The argument in one line.
The productivity inflection from AI agents is not coming — it is already confirmed by Anthropic's own internal metrics, and the companies that fail to restructure their workflows around it will repeat the mistake companies made when personal computers arrived and productivity did not follow.
Read if. Skip if.
READ IF YOU ARE…
- You are a software engineer trying to understand how far agent-based workflows have actually matured.
- You run a team and want a grounded benchmark: how much code is really being written by AI at top-tier companies?
- You are evaluating whether token usage in your organization is real demand or incentive noise.
- You want a first-person account of what running hundreds of parallel Claude agents actually looks like day-to-day.
- You are watching the SaaS disruption thesis play out and want a framework for which moats survive.
SKIP IF…
- You want technical deep-dives into model architecture — this is a product-and-business conversation.
- You are already up to speed on Claude Code's capabilities and are looking for new feature announcements.
The full version, fast.
Claude Code grew faster than any product the team had seen across prior careers in tech, with each model release (Opus 4.5, 4.6, 4.7) producing a new exponential inflection. Code-per-engineer at Anthropic is up 250% since launch, and 100% of Claude Code is now written by Claude Code itself. The token-maxing debate misses the real dynamic: companies that restructure their workflows around agents — as some did with PCs in the nineties — capture enormous productivity gains; those that bolt AI onto existing processes see little. The near-term road map centers on longer-running tasks, parallel agent fleets, and the auto-mode safety layer that routes tool-use decisions to a second model rather than a fatigued human.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Who's talking.
00:39guestBoris Cherny
00:00hostAlex Kantrowitz
Where the time goes.
00:00 – 00:39
01 · Cold open
Coming-up teaser: growth, tokenmaxxing, sustainability.
00:39 – 05:40
02 · Claude Code's explosive growth
Internal release to exponential inflection; each model drop (4.5, 4.6, 4.7) re-inflected; team had never seen growth like this.
05:40 – 09:22
03 · What is Claude Code?
Agents vs. chatbots; the 'fancy text editor' bet; tools as the key differentiator.
09:22 – 12:45
04 · Using AI agents to book flights
Boris's CoWork story: 8 flights, 5 hotels, wrong hotel corrected; trust ratchet analogy to Waymo.
12:45 – 21:27
05 · Token-maxing and real demand
Amazon FT report; HBR PC paradox analogy; 250% code volume at Anthropic; advice: give tokens + psychological safety.
21:27 – 30:45
06 · Are AI agents too inefficient?
PDF spiral example; effort controls; intelligence vs. efficiency tradeoff; commenter's 'inherent to LLMs' claim rebutted.
30:45 – 36:03
07 · Rate limits and user frustration
Doubled rate limits; weekly limit increase; Colossus capacity; power users running hundreds in parallel.
36:03 – 41:48
08 · Beyond coding
QuickBooks, auto-mode safety layer, parallel agents as the next UX frontier.
41:48 – 44:37
09 · Claude prompting other Claudes
Boris's own workflow: a Claude that talks to his Claudes. Engineer leverage up; still bottlenecked on good people.
44:37 – 49:26
10 · The SaaSPocalypse
Seven Powers framework; network effects gain importance, switching costs collapse; one-app-for-all-software thesis pushed back.
49:26 – 51:54
11 · Self-improving AI
100% of Claude Code written by Claude Code since Opus 4.5; Jack Clark's 60%/2028 estimate endorsed; AI safety as the reason Anthropic exists.
51:54 – 54:45
12 · Do AI agents need world models?
Jan LeCun vs. Greg Brockman; Boris sidesteps the theory but cites surprising emergent planning (poetry experiment).
54:45 – 57:24
13 · Is this the future or a fever dream?
Opus 4.7 hackathon: doctor, electrician, carpenter. People jumped through terminal hoops to use it — the ultimate market test.
Lines worth screenshotting.
- Code-per-engineer at Anthropic grew 250% after Claude Code launched, without any regression in code quality.
- 100% of Claude Code is written by Claude Code — a self-authoring loop that has held since Opus 4.5 in November 2025.
- The HBR computer paradox of the nineties is repeating: companies adopting AI without restructuring their workflows will see no productivity gain.
- The productivity gains from AI tools come from people you never predicted — accountants, marketers, new grads — not your top engineers.
- Token-maxing is a real phenomenon but a small share of demand; the larger signal is that most Claude Code users never hit their rate limits.
- Switching costs as a business moat weakens as AI makes migration easy; network effects strengthen because the code author becomes irrelevant to the network's value.
- Running one agent at a time is already obsolete — Boris Cherny runs hundreds overnight in parallel, sometimes thousands.
- Auto-mode routes tool-use decisions to a second Claude instance, and that second Claude catches unsafe commands the fatigued human would have approved.
- Non-engineers installing Claude Code in a terminal — their first terminal experience — was the early signal that the product had broken out of the developer niche.
- The Opus 4.7 hackathon produced a winner who built and sold a startup; other winners included a doctor, an electrician, and a carpenter with no coding background.
- Claude is starting to generate its own feature ideas for Claude Code, though they are not always good ones yet.
- Effort controls (low/medium/high/extra-high) let you trade token spend against intelligence without switching models.
- The model improves faster than users update their mental model of its capabilities — people who last tested it a year ago still think it is unreliable.
- Jack Clark's 60% estimate that models will start self-improving by 2028 was described as 'seems right' by the person running the product that already writes itself.
The agent inflection is already past tense.
WHAT TO LEARN
When the team building the product runs hundreds of instances of it overnight, and the product writes 100% of its own code, the question is no longer whether AI agents work — it is whether you have restructured your work around them.
02Claude Code's explosive growth
- Each model release (Opus 4.5, 4.6, 4.7) produced a new exponential inflection — a pattern the entire team described as unlike any hypergrowth they had seen before.
- Code-per-engineer at Anthropic grew 250% without quality regression, a benchmark worth measuring in your own organization.
03What is Claude Code?
- The jump from chatbot to agent is a single property: the ability to use tools — edit files, run commands, access browsers — rather than just talk.
- The original bet was explicit: the existing model for writing code (a fancy text editor) was so suboptimal that something radically different was worth building.
04Using AI agents to book flights
- The trust ratchet with AI agents mirrors the Waymo experience: white-knuckle approval of every action, then five minutes later you are on your phone while it works.
- CoWork found two missing travel stops and incorrect dates that Boris had provided — catching errors in the human's input, not just executing instructions.
05Token-maxing and real demand
- Give everyone tokens and psychological safety before you optimize — the productivity gains come from people you never would have predicted, not your top engineers.
- The HBR PC productivity paradox is repeating: companies that do not restructure their entire business process around the new technology will see no productivity gain, even with access to the tools.
06Are AI agents too inefficient?
- Effort controls let you dial token spend vs. reasoning depth without switching models — low/medium for speed, extra-high/max for the hardest tasks.
- The 'loops and spirals' inefficiency is a current-model problem, not an inherent LLM problem — Claude Code itself 18 months ago had the same behavior before the model improved enough to self-author.
07Rate limits and user frustration
- Very few users actually hit rate limits — the most vocal complaints come from a small edge of power users running parallel fleets, not the median user.
- Running hundreds of agents in parallel overnight is now a real use case, and the API offers unlimited tokens for those who outgrow plan limits.
08Beyond coding
- Auto-mode routes tool-use decisions to a second Claude instance; the second model is safer than a human at the permission prompt because it does not rubber-stamp from fatigue.
- The next frontier is not new categories but longer-running tasks and better UX for orchestrating parallel fleets.
09Claude prompting other Claudes
- The leverage stack now runs human prompts a Claude that prompts Claudes — the human is no longer bottlenecked by model throughput, only by the quality of their steering.
- Even with per-person leverage multiplied dramatically, Anthropic is still bottlenecked on good people because the demand growth outpaces the leverage gain.
10The SaaSPocalypse
- Network effects are the moat that strengthens as coding becomes cheap — the value of a messaging app is who is on it, not who wrote the app.
- Switching costs are the moat that collapses — Claude Code can already migrate stacks between vendors, and the model will only improve at this.
11Self-improving AI
- 100% of Claude Code has been written by Claude Code since Opus 4.5 (November 2025) — the self-authoring loop is already live, not a future milestone.
- Claude generates its own feature ideas for Claude Code, but the ideas are not always good yet; the human is still responsible for steering, not just prompting.
12Do AI agents need world models?
- The empirical counterargument to the world-model critique is that models trained only to predict the next token exhibit surprising planning behaviors — including composing the second line of a poem before finishing the first.
- Boris's practical rebuttal is an open invitation: sit down and use Claude Code for one hour, then decide whether world models are missing.
13Is this the future or a fever dream?
- People without coding backgrounds installing a terminal tool for the first time because they needed what it did is the most reliable product-market fit signal — not press, not benchmarks.
- A hackathon winner built and sold a startup with Claude Code; other winners included a doctor, an electrician, and a carpenter — the non-engineer breakout is already happening.
Terms worth knowing.
- Token-maxing
- An organizational practice where employees are rewarded or measured by the volume of AI tokens consumed, sometimes leading to artificial usage that inflates demand metrics without producing real productivity gains.
- Auto-mode
- A Claude Code feature that routes tool-use permission prompts to a second Claude instance instead of a human, reducing approval fatigue while improving safety by using a model that is not subject to fatigue-driven 'always allow' approvals.
- Effort controls
- A per-session setting in Claude (low/medium/high/extra-high/maximum) that trades token consumption against the model's reasoning depth, independent of model tier.
- Seven Powers
- A strategy framework identifying seven durable business moats: scale economies, network effects, counter-positioning, switching costs, branding, cornered resources, and process power. Boris uses it to reason about which software businesses survive AI-driven commoditization of coding.
- CoWork
- Anthropic's computer-use product — a Claude-based agent that can take over a user's browser and desktop to complete multi-step real-world tasks such as booking flights or configuring software.
- World model
- A proposed AI architecture component that would give a model an explicit internal representation of physical and social causality, allowing it to predict consequences before acting. Yann LeCun argues LLMs lack this; Greg Brockman and Boris Cherny both push back empirically.
Things they pointed at.
33:20bookHBR article on PC productivity paradox (1990s)
54:10linkJan LeCun world model argument
54:20linkGreg Brockman LLMs-to-AGI perspective
43:20bookSeven Powers (Hamilton Helmer)
Lines you could clip.
02:36
“I've just never seen growth this steep, and then it just kept going more and more exponential.”
1:12:48
“I don't write code. I prompt Claude. And actually nowadays, mostly what I'm doing is I have a Claude that prompts other Claudes. So I don't even talk to Claude.”
41:24
“The amount of code written per engineer at Anthropic has grown something like 250% since we introduced Claude Code.”
51:48
“Most nights I run hundreds of Claudes in parallel. Sometimes thousands.”
56:24
“People were jumping through hoops to use it because it was so useful.”
51:49
“Claude Code is a hundred percent written by Claude Code. CoWork is a hundred percent written by Claude Code.”
Where the conversation goes.
00:39 – 05:40denseGrowth metrics and trajectory
05:40 – 09:22steadyAgent definition and product framing
09:22 – 12:45steadyReal-world agent use cases
12:45 – 21:27denseToken-maxing and organizational change
21:27 – 30:45denseAgent efficiency and model controls
30:45 – 36:03steadyRate limits and infrastructure
36:03 – 41:48steadyRoad map: beyond coding
41:48 – 44:37denseLeverage stack and jobs
44:37 – 49:26denseSaaS disruption and moats
49:26 – 51:54denseSelf-improving AI and safety
51:54 – 54:45steadyWorld models debate
54:45 – 57:24denseMass-market adoption evidence
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
00:00I've just never seen growth this deep, and then it just kept going more and more exponential. QuadCode is a 100% written by QuadCode. Cowork is a 100% written by QuadCode.
00:09An increasing number of features
00:11are fully written by quad code across Anthropic and and products. So I wanna hear your perspective on on token maxing and whether you think that makes up a large portion of the usage of the products that you're building. I don't write code.
00:26I prompt Claude. And actually nowadays, mostly what I'm doing is I have a Claude that prompts other Claude's. So I don't even talk to Claude.
00:36I have a Claude that's talking to my Claude's. Let's talk with Claude code head Boris Churney about the product's explosive growth.
00:43What's next on the road map? And whether all this is sustainable? That's coming up right after this.
00:49Welcome to Big Technology Podcast, a show for cool headed and nuanced conversation of the tech world and beyond. We have a great show for you today. Cloud Code head Boris Churney is here with us in studio.
00:59We're gonna talk all about the product, the way it's taken off, what's next on the road map, and of course, whether it's sustainable. Gonna go into things like token maxing, token inefficiency, and then of course, the future of knowledge work.
01:11So no lack of topics to cover. Boris, it's so great to see you. Welcome to the show.
01:16Yeah. Thanks for having me. So let's talk a little bit to begin with about the growth of Cloud Code.
01:23It's been massive. Right? I think at a recent event, Daria Amadeh, the CEO of Anthropic, talked about how demand for Anthropic's products has been up like 80 times year over year.
01:34I remember speaking with him last year around this time and he was thrilled that Anthropic was at $4,000,000,000 ARR. That seems quaint right now.
01:43The numbers right now say maybe it's 45,000,000,000. Right? So a 10x there, ADX demand, and the question is how fast the company can serve the demand here.
01:53But talk about the portion of demand that Cloud Code makes up and what you've seen in terms of demand growth and the amount of people using this thing.
02:02For an increasing number of people in the world, I think the way that you use agents and the way that you use AI, it's not just Anthropic products, but it's QuadCode in particular. And, you know, of course, for Anthropic, there's a lot of different products.
02:15There's, you know, there's QuadCode, there's QuadAI Chat, there's QuadDesign, there's there's CoWork, there's, like, the API products. There's a lot of ways to experience Anthropic.
02:23But for a lot of people, QuadCode is their first introduction. And, yeah, the growth has just been insane. It's, you know, when we first released it internally, it just skyrocketed immediately.
02:34And so before we even released Quad Coat to anyone outside of Anthropic, we felt that it's pretty likely that this is gonna be a hit. And around the time that we released Opus four and Sonnet four, this was in May of last year, the growth just went exponential.
02:52And I've just never seen growth this steep, and then it just kept going more and more exponential. With with Opus 4.5, that was November, and then 4.6, that was February of this year, and then 4.7, it just keeps inflecting over and over.
03:05And, you know, there's a lot of people on our team that have worked in tech for a long time. And, you know, we worked on all sorts of hypergrowth products. Like, this is something you talk about in tech all the time.
03:15He's like, you know, Quarren's in hypergrowth, but even on the team, we've never seen growth like this. Uh, and so we're we're just trying to figure out how do we how do we make it so everyone can continue to experience this.
03:28How do we make it so we can continue growing at the space and the pace that we expect in the future, which might be even steeper than it is today? And we're learning a lot about about how to do this and how to how to keep scaling the services.
03:43So a year ago, it was clear that the bulk of usage of Anthropix
03:47AI models was happening through the API. Right? That would be like a company, like a consulting group, for instance, putting it into action at a bank and the bank using it to summarize some calculations.
03:57I'm just throwing an example out there. That compared to the Cloud Chatbot, it was far and away the API was the lion's share of usage, revenue, all these things.
04:06Does that still the case today or is Cloud Code overtaking that? We have a mix. So, you know, like, products play a much bigger role for Anthropic than they did a year ago.
04:16That's that's definitely the case. Product growth is accelerating. It's growing very quickly.
04:20API is also accelerating and growing very quickly. And for us, we are investing in both.
04:27We have to be a product company because there's kind of a lot of reasons for a lab to build products. And, you know, this actually wasn't clear early on. Like, very early on in Anthropic's history, this is before I joined, this was actually like an active debate.
04:40Should we even build products? Like, is this actually, like, a a useful thing to do? And it turns out it's very useful, you know, for mindshare, but then also for safety.
04:48Fundamentally, we exist to study AI safety. This gives us better tools to do that.
04:54We're also a small number of people, and so most things in the world, we will not build. Right. And so this is why we also have to provide a platform, and we have managed agents and API and SDK, all of these products, so people can build on top.
05:07And, you know, thousands and thousands of businesses choose to do that. Yeah. It's it's interesting to hear you even answer the question saying that it's a mix.
05:15So I take it you're not going to share which is bigger right now.
05:19Maybe not right now. Okay. But the fact that it's not a clear cut, the API is bigger, maybe it is, but the fact that you even say it's a mix just shows the fact that Anthropic's owned and operated products are just growing massively.
05:35And now so, you know, we've set the we've set the stage here that this is a thing something that's growing exponentially. We've obviously we obviously have seen the Anthropic revenue grow exponentially kind of alongside this product.
05:48This is a product that you conceived of and built and run today. I think that there's probably some people watching who are like, well, is Cloud Code? Most of our our viewers obviously know what it is and I was like how do I write this like in a simple one sentence definition?
06:05And I wrote that it's a way to build websites and software in plain English and then on the way over here I was like well that kind of sells it short a little bit. I mean, what would you describe it as?
06:17I think that's actually a pretty good description. It's Alright.
06:21We'll take it. I I think when a lot of people think about AI, think about chatbots. And, you know, for engineers, that's what AI was, you know, maybe like a year and a half ago before we started quad code.
06:32That's what AI was for most people. And we realized at some point that the model was actually getting really good at coding, and it's getting really good at using tools. And these are things that we've kind of always trained the model to do, and, you know, this has kind of been the research direction for a while.
06:48It started to become commercially useful about a year and a half ago. And so for Cloud Code, we took this bet, and we deviated from the way that everyone wrote code at the time.
06:58Because the way that everyone in the world wrote code was using essentially a fancy text editor. And we just thought maybe we can do much better than this, and we could do something really, really different than what's been done before.
07:11It was very much a bet. And so we introduced, you know, QuadCode, and the thing that made QuadCode different from chatbots at the time was QuadCode can use tools.
07:21And this is it. Like, the this is just the difference. It's with the chatbot, you're going back and forth and you're talking, but an agent and QuadCode is an agent.
07:29It can use your tools. Right. And can we just quickly define the tools?
07:33So tools could be anything and you tell me if I'm wrong from using a browser
07:37to like logging into Cloudflare and then setting up some agent that way. Right?
07:42So it becomes less of what does this product do itself and more of like what can this product log into and then sort of do with a multiplicity
07:53of products. Yeah. That's right.
07:55That's right. It can it can connect all your different tools. It can use your browser.
07:58It can use your computer. Even something as simple as, like, editing a file on your computer. You know, like, a year and a half ago, there was no AI product that could actually do that.
08:08But this is the first thing that QuadCode was able to do. It could edit a file on your desktop. If you have a bunch of files on your desktop, it can organize them.
08:16And so, like, QuadCode and Cowork have this access if if you choose to give it. They grant it. Yeah.
08:21And and, you know, it it can it can do this. And this is magical. It's this tiny difference completely changes the way that people can use this product and it totally changes what this product can do for you.
08:32Yeah. I mean, the fundamental thing, I think, just to drill down here, is that
08:38it seems like AI has shifted from sort of like as great at auto complete. Right? Because at the fundamental layer, AI is just predicting what comes next.
08:46Predicting, you know, if you if you're using machine learning and applying it on a large dataset, predicting whether you might default on your mortgage and whether a bank should grant a mortgage. When it comes to a sentence predicting the next word with code predicting the next bit of code in the sequence.
09:02Right? So I think that was gen gen one. But what you're talking about now is the machine is actually just able to go and after you give it this natural language prompt code itself, hook into tools, and then do things for you.
09:15And so correct me if I'm wrong, the the use cases here have gone from developers hooking into it and writing code with Cloud Code and we've seen this explosion, I guess largely driven by them, but then by a secondary force by non technical folks, people like me who can build software by directing the AI agent, is Cloud Code to build a piece of workflow software for them or a website or to take control of your computer via something like Cloud Cowork, which is sort of the maybe I would call it the easier sister product and saying, well, you have access to my to my browser now.
09:56You know what type of flights I like to book. I need to be in India in a couple of weeks. Book the flight.
10:01Yeah. Yeah. Exactly.
10:02I I actually just used Cowork to book a bunch of flights. I'm I'm gonna be flying a bunch this month for you know, we have Code with Claude coming up in in London and Tokyo, there's some other stops along the way.
10:14And I went back and forth with Cowork, and I was like, okay. I need to be in these in these places at this time.
10:20And it was five stops. It was like a lot of cities. And here's roughly the schedule.
10:25Look through my email, look through my calendar, and just double check it, make sure I'm not missing anything. It found actually two stops that I was missing and also a couple dates that I told it wrong. And it just found this by looking at my email after, you know, I I asked it to do that.
10:39And then I told it to book the flights. And I went and, you know, was coding on something, and I I was I was just doing work. And I came back an hour later, and it booked eight flights and five hotels.
10:51And one of the hotels was kinda incorrect. It was in the wrong area. I asked her to rebook it and change it, and it was done.
10:57Those are the I actually this is something that I try every time with Cowork and with QuadCode. I have these sort of, like, test cases. So these sort of like, a common thing that I would do, and I just retry it with different models and, you know, as the model improves.
11:11This is the best result I've ever gotten, and there's something about CoWork combined with OPUS 4.7 where it's able to do this. And I think one of the hardest things for me has been as the model improves, you constantly have to readjust your expectations of what it can do.
11:30And if you talk to people, especially engineers that use the model a year ago, they might and and they didn't use it since. They might say something like, oh, well, you know, it's not very good at coding, and, you know, I don't trust it to write more than a few lines at the time at a time because that's what the model was a year ago.
11:46It wasn't very good yet. And if you fast forward to today and you sit down these people and, you know, they they they try the new model and as as, like, a lot of people have been doing an increasing number of engineers, it's just a completely different experience.
12:01The the capability is completely different. And I I think this is the first technology I've used like this where every month, there's a step change in what it can do.
12:12And as a user of this technology, it's just quite hard because you have to kind of keep retraining. You have to keep retrying. You always need this, like, beginner mindset to to retry the technology and use it for a thing it was not good at before because the next model might might just do it perfectly.
12:26Right. And so I think this is the vision. The way that you're outlining it is
12:31effectively, previously, when you would use technology, you would be subject to the interface. You would have a software company that built for scale but you would get a lot of features that may be more inapplicable to you.
12:44You would have to go through all these bells and whistles whenever you were trying to book something even though you knew what you wanted and you wouldn't have a website that would know your preferences. Now it sort of shifts the paradigm where you have again, it's an agent.
12:59It's something that goes out and does things for you and can potentially shape your experience online the way that you want it. And that's that is I think what people are seizing upon and that's why we're seeing why you're seeing really the explosive growth. But now I wanna pressure test the thesis a little bit and bring up some things that make me curious how much of this is real and how much of this is just unbridled enthusiasm at the potential, but maybe stuff we should have a reality check on.
13:30And the first thing is that there is such great demand, But the question is how much of that demand is pure demand versus demand that's gamified.
13:41And there is a practice that's going on within Silicon Valley and outside of it that's called token maxing. I'm sure you've heard of it. It's where companies are have a mandate where people are supposed to use lots of AI tokens by running their AI agents as much as they can and then those who run the you know, use the most tokens are, like, rewarded on a leader or on a leaderboard or meet a goal of AI actions that they have to take as opposed to physical actions.
14:13So I wanna hear your perspective on token maxing and whether you think that makes up a large portion of the usage of the products that you're building.
14:23Yeah. I don't think token maxing is a large percent. Way that I would think about it is, you know, before Anthropic, actually, I used to work at a big tech company.
14:34You were on Facebook? I was on Facebook. Which is one of the companies that's token maxing first.
14:37That's right. That's right. Yeah.
14:38And one of my responsibilities was the health of all of the code across, you know, the across Meta's apps.
14:46So this is like Facebook, Instagram, you know, WhatsApp. And one of the reasons that we care about the health of the code, and this is essentially things like code quality, is if the code is really high quality, engineers are more productive.
15:00And there's, like, a big team of people that worked on productivity. And before models, before Claude, you would work for a really long time, and you would see maybe, like, a one to 3% improvement in productivity per engineer over the course of a year, like, something like that.
15:17And that was, like, a pretty big improvement, and it was, like, a very hard one. You essentially had to try a lot of ideas, and eventually you find something that improves productivity like this.
15:27And what happened with Claude is now many companies, including Anthropic and all of our biggest customers, are reporting gains on the order of hundreds of percentage points.
15:38And I think the the last number that we reported is the the amount of code written per engineer at Anthropic has grown something like 250% since we introduced quad code. And this is with while keeping code quality and reliability and all these things kinda stable.
15:55So without those things regressing, the volume of code has has grown a lot. And so this kind of productivity impact, think, is just, like, very new.
16:04And I think people are trying to figure out how do we get this. There's a lot of companies asking, like, how do we how do we get these kind of benefits? Because a lot of companies are seeing it, and then some are still figuring it out.
16:15And I think my advice is almost always the same. The first thing is just give everyone tokens. Let people experiment.
16:22I wouldn't necessarily recommend token maxing, but I would recommend let people experiment so they don't have to ask for approval for every token. The second thing is give people psychological safety. Because a lot of times, when people are innovating and they're building tools that make them more productive, they're changing their own workflows to make them more productive.
16:41They try a bunch of ideas. Some of them might not work, and then some of them work. So you want to give people this kind of psychological safety so they feel okay experimenting with it and finding these new processes.
16:51And then the thing that a lot of companies see is the productivity improvements and the innovations do not come from the people you expect. Back in the old days, you know, everyone could point out, like, these are my most productive engineers.
17:05But I think nowadays, a lot of the improvements are coming from people you just never would expect. It could be like an accountant somewhere in the corner of your org that just automates, like, accounting in a way that no engineer would have thought of. Um, it could be some marketer automating, like, marketing in a way that you never would have thought of.
17:21It could have been, like, a new grad software engineer that just built something amazing. And this is something that just, like, didn't happen before. The challenge is you can't identify these engineers and these people ahead of time.
17:33You don't know who they are, and it's almost always gonna surprise you. And so the thing you wanna do is let people experiment, give them safety, and then once there's some kind of use case that scales up, that's when you think about optimizing it.
17:45But you don't wanna optimize ahead of time. So I don't know. If doing it in a competitive way works for some companies with their culture, then I think that's great.
17:56If for other companies, the way they want to do it is just kind of create safety and create space for engineers to experiment, which is what we do at Anthropic,
18:03then I think that's great too. It really depends on the company. Yeah.
18:06And and I'll say, look, I use a lot of tokens. I'm in the tools all the time. I think Cloud Code and Cloud Cowork have both been pretty great for my business.
18:16I'm a solo operator although that kind of sells it short because I have a team of people behind me that help me mostly in a part time basis but that's for a different show.
18:26But but I I do wonder, you know, when I read these stories, the large corporations are largely making up big big percentages of these budgets And the incentives, you know and and again, like, I started the show saying how sustainable is this?
18:41The incentives are are bad in some of these places. This is from the Financial Times recently. Amazon staff use AI tool for unnecessary tasks to inflate usage scores.
18:51Some employees said colleagues were using the software to automate additional additional unnecessary AI activity to increase their consumption of tokens. They said the mood reflected pressure to adopt the technology after Amazon introduced targets for more than 80% of developers to use AI each week.
19:09I've I've got checked this with an Amazon employee. They're they're like, yep. This is what's happening.
19:14They told me I triggered an automation that runs for hours and then gets deleted every day in order to meet these targets. So you said you don't think that this token maxing stuff is is a big part of demand.
19:27Is there anything that you can see on your end to indicate that it's not? That this is an outlier and not the rule in most places?
19:35Yeah. This is I don't know how many companies are doing this token maxing thing. I've I've heard of it as a trend, you know, a little bit.
19:41If you look at QuadCode's customers, we have just many, many, many customers. So it's not like, you know, there's, one company driving the usage.
19:50It it's not like that. I I I do wanna kinda step back a little bit and just think about, like, how does this kind of change happen? Because I think the goal of what these companies are trying to do I don't wanna speak for them, and I would recommend just talking to Yep.
20:03But the the goal of what they're trying to do, I think, is probably, like, organizational change and business process change. How do you make it so your company benefits from AI? And this is often unclear.
20:13It's very dependent on the company because every company has a different business, a different culture, a different org, a different way of doing things. There there was this old Harvard Business Review article from the nineties, which I just love.
20:25And I I forget the title, but it but it was something like, computers are here. Why is no one seeing the productivity impact?
20:33And this was a big question. Right? It's like, to us, it's obvious.
20:36Computers make us more productive. This is just incredibly obvious today. But in the nineties, this was not obvious.
20:42And what was happening is personal computers were being adopted. They were replacing mainframes, and now they're affordable.
20:48So the average company, the average start up can can buy one. You don't have to spend, you know, millions of dollars on a mainframe anymore. But there was this challenge and there was this paradox.
20:58Companies were adopting it, but they were not seeing productivity improvement. What's going on? And so this Harvard Business Review article, made the case that in order to get a benefit from computers, you have to restructure your prod your your your whole business process around computers.
21:14They have to be at the center of the way that you do things. And if you still have, like, paper, you know, filing cabinets and you have a bunch of drawers full of stuff and it's still a paper and pen kind of physical process and there's a computer somewhere on the periphery, you're really not gonna benefit.
21:29But if you throw away your filing cabinets, you throw away your, you know, desk drawers full of, you know, papers, and you put a computer at the center of it, and that's the way that you do you all your business process, then you benefit. And there was this split between companies.
21:43Some were doing this, and they were doing this very painful change, and they benefited from it, and then others didn't. And I think it's kind of the same thing now. A lot of companies are trying to figure out how to benefit from the productivity impacts of AI, and there's just a lot of experimentation.
21:59And everyone is trying different approaches to to figure out how to how to benefit from it. I don't I don't think there's one right approach.
22:06Okay. And look, I I think that when we see something grow as fast as Cloud Code has grown and as fast as Anthropic has grown, it's good to just kind of talk this stuff through and it's good to hear your perspective.
22:22So okay, that's token maxing. Now tokens, of course, are the output of the model, like the words or portions of words that the model outputs and the words and portions of words that go into it.
22:35Right? And that is how these companies charge. The more you have, the more data centers you need, etcetera, etcetera.
22:43As these models get better, they haven't well, let me put it to you this way.
22:49Sometimes I wonder whether they're as efficient as they can be. These big models can sometimes do a lot of work, use a lot of tokens, even if the output is great. People wonder, well, is this sort of just driving up token demand where it could have been a really easy process and and the models are are expending many, many tokens and not getting there as as efficiently as they could.
23:12Let me give you an example. I've been using Claude Cowork to make PowerPoint presentations.
23:18It's really good at it. And I've been using the Opus 4.7 model. And a couple of times I've said, alright, you know, you're working on this this ship it as a PDF.
23:30And it just starts to lose its mind. It cycles and it uses as many tools as it possibly can and you know it just seems unable to ship the PDF and eventually I kept telling it, no, you're making this PowerPoint.
23:47You know where it is. Ship it. And it goes, owe you an apology.
23:51I went down a rabbit hole worrying about a constraint that wasn't actually blocking us, the files there. Mhmm. And then it shipped it.
23:58I mean, a little bit about the efficiency of these models and whether that is a legitimate worry that, you know, as we've seen the growth, part of it is these, like, loops that a model like OPUS 4.7 might find itself in to do basic tasks.
24:15Yeah. Generally, when we think about models, there's a few different aspects of it. One is just how intelligent is it.
24:22Another one is how fast it is, and another one is how efficient it is. And we generally try to move all these together. Between these, I think we should probably optimize for intelligence.
24:32That's the most important thing. So even if it's, like, a little bit less efficient, but it's more intelligent and it lets you do more things, that's really useful because the efficiency optimization comes after. After we make it more intelligent, then we can make it more efficient.
24:44So it's sort of kind of we do one, then we do the other. We've been experimenting a lot with how exactly we give people control over this because we don't always know the right default. Sometimes, like, when you're using it, you know better you know better.
24:58And so one mechanism that we had for this is picking a model. So you can pick, you know, opus or sonnet or haiku. Another mechanism that we've been experimenting with is think opus is, like, the biggest sonnet middle haiku smallest.
25:12That's right. That's right. That's right.
25:13And this is just like the size of the model. Right. And then there's effort.
25:17And effort is essentially how it you know, I think the word is actually really descriptive. It's how much effort do you want to put into it.
25:24And you can set this. We have a recommended effort. So, you know, for example, to maximize intelligence for OPUS 4.7, you want to use extra high or maximum effort.
25:34But if you wanted to use less tokens, can pick, like, medium or low effort, and this is a control that you have.
25:39Yeah. I talked about this on the show recently and we had a commentary that came in and I was of the opinion that this will these, you know, bigger models will find a way to become more efficient on like the export, the PDF thing. We had a commenter come in that wrote, Alex, they can't fix things like that PDF problem.
25:56It's inherent to LM technology and it's the biggest barrier to useful widespread dissemination and usage of agentic AI. I think I'm gonna try to translate that.
26:06What they were trying to say is we talked about predictions earlier that this is all probabilistic. It's sort of predicting the next word. You don't get the same answer from an AI agent twice.
26:17And so therefore, this type of thing is a feature of the way that they work and not fixable. What do you think? No.
26:24I I don't think that's right. When when you think about, like okay. Let's zoom out a little bit.
26:29Yep. So engineers are the first adopters. Right?
26:31Like, engineers started using quad code, like, a year and a half ago. And, you know, this is before non engineers were using agents in a meaningful way.
26:40This is, you know, before co work and so on. If I think back to what Cloud Code was a year and a half ago, it wasn't very good. I could use it to write a little bit of code, but if I really trust it to build an entire feature or an entire product, it wouldn't turn out well.
26:54It did the same thing. Like, it would go in spirals and the quality wasn't good or, you know, it built it and either the code was bad or it didn't work. And at some point, it just started to get better.
27:05And as the model improved and as quad code improved, the result just got better and better and better. And so you fast forward to today, quad code is a 100% written by quad code.
27:15Cowork is a 100% written by quad code. An increasing number of features are fully written by QuadCode across Anthropic and and products, and this is something that we hear from customers also.
27:27I did a I did a talk at Y Combinator, you know, the the the the startup incubator yesterday. And I asked people to raise their hands.
27:35You know, everyone everyone's using quad code, and I asked them, you know, raise your hand if a 100% of your code is written using quad code today. About half the hands went up. And then, you know, I asked people, raise your hand if 0% of your code is is, you know, written written with AI.
27:48There's, There's, one hand that went up, and this will remove, like, a few 100 people. Power to that person. That's right.
27:53That's right. And, you know, there's still room for this, obviously. And then everyone else was somewhere in the middle.
27:58You know? It's like most of their code is written with quad code, but not all of it. But that's kind of the place where the model is at today.
28:03It was not there a year ago. A year ago, it was not good enough for this. And so this is exactly what you're saying play out with Cowork right now.
28:10It's still early. You know, we released it, well, like, a a few months ago. It's it's gonna keep improving.
28:16It's gonna keep getting better as the product gets better, as the model gets better, but this is early days. I think still everyone using Cowork today is an early adopter.
28:25Everyone even using AI today is an early adopter. There are so many people in the world, and most people have not tried AI in a meaningful sense.
28:33So there's just like there's a lot more room to improve this. Yeah. We're hosting an event here in San Francisco on June 18 and a lot of the marketing material I've turned out with Cowork.
28:43Now I go back and forth. I don't let it one shot it, so I'm looking at the copy. But I do things like upload our download statistics to show the growth of the podcast, and I give it the names of the speakers and it like is amazing at saying building a prospectus.
29:01Here is what the event's gonna be, here's who's gonna be in the audience, here's who's speaking, here's why you should be there, here's how to get in touch. Mhmm. Insane.
29:08It's so good. What what was your what was your feeling like the first time they used it and the first time that you saw, like, the agents use your tools?
29:16Well, I I mean, obviously, I've sort of enabled everything. So and I think this is kind of an experience that many people have had where you there's a browser extension for Claude and you realize that you can only get the benefit of this or you're you'll get most benefit by letting Claude take over your browser and do things for you.
29:35And the experience is it's almost the same as I had with the Waymo where those first couple turns I was like white knuckling and like watching them like should I approve reading everything? And then you start to trust it a little bit and you just hit approve approve approve.
29:49Right? And the Waymo the same thing, you're like, okay, this looks like it's not gonna kill me. And then five minutes later, you're on your phone as the AI does the work and that was my experience with code and co work.
29:59Yeah. Is that is that sort of track? I mean, this is like my experience too.
30:02It's like it's I think it's like any technology. Mhmm. I was watching someone that's it's it's like a friend that's that's been learning to use a co worker over time, you know, she she's not an engineer.
30:12And there's this use case the other day, like, her there was, like, a language input on on the computer where you can kinda choose between languages on a laptop, and there was some issue with it, and she couldn't figure out how to fix it. And so before, what she would have done is go to Google and ask, like, hey.
30:26How do I fix this, you know, this issue that I'm having with my computer? And this time, she just, like, asked coworker. And the coworker was like, cool.
30:31Let me take a look. Can I can I use your computer? And she said, and I took over the computer.
30:35And I guess this kind of, like, orange glow, and you get to watch as co work open settings, and it sees what's going on with the language picker.
30:44And it diagnoses it, and it fixes it. And, you know, you're still in the driver's seat, so you you can see this happening. You can monitor it.
30:51It's not happening in the background or anything, but it's just it's magical. And I actually did like, my instinct was to open Google.
30:58And it so so it's funny that, like, for her, she went to using Cowork for this. And this is actually something I feel all the time.
31:06I I think for people that have kind of grown up with these products and they've seen previous versions, they might not be as ambitious as they could. But for people that are new to the products, I often see them using QuadCode and CoWork for things that I wouldn't have even thought of, and it's just, like, amazing. It's it's so creative.
31:21And I I learn a lot every every time I see it. Yep. Now the biggest drawback right now, I would say, and I've seen you reply to people on x about this, is the rate limits.
31:32Like when I see people say I've given Cloud Code a shot but I'm I'm kinda done with it. It's typically because they've hit their token allotment and it only works for like an hour for them and then they have to wait four to use it again and they look for alternatives.
31:52What do you think the rate limits have done to the ability for your product to grow And what is the plan, if there is one, to make people be able to use this without those rate limits?
32:06There's this is something we're actively working on. The reality is a very small percent of people actually hit their rate limits, which is surprising for That is surprising.
32:16For pro users, it's a little bit higher. For Macs, it's actually quite low. And I think the thing that you're saying when people talk about it is there's a couple of things happening.
32:26One is that we actually reduced the peak rate limits, and that's now rolled back, and we've actually doubled rate limits.
32:33So we're giving people more rate limits, but there was a brief period where we reduced them, and so people were running into that. The second thing that's happening is Cloud Code is actually quite extensible, and so people can use plug ins.
32:46They can use all sorts of integrations, and some of these use tokens in a pretty inefficient way. And so the thing that we've been working on is surfacing this to you so users can decide, do you wanna use this plugin or do you not?
32:58So you can see kinda what percentage of your tokens goes to it. And then I think the third thing is there's a lot of people that have just increasingly become power users. Like, first, when we released quad code, you know, you ran one quad at a time.
33:11Nowadays, I'm running, you know, like, on my computer, I run maybe five at a time. And then every night, I run, like you know, not every night, but most nights, I run, like, hundreds of quads at a time. In parallel.
33:21Yeah. Hundreds, sometimes thousands. And this is something that I just, like, wouldn't have imagined a year ago.
33:26And, obviously, this uses a lot of tokens. And there's a lot of people that are figuring out these new workflows that are using a lot more tokens, and this is sort of, like, at the edge of what you can do with the max plan.
33:37And, you know, this is why you can just, like, pay using API also. So if you just wanna have as many tokens as you need, you can do this too. And this is what a lot of enterprises do.
33:45Right. Now it wasn't long ago where I'm pretty sure Dario,
33:50Anthropix CEO, was referring to OpenAI and talking about the spending on the build out.
33:57And he and he's talked about this afterwards. He said, I'm trying to be disciplined in the way I spend, which is still spending many billions of dollars on data centers to enable this stuff like you're talking about and others, which we think is OpenAI, are yolo ing. Right?
34:11But now OpenAI is is doing this too with Codex. And you could call it YOLO ing but they have a lot of data center capacity that they've built.
34:24How do you think about that? Because you know when people do hit these rate limits they may just go over to Codex. It's pretty intense competition.
34:33So how do you think about that? How does Anthropic think about that internally that, know, at least from the outside perception is that this added discipline on data center build outs might end up losing users in the most important product battle that your two companies are engaged in?
34:51Yeah. So, know, first of all, our growth has never been faster than it is today. So, you know, for Quadco, the growth is accelerating.
35:01And I think because most people don't actually hit rate limits very often, it's it's actually not not a huge issue.
35:09For the people that are, we are laser focused on improving the experience. And so we doubled the five hour rate limits. We are announcing today that we're increasing the weekly rate limits.
35:22And, of course, we announced the new Colossus capacity, which, you know, we brought online to serve all these new users. Via Elon Musk. Via Elon Musk.
35:28Yeah. Because this I mean, this growth is just no one no one would have predicted this. Was just beyond our wildest forecasts.
35:35And so, you know, I think for us what matters the most is we we need to serve our users. We want to make sure our users are really happy, and we're doing everything we can to to make that happen.
35:45Are you surprised by Codex?
35:47How do you view them as a competitor?
35:49I think there's always, you know, there's always copycats. There's always competitors. For me, it's it's flattering, and I think it just forces everyone to do better.
36:01So, you know, I the for me, the thing that I care about the most is just doing the best job that we can to serve our users, and we encourage everyone on the team to, you know, talk to users every day and keep making the product a little bit better every day.
36:16So this is what I care about the most. Okay.
36:20I want to take a break, but we have so much more to cover. I want to talk about how this extends beyond code, the future of the chatbot, and then maybe talk a little bit about we have I mean, I could go through our agenda.
36:30We really need two hours. So why don't we take a break and come back and get to as much as we can right after this? And we're back here on Big Technology podcast with Boris Churney, the head of Claude Code at Anthropic.
36:42Boris, it's great having you here. Like I said, I'm in your product daily, so it's really fun to speak with you about it. We talked a little bit about this, but I think one thing we should highlight is that this is really gonna extend beyond the chatbot.
36:56We talked about booking flights. I talked about it with marketing presentations. And, you know, the week that we're talking, you have a a new use case out where Claude Cowork can be used for small businesses including taking over QuickBooks and doing some bookkeeping.
37:15Where where does this go? I mean, what do you think the broad road map where does where does the broad road map take you?
37:22We're thinking about a few things for quad code and for co work. There's a few few big themes. One is improving intelligence.
37:28And, you know, I I think almost all of this is just the model. As the model improves, we can do more and more ambitious work. For coding, it used to be writing a line of code at a time.
37:37Now it's building entire features or entire products. For a coworker, it used to be, you know, like you know, it it started pretty recently, but it was, like, you know, making a document, and now it's things like booking flights, combining many tools, doing doing your QuickBooks. So this this frontier is improving and and moving just very, very quickly.
37:56We're also thinking about how to do longer running tasks. For Cloud Code, we recently shipped this thing called auto mode, and auto mode is essentially a replacement for permission prompts. Before, what we used to do is whenever the model uses a tool, Claude would ask you, is it okay if I use this tool?
38:13And, you know, usually you just say yes, and you get kind of tired of saying yes over and over. Always allow. That's the button to hit.
38:19That's right. That's right. But it's actually very important for security that you're very thoughtful about this.
38:26And the thing that we're realizing is actually instead of being thoughtful about, you know, every prompt, because we're showing people so many of these dialogues, they just kinda got fatigued. They would just say yes or, you know, always allow, and so auto mode is the answer.
38:39This is a new way of routing these tool calls, and the way that it works is whenever a quad wants to use a tool, it asks another quad, is it safe to use this tool?
38:51Claude has some of the context. It doesn't have all the context. And there's also a number of layers of safety checks, and we spent months iterating on this to make it really safe.
39:00There's thousands of different benchmarks and evals that we use to make sure that this is safe. And, essentially, we found both in the laboratory setting and now we're finding in the wild, this is safer than what we had before.
39:12So as a user, it's a really nice benefit because you don't have to sit there and say yes over and over. And, actually, the result is better because if there's one unsafe command buried somewhere in this big list of things that Quad asked you to do, you might have accidentally said yes.
39:26But, actually, if you ask a second Quad using auto mode, it's not gonna say yes. So this is kind of one big investment. Maybe the third big one is just running more quads in parallel.
39:39One of the cool things about quad, and this is something that we started to see pretty early with quad code users, is actually very few people nowadays run one quad code at a time. Most people run many, many quad codes, you know, ranging from, you know, a few to thousands. And with co work, we're starting to see the same exact thing.
39:56As you get more comfortable letting co work run, you start a task, then you start a second task, and you move on, and you just do more in parallel. And I think there's just a lot of opportunity to make this experience very nice and to make it more obvious for people.
40:09How how do you do this? When do you do it?
40:11Right. And and it probably extends to the way that you use a chatbot.
40:16Right? And it's interesting because Anthropics had this kind of interesting relationship with the chatbot.
40:22Started out as technology first, decided to build the chatbot, ship Claude, and then just kind of moved more towards enterprise like you looked at all the charts and and Claude was always at the bottom.
40:37But now you're seeing Claude's usage rise. And I have a thought and I'd to check this by you, that the future of the chatbot is is not like I'm gonna give you a question and you'll give me an answer.
40:49It's I will give you a question or, you know, talk to you about a problem, and you the chatbot will then suggest some sort of action you can take on my behalf.
40:58Like, now I'm talking a lot about a trip to India and what I think I'm gonna get back in the future is this thing being like like what you said, not having this like secondary step between having to go there and and book the flights, a more proactive chatbot that's going to say, okay, let me take
41:17care of this for you. Is that the right direction? Am I thinking about that?
41:20I could see that. I could see that. Yeah.
41:22Are you working on it? Asians are the future and we're trying all these different experiments. There's some stuff that we're trying that's like this.
41:30But there is a limit here to what this can do. A funny way people have talked about the limits of the thousands of clouds that you can run-in parallel is kind of looking at who Anthropic is hiring. My favorite job listing on the Anthropic site is that you're hiring Salesforce administrators.
41:49You're also hiring consultants to help enterprises deploy this technology.
41:56And many are viewing that as like a sort of tacit admission that this stuff can only take you so far. Here's Wharton professor Ethan Moloch on it.
42:05He says, you will know that the AI labs believe in artificial superintelligence when they disband their newly formed consulting, sorry, forward deployed engineering groups.
42:17As long as people are required to figure out how AI is useful and do organizational change and systems integrations, jobs seem pretty safe. What do you think about that?
42:27Yeah. When you look at the kind of engineering that I do, I don't write code.
42:34I prompt quad. And, actually, nowadays, mostly what I'm doing is I have a quad that prompts other quads.
42:43So I don't even talk to quad. I have a quad that's talking to my quads. And I think in engineering, you've seen just this explosion in the amount of leverage that a single person has.
42:54It's about how how how big of a business can a person build, how many products can one person support. The leverage that one engineer has now at Anthropic is just insane.
43:04And I think we're starting to see this across other disciplines too. So we're starting to see this with the marketers that are, you know, using Cloud to do things. We're starting to see this also for forward deployed engineers that are using Cloud Code to build implementations.
43:17We're seeing this for our sales team because, you know, actually at Anthropic, I think, like, half the go to market team uses quad code, and the other half uses core. You know, I I think everyone's using all these products.
43:28And so the thing that we're seeing is the amount of leverage an individual has goes up, and we are still bottlenecked on the number of good people. And so even if the leverage per person goes up, you still just can't hire enough good people because the demand is so insane, and there's so much more to build.
43:45So that that's still the bottleneck for us.
43:48But I would say like, if people would people would argue that if this stuff was so powerful, you could say, take a look at the way my sales organizations operates and then configure Salesforce that way with a prompt.
44:02Is this another and people another example people give is I'll believe that Anthropic has very powerful AI if they let it let it handle the the IPO paperwork and don't hire an investment bank.
44:16Are these unfair tests?
44:18Well, we're starting to see
44:20there there's one person on the TLOs using Quad to do their taxes. Yeah. You know, I I would not necessarily recommend this, but I'll admit they did.
44:26I've run my taxes through Quad and compared it against my accountant, and it was pretty close. Yeah. I did the same thing.
44:32Folks, not not saying you should do that, but it is
44:36it's an interesting use case. That's right. But it I think fundamentally what people are missing in this conversation is in the end, it's a person that has to talk to Claude to ask Claude to do this thing.
44:47So even if Salesforce is automatically configured and, you know, it's not a person pressing all the buttons, it's Claude doing it, someone has to ask Claude to do that. And if you have to configure Salesforce in, you know, a bunch of different ways, it could actually be a full time job to ask Claude to do this. And at some point, Claude is gonna become really good at asking Claude to do this, and that person is gonna be asking Claude that asks Claude to do this.
45:07And this this chain will just keep getting deeper, but in the end, you still need people that that are piloting this. But maybe their job is just asking one question then in the future. Yeah.
45:16But imagine how much leverage that has asking the right question. That's true. That's a good point.
45:23So, you know, we we talked about Salesforce, so we have to talk about the SaaSpocalypse. You have some interesting views on the type of software companies that will be safe as we get more automated programming and those that that might be in trouble.
45:38And you talked you've talked previously about the different moats that exist and which moats are more important and which moats are less important.
45:46Can you just share that briefly,
45:48you know, while we're talking about it? There's this really good framework called the seven powers for talking about moats in business. You know, there there's so many of these frameworks for this, but this is my favorite.
45:57I actually studied economics in school. I I didn't study computer science. So this is still kind of the way that I think is in terms of these kind of frameworks.
46:04And there's a lot of these different modes in business, and some companies have one mode, some have a few modes. You know, they're they have, like, a portfolio of modes.
46:14There's a bunch of these modes. So, like, one is scale economies. So as you scale up your production, then there's increasing returns to scale.
46:22Another one is network effects. So this is like a, you know, like a messaging app or something like that. The more people that are on it, the more valuable it is for any person.
46:30Another one is switching costs. There's another one that's process power. I I think most of these modes are still gonna matter, and relatively, some are gonna increase in importance over the next year, and some are gonna decrease in importance.
46:42One that I think will increase in importance is something like network effects because it doesn't matter who's writing the code. It doesn't matter if it's an agent at the core of your product or or something else or if there's intelligence in your product.
46:55If there's a network effect in your product, that's still gonna matter. Some modes get less important, and this is, for example, switching costs. Because if you wanna switch from vendor a to vendor b, you can, you know, you can just ask Quad to do that, and Quad is gonna get better and better over time at it.
47:08And so I I think as a company, a thing that you should be thinking about is what are your motes? And I think a lot of the largest companies just have many, many motes. It's not it's not just one thing because the way you get to a scale and the way you build a defensible business over time is you accumulate these moats.
47:23You need a number of them. But, yeah, I would I would just think what's gonna be more valuable in the year and what's less valuable.
47:30I think that when you think about these different software companies, though, you're using a cloud code, do the most almost kind of blend away because you could potentially be in this like one app that is interfacing with all software, which means therefore there's really only one software company.
47:50Yeah. I mean, there there's just, like, a lot of ways that this could play out. I think something like this is possible, but it it seems a little far fetched to me.
47:58Because if I think about, for example, like, let's say I'm using a messaging app, how do I decide which app to use? I use the app that my friends are on that I can that I can reach. So it doesn't matter if I can build a really awesome app for myself, which I can do today.
48:10Like, I can build a great messaging app with quad code in, like, a few hours.
48:14It's still not useful because I can't talk to my friends. But this is the example exactly. You'll have just you can you can fact check me on this.
48:22You're gonna have an agent in your messaging apps that's going to let you know when your friends have messaged you. I know you use Cloud Code on your iPhone a lot. Right?
48:32So then you will just see the notification and you'll speak back to people. All your communication could potentially be centralized.
48:40And as long as the companies play ball
48:42Yeah. I mean connect. It could be kind of the agent in the end, but how how does the communication actually happen?
48:48So, like, you know, for example, if you look at a messaging app like, you know, like Signal, there's a protocol that it uses to communicate. And, you know, I can build an app.
48:57It can maybe use that same protocol, but I think it actually can't message other people that are on signal. But, yeah, like, I can have an agent that uses my app to do to do that messaging using an existing app that that supports this.
49:09Yep. So, yeah, it's it's not obvious how it's gonna play out. I think today, people use a mix of, you know, apps and and agents.
49:16But, you know, I I do fundamentally think that a lot of these modes are actually still gonna increase in value over time. You can think of another example. Let's say, you know, like a TSMC or some kind of, like, chip manufacturer.
49:28If you think about the amount of work that they put into making a process, and in making a process where the costs go down with scale, this is a fundamental economic force.
49:40And there's a lot of companies that that that do this kind of thing where, you know, especially in manufacturing, where with scale, the cost goes down. With tech companies, this is the case for infrastructure. So if you build a really great infrastructure, you can support more users, and the marginal cost per user goes down over time.
49:56So if you have this kind of effect, it doesn't matter if you or I can build apps.
50:00That's still a really powerful moat. But I I do think for sure both things are in play. Okay.
50:05I got three more in ten minutes. Let's see if we can get to them all. Jack Clark, one of the Anthropic founders, recently said I think that he believes there's like a 60% chance that these models will start improving themselves by 2028.
50:19It could be off by a percentage or a year, but ballpark, that's accurate. You're in the app where coding happens autonomously.
50:28You're running this app. Do you agree with Jack?
50:31Seems right. Yeah. I look at the way that quad code is written, a 100% of quad code is written using quad code.
50:41This has been the case since, I think, November of last year, since Opus four point five. It's like a fast takeoff scenario then. Do you anticipate that?
50:50Yeah. I mean, it's it's it's possible. And, like, this is this is why Anthropic exists.
50:54If you ask anyone, any any engineer, any researcher why they joined Anthropic, they're gonna tell you it's for AI safety. And it's because for us, when we think about the future, you know, years from now, the thing that's the the most important and the thing that we wanna get right, you know, for our kids is we wanna make sure this thing is safe, and we wanna make sure it goes well.
51:12Because, yeah, like, that is one of the possible outcomes. I think that's not yet what we're saying. You know, right now, QuadCode is writing itself, but it's still a person that's doing the prompting.
51:22Quad is starting to generate its own ideas for what to build next for QuadCode, but it's, you know, it's not always good ideas. And I still would generate most of the ideas.
51:31And, you know, at some point it's gonna change. The model's gonna improve, and it's gonna become more of a of a self reinforcing loop.
51:37Okay. I definitely wanna get your thoughts on the world model argument here, where people who are pro world models say that a large language model has no understanding of the consequences and you need to build a world model into it to have effective agents.
51:55Here's something from Jan Lecun. He says, you cannot build a reliable agentic system without a world model. LLMs don't have world models.
52:03They can't predict the consequences of their actions before taking them according to Jan. They just act and whatever happens next is someone else's problem. I was speaking with Greg Brockman from OpenAI recently, and he said, basically, he doesn't accept that argument, and he thinks LLMs are the way directly.
52:21These text models are the way to AGI. Which side are you on? Are you a believer that that world model intelligence needs to be baked in, or do you think that LLMs alone are good enough?
52:33I would put out an offer to Jan. If he wants to sit down and quad coat together for an hour, I'd love to show him. You guys should do that on this show.
52:40Yeah. And then I'm curious to hear what he thinks. Maybe he'll change his mind, maybe he doesn't.
52:44Right. But your perspective though? You know, I'm pretty firm on the product side.
52:49So, you know, I I don't I don't really have a have a perspective on it. Okay. Let me let me drill down a tiny bit deeper, if you don't mind.
52:58You know, you're you're on the product side, but I've heard multiple people bring out this idea that without a conception of the way the world works, like in a world model, LLM just doesn't have an understanding of the way that the world works and consequences and stuff. You use co work to book how many flights?
53:16Eight flights in hotels? Like, you must think that it has some understanding of consequences, otherwise you wouldn't have given it your credit card, which I presume you did.
53:26So what do you think about that argument in particular? I think from what I've read from folks working on on research at Anthropic,
53:32it is surprising the degree to which these models are intelligent. Because like you said at the beginning, the the thing that they fundamentally do is they predict the next token.
53:41Mhmm. And so you think, like, this is kind of like a stupid thing. Like, how can this possibly lead to intelligence?
53:45But, you know, we we've actually published a lot of work about how the models are able to plan, and they're able to actually reason. There's all these very surprising behaviors that you actually wouldn't expect from a model that just predicts the next token.
53:58So I don't know. I wouldn't discount it. I think my favorite is when they write poetry.
54:04As they're writing the first line, you can see in the model, this is anthropic research, that they're already thinking about the next line.
54:11That's right. Which is like, how is that even possible?
54:15That's right. I mean, that's kind of how I think about it. Like, if I if I were a poetry, that's how I would do it too.
54:19It's it's crazy. Like, you teach this thing to predict the next word and somehow, if the next word is hard enough, it has to learn to really plan ahead and it has to learn how to do all of this.
54:30Okay. Last one for you.
54:32Sometimes I wonder when I see big tech changes underway and in my career covering this stuff some have worked out and some haven't.
54:41I always have to ask myself how are we sure that this is the future and this is not a fever dream? And I think the data indicates that this is a real thing but I also wonder you have to sort of you have to question how much you can extrapolate towards the future in terms of how will this continue to progress.
55:00The argument that this is a fever dream is that maybe people just want simple interfaces and they don't mind tapping through things.
55:09And you know speaking in a cloud code feels a little bit too techie and it just won't appeal to the everyday user as much as it's really taken off with developers.
55:21I mean, how would you answer that?
55:24We had this had this hackathon for Opus 4.7 recently, and one of the winners was a doctor that built an app.
55:32There was there there was an electrician. There was a carpenter.
55:36And a lot of these people didn't have coding experience, but they used quad code to build something useful. There's one person that built and sold a startup as a result of one of these hackathons that we put on. Undoubtedly, when we first built QuadCode, it was for engineers, and engineers kinda figured out how to use it.
55:56But very quickly, people that were not engineers figured out how to use this to build economically useful things. And, actually, if you look at a lot of the usage today, it's like it's not engineers, and it's just so useful for people that they are going out of their way.
56:09They're jumping through hoops. Even before co work, People were, like, installing quad code in a terminal.
56:16For a lot of people, this was their first time using a terminal. And, of course, now, you know, for quad code, we have a desktop app. We have iOS app.
56:22We have a Slack app. You know? There's many ways to interact with it.
56:25But people were jumping through hoops to use it because it was so useful. And so for me as a product person, this is the ultimate market test of is this thing useful, is are there a lot of people that use this every day and that keep using it every day? And, yeah, it's a lot of people, and it just keeps growing.
56:42And I'm just constantly surprised by the way that people use this.
56:46Yeah. I I will say I've been surprised by the way that I found myself using the tools and I don't know.
56:52We'll we'll see what comes next. So excited to keep using it and and thrilled to have a chance to speak with you. I hope we can do it again.
56:59Yeah. Thanks for having me on. Alright.
57:00Thank you, Boris. Great speaking with you. Alright, everybody.
57:03Thank you so much for listening and watching and we'll see you next time on Big Technology Podcast.
The Hook
The bait, then the rug-pull.
Boris Cherny has a front-row seat to the fastest-growing product in Anthropic's history — and a disarming willingness to say what he actually thinks about the parts that are not working yet.
How they asked for the click.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
