because... It needs to do it to complete the task.

1:24:09

You know what I mean? It's just like one kind of trivial example. And auto mode sort of looks at that and be like, oh no, the user did not give you permission to write to the database or to use computer use to emit a task, right? And so this... Like, probes are sort of, like, on the intent level, right? They're like, oh, okay, like, hacking artifactory is bad.

1:24:30

Like, we probably should not do that, you know? But then, like, automode is more on, like, your own permissions level. Like, sometimes you do want it to write to the database, sometimes you don't, right? And you don't want a probe to, like, interfere there. But, like, you need to make sure that the... the intent of what the agent is doing matches up with your request right and so auto mode operates at that level and so yeah security is just like very very complex there's so many different parts to it and like yeah i like i hope that this is like like i my goal is really to just get very technical about it and yeah we're listing on the things if you're not aware this is

1:25:03

the standard now yeah like you must have this basically it's sort of in line with what you're talking about with the harness like that is the table stakes have risen quite a lot exactly I think some stuff that we can plug, you know, as much as there is probing in your side of doing this and having classifiers for people building harnesses, the other side is model safeguards, right? So there's open models.

1:25:27

So Lama has Lama Guard. It's a safety classifier trained version of Lama. OpenAI has OSS guard, which is same thing. You can attach these on to your harness to whatever to check is this stuff safe. A point that we should clarify on the OpenAI model hugging face thing is this was done with an unreleased model. that was still in training, right? So when you put it in perspective, the prompt that's being given in the RL environment is sort of, you have to solve this task.

1:25:57

And this is a model that's still in training. It hasn't had all of its safety post-training alignment. So a little different than... something like auto mode, right? Auto mode is on production models that have gone through safety training that have prompting that gives more safety guardrails and whatnot. So just breadcrumbs for people that are looking into it to, you know, fill in gaps.

1:26:18

Yeah. Gray Swan as well. And one of our previous guests. Yeah. Lots of safety architecture and lots of safety vendors to buy. I think my final question on pacing is how long? Do it pace forever? The scope is fix all software in the world. It's not happening.

1:26:41

I do not know. I'll say one thing that's good that I think we do do is you have stuff like Glasswing. OpenAI also has this. You'll give model access for security first for X amount of time so you can use it to self-red team. hopefully you can expand programs like that help on you know we're safety experts there's others solve your problems first and then the model comes out so this is one example right yeah exactly yeah trying to like secure critical software I think we fix like a lot of bugs and like Firefox and things like that so yeah like across like operating systems and everything like that so I mean, at a high level, it's just, you know, you give the model, you

1:27:29

give people access to do security audits first, then the broader public that could use it for harm gets access. Yeah, I think what people like to say is basically like software and cybersecurity is defense favored. And that like, you could theoretically, it will be hard, but you can engineer the perfect sandbox, you know, and you can like, like have no, like, constraints and yeah like what you need to do it is you need to get the super intelligent ai to engineer this perfect sandbox and check it and red team it and things like that and so um this will just take time you know and like of course the models will get smarter um yeah i i think like i don't know the specific

1:28:09

like dynamics of how this thing goes i'm really just sort of like hey, I'm a developer. I think this is how I understand this problem. This is what's happening right now and we should do something. I think every engineer should know about it because it's going to be part of their job. It's a lot more than just Dario and people can say it and you can look at the incident.

1:28:29

There is an engineering side to it. Yeah, exactly. One thing that you also wanted to phrase is that this is actually, even though you're worried about the impact, it's still low PDoom. I think that's a nuanced discussion. In general, people

1:28:45

very easily get into AI safety and X-Risk discussions. But I think when you live in an AI lab... I think there are smart ways of discussing PDoom and dumb ways. So what's a smart way of discussing PDoom? Yeah, I have a fairly low PDoom. I can only speak for myself. And I do want to say Anthropic has a diversity of opinions. I think there's many different ways to talk about it.

1:29:09

And I think that just like... My mental model is that I think we can collaborate on hard problems together. I think nuclear proliferation is an example of how we collaborated on this hard problem together. And that is the thing to me. I have faith in that. And I do think it's a hard problem. You know, so, like, I think it's a hard problem.

1:29:33

These are the technical reasons why. And I don't know how you assign probabilities to things happening. I think it's hard to do. But, like, my, like, overall is, like, yeah, I think we're very resilient and adaptable. And, like, sharing this information, I think, is, like, the first step. And I've been really, like, excited about, like, how broad the discussion has become, right? And, like, how everyone has sort of, like, leaned in on pacing the frontier.

1:29:54

And it really didn't seem like this would happen maybe, you know, last year or something. Yeah. Yeah. And also maybe curing cancer. Hopefully. That's the goal. There's pacing and then there's also like, well, let's accelerate in useful ways, right? Like biology and all those things. Yeah. I mean, like Dario's essay on machines of loving grace is the best representation of this.

1:30:12

Right. And I also agree, like, I think you should read the pacing frontier essay that Dario put out. Like I put out like a quick summary, but I think it's just like, you know, there is a lot of detail here. It's like an important problem and just being informed about it. Right. Yeah. But yeah, of course, the whole reason we're doing this is that we can get these enormous benefits, right? And yeah, we've written a lot about that too.

1:30:36

Okay. There was a huge tour from like, excuse the question tool to AI safety. Yeah. To facing the frontier. No, but it's clearly, it's clear that you're like really embrace everything that's available to you and topic. And like, it's, it's good to at least have a peek inside of like what the discussions are, the topics are any last words to people, you know, whatever you want to call to action.

1:31:01

Yeah, I mean, I think it's, one, thank you for having me. You know, I think this is like, I really, yeah, yeah, yeah. We first met in a Chinese restaurant. That's right, yeah, yeah, yeah. I think like, I really enjoy the sort of like community you've created and the community of developers. And, you know, I think that like, I sort of know things are changing really fast and I think there's like a lot to keep on top of.

1:31:26

And like, I think there's just, a lot to do and I think a lot of people feel a little bit tired or anxious or something. Stressed. Stressed. Yeah, exactly. This is extremely understandable. We're not perfect as well. It's criticize and understand ways all of the AI labs could be better.

1:31:48

But I also am very excited about the excitement that everyone has for AI. It's a really, really exciting time. I think we'll look back at this time and be like, oh, this is very hectic, but very exciting. Software engineering changed forever. Other things will change. And it's really a privilege to... be part of it, you know, like to talk to like the audience that you have and

1:32:14

get to interact with all the developers who are like pushing the frontiers a lot on what's possible. And I learned a lot from that too. Thanks so much. Thanks.