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

The Genesis and Impact of Claude Code [00:00]

• Key Insights:

  • The speaker states that 100% of their code is written by Claude Code, with no manual edits since November, shipping 10-30 pull requests daily.
  • Productivity per engineer has increased by 200% due to AI assistance.
  • The prediction is that in a year or two, coding as a skill may not matter as much, with the assertion that "coding is largely solved."
  • The future vision is a world where "everyone is able to program" and "anyone can just build software anytime."
  • Claude Code is beginning to generate ideas, review feedback, and analyze bug reports, acting more like a "co-worker."

  "The title software engineer is going to start to go away. It's just going to be replaced by builder and it's going to be painful for a lot of people."

Boris Churnney's Return to Anthropic and the Mission of Safety [04:17]

• Key Insights:

  • Churnney briefly left Anthropic for Cursor but returned within two weeks, driven by a strong resonance with Anthropic's mission.
  • The core mission that originally drew him to Anthropic was safety, which he finds is the primary motivator for employees.
  • He emphasizes that even exciting product development cannot substitute for a mission-driven purpose.

  "And the thing that drew me to anthropic was the mission. And it was, you know, it's all about safety. And when you talk to people at Enthropic, just like find someone in the hallway, if you ask them why they're here, the answer is always going to be safety."

Claude Code's Explosive Growth and its Early Development [05:45]

• Key Insights:

  • A report by SemiAnalysis shows Claude Code authored 4% of all GitHub commits, with predictions of reaching 20% by year-end.
  • The growth rate of Claude Code is accelerating, not just increasing.
  • Initially conceived as a "little hack," Claude Code was part of Anthropic's broader strategy of developing AI capabilities in a sequence: coding, tool use, and computer use.
  • The transition for engineers is from AI as a conversation partner to AI that acts in the world using tools.
  • The project began with Churnney hacking for a month and then spending a month on post-training to understand the models deeply.

  "The craziest thing for me isn't even the number that we're at right now, but the pace at which we're growing because if you look at Quad Code's growth rate kind of across any metric, it's continuing to accelerate."

The Terminal-Based Innovation of Claude Code [09:37]

• Key Insights:

  • The very first version of Claude Code was called "QuadCLI" and was terminal-based because it was the easiest way for Churnney to build it as a solo project.
  • This approach initially received little internal reaction, as the expectation was for more sophisticated IDE-like environments.
  • The team stuck with the terminal form factor for a while, believing it could better keep pace with the rapidly improving models.
  • Claude Code became a hit internally at Anthropic, with daily active users going "vertical" after its release.
  • External release in February was not an immediate hit, taking months for widespread understanding due to its novel nature.

  "And the very first version of it, I I have like a there's like a video recording of the summer because I recorded this demo and I posted it. It was called QuadCLI back then. And I just kind of showed off how it used a few tools and the shocking thing for me was that I gave it a batch tool and uh it just was able to use that to write code to tell me what music I'm listening to..."

The Evolving Role of AI in Software Development and Beyond [13:33]

• Key Insights:

  • In just one year, the profession of software engineering has dramatically changed due to AI, with predictions of 100% AI-written code becoming a reality.
  • A talk in May predicted that by year-end, an IDE might not be necessary for coding, a prediction met with gasps but now seemingly validated.
  • Anthropic's thinking is rooted in exponentials, a concept driven by its co-founders who were early authors of the scaling laws paper.
  • AI is transitioning from just writing code to acting in the world, using tools, which is now extending to non-technical users.
  • Coding is considered a "largely solved problem," with AI like Claude now capable of handling it.

  "And so all I did was trace the line and yeah, in November that, you know, that happened for me personally and that's been the case since and we're starting to see that for a lot of different customers too."

The "Bitter Lesson" and Building for the Future of AI [14:52]

• Key Insights:

  • Innovation cannot be forced; it requires space and psychological safety for experimentation.
  • Even in February, Claude Code wrote only about 20% of the speaker's code, reaching 100% in November.
  • The core principle is to bet on more general models rather than highly specific or fine-tuned ones, aligning with the "bitter lesson" concept.
  • Claude Code was built anticipating the capabilities of models six months in the future, not just the current state.
  • Key future capabilities of AI models are expected to be enhanced tool and computer use, and the ability to run for extended periods without constant oversight.

  "The very first version of it, I I have like a there's like a video recording of the summer because I recorded this demo and I posted it. It was called QuadCLI back then."

Shifting Paradigms: From Coding to Building and Beyond [17:32]

• Key Insights:

  • AI (Claude) is beginning to generate ideas for bug fixes and features, acting more like a co-worker.
  • Coding is considered a solved problem; the focus is shifting to adjacent tasks and general capabilities.
  • Claude is now used for a wide range of non-coding tasks, including project management and paying parking tickets.
  • The next frontier involves AI agents handling complex, multi-step tasks across various applications.
  • The speaker predicts that roles adjacent to engineering, like product management and data science, will be significantly impacted by AI.

  "I think something that's happening right now is Quad is starting to come up with ideas. Um so Quad is looking through feedback. It's uh looking at bug reports."

Latent Demand and the Evolution of AI Product Design [19:16]

• Key Insights:

  • AI's ability to help identify what to work on is valuable for product managers, who can leverage tools like Claude Code by pointing it at feedback channels.
  • Claude Code now handles much of the work of processing feedback, suggesting PRs, and even performing code reviews, improving the bottleneck of code review.
  • The model's capabilities have improved significantly, partly due to specific training for coding and also general advancements in models like Opus 4.6.
  • Engineering team growth at Anthropic has quadrupled, but productivity per engineer has increased by 200%.
  • The rapid pace of change driven by AI is unprecedented, yet has become normalized, highlighting the need to recognize its magnitude.

  "The next big open question is just like, okay, now we need to now now humans are necessary for figuring out what to build, what to prioritize. And you're saying that that's where claude code is starting to help you."

Embracing AI's Potential: Underfunding and Speed [24:23]

• Key Insights:

  • A principle for success is "underfunding" projects slightly, forcing greater reliance on AI and encouraging efficiency.
  • Encouraging speed and shipping quickly is a key motivator, especially by leveraging AI to automate tasks.
  • The advice for companies is to initially provide engineers with as many tokens as possible for experimentation, rather than focusing on cost-cutting.
  • This approach allows for the exploration of innovative ideas that might otherwise be deemed too costly or complex.
  • Token costs are currently low at a small scale but become a point for optimization once an idea proves successful.

  "So, you know, for work where sometimes we just put like one engineer on a project and the way that they're able to ship really quickly because they want to ship quickly. This is like an intrinsic motivation that comes from within is just wanting to do a good job."

The Personal and Societal Shift: Coding as a Tool, Not an End [28:00]

• Key Insights:

  • The speaker no longer misses writing code manually, viewing programming as a practical tool for building rather than an end in itself.
  • Early programming experiences were driven by practical needs, like cheating on math tests and building solvers.
  • The beauty of programming, functional programming, and type systems are appreciated but not the primary focus.
  • There is still space for individuals who enjoy the "art" of coding by hand.
  • The evolution of programming, from hardware to software to AI, is a continuous progression, and understanding underlying layers can enhance engineering skills, though this may become less critical over time.

  "But from the very beginning, it's it's always just been very practical for me where programming is a way to build a thing. It's not the end in itself."

The Printing Press Analogy and the Future of Programmers [32:42]

• Key Insights:

  • The transition to AI in coding is compared to the impact of the printing press, which democratized knowledge and literacy.
  • Just as the printing press dramatically increased the volume of written material, AI is doing the same for software.
  • While learning to read and write took time, literacy rates eventually soared, suggesting a similar, albeit potentially faster, societal shift with AI.
  • The speaker believes coding will be largely solved within a year or two, making traditional CS education less relevant in its current form.
  • The future might see roles like "builder" emerge, or a convergence where "everyone's a product manager and everyone codes."

  "And so the the volume of printed material just went way up. Uh the cost went way down. It went down something like 100x over the next 50 years."

Advice for Thriving in the AI Era: Experimentation and Generalism [40:42]

• Key Insights:

  • Key advice for succeeding with AI includes experimenting with tools, embracing them, and staying on the bleeding edge.
  • Becoming a generalist across disciplines is encouraged, as effective engineers often cross boundaries (e.g., product, design, business).
  • The speaker's team at Anthropic includes individuals from diverse backgrounds who all code, reinforcing the generalist approach.
  • Hybrid roles, such as product and infrastructure engineers or engineers with design sense, are highly valued.
  • Future success will reward those who are not only AI-native but also curious, generalists, and capable of thinking about broader problems.

  "So on the quad code team, everyone codes. You know, our product manager codes, our engineering manager codes, our designer codes, our finance guy codes, our data scientist codes."

The Democratization of AI and Latent Demand in Product Development [47:25]

• Key Insights:

  • The principle of "latent demand" is crucial: building products that meet users where they are, making existing workflows easier, or enabling unforeseen uses.
  • Facebook Marketplace and Dating emerged from observing users "abusing" existing platforms for new purposes, indicating unmet needs.
  • Claude Code was used for non-coding tasks like growing tomatoes or analyzing genomes, prompting the development of Co-work.
  • The inverted approach of making the model itself the product, rather than boxing it in, leverages latent demand from the model's perspective.
  • Releasing products early, even if rough, is vital for learning about user needs and AI capabilities in the "wild."

  "And so the the first thing was buy and sell groups. So kind of special purpose groups to let people do that. And the second product was marketplace."

Safety, Observation, and the Pace of AI Innovation [54:30]

• Key Insights:

  • Safety in AI development involves three layers: alignment/mechanistic interpretability (lowest level), evals (lab setting), and real-world behavior (in the wild).
  • Early release of Claude Code and Co-work allowed for studying safety in real-world scenarios, which is critical as models become more sophisticated.
  • Mechanistic interpretability, pioneered by Chris Olah, involves studying AI neurons to understand how models think and plan.
  • The concept of "race to the top" encourages open-sourcing research and tools to ensure safe and beneficial AI development.
  • An open-source sandbox for agents was released to make it easier for others to implement safe AI practices.

  "And so for cloud code, for example, we released an open source sandbox and this is a sandbox they can run the agent in and it just makes sure that there's certain boundaries and it can't access like everything on your system."

The Future of Work and the Enduring Value of Common Sense [17:32]

• Key Insights:

  • AI agents are becoming increasingly sophisticated, with Claude Code now capable of generating ideas, bug fixes, and even handling project management.
  • The speaker believes coding is largely solved, with future advancements focusing on tasks adjacent to coding and general AI agent capabilities.
  • The transition from "software engineer" to "builder" is anticipated, blurring traditional role distinctions.
  • A core life motto is "use common sense," emphasizing first-principles thinking and avoiding blind adherence to processes.
  • The speaker's personal journey, including living in rural Japan and making miso, highlights the value of long-term thinking and skills beyond engineering.

  "I think the best results that I see are people thinking from first principles and just developing their own common sense. Like if something smells weird, then you know it's probably not a good idea."

Mastering AI Tools: Best Practices and Future Vision [49:12]

• Key Insights:

  • The "bitter lesson" principle suggests always betting on more general models and avoiding overly specific or fine-tuned solutions.
  • Building for the model of the future (e.g., six months out) rather than the present is a strategy for staying ahead, even if it means initial product-market fit challenges.
  • Models are expected to improve significantly in tool use, computer interaction, and the ability to run for extended periods autonomously.
  • For Claude Code users, tips include using the most capable model (Opus 4.6), enabling "plan mode" for structured task execution, and experimenting with different interfaces (terminal, desktop, mobile).
  • The ultimate goal is to empower anyone to build software, reflecting a vision similar to the democratization of knowledge brought by the printing press.

  "From the very beginning, we bet on building for the model six months from now, not for the model of today."

The AI-Driven Workflow and the Unfolding Future of Work [17:32]

• Key Insights:

  • The speaker's personal workflow now involves 100% AI-generated code, with no manual editing since November, and a high volume of daily pull requests.
  • AI agents are becoming more integrated into daily tasks, from coding to project management and even personal errands.
  • The "latent demand" principle drives product development, identifying how users are creatively repurposing tools for unforeseen needs.
  • The rapid advancements in AI are transforming professions, with coding becoming a solved problem and roles like "builder" emerging.
  • The future promises a world where AI agents can handle complex tasks autonomously, requiring a societal conversation about job displacement and adaptation.

  "The thing that I love about it is just thinking in these longtime skills. Uh, and yeah, I think postGI or if I wasn't at anthropic, I'd probably be making miso."