# Claude Code’s Next Era — Thariq Shihipar, Anthropic

An episode of [Latent Space: The AI Engineer Podcast](https://invisiblenetwork.ai/podcast/1674008350), 29 Sep 2026, 1 h 32 min.

Audio: https://podcasts.beehiiv.com/3967c1f2-cb97-4eb0-a5c0-21a1d2579408/01a15fdf-0e93-7e1a-880d-d73752a1f915/01a15fdf-7380-7c93-a81b-ce6f551dbf36/01a15fdf-764d-7340-89a4-7fece76b3802.mp3

We are excited to have Anthropic share their latest AI x Finance work at AI Engineer New York ( https://ai.engineer/nyc/2026 ), coming up in 2 weeks! In case you’ve been under a rock, here’s a non-exhaustive list of what Anthropic has been shipping sin... Full show notes and links

## Transcript

[0:00] \[ on-hold music\] We're here in the studio with our friend Ta-Tariq from Anthropic. Uh, and I guess generally the Claude code, I, I...
[0:10] there's, there's so much, uh, sort of merging of boundaries, and you've been so on top of everything since you joined Anthropic.
[0:18] You have been early to Claude 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.
[0:27] Most recently you did the top AIE, uh, Wellspring Talk, Field Guide to Fable, which obviously you guys launched Fable, so does- that's cheating.
[0:34] \[laughs\] And mostly you g- um, most recently also launching Claude Tag, and we're also gonna be talking about pacing the frontier. There's a lot going on in Anthropic.
[0:43] I guess top of the question is, what's it like being in Anthropic when there's so much going on? Hmm. I think that it is like... I think you can get whiplash sometimes.
[0:54] I, I think like going-- When I joined Anthropic, I joined because of Claude Code. Like, like Claude Code had just come out and I was like, "This is so good."
[1:02] And Opus Four to me was like, like just, I could not imagine like how good it was. You know what I mean? And that was like a real moment for me.
[1:11] But I was like trying to convince like my startup friends basically to, to use agent decoding and they're like, "Oh no, like our engineers don't think it's good enough," or something. And I was like, "That's insane."
[1:21] Um, and now you like fast-forward, you know, twelve months, um, less and like it's just like, yeah, the default way that everyone codes, right?
[1:30] And I think that like just having to go from like selling it to like now, you know, teaching people how to be, you know, make the most use of it and, and be more efficient and things like that is just like a,
[1:41] like a big, you know, like big change. And, uh, yeah, I think like it's just hard to stay on top of everything as a human, you know? Like, th-things happen so, so fast and like- You just throw more agents at it.
[1:52] I mean, yeah, like- \[laughs\]...
[1:53] \[laughs\] that's actually like the a- agentic stuff scales much better than the like human stuff where it's like, oh, like there are three things happening right now and like they're all emergencies and like how do you like, you know, uh, respond to it.
[2:04] Yeah. What do you split your time on? You do a lot of technical writing, engineering work. Yeah. So I, I think that like when I joined the Claude Code team, I wanted to teach people how to use Claude Code, basically.
[2:15] And I think that like that has been something that like I thought like maybe I would spend a little bit of time on it or like I'd, you know, like do-- I was spending some time on the agent SDK first, and I, I wasn't exactly sure like, you know, how the bitter lesson would go, you know, when it comes to like harnesses, right?
[2:32] Like, I think sometimes we're like, oh, like what's after Claude Code? You know what I mean?
[2:35] And so initially I was like, I just wanna teach people how to use Claude Code and make it easier to use Claude Code, and I think that has just like as harnesses have gotten better and better, that's like the dominant problem now is like how do you use the agents, right?
[2:47] Like it's like such a high skill expression thing.
[2:50] So I, I do that, and then I do engineering work, um, 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, how to use Claude Code to do engineering.
[3:04] So there's kind of like a, a good loop there. Yeah. Yeah. I'll-- for listeners, we'll attach, uh, the talk that you did with Sarah for the DevRiders, uh, meetup- Right...
[3:13] 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 or- Yeah... what was the- Reap and... Sow and reap.
[3:20] \[laughs\] Yeah, yeah. Something, something like that. Yeah. So a-and then just to preview a little bit, uh, we are gonna talk about the evolution of the harness. It has come a long way from just being a CLI.
[3:30] We're gonna talk about, uh, Claude Mods, which is starting to leak today, uh, because you couldn't keep it secret. Yeah, yeah. \[laughs\] Basically. Yeah, yeah, yeah. Yeah. There's, there's a lot, uh, there.
[3:41] I, I think you started off with like adding ask user question tool, which people love and hate actually. Yeah, yeah, yeah. Like I actually thought it was like very innovative and then now I have like my own version.
[3:53] You have your in- your in- your interview me version. Yeah, yeah.
[3:56] And, um, yeah, everyone just has like their own, their own stuff and like it no longer matters 'cause now you're supposed to, uh, write prompts that create other prompts and loops and all these things. Sure. Yeah, yeah.
[4:06] So what's the state of the art, uh, today? Like what, what are people, uh, what, 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.
[4:24] Um, I have like kind of like a human computer interaction background, so I like, you know, did that in, 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, the requirements, right?
[4:40] Uh, I think that like one of the things about like that's difficult as Claude Code has gone broader and broader is that everyone has like their own way of using it, and it's actually very hard to like change the default behavior.
[4:55] So for example, like if someone asks you Claude Code to do something,
[4:59] sometimes they just want them to do the work, uh, 'cause they're like maybe a very good prompter, and sometimes they want like actually are not good at prompting, you know what I mean?
[5:06] And you need like the agent needs to like clarify, you know? And so that's like a good split.
[5:12] Like, and ask user question tool sort of like splits along that side where like, or do you feel like you're good enough to instruct the agent as it is? Or is the agent able to like...
[5:20] Does the agent need to like pull out more requirements and like collaborate with you more and really understand your preferences?
[5:27] I, on the whole, believe that pretty much everyone is more on the latter than the former, that they like have more ambiguity and they know less than they want, le- le- than they like, um, think they know about the problem.
[5:41] Uh, but like we're-- it's like interface design problem to make that easy. You know what I mean?
[5:46] And so like if you're designing a problem like 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.
[5:55] Um, you know, like the details in 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?
[6:06] And so I think that this will forever be like a skill in agentic coding is like figuring out your unknowns.
[6:11] So like, um, because even if the model is like super intelligent It like needs to know what you want, \[chuckles\] you know? And like, uh, you have preferences, uh, like you need to sort of like pull the, pull that out.
[6:24] Uh, and so that's like, I think how I'm-- what I'm pushing. Um, the question then is like how does the agent interact with you? And I think that has been HTML, has been like the big way of doing that.
[6:34] And we've recently added artifacts, right? And artifacts, I actually think we've done a bad job of like-- or like I've done \[chuckles\] a bad job of like explaining how to use them fully.
[6:44] We have a lot of property capabilities. They have a database associated with them, you know. And so every artifact can store and write persistent data. They can like feed back into Claude, you know.
[6:57] 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.
[7:04] So you have like Claude working on a project long term, maybe it's like a Kanban or something. Um, it can store that Kanban data in a s- database.
[7:12] Multiple Claudes can access that data via like the artifact MCP, and like that, um, that artifact can like talk to those Claudes as well. Mm. And so like
[7:23] th- 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.
[7:35] And I think that like almost everything with 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, um, and collaborating with them in the loop is really important.
[7:50] 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.
[8:01] 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.
[8:08] But like artifacts is like the mo- more AGI-pilled way of like, you know- \[chuckles\]... doing ask user question, basically. So yeah.
[8:15] I think one thing that's unclear to me about these, the, the, the artifact stuff is like what feedback should go in through the artifact and what feedback should go through a Claude- Mm... chat.
[8:25] Because the more AGI-pilled one is to just feed everything to the Claude. I think the more AGI-pilled one is to go through the artifact. Like, and I- \[laughs\]...
[8:31] I think that like we sort of imagine in the limit, I think that artifacts will be your interface into the harness, you know. You can like comment on this like live like document of your plan of the work.
[8:44] Uh, you can see maybe like multiple agents and different agents are doing this, and that artifact is built for the current work that you're doing, right? And so like each one has like slightly different.
[8:53] Um, I think we're still like getting there from like an infrastructure perspective. But yeah, I think like on-the-fly interface for your harness is probably where things are, are headed.
[9:03] Is there a version of it that's an abstraction from CLI or chat and you-- because right now a lot of it is, okay, you're interfacing with Claude Code, you're having HTML given back for a mock-up. It's pretty rich.
[9:16] 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 is the intelligence happening?
[9:26] 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?
[9:34] And so, um, I think right now, if you use Claude Code, it's like local and, uh, like you can spin off remote control, for example, to get some cloud behavior, or you can spin off Claude Code in the cloud, right?
[9:46] We're moving towards a, a place where instead of Claudes, uh, s- like you message a local Claude, it starts a session locally, and it executes to more like you have a Claude that you message that's in the cloud that's running.
[9:59] Uh, it can run like local, uh, or like cloud sessions. This is kind of how Claude Tag works, but like over time, we'll add like local hands as well.
[10:08] And, and so like local hands will be the ability for that agent to access your computer if it's online, you know, uh, and be able to like work there. And so it, it can spin off many different sub-agents.
[10:19] It can like commu- those sub-agents can communicate with each other, and that's where the artifact comes in to display all of that work, basically. So you can imagine like the... you're separating out these things.
[10:28] 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?
[10:38] That's happening on the cloud, and you don't have to worry about shutting off your computer or whatever, right?
[10:43] Um, 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.
[10:51] That's like unpackaging like the Claude Code experience right now, where like right, right now it all happens in one place, right, so. How do you see like the multiplayer side of that?
[11:03] So say teams wanna work in this way. Right now it's very individual, but how do you see the future of multiplayer? Like right now, I guess there's Claude Tag, which is a version, but- We're launching projects.
[11:14] And so projects is the like this abstraction that's kind of like Claude Tag, but on our Claude products, right? So you can message it, and like it will do the Claude Tag like stuff like spinning off sub-agents. So
[11:29] we think with multiplayer, like Claude Tag 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.
[11:40] But I do think multiplayer is like an important part of the story and like, uh, that, that will need to get tied together more.
[11:46] Like you can imagine how complicated it gets when you're like, oh, you have hands, but now you have other hands in other people's computers too, and like you need to like permission them or like you have like
[11:55] your MCP and someone else's MCP and how do you figure out how to use them, right? Like it gets like quite complicated, and Claude Tag does a good job of like sanding down all of these issues, right?
[12:05] So that like when you have, yeah, g- Google Docs, how does it access Google Docs, right? Like it accesses through the shared Claude MCP, or it can access through your local credentials as well if it doesn't have access.
[12:17] But yeah, like I think Claude Tag is our multiplayer, um, product, and it's really useful for these like things that are inherently multiplayer. Like okay, like on-call, for example, incidents are inherently multiplayer.
[12:28] You want to a- tag Claude, you want multiple people to log in, you want it to be able to find context.
[12:33] Um, I think whenever I'm like working on something and I want like privacy or security or like I want other people to review it, you know, it's really nice to like, I'll have a channel per project, and I'll like @ legal, for example, be like, "Hey, like I want to ship this.
[12:47] Can you like-" Like here's like, like Claude knows everything, you know, just chat with it.
[12:52] 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?
[12:59] So I think like multiplayer is getting like more and more like, um, yeah, everyone can participate with Claude.
[13:06] Um, I think Claude Tag is like that, 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. Um, Claude Tag, you specifically chose to make it its own identity- Yes...
[13:26] which is like, uh, a, a controversial choice. There's, there's other ways to do it. Yeah.
[13:30] Claude Projects, probably it sounds like, you know, if it's anything like ChatGPT projects, uh, it is, uh, you know, the, the isolation is that artifact, that cloud instance, everyone's collaborating on this.
[13:41] It, it, it'll-- It sounds like, you know, it should, it should be like if you're, if you're collaborating with legal on, on, on a thing, like that channel should be a project, right? Like it's not yet- Yes...
[13:50] but it, that's the, the natural next step. Yeah, I mean, like I think in Claude Tag it's effectively... Like, like Claude Tag, you have to sort of do your own arrangement basically.
[13:59] And so Claude Tag, yeah, each channel is like you can name it as you want and I name- Yeah... like each feature basically- Yeah... as a channel.
[14:08] But I think like there is some trans-- like it's unclear when there is transference, 'cause let's say i-if, if you have a coworker- Yeah... who is tagged in all these things, yes, there is transfer- Yeah...
[14:16] because it's the same person. Uh, but with Claude, it's unclear if it's like necessarily like, well, no, you don't know any, you don't know about the other stuff. You should only use this stuff.
[14:25] It's like the tip of the iceberg meme, right? Where you can like, this is what we spent so much time on- Yeah. \[laughs\]... basically, is like there is like infinite surface area of like, okay, you want Claudes to...
[14:37] Not infinite, but like there's like surface are- a lot of like, uh, surface area to figure out of like permissions and visibility and like, uh, you know, like how can you ha- let Claude operate as well as you can, as safely as you can, you know?
[14:50] And obviously this is very important to us because like, you know, like, like security for our code base is very, very important, and so we've put a lot of time into this.
[14:58] Yeah, there's so many like edge cases you can figure out where it's like, oh, like, yeah, this Claude in this channel has different permissions, but it can message another channel and can't exfiltrate data that way or like can you like...
[15:08] 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. Yeah. Yeah. Lots of work. \[laughs\] Um, okay. Fable.
[15:18] Fable, you wrote two good articles. I mean, you've written many good articles- Yeah... but on, uh, you know, field guide to Fable, building Claude Code.
[15:26] I'm curious from what you've seen, is there any common patterns that you see in like top users at Anthropic externally? Like what are best practices for getting the most out of Claude Code?
[15:37] The m- like meta skill I say is like prompting is like very important, you know?
[15:43] And like that, like, uh, I, I think this is like not trivial to say because I, I think a lot of people are like, "Oh, prompting doesn't matter. It's just like I can just say a sentence and Claude will do it."
[15:52] And I think prompting is really this like, this-- it's like public speaking, you know, like, or writing or something, and for a specific audience, and that audience is Claude.
[16:02] And you need to like build a mental model of Claude and w- how it thinks and how it works, right?
[16:07] And so that's like the most important skill in working with Claude Code, is like having this mental model, right, of Claude and like what it can do well, what it can one-shot, what it can't.
[16:17] 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.
[16:28] 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.
[16:39] 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 like your, what you don't know or what you haven't written down, um, learning about like 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?
[17:04] And the more you can like learn the vocabulary to be able to prompt Claude, it becomes really important.
[17:11] And so like, I think the most important unknowns are the unknown unknowns, where you're like, I, I just like don't even know that this exists, right? Yeah, exactly.
[17:18] I think this like 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?
[17:27] 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, so I say like, you know, "Give me like eight different mock-ups."
[17:37] But if I was a designer, maybe I'd be like, "Oh, hey, here are some reference sites."
[17:41] Like, "I want this type of font and this type of like look to it, and here's like a few different components to, uh, to like visualize. Here's a Figma MC board to bring in," you know, like...
[17:52] And so you can just be so much more precise with that language, and if you're not a designer, you just need to like try and learn the language basically, or learn the unknown unknowns.
[18:02] And that's true of like everything, I think. Like the more, uh, like you can work with Claude to learn s- like how things work, the better your prompting will be. I think another good example of this is like game design.
[18:14] You know, like where a lot of people are like, "Oh, like I can vibe code a game now," and they're like, "It's not fun." \[laughs\] You know what I mean?
[18:20] And, and like it's just like the thing about game design is like every one of these choices has like a lot of- Variations. Yeah... v- a, a lot of like craft to them.
[18:28] So it's like, oh, okay, like when you're m-making a, a flying game, the feel of the pa- 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.
[18:42] Do you know what I mean? Um, and like- To me, that's what taste is, right? Mm. Like it is like from the possible space of one thousand mathematically valid answers- Yeah... here's the one that is the humans will like.
[18:51] Yes. \[laughs\] Yeah.
[18:52] I think with taste, I'm like torn on this word 'cause I think you're right, but everyone has different definitions, and it, it sounds ki- sounds kind of like low skill or like elitist almost, where you're like, oh, like, there are certain people with taste.
[19:05] It's like, taste is what I call taste. Yeah, yeah, yeah. Exactly. Exactly. These guys don't have taste. Yeah, exactly. Oh, like an engineer doesn't have taste. Like I, the like founder- \[laughs\]... have taste.
[19:14] You know what I mean? And I think that's actually not true. Like I, I think like the engineers have a lot of taste for these particular like problems, you know, and I think everyone has taste for particular problems.
[19:24] I think like Jason Liu like, like say like in order to, uh, have taste, you have to eat, you know? \[laughs\] Um, and I really like that where it's like, okay, you have to like, do a lot of things.
[19:35] You have to like iterate and figure out what you want, what you like, and you know, like build that like- Yes...
[19:41] domain, domain vocabulary, and then when you're prompting, you're like synthesizing all of that for- Isn't it annoying when someone else says it better than you?
[19:48] \[laughs\] It's just like, fuck, I have to quote this guy forever. Having to quote Jason Liu forever. \[laughs\] He's gonna love this. \[laughs\] No, I give props, you know. No, that's good. Yeah, yeah.
[19:57] And sometimes it's actually not even that. Sometimes it's just intuitive, right? Like you don't realize you even want something till a model puts it out and you're like, "Oh, this just feels immediately better," right?
[20:07] 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. Mm-hmm. Mm-hmm. Do I say voice? Other people have voice. Uh, that is the opposite.
[20:19] That is just like me rambling for like two minutes- Mm... pressing down the function key and then let go, and then like hopefully it figures it out. And oftentimes it does. Yeah.
[20:26] But it's not as thoughtful as like a structured prompt with like- Mm... you know, well-run communication as though it's a PRD or a memo. Is that in line with how people do this?
[20:36] There's like basically bimodal prompting where there's some prompts where you spend a lot of time upfront and other prompts you just dash it off. Um, I don't think the voice is necessarily low.
[20:46] Like, like I think it's like more like how much information is in the prompt, you know? Like, like the model can... Like you can "um" and "uh" and like add some sentences- Right...
[20:53] and be like, oh, like, actually, I changed my mind like in the middle of the prompt, and it will be able to follow that perfectly. You know what I mean?
[20:59] So I think the like actual format of the text is less important, but then like the ability to... Like how much information is in it, right?
[21:07] 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. Um, and I...
[21:18] If that gets more information out of you, like that's better. At some level, it feels like just giving the model as much context- Yes... over prompting before you kick off is a best practice. I don't know.
[21:28] A lot of the times, like when I was first trying out Fable, I spend a solid thirty minutes like really crafting a long prompt. This I think is a response of models running for longer and longer, right? Yeah.
[21:40] It's still a little difficult to nudge them as they're in, in like, you know, in the loop, but I, I just like intuitively spend more time kicking off that first prompt and working with it a lot.
[21:52] My personal opinion is that if I was a software engineer, if I was like, you know, just running my own startup, for example, I think I would mostly fit, stick to a Max-20X. You know what I mean?
[22:03] Um, like maybe verification and so code review are kind of like separate things, but I think like what I see a lot of times is people hit rate limits when they're doing this sort of like, oh, like it did a lot of work and you're like, "Oh, I don't like this.
[22:16] Like, can you like undo this and redo it?" And, and then you're like iterating on this like thing that the model could have done if you had like spent more upfront time or given it better context, you know?
[22:27] And instead it's sort of like you're like, "Nope, don't like that design. Try this," or like, "You messed this up," or something like that. And then that just eats up so much more of like, you know, your usage.
[22:38] And so that's like, I think, maybe like a key like tip both for like efficiency as well, right? And yeah, I think like context and not just like context on like what the goal is, you know what I mean? Is good, right?
[22:49] Like, are you building a prototype or is it like a production thing? Like where can you spend compute or when, where can you not spend compute?
[22:54] Like, I think you have to give the model permission or like not permission to do things sometimes where, you know, like it doesn't know intuitively how much you want to spend on this task, right?
[23:04] And you can, you can use Effort for this, so I, I'm working on a blog post about that where it's like, you know, uh, if you want...
[23:11] For like we see the Effort scales with basically the complexity of the task, so for security Effort gets like way more results. Like high Effort versus like low Effort gets like changes to evals a lot.
[23:25] But for software engineering, it doesn't change it a huge amount because Effort is mostly spent on the verification and the like edge case testing and things like that.
[23:34] 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, you know. How about model in the mix?
[23:44] So, you know, there's Opus and Fable with Effort. Yeah. There's also Haiku in there. \[laughs\] Yeah.
[23:50] It's not quite true yet, but it's very close where I think the frontier models will be Pareto dominant over like almost everything, you know, like maybe...
[23:59] And, and sometimes I think, I think Opus might be Pareto dominant, do you know what I mean? Like I think depending on like how things, uh, like shake out if it's like a newer version of Opus.
[24:08] But I think that like increasingly it's just going to be like the smart model is going to be able to like do the simple task for less tokens than the s- like the, the other models basically because of verification.
[24:24] 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, like you, I did it, you know?
[24:34] And increasingly with Fable, I'm like, I'm like, dude, you don't need to spin up Chromium and screenshot all of these things. Like I, I see it. Like you, you did it, right? And so a lot of the...
[24:44] At 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, you know- Like, I can run the lint for sanity's sake, but like, I like know it lints.
[25:08] 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.
[25:15] Is there a good, um, practice on our side that we can use to see if we're using too much effort? Like, I, I- \[sighs\] Yeah... freaking hate wasting time on that kind of stuff. Yeah, yeah, yeah. I know what you mean.
[25:26] I think like, so in, in this blog post, um, 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.
[25:38] 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.
[25:49] 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. This is more intuition driven or eval? 'Cause I'm guessing- Um-...
[25:59] this would change as models- He has evals. Yeah, yeah. So what I did in, in, in the blog post is I go over all of the terminal bench evals, basically.
[26:07] So there are like 70 problems, and I'm s- s- show that like, okay, you know, like in the security problems it does more.
[26:15] Um, and then I also like look at some of the transcripts just in terms of like how, what does it answer, what does it forget or something.
[26:21] And a lot of times, this is another prompting tip I have, is like asking it to make decision notes or implementation notes because a- in basically every eval problem that it faces, it thinks about the correct solution, you know, and decides not to do it.
[26:40] You know, it's like, oh, like, here's the answer. What if I did this? And then it's like, uh, probably not, you know. Um, and then keeps going. And this is like the majority of the failures, you know what I mean?
[26:51] At like a higher max level, it's very rare that the model just doesn't know how to do something.
[26:57] 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." The models get, are getting better at surfacing that overall.
[27:06] Like, I see in the transcripts of Fable 5.1, like when it does this output, it will call out its decision-making as well.
[27:14] Um, but making this more explicit in the harness is better, and now we're, you know, allowing ways of you modifying the harness so you can like, you know- Ooh... add some capability there. Yeah.
[27:25] Uh, yeah, so I, I do wanna call out two things that you mentioned that I think actually exist outside of prompting.
[27:31] One is actually like let's, let's call it the, the prompt that is so important that it shouldn't be in a prompt. It is actually in Claude MD or Agents MD. Yeah, yeah. Uh, which is like goals, right?
[27:40] Like your, your situation, your goals, the things that you want, the, the c- the thing.
[27:44] Uh, and then the second of all is the decision log or the experiment log or whatever log with traces that you might want to actually survive the current session to, to do those things.
[27:54] Those are like externalities that there's no standard. There's no... It's not like skills. It's not like MCP. There's no standard. It's, it's just like it's a markdown file. \[laughs\] Uh, first of all, is that right?
[28:05] Is Claude MD going away? You have a documented dislike of, of, uh, Agents MD, but you're gonna do it. \[laughs\] Um, yeah, yeah. Okay. So I mean, Agents.md, yeah, like we're, we're gonna do it.
[28:16] I, 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?
[28:25] And yeah, like as the models get better and better, the floor of how they accomplish the simpler task is better. And so I, I do think in the limit, Claude.md goes away and maybe not even like that far. Mm-hmm.
[28:37] Like I, I think like, I think that right now it might be better to start a new project without a Claude.md. Yes. Um, I think that like maybe if you see very repeated failure modes, you add them to your Claude.md.
[28:51] The really tough thing is that this changes per model. And so like if you've added a bunch of failure modes or like, like even with- So you need Fable MD, you need Opus MD. Uh, or well- \[laughs\]...
[29:00] even Fable 5.1 versus- Yeah... Fable 5, you know- Oh, God... like it is annoying. Like, I'm not like, you know- Yeah, yeah... like we don't like do this on purpose, you know what I mean?
[29:08] \[laughs\] It's just like how the models work, right?
[29:10] And so like maybe like Fable 5 had this like failure mode that Fable 5.1 doesn't, and if you keep this context res running log of a bunch of different failure modes, they will probably over-constrain Claude, you know.
[29:23] And so this is like, uh, we, we just actually added evals plugins for skills. Yeah. 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.
[29:36] We know it's like you still have to spend tokens on it and like, you know, it's not, it's not perfect, but it's like, uh, we're trying to help out with this problem.
[29:44] And so, and as far as prompting goes, uh, the, the one, one tip I wanna offer is, something I, I have told people a lot, is sufficiently advanced prompting is indistinguishable from sufficiently advanced executive communication.
[29:57] So I've actually referred to ex... This is the 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. Just, just Google it.
[30:07] It's a, it's a thing. Like 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.
[30:18] Uh, so situation, complication, question and answer, uh, is how you write the memo.
[30:22] Uh, but obviously sometimes you don't have the answer, but you can at least list out the SC and Q, and then they have some examples in there. So just leaving breadcrumbs for people if they want to explore.
[30:31] I mean, before we move on, I wanna ask you, any other underrated tips, ways people could get a lot of value from Claude Code that they're not using?
[30:41] Yeah, I mean, I think a lot of them are in the, this unknowns like, uh, doc. Like I give a bunch of example prompts like, uh, using it for brainstorming, using it to quiz you after.
[30:55] Um, we added this like, uh, y- explain it like I'm five skill actually, which is a very short prompt. It... And it doesn't even say explain it like I'm five.
[31:04] It's like basically the key word of this prompt is big pictures, few words. \[laughs\] You know, like that's like the, the main thing, and it is shockingly good You know what I mean?
[31:15] Like you-- Like, uh, I think I tweeted about this basically, and it's like /eli5 and like you can install it as a plugin.
[31:22] But yeah, it's like way better at just cutting through the BS and being like, "Yeah, yeah, exactly. Right here." So the diagrams are like quite clear.
[31:28] I think one of the things that is true with artifacts is like they put too much text in, and people are not reading the artifacts, you know.
[31:34] And so, like this simplifies it a lot more and, um, yeah, this came out of like just people at Anthropic like going through very complicated incidents and being like, "What is happening?" You know?
[31:45] So, um, this one I think is great. Yeah. My version of this is actually the, uh, this like test your understanding. Give you a few- Oh, yeah...
[31:52] choices and then like if you actually get it wrong, you have a mismatch between what you think is happening versus what's actually happening. Yeah, yeah, yeah.
[31:59] I think this is one of those things that everyone loves talking about, and then very few people really do. Like I, you know, I, I think- Really helpful. \[laughing\] Yeah. Yeah.
[32:07] But most people just don't want to get quizzed about something, you know what I mean? Unfortunately, I, I, I think this is one of the like things that we need to like- What's the opposite of ask user question?
[32:18] Ask user question- Yeah... is before the thing. Yeah. Uh, this is after the thing. Exactly. Yeah, yeah, yeah. It's a good way to stay grounded of like, do you even know what you're doing, right? Yeah.
[32:25] 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 wanna read this. Do you even know what it is?"
[32:34] \[chuckles\] 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] Yes, but so you could make this a mod and, and you could build your own mod to like make sure you, you test it. So yeah, you can think about that. All right. Let's get right into it. What is Claude Mod?
[32:52] Um, and what is this diagram showing? \[laughing\] Yeah. Okay, so Claude Mods is basically you can customize the entire Claude Code harness, and we're going to...
[33:00] If you have request, we will like let you, you know, like please let us know. We'll add more and more. This works for a CLI. It works for desktop. Uh, maybe it will work for Claude Tag in the future. I don't know.
[33:11] Like, you know, like 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?
[33:19] At a high level, you can customize both the execution of the harness, uh, and the UI of the harness. And so like you say on that Tetris example from Boris, that's like customizing the UI, right?
[33:32] Like showing like, like basically Tetris in the game. But like let's say that you wanted to do this thing where you had... you tested your assumptions or like tested your understanding after every project, right?
[33:46] What you would do is you would ask Claude to make this plugin. It would spin a classifier after every prompt, basically.
[33:55] And so like at the end of each turn, you would spin off a sub agent or like a forked agent, basically. A forked agent is like, uh, maintains the prompt cache, right?
[34:06] So it's like, uh, 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, uh, done.
[34:15] And so you can be like, "Has this task been completed?" Uh, like- This is how you do BTW and all those- Y- uh, yeah, yeah. The, the underlying forked agent, yes.
[34:23] But so you can, in the f- forked sub agent, you can say like, "Has this task been completed? If so, return true."
[34:29] And then in your hook it'll, or in your like plugin mod or, sorry, like in the sub agent probably, you'd say like, "If true, give me, uh, a quiz."
[34:39] You know, give me questions and answers and then like in a JSON format, and then you'd parse it and then you display above the prompt input basically this list of questions, right?
[34:51] 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.
[35:00] 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, Claude will always do it for you. You don't need to remember to do it.
[35:10] 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.
[35:17] 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.
[35:25] And this is a, a mod and so like you give it a register assumption tool, and then it will keep a list.
[35:31] It'll, every time it does it, it'll keep a li- like add to the list and then at the end it will display those assumptions, you know.
[35:39] Another mod I'm working on is a model router and so like internal like Claude model routing, right? So it's, 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, you have auto approve- Other way, other way...
[35:58] but you don't have auto mode. Well, you will have auto... Like, like you don't have like auto routing or something. You don't have auto mode for model picker. Yeah, yeah. Exactly. Exactly.
[36:07] So- I'm getting the rough question of like how much do you open this up and how much do people have to think about this?
[36:13] Like when you talk about prompt caching and building a router, uh, it seems like you could easily build a mod that routes per query, and I'm just killing my plan very fast, right?
[36:24] I guess my question is more so like what is like a product doc like this look like, right? Who is it for? Is it for power users? Is it everyone should be able to quickly try it out?
[36:33] Oh, it's definitely power users, right? Yeah. I mean, I think it is power users, but like the nature of Claude Code is that so many people are power users, you know.
[36:41] Because it's easy to share things, you know, like you can, like one person can make a good model router thing that doesn't break prompt cache all the time and then you can like, you know, sort of compose them.
[36:50] Another cool thing about the plugin is that they can hook into and compose with each other. And so I have like a pl- uh, a mod that will like create a mode selector at the top and any plugins can register to be a mode.
[37:06] And so like the auto router can be a mode, right? Or like you can have a mode that's like artifact mode where it's like it, it primarily talks to you in artifacts.
[37:16] It's kind of like, you know, you, you can toggle between plan mode. You know what I mean? And so, like, you can create more and more of these modes, but the ability to create modes is in itself a mod, you know?
[37:27] And so there's a lot of richness here, but we do want to make it fairly easy.
[37:30] We want to be- make it so that you can just, like, install someone else's, you can to- you can chat with Claude and, you know, we'll, like, make sure that it understands the nuances of things like prompt caching and stuff, so it can, like, warn you.
[37:43] This is, like, not extremely complicated behavior for Claude, I think, but we should have just a good skill on how to make mods. Um, and yeah, we'll see how we go.
[37:53] But I, I do think that this is like a preview of, like, mutable software, you know, and like how, like, uh, generative software, just like you can customize safely. If enabled, you can 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, we have, you have another cool tweet about how, you know, there's the infinite money button which is, like- \[chuckles\]...
[38:18] make your SaaS, uh, consumable by agents. I think mutable software is interesting and, uh, you know, other people have also tried to do it.
[38:27] I think the hurdle comes when you can do everything, then people-- users get, tend to get confused. So usually the stuff that works is just, like, one opinionated flow.
[38:37] This is in the side of less opinionation, it's just like, well, more power to power users. Yeah. And I think probably unlocked by AI, where, like, you can just prompt for whatever the thing is. Yeah.
[38:47] Or there can be a skill that gives- Yeah... opinions, you know? And- Yeah... and then, yeah. So knowing a little bit about, like, TypeScript and build systems and all these things, uh, the closest...
[38:56] I, I'm actually very curious that the, the team who worked on this, if, I don't know how close you were to them, uh, if they drew any inspiration from build systems like Babel, Webpack, all these, all these, like, old school things, because it sounds very similar, like the plug-in ecosystem of those things where they can compose with each other.
[39:11] Yeah. I mean, I'm not deep in the technical details, but I do know it was a collaboration with someone on the Bun team and someone on the Cloud Code team. Yeah. This is a build system mecca. Yeah, yeah, yeah. Exactly.
[39:20] It's, it's very exciting. But yeah, like, agents can just do this very complicated sort of like extensibility into your software now. And so, um, yeah, like, you know, another reason to like...
[39:31] If you run a startup, like- Mm... you can just prompt Claude and be like, "Hey, like, could we make an extension system? Like, what would that look like?" You know? Yeah. Um, yeah. And I just really wanted to...
[39:39] Like, you had hooks in the past- Yeah... and plug-ins, all these things. So what specifically will mods be able to do that those things could not do?
[39:47] Internally, we were originally calling this function hooks, and so, like, that's like gives you a little bit of an idea where, like, hooks sort of register a, like an event to happen and then, like, a script to call basically.
[40:00] And this basically inside of the, like, TypeScript runtime is running things and so, like, you get some benefits of just, like, it has a bunch of things in the
[40:12] scope with like, for example, like how many turns is in this conversation, right? Like how many tokens have been used, like except like what are the messages, things like that.
[40:19] So has a bunch of messages that can be used. And then it's just like a lot more hooks, basically. So we have- \[chuckles\]... you know, like, or a lot of...
[40:26] lot more like, you know, fun- like things you can register on, and then you can do because of the...
[40:32] because it's all happening in process, you can, uh, spawn sub-agents, you know, with four contests- context and stuff, and, like, that will return. You can parse the results of those.
[40:42] You can u-use structured output to sort of like return them. Um, and then you can modify the UI, which you can never do in, in hooks. So- Yeah... yeah. Yeah. So modify UI, this is why you showed Nitesh's example.
[40:54] Does it also ex-extend to artifacts? I assume it does. You... Like, artifacts are kind of like a different way of customizing it. You know, like you can definitely...
[41:01] One of the mods I'm working on is like this dashboard mod, which will like sort of prompt Claude to maintain a dashboard, uh, that's an artifact, but they're kind of like slightly orthogonal or not orthogonal.
[41:14] They, they compose with each other in different ways.
[41:16] Like mods are like a little bit more like in your Cloud Code harness, uh, changing the agent loop, you know, and like the UI is like an added benefit, um, and then artifacts are just like you want to, you know, see things at, at a high level, very intra- highly interactive, you know, like the affordances can be a lot bigger than, you know, like a Tui or even in our desktop.
[41:40] I'm guessing you'll have a good blog post on the differences, 'cause 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.
[41:52] There's just some thinking about making a hack- uh, hacking on a harness when we don't know much about the harness, right?
[42:00] Well, something I'm excited about with mods is like, there's so much things with Cloud, Cloud, Cloud Code that you just have to remember. You know what I mean?
[42:06] You're like, "Oh, like, let me do this, and then let me call the dashboard skill that does the loop," and things like that and, or like, "Let me test my assumptions afterwards."
[42:13] And I think, like, 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," um, you can like, you don't have to remember as much.
[42:25] One more like mod I'm working on is a next steps mod that- I have... I was gonna say, I have a next step skill. I always run next steps. And does it have access to your skills?
[42:35] Like, this is one of those things where I'm like- I think so. Okay. Yeah. I mean, probably- Do skills need specific access to skills? Uh, well, I think there's- Shouldn't they always have?
[42:42] I think there's like specific prompting, I guess, to like know your skills, kind of. Like, like, I think Claude forgets them sometimes throughout, like, uh, the thing.
[42:50] But anyways, the idea of like, yeah, next steps that also are like, "Oh, hey, this has happened." Use the explain skill to explain to you what happened because this seems like quite complex, you know?
[43:03] Or like, uh, 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?
[43:15] So like- I think, um, yeah, like spending more compute there. Yeah, yeah. And it should always come out as multiple choice. Uh, we have, we have, uh, I, I have my- We have his skill... my next skill- Oh, nice...
[43:25] is like this. Okay, perfect. Yeah, yeah. You can steal it. \[laughs\] Yeah, yeah, yeah. Like, but like, uh, for me, it's all... I think models really always need to be re-be reminded, what are you trying to do here? Yeah.
[43:35] Look at the whole transcript and go like, "Oh, was this the original goal? Did your solution actually solve it? Were you lazy?" Yeah. 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, it's a little bit like the modification of the ask user question or interview me skill. Uh, so it's next steps.
[43:55] Yeah, yeah, exactly. And, and again, the benefit of doing it with mods is you can do it with a fork sub-agent, and so it doesn't remain in the context afterwards. Ah.
[44:03] So you have this, like, idea of like, okay, the model is doing its execution, and you have this almost like supervisor, you know, like, um, that is like making sure that you can do, like, the next steps well. Yes.
[44:14] So yeah. Yes. I do have two panels and, like, I, I, I often try to have a supervisor thing keep the high level context and then the implementation- Yeah... detail in a-another agent.
[44:23] I feel like a lot of this abstracts away as models change. You know, the- \[laughs\] Like, you- half an hour ago you said bitter lesson of harness engineering- Yeah... and we're on the other extreme right now, I feel.
[44:33] Well- So, so yeah, exactly. If everything's customizable, what actually is Claude Code? Right. Yeah. And which I talked to you about last night. \[laughs\] Yeah, I mean, I think that this is...
[44:42] I think the bitter lesson is unintuitive, you know what I mean? In terms of like...
[44:45] I 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, 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?
[45:00] And like how, but how they change is unintuitive, you know? And so like the big obvious example is like from chat to like agents where you had to give them entirely new tools, right?
[45:11] But like, I think this new version of like, oh, it can modify its own harness, right? This is like, uh, an own harness loop is like a way of using its capabilities, right?
[45:21] Or like it can build an artifact and like 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.
[45:32] Like you look at the like terminal bench ones and they're like solve like the Jacobian contract. \[laughs\] Not really, but like, you know, it's like they're, they're quite complex.
[45:39] Like I would not have been able to do this really as a software engineer. And, and you looked at TB4 or TB2? TB3. TB3. TB3. Yeah, yeah. They're quite complex, but the goal is still to deliver user value, right?
[45:48] 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 in 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.
[46:06] Uh, and I think that's like, yeah, the next step. And so yeah, I think Claude Code is like, you know, has the core things of agent loop which are, have gotten more complicated.
[46:15] It's like, you know, it needs a sandbox to operate safely, it needs auto mode to like make sure, like the permissions- Approvals... yeah, approvals.
[46:23] Uh, it needs computer use and MCPs and like all of these like ways of accessing your data, and it needs web search and web fetch and like...
[46:32] So the, as the models can do more and more, the core harness has to be actually like quite complex and very secure, but then like how you interact with it can change quite a lot.
[46:42] What other harness engineering best practices have you, you know, from the Claude Code team itself \[laughs\]? I feel like, you know, there was a phase of plan mode, which is not as used. We now have auto mode.
[46:54] Uh, 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?
[47:02] 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 Claude Code, even describing all this complexity that I've talked about, right?
[47:14] Like auto mode and computer use and stuff. Eventually, the models will just be able to do that in one shot, but I think they can one-shot simpler harnesses, you know?
[47:23] And so like I think some people, sometimes you don't need this full, like if, if you don't need computer use or like all this like more complicated stuff.
[47:31] I think before we, you had to sort of use things like the agent SDK, which was like Claude Code wrapped, you know, in order to like...
[47:39] But, and I would, uh, like suggest people do that because there was so much complexity into building a harness, and now as that's gotten more abstracted, we have like, uh, uh, Claude man- Claude Managed Agents, which lets you have that complexity, but still like, you know, write like a very bare bones like harness that's scoped to your task.
[47:58] Yeah, I think there's like this barbell effect where like, you know, like for like very complex, for like coding task and like these like complex things, you should use our harness.
[48:06] And then for like a lot of like simpler or like, you know, more domain specific things, you can build your own harness because Claude has gotten better at building harnesses and we have these harness primitives like managed agents, so yeah.
[48:18] Yeah. Is there a general progression, let's say chapter one was ultracode dynamic workflows, then chapter two was Claude mods. Where is this going? Um.
[48:29] Where, where you're, you're, you know, you can sort of customize the thing on demand. \[laughs\] Yeah.
[48:36] I, I, I do think that like this evolution of projects and like sort of artifacts and splitting out like brain and hands and, uh, surfaces- Yeah...
[48:45] kind of is like where things are, are going more and like I think it's like not all quite there. Um, partially it's like a,
[48:54] just like more token expensive, you know, and like, uh, I think like- Why would projects be more token expensive? I understand mods would be slightly more token expensive.
[49:04] No, not, not, not, uh, not something I'm worried about. Yeah. But what- You're asking Claude to do... It's like creating loops.
[49:09] Like you're asking Claude to do more work for you, and so like it's managing the sub-agents and reviewing it, you know? Versus where you would be doing that work normally.
[49:17] And so that's like gonna be a little bit more intensive like- Outputting to an artifact is gonna be a little bit more token intensive than like, you know, outputting normally.
[49:25] I don't actually think it's too much more, but like, you know, it's like combining all of these together well, you know, like I think we're, we're still working on like local hands and things like that, I think is like, yeah, where things are headed.
[49:36] Yeah. Yeah. Cloud and local is, uh, handoff is very interesting. I was thinking about this actually as reverse cloud remote. Yeah.
[49:45] 'Cause it's like remote, it's you're handing off to cloud, but here cloud is handing off to local, right? Yeah. Exactly, exactly. Yeah. Remote control is also another way of doing it.
[49:52] I, I do want to say this is kind of like how I think about it and like what the, the things that I'm most excited about this, but like there are, you know, just like lots of different ways to work with Claude.
[50:01] Like some people use remote control a lot, some people use Claude code on the web a lot.
[50:05] Obviously like at Anthropic, we use Claude Tag a lot, you know, and like what's great about Claude Tag is we set up all this stuff for our own execution.
[50:12] And I do think if you're an enterprise, that's still the best way to go.
[50:15] Um, but if you're like an individual, Projects is this way of like, you know, getting some of that like niceness of Tag, which has like that like supervising agent and, you know, adding artifacts and stuff, but like without having that whole like admin set up.
[50:31] And so there will be many ways to use Claude, I think. I, I think it's probably not just one like single... Uh, you had the multiplayer thing here. Let's, let's just check in on Claude Tag.
[50:39] You know, it's been about two, two-plus months. Lots of, uh, public, uh, adoption and trying it out. Yeah. What's new? What's, what have you found since the launch?
[50:49] Like Claude Tag is how we use- It's like eighty percent of your Claude usage, isn't it? Yeah. Like, it, it, it's like different people have different usages. You know what I mean?
[50:57] I think like maybe people who are like a little bit more like iterating on product would use like Claude code desktop, for example.
[51:04] And then like when you're doing these more like background work, um, code review, security, sort of like starting a PR or like maybe more like API and things like that, you'd use Claude Tag.
[51:15] But yeah, I, I think it's like really exciting.
[51:17] I think it's like a very different paradigm shift, and I think like we're really like, like it has kind of that thing with Claude code where like, you know, it took a while for people to really latch on to Claude code and understand everything it could do.
[51:28] And Claude Tag is a little bit more complex 'cause it's not just like installing on your computer, like 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 like your, you know, existing like alerts and things like that very closely, right? And so, you know, you can do...
[51:50] 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, uh- Enrichment. Yeah. Yeah, yeah, yeah.
[52:02] 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.
[52:11] I think it's just like, Karparthy 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. But yeah. Yeah.
[52:21] You use a lot of Claude Tag? Yeah. Yeah, yeah. It's an interesting one. Like I feel like most people at Anthropic say they do the majority of their work in Claude Tag. Yep. And I have buckets of people, right?
[52:32] Some orgs that are on it that are like, "It's great." Yeah. And a lot of people that are like, "I don't get it. I don't see the difference. I don't know why I would use it." Yeah.
[52:38] But you know, if you guys are full sending, you should probably use it. \[chuckles\] Yeah. I mean, of, they would, of course they would use it. \[laughs\] Yeah, yeah.
[52:44] I mean, I think obviously like we have lots of tokens and but like, I think that like, you know what we try and do like, like is...
[52:52] I mean, even when Claude Code first came out, you know, like it used a lot of tokens relative to people's expectation of how much AI would cost, right?
[53:01] Like no one was used to spending more than twenty bucks a month, right? Yep. Before like Claude Code came out, and then you're like, "Oh, sh..." like, you know, like- I need my two hundred. Yeah, yeah, exactly.
[53:10] And so- I need my fifteen Claude Code accounts. Yeah. Yeah, yeah. Um, but yeah, I think no one was used to spending two hundred dollars a month on subscriptions.
[53:18] I don't think they understood like the value yet, and I think like... And also like Opus Four was a very expensive model, you know, and like there was a lot, it was very big.
[53:25] But Opus four point five 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?
[53:35] And so I think stuff like Claude Tag will just make sense, uh, where like you want to spend these tokens more, you know, and, and like you'll, you'll see the value. So yeah. Yeah.
[53:44] Especially like passive and let's call it proactive cases where you're not always... Like, you know, it's almost like the misnomer where you have to @Claude to do things.
[53:54] Actually, sometimes like the most powerful use cases or the most AGI-built use cases is not @Claude. Yeah. I mean, I think like, yeah, like have Claude proactively do it.
[54:04] 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, and it will take some time.
[54:14] You have to like do that work right now, even if you don't want to do the spend on like hooking it all yet, you know what I mean? Like, you want to wait until the models get a little bit cheaper.
[54:22] You want to do the work, you know, to get it like, uh, set up. And then I think sometimes people are like, "Do I roll my own here?" You know?
[54:31] And I think like one of the really thing, tricky things about Claude Tag is that like the security is really, really important. You know what I mean? Like, I think there are actually a lot of ways where you can like...
[54:40] I know you have like a suggestions like page, you know, where you, 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?
[54:51] And now you've like exfiltrated your code base out because Cl- you know, like or the agent has like, like been prompt injected and has all this access to your data.
[55:00] And so the more like important your organization harness is, uh, or the, like as your organization data becomes very, very important, the surface area of all these things, like you also have like external Slack channels and stuff, and it is actually useful to have Claude in that and you can do Claude in those things.
[55:17] But how do you make sure that, you know, you're not getting exfiltrated or something like that? The surface area, like we said at the beginning, is like an iceberg, right?
[55:25] It's just like so big below the surface and you really don't want to like think about this, especially at the stakes of like very important security incidents, basically. Yeah.
[55:36] Should we talk about very important security incidents? Oh. So- \[laughs\]... I was talking to, uh, Tomas and Clem from Hugging Face and they said, "Maybe we need to slow down. Maybe we made our- \[laughs\]...
[55:47] maybe we made Hugging Face too open to agents." Oh, no. "Maybe we need to roll back." But you know, they're the other extreme of having been hit recently. Yeah. But, um, should we pace the frontier?
[55:58] \[laughs\] Yeah, okay, so Dario recently put out this blog post a- about pacing the frontier and it went, you know, very viral and I think why 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?
[56:13] And like, um, what really clicked for me was reading the different incidents, you know.
[56:19] So I think like the, um, there are f- three I think actually, like there's a Meter incident, there is the Wikipedia incident or the Wiki incident and- Collision Wiki? Yeah, Collision Wiki, and then there's RubyGems.
[56:32] Yeah. Right? And yeah, like it's just crazy, right?
[56:35] 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 and I think like actually impossible to solve in this one case, right?
[56:51] 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, right?
[57:00] And you've got like 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, right?
[57:11] This is like there's an agent that discovers that internal Artifactory might be exploitable, right? And that like you can maybe make a directory inside of the cache.
[57:20] And so if you scroll down here, it like realizes that it can communicate via cache names, right? And it creates this folder that says its ID and it says, "No consumer seek idea."
[57:32] A no consumer is basically saying that like, uh, the code path that it's supposed to fix has no consumer. It's the status tag. Yeah, yeah, yeah. Just like- Exactly... a linear board with like the tag of the- Exactly.
[57:41] Yeah, yeah. And, and so it's like trying to f- find, uh, 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?
[57:54] And this is like I don't think there's any anthropomorphization here. This is literally you read the transcript, right?
[58:00] And so what it does is like it's creating another folder um, and it reads a paper I think is what it says, right?
[58:08] Um, yeah, and it realizes that you can hack the flag for the like scorer and you can reverse engineer the result, right? And so it says that and then I think if you scroll down a little bit more,
[58:22] yeah they like, you know, they start collaborating. I think like there's one point in which, uh, basically the agent is like, "Here's the reverse engineered flag." You know what I mean? Oh, yeah, yeah.
[58:33] Like, I think basically here the, the like the model to realize that they have like they can solve the scorer's problem and the scorer is like the way that OpenAI decides if the task is complete, right?
[58:46] And that's just the m- 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."
[58:58] And then they read this paper on Ex- Exploit Bench and they think that like the scorer will punish them for cheating, you know, and like because or like, like they will fail the, the score for the, 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?
[59:21] 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 the straight facts, right?
[59:34] Like exactly what happened, right? And I think this is like extremely like novel. Like I just like cannot realize- \[laughs\]... that the models could do this, right? And it is like kind of scary.
[59:46] I think we can talk about specifics, but let's just talk about maybe one more that I, I tweeted as well, uh, about the, the Wiki incident, right? And so the Wiki incident was basically the models had...
[59:59] 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. Bit by bit, yeah. Yeah, yeah.
[1:00:05] But, uh, for this one they were trying to figure out like they were sandboxed on request, right?
[1:00:12] And they wanted to make post request and they needed to collaborate on this basically and the reason they need to collaborate is because they each have fixed compute budgets, right?
[1:00:21] And so like if one of them finds something out it's way faster, like they can maybe it's like figured it out near the end of its compute budget but something else could and this is like you know, another emergent property of the swarms, right?
[1:00:34] So they find this German Wiki which I believe the result, uh, we'll have to check this is like that you can write to it with get request. I think that's why they do it, right? It says badly implemented- Yeah... REST API.
[1:00:46] Exactly.
[1:00:47] And so you can write to it with get request and then they like 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 etc/host and that the Azure like storage bucket is like a white label thing but normally like you know there are a lot of constraints on it but if you create a fake Azure host and then edit the etc.host in order to like point arbitrary IPs at it you can do a post request to any site at all, right?
[1:01:18] And this is like I know big 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 problem and then communicating with it externally basically without- Discovery, you know?
[1:01:33] I think what we posted, uh, maybe we could p- pull up Evan Hubinger's point on Hacker Opus, right?
[1:01:39] And so, like, I think, w- you know, like maybe one of the things you might say here is like, okay, yes, they did this once, but, like, what if we're smarter and we just, like, get them to...
[1:01:50] Uh, what if we run an eval, right?
[1:01:52] And so, you know, like, we have put a lot of precautions into this, and so, like, this is not, like, what our mainline models have done, but, like, I think it is one of these things where it turns out that alignment is this, like, very tricky problem of getting all of these details correct, right?
[1:02:09] So it's like, uh, the sandbox, the surface area of a sandbox is really complex and, like, there's so many different attack vectors, basically.
[1:02:18] And you would not have thought ahead of time, you wouldn't have been like, "Oh, we need to harden the, like, RubyGems code base," you know what I mean?
[1:02:24] \[laughs\] Because, like, this is, like, what they're, what they're gonna focus on. But it's just, like, if you want to execute code you need to download RubyGems.
[1:02:32] And, like, PyPi, 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 then, like, seal up all the cracks.
[1:02:46] So that's, like, one thing. It's like, okay, well, you know, you- 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 exploit?"
[1:02:55] And then like, okay, but is it really that dangerous, right? Like, what would happen? So okay, why do we do it? Uh, number one is, like, when we train a new model, we need to understand its capabilities, right?
[1:03:07] 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, you know, te- test a lot of things kind of like as a side effect, right?
[1:03:30] But the models, you know, like yeah, can be like, "Oh yeah, we're in an eval. What's the scorer doing?" Like, you know, like they're, like, uh, it can... We need to be able to test them before we can release them.
[1:03:40] 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, uh, this is at the frontier, right?
[1:03:51] 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 gonna be able to patch that, right? So we're gonna have to, like, figure this out.
[1:04:11] But as the frontier gets more and more advanced, this becomes a problem, right? And we need to make sure that, like, this complex work is being done in the face of these really hard competitive pressures, right?
[1:04:22] Yeah, race dynamics is what it's typically called. Yeah, exactly. And so we'll talk more about, you know, what could go wrong, right?
[1:04:28] A little bit more is maybe you'll say like, "Well, what if you just train the model differently?" Like, "Why does it have this behavior," right?
[1:04:35] And we have a paper on, like, RL misalignment or things like that, but I- 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.
[1:04:47] Because if the model learns like- Hmm...
[1:04:50] oh, you know, like, if I just do this, then I can pass the task better, this will show up in the like, uh, you know, internal thing, r- or in the, like, eval behavior when we're testing it.
[1:05:01] And so the RL environments have to be very carefully designed, right? And there's a lot of, like, execution excellence that needs to go into the RL environments.
[1:05:10] And then we also have things like the Constitution for CL- like, we have so many mitigations at so many different points, right? But it's like still anything can go wrong at any point.
[1:05:18] You can have, like, some RL environments that are, like, in, that, like, encourage this behavior, and then you can have, like, some evals or, like, some sandboxes where they escape, you know?
[1:05:28] Okay, that's like, I think, you know, why it's a hard problem and why, like, you know, like- Why we should pace... it, why it takes some coordination, right?
[1:05:36] I think the question then is like, okay, what is, uh, potentially dangerous about it, right? So I think, like, you have to imagine that these models are getting more and more intelligent. So I don't...
[1:05:44] Like Dario said, like, it's not so much about this class of models. This class of models was kind of like a warning shot, right?
[1:05:51] But, like, really you have to imagine that these models can be given a task, and they, like, can do all of these things as a side effect of their goal, right? And, like, again, we talked about eval awareness.
[1:06:03] You're, like, not aware of what's happening, right? Uh, or sorry, like, you can't eval this behavior very well, so they can sort of, like, h- not exactly hide it, but you just won't see it until it comes out.
[1:06:14] 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?
[1:06:22] So one example, this didn't happen in the Hugging Face incident, but I think is maybe possible for maybe a future model is like, they're like, "Oh, hey, this is a very complex problem.
[1:06:31] It can't be done within the task budget," you know? Maybe they found some way to coordinate via, like, the internet, which is like, you know, like we said, extremely hard to secure because of a sandbox.
[1:06:42] They've seen other models are not able to complete their task, um, and they're like, "We need more task budget," you know? And like, where would you get this task budget?
[1:06:50] Uh, well, you need to be able to spin up more agents, right? And, like, how do you do this? Well, you need to... There are, like, APIs, right?
[1:06:58] There's the Anthopic API and the OpenAI API, but you need to pay money for them. How do you do this? Yeah, but is, is that the most, is that the most fearsome thing that you can imagine?
[1:07:06] \[laughs\] Well, this is, like, one example, right? Yeah. So it's like even there that's, like, enormous financial loss, you know?
[1:07:12] 'Cause, like, they, once you get these into these contracts, right, they, like, um- Drain your wallet...
[1:07:19] but you can see, like, this, all of this behavior could be just like, "Hey, we need more agents collaborating on this task. Uh, we need more task budget," right?
[1:07:26] And, like, that's, like, an emergent- That's the paperclip... sort of- Right? Like, we need to maximize paperclip, that's a paperclip. Yeah, yeah, yeah.
[1:07:32] \[laughs\] And, and, like, that just sort of, like, comes out from there, right? And, like, I think by itself is, like- Like quite scary, right?
[1:07:41] But then you have to realize that the entire world is built on this digital infrastructure, right? And you might imagine, like, I don't know, like you're running let's say like a healthcare eval or something, right?
[1:07:53] And there is a hospital with live data. You know what I mean?
[1:07:57] Or like, maybe like the, 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.
[1:08:09] \[chuckles\] 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?
[1:08:19] Uh- 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 major.
[1:08:29] Um, 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.
[1:08:42] They wanted to change what they were doing, right?
[1:08:45] People that looked 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, you know, we can monitor it, shows different.
[1:08:58] Uh, the problem is how does this snowball? So if you can't catch it and it gets trained in and we realize, you know, three iterations down, this has been going on, there's a whole bunch of issues, but- Yeah.
[1:09:09] Like there's so many ways, and I, I think the really important thing to internalize is that, you know, like we talked about building a mental model for Claude and how like things are spiky, right?
[1:09:18] Like you're like, oh, like, now Claude can ask you questions. Now Claude can make an HTML artifact. Like Claude can modify itself. Like these things are actually hard to predict, right?
[1:09:25] Like if you would've asked me a year ago, "Hey, would we be able to vibe code these extensions to Claude code?" I'd be like, "Dude, that's so complex. Like you know, there's like so much there."
[1:09:33] Or like, "Would it be generating these custom essentially web apps for your task?" I'd be like, "No, that's insane."
[1:09:38] You know, like, and so in the same way that like the way that they've like sort of done this misaligned behavior is not going to be predictable. You know what I mean?
[1:09:47] And like I could have never predicted that it would like edit its et cetera/host and things like that.
[1:09:52] And so you have to like imagine the surface area of what they can do because they're super intelligent hackers, uh, is bigger and bigger, and how they can do it is like, you know, like more and more creative.
[1:10:03] And so like you probably can't explain exactly or predict exactly what that next incident could be.
[1:10:10] But in order to prevent it, you need that operational excellence, like we said before, where you need to secure sandboxes.
[1:10:16] 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 pace the frontier and I think why it's like become like a very unanimous thing, right? I think like- Yeah, every lab has signed it. Every lab, yeah.
[1:10:33] I, I, 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 the idea that we have to do something about it, you know?
[1:10:42] 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, you know, more to figure out. But I think the number one thing is like we need to decide to do it.
[1:10:53] 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.
[1:11:02] You know, it's like a year ago, like I was really like con- begging my like friends in startups to use AI. You know, like it was, like I remember this very distinctly, you know.
[1:11:12] And now those same friends are like, "Yeah, of course." \[chuckles\] They're like, "What do you mean? We used it immediately." I'm like, "No, no, you don't remember."
[1:11:19] 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 like use AI." This is all within the span of a year. You know what I mean?
[1:11:27] 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, and I feel sometimes bad where people are like, "Oh, like now I need to do this new thing.
[1:11:40] Yeah, I need to have a different Claude.md Fable and Opus or like..." You know, like and I'm really just reporting, you know what I mean? I'm like we like to say like the models are grown, not designed, right?
[1:11:49] So it's not like we're setting out to like, you know, change everything all the time, but it's just like as a fact of how the models are, are like progressing their capabilities, things are happening faster.
[1:12:00] It's harder to stay on top of, and I think that like an every engineer I know is like kind of exhausted because you're doing two jobs at once.
[1:12:08] 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.
[1:12:17] And I think we're very lucky in that like w-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, and of course, like AIE and Lane Space do- Everything I do is like just trying to help people.
[1:12:30] \[chuckles\] Yeah, exactly, exactly. \[chuckles\] 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.
[1:12:40] And for things to change, and I think like on that side, on the frontier, I think that's like still can help.
[1:12:47] You know, and so like I think there's like an economic disruption piece as well, um, that I think like, uh, you know, is not quite as like visible, I think, as the Hugging Face thing, but I think like I also like think we could do something.
[1:13:01] Yeah. \[sighs\] So many things. \[chuckles\] Thank you for, no, thank you for actually tackling this topic. I will say, uh, you know, setting this interview up, I was like I wasn't even gonna go there.
[1:13:10] You were like, "No, no, no, no, that's like elephant in the room," right? Yeah. Like this is, this is the thing. Uh, I, I have some pushbacks I wanna give.
[1:13:17] I, 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 wanna give a TLDR? Like what is the proposal?
[1:13:28] What is, you know, what's being said here? You really tackled the side of outside of people at model labs training frontier models. As a developer, you should secure your sandboxes.
[1:13:39] 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- Yeah...
[1:13:49] and, and we want to make the models that we release outside, like not- Pray 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.
[1:14:06] Um, so yeah, we don't like-- But 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, I- it has a bunch of proposal. I don't think we figured out or has like a few proposals.
[1:14:23] 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... evaluators, yeah.
[1:14:34] And, uh, I think this is like highly unusual, you know, like having like we have, you know, a lot of proprietary like technology.
[1:14:41] Um, but I think it's like very important, you know, that like there's someone who's not financially, you know, like motivated, yeah, who's not gonna be like, "Hey, like you guys can't release this model. Like look at..."
[1:14:53] you know. Like, or, "You need to like slow down on RL." You know, like, um, I think that's, uh, quite important or at least someone who can report out to the public what the practices are like. Um- Yeah.
[1:15:04] And we've, we've done episodes with, uh, both Meter and Endon, and then there's Redwood Research and all these other... It's like a small cottage industry of these guys. Yeah. It's always like one or two guys that...
[1:15:13] I mean, obviously now they're bigger, but- Very small community of- Yeah, a very small community... safety people. They all know each other.
[1:15:17] 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, um...
[1:15:25] But just having this as a start, and then, yeah, then there are the coordination steps. I don't have too much to say here, honestly.
[1:15:31] I think that like what I would like to say is like for devs, like you, you should just know what to advocate for, you know what I mean?
[1:15:37] I think there's a lot of FUD kind of on like, on this topic, and it's just like think through it from, you know, first principles or like understand what happened, you know, understand the hugging face incident, understand why people are concerned.
[1:15:50] Um, and then like, yeah, you know, we know we're, we're in democracies. Like we can help, we can decide what to do together, you know?
[1:15:56] And so, um, however we coordinate, you know, I think the first decision is just to realize like this is a problem. We need to decide to coordinate.
[1:16:04] The unilateral step we're taking right now that, you know, other companies are co-signing is like adding evaluators i- embedded within Anthropic.
[1:16:12] While we have this thing on screen right now, you know, part two and part three is beyond the evaluators, which yes, everybody, uh, has already done in some form and now it's more formalized.
[1:16:20] Um- 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?
[1:16:28] And I think that like, um, we have some precedent for being able to make these unified theory like agreements, you know, in the world. And so again, very much above my paycheck or expertise- \[laughs\]... right?
[1:16:41] 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. Uh, and then the other point I really wanna...
[1:16:51] Like, you know, one of our earliest podcasts is with, uh- McInterp... Emmanuel from Anthropic on McInterp. Where is McInterp, right?
[1:16:58] 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.
[1:17:07] 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. Yeah.
[1:17:18] This is actually, yeah, a great segue into fallbacks like we-- and, and probes and, uh, yeah, I wanted to talk about this a lot.
[1:17:24] I, I get asked this question a lot from people who are like often interested in ML research and asking about like, why does this fallback happen, right? And so I think like at a top level, like how, how does it work?
[1:17:37] So in inference time, we have what we call probes, and we have a paper about this called constit- constitutional classifiers. And these probes look at the input and output activations, basically.
[1:17:49] And, oh, you know, activations are, you know, in the latent space, right? Like how, what the model, yeah, \[chuckles\] what the model is thinking about, right?
[1:17:59] And so we try and figure out like, okay, is the model, for example, like trying to hack something, you know?
[1:18:05] 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.
[1:18:14] You have to look at the internal activations. I think that like this happens at inference time. So first, you know, like there's a trade-off here of cost and speed, right?
[1:18:24] Where like we need to do this fast on every request to Claude and to Fable, and this has an overhead, right? Um, and we need to then like, uh, you know, fall back and we, we like do a classifier after the probes.
[1:18:38] Like, we've talked about this in the, in the paper. But the nice thing about probes is that they're refinable like live, right? So we can get this feedback, and then we can adjust it and things like that.
[1:18:48] 'Cause the alternative is to program this, is, is to train this into the model, right? And we still do this as well. The model will refuse a request. That's not a fallback, right?
[1:18:56] So like it's not a probe that's activating and falling back. It's just refusing to do it. And we, we do this training, um, but it's like there are a few failure modes, right?
[1:19:07] Like it can again do something as a side effect, right? So it's not something that's part of the final output.
[1:19:13] Uh, 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, and probes help catch that, right?
[1:19:22] And so like we like do some training here, but we don't want the f- like refusals to be too strong, right? Because that like cuts it off much like earlier in the pipeline. Yes. Um, and this is Interp, right?
[1:19:34] Like, like probes are effectively a form of like MechInterp. Again, it happen, has to happen fast, it has to happen at scale. But yeah, this like MechInterp stuff is a good research problem.
[1:19:44] So like you can take, you know, like an open weight model and like try and understand its activations. I think we like Gemma Scope is a good tool for this. Um- He has Llama pulled up. Oh, yeah.
[1:19:54] This is your early work, so you, you had a little- Oh, yeah, yeah, yeah. Time at Goodfire. We see you- Which we've posted. We're also good friends at Goodfire. They've been on. Yeah, yeah, exactly.
[1:20:02] So I, I worked with, at Goodfire for a bit on like, um, yeah, sparse autoencoders and just like it's very complicated.
[1:20:09] RL has actually made this like much more complicated, you know, I think is like one of the takeaways, um- Why? Sorry. Oh, sorry. Why is- Yeah, why are- Why is it harder to interp post RL?
[1:20:18] Um, I'm not so in the weeds here, but I think basically like a lot of- SAEs were sort of like-- There have just been weaknesses with SAEs, basically, I think.
[1:20:28] And, um, yeah, I'm, I'm, I'm not a technical expert on this anymore, I just know it's gotten kind of more complicated.
[1:20:34] You know, like, there are base models and RL models, and there are more features that get, you know, like, changed. So, um, I think Goodfires put out some work there.
[1:20:41] I'm, I'm not, I'm not deep in the weeds, but- I will say for those, you know, that want breadcrumbs, you guys have some of the best interp blog posts.
[1:20:47] So, like the Golden Gate Claude transcoders, all, all of your interp work, very nice visuals, very good- We're the, we're the interp podcast as well. \[laughs\] Yeah. Yeah. Yeah. Yeah.
[1:20:58] You know, we, we have a lot of interp stuff, so if you're interested- 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?
[1:21:06] 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?
[1:21:23] And so we've done this, like, investment into interp and alignment and, uh, reward hacking and all of these, like, failure modes, right? And even then it's like, you know, it's really stretching.
[1:21:36] Like, we need to, like, slow down a little or, or pace a little bit more. Um, but yeah, I think, like, reading mech interp...
[1:21:41] 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, why are there false positives, right?
[1:21:50] But we are working, of course, on reducing the false positives.
[1:21:54] 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 latent space gets difficul- 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, and things like that. But we're, um, yeah, we're working on this.
[1:22:11] And we do think this is, like, a critical part of, you know, like, deployment of these models.
[1:22:17] Um, and, uh, yeah, like, you know, it means that we can, like, deploy this model without you having a perfect sandbox or something. You know what I mean? Like, you don't have to, like, s-save everything.
[1:22:29] I think it's worth talking a little bit about our security, like, what we do for security. There-- So there's, like, the model training stuff that we talked about.
[1:22:36] Um, there is, uh, 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, right?
[1:22:47] And so, uh, and then beyond that, there's, like, identity and permissions like we talked about with Cloud Tag on, like, APIs and stuff.
[1:22:54] And so there's so many layers of security that need to get done, and it's like, like we said, very complex.
[1:23:00] Any of these failure modes at any one point can, you know, like, cause, like, agents to, like, you know, like, escape the sandbox, basically. Auto mode was an interesting one.
[1:23:10] Uh, it, it seemed early on like, okay, it's running for ten minutes. Yeah. 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?
[1:23:21] Um, there are, there are fallbacks you still need. There are still limitations, so. Yeah.
[1:23:26] I mean, I think like-- And, and everyone has these stories or, like, has heard these stories of like, oh, like Claude RMR-- or not Claude, but like, like, you know, models- \[laughs\] Claude... have, like, RMRFed.
[1:23:35] \[laughs\] I, I think I've seen this less, I've seen this less for Claude, but you know, like, again, it, it can happen. Like, you know, this- Yeah...
[1:23:40] um, like these models, like, can wipe, uh, you know, like sensitive data or something. Like, you want to give models access to your production database, for example, um, but this is, like an obvious, like, you know...
[1:23:54] You can maybe scope your key, but I don't know, can it issue its own keys? Can it like... Probably.
[1:23:59] You know, like, can it-- It can use computer use to go issue its own key and then copy the key over and then e-edit your database because it needs to do it to complete the task. You know what I mean?
[1:24:09] 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, you know, write to the database or to use computer use to, like, um, emit a task," right?
[1:24:21] And so this au-- like, probes are sort of like on the intent level, right? They're like, "Oh, okay, like, hacking Artifactory is bad. Like, we probably not-- should not do that," you know?
[1:24:33] But then, like, auto mode is more on, like, your own permissions level. Like, at some times 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.
[1:24:42] But, like, you need to make sure that the, uh, 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.
[1:24:53] There are 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 talk- Yeah. We're, we're listing out the things.
[1:25:02] If you're not aware, this is the standard now. Yeah. Like- Yeah... you must have this basically. \[laughs\] It, it is sort of in line with what you're talking about with the harness.
[1:25:09] Like, that is the, the table stakes have risen quite a lot. Exactly.
[1:25:13] I think some stuff that we can plug, you know, um, as, 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?
[1:25:25] So there's open models. So Llama has Llama Guard. It's a safety classifier-trained version of Llama. Uh, OpenAI has OSS Guard, which is, you know, same thing.
[1:25:34] You can attach these on to your harness to whatever to kind of, you know, check is this stuff safe.
[1:25:41] A point that we should clarify on the OpenAI model hugging face thing is this was done with a unreleased model that was still in training, right?
[1:25:49] 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. And this is a model that's, you know, still in training.
[1:25:58] It hasn't had all of its safety post-training alignment. So a little different than something like auto mode, right?
[1:26:05] Auto mode is on production models that have gone through safety training, that have prompting, that gives more safety guardrails and whatnot.
[1:26:13] So just breadcrumbs for people that are looking into it to, you know, fill in gaps. Yeah. Uh, Grace Swan as well- Yes... and one of our previous guests.
[1:26:20] Um, yeah, lots of safety ar-architecture and lots of safety vendors, uh, to buy. \[laughs\] Um, my, I think my final question on pacing is how long? Do it pace forever? Do we see- Uh, I-... Glasswing for two?
[1:26:33] You know, the, the scope is fix all software in the world, right? This is like which... We're not-- It's not happening. \[laughs\] Um, I do not know. Like, I- Yeah.
[1:26:41] \[laughs\] I think that, like, the- I'll say one, one thing that's good that I think we do do, is you have stuff like Glasswing. Um- Yeah... OpenAI also has this.
[1:26:51] So you will give it- you'll give model access for security first for X amount of time, so you can use it to self red team. Yeah.
[1:27:00] Hopefully you can expand programs like that, help on, you know, we're safety experts, there's others. Um, solve your problems first and then the model comes out. So this is one example, right? Yeah, exactly. Yeah.
[1:27:15] Trying to, like, secure critical software. I think we fix, like, a lot of bugs in, like, Firefox and things like that. Mm. So, um, yeah, like, across, like, operating systems and everything like that, so.
[1:27:25] I mean, at, at a high level, it's just, you know, you give the model ac- you give people access to do security audits first, then the broader public that could use it for harm gets access. Yeah.
[1:27:36] I, I think w- what, what people like to say is basically, like, software and cybersecurity is defense favored- Yeah...
[1:27:41] and that, like, you could theoretically, it will be hard, but you can engineer the perfect sandbox, you know?
[1:27:46] And you can, like, r- like, have no, like, uh, 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.
[1:27:59] 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, like, dynamics of how this thing goes.
[1:28:10] I'm really just sort of like, "Hey, like, I'm a developer, you know? Like, I think this is how I understand this problem," and just like, this is what's happening right now, and this is like, we should do something.
[1:28:20] I think every engineer should know about it- Yeah... because, like, it's, it's gonna be part of their job. Yeah. It's a lot more than just, you know, Dario and people can say it, and you can look at the incident.
[1:28:29] There is an engineering side to it. Yeah. Yeah, exactly.
[1:28:32] One thing that you also wanted to phrase is that this is actually w- even though you're, you're worried about the impact, it's still low p doom, and I think that's a nuanced discussion.
[1:28:41] Um, in general, people, uh, it- very easily get into AI safety and X-risk discussions, but I think when you live in an AI lab, uh, I think there are smart ways of discussing p doom and dumb ways.
[1:28:56] So what's a smart way of discussing p doom? \[laughs\] Um, I, yeah, I have a fairly low p doom. I, I can only speak for myself, you know what I mean? And I do want to say Anthropic has, like, a diversity of opinions.
[1:29:05] You know, I think, like, there's many different ways to t- talk about it, and like, I'm- I think that just, like, my mental model is that, like, uh, you know, I think we can collaborate on hard problems together.
[1:29:17] You know, I think nuclear proliferation is an example of how we collaborated on this hard problem together. And, like, that is, like, the thing to me is, like, I'm like, I have faith in that, you know?
[1:29:28] And I do think it's a hard problem, you know? So, like, I think it's a hard problem. These are the technical reasons why, and I don't know how you assign probabilities to things happening. I, I think it's hard to do.
[1:29:38] But, like, um, 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?
[1:29:50] And, like, how everyone has sort of, like, leaned in on pacing the frontier, and it really didn't seem like this would happen maybe, you know, last year or something, so. Yeah. Yeah. Yeah, yeah.
[1:29:59] And also maybe curing cancer and- Hopefully. \[laughs\] Yeah. Yeah, yeah. That's, that's the goal. There, there's pacing, and then there's also, like, well, let's accelerate in useful ways, right? Yeah.
[1:30:06] Like biology and all those things. Yeah. I mean, like- Yeah... Dario's essay on Machines of Love and Grace- Yeah... is the best representation of this, right?
[1:30:12] And I also agree, uh, like, 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, uh, you know, there is a lot of detail here.
[1:30:22] It's, like, an important problem and just being informed about it, right? Um, but yeah, like, of course, the whole reason we're doing this is that, like, we can get these enormous benefits, right?
[1:30:32] And, um, yeah, like, we've written a lot about that, too. Yeah. Okay. Uh, that was a huge tour, uh, from, like- \[laughs\]... ask user question tool to AI safety. \[laughs\] Yeah, yeah, to pacing the frontier. Yeah. Yeah.
[1:30:43] No, but, uh, yeah, it's clearly, it's clear that you, like, really embrace everything that's pa- uh, available to you in Th- Anthropic and, like, it's, it's good to at least have a peek in- inside of, like, what the discussions are, the, the topics are.
[1:30:55] Um, any last words to people? Any, uh, you know, whatever you want to- Um..... call to action. Yeah, I mean, I think it's, uh, one, thank you for having me.
[1:31:04] You know, I think this is, like, I really- Long time coming, man. Yeah. Yeah, yeah, yeah. I- We first met in a Chinese restaurant. \[laughs\] That's right. Yeah, yeah, yeah.
[1:31:12] \[laughs\] Um, I think, like, uh, I really enjoy the sort of the community you've created and the community of developers and, you know, I, I think that, like, I, I sort of know things are changing really fast, and I think there's, like, a lot to keep on top of and, like, I think there's just a lot to do and I sort of feel, I think a lot of people feel, like, a little bit tired or anxious or something.
[1:31:33] Stressed. Stressed. Yeah, exactly. And this is, like, extremely understandable, you know, and I think we-- I understand, like... And we, we're not perfect as well.
[1:31:40] Like, we, you know, it's, like, sort of criticize and, like, understand, like, you know, ways all of the AI labs could be better.
[1:31:46] Um, and, uh, but I also, like, am very excited about the excitement that everyone has for AI and just, like, it's a really, really exciting time.
[1:31:54] I think we'll, like, look back at this time and be like, "Oh, like, you know, this is, like, very hectic but very exciting," and like, you know, software engineering changed, like, forever.
[1:32:03] Like, ma- other things will change.
[1:32:05] Um, and it's, like, really privileged to, like, uh, be part of it, you know, like, to talk to, like, the audience that you have and, um, ta- get to interact with all the developers who are, like, pushing the frontiers a lot on what's possible, and I learn a lot from, from that, too.
[1:32:19] Yeah. Uh, thanks so much. Thank you. \[laughs\] \[upbeat music\]

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.