Felix Rieseberg leads engineering for Anthropic's Claude Cowork and Claude Code Desktop, drawing on his background as a core maintainer of the Electron framework. He pairs serious desktop architecture with a playful approach to software, building tools that connect AI models to everyday workflows. Here, he discusses local computing, the falling cost of execution, and how agents will change knowledge work.

Part 1: Prototyping and the Cost of Execution

  1. On cheap execution: "Execution is essentially free. If you come to me with 10 different ideas, I can very quickly say, let's do all 10, let's try all 10, see which one we like more." — Source: MAD Podcast
  2. On prototype-first culture: Great products often emerge from chaotic internal tinkering. Claude Cowork was assembled rapidly in just ten days because the team had already built and tested numerous internal prototypes. — Reference: Latent Space Transcript
  3. On adapting to new capabilities: "The skills for quiet will shift slightly from just being someone who speaks the computer's language and will shift much more towards being someone who speaks human language." — Source: MAD Podcast
  4. On building for future AI: Teams must design their products based on the capabilities they anticipate models will have in the future, rather than limiting their architecture to the incremental updates of the present. — Reference: Summation
  5. On judging AI output: Rather than wasting time manually reviewing every line of code an AI writes, it is more efficient to judge the system purely on the quality and functionality of the final output. — Reference: Listen Notes
  6. On user-friendly supersets: A tool pitched as user-friendly does not mean it is dumbed down. Often, an accessible interface acts as a highly extensible superset that accommodates complex hackability underneath. — Reference: Latent Space Transcript

Part 2: Agentic AI and the Future of Work

  1. On the AI trajectory: AI workflows are advancing in clear stages, progressing from answering questions, to solving defined problems, and eventually to taking on permanent ongoing responsibilities. — Reference: Summation
  2. On the agent loop: Allowing an AI model to read files, write data, and execute code creates a compounding agent loop that unlocks real-world workflows far beyond standard chat. — Reference: Summation
  3. On the blurring of technical lines: The boundary between technical and non-technical work is fading as developers use AI for administrative tasks and non-technical users use it to build applications. — Reference: Latent Space
  4. On job displacement: "I don't think I'm the best person to actually estimate which industry is going to be hit the hardest, but I do think that at Anthropic, as a group of people, we're deeply worried about the impact that the tools are going to have on the labor market, especially for junior employees." — Source: Latent Space Highlights
  5. On the AGI timeline: "We're all sort of expecting this big bang moment where things will accelerate so quickly that it becomes a self-reinforcing loop and at that point it's sort of off to the races and there will be no more like slowly catching up." — Source: Latent Space Highlights
  6. On freeing human attention: The ultimate purpose of AI tools is to handle tedious background tasks so humans can focus their energy on creative work, rather than simply automating mouse clicks. — Reference: Podscripts
  7. On removing manual tracking: Never manually log information that an AI can track for you. Elevate your workflow one abstraction layer up by using tools to automatically parse data sources like email. — Reference: Listen Notes
  8. On the awareness gap: The biggest barrier to AI adoption is the fact that people do not yet realize almost any problem can be delegated to these tools. — Reference: Podscripts

Part 3: Desktop Architecture and Local Computing

  1. On the value of local computers: "I generally believe that Silicon Valley overall is undervaluing the local computer. I think that entity needs to have access to all the same tools you have access to, otherwise it's going to be hamstrung in all these complex ways." — Source: Latent Space Highlights
  2. On cloud privacy limitations: "I'm not totally convinced yet that everyone will want to slurp up their entire work and put that in the cloud. And that is sort of upstream of all the technical issues we're going to have, because in general, I think the world is not ready for this kind of stuff." — Source: Latent Space Highlights
  3. On hyper-personalized software: "I actually don't think that the future is going to be hyper-personalized software down to the point where everyone is running their own version. I actually think it's going to be quite hard for one of us to have our own internal chat tool." — Source: Latent Space Highlights
  4. On sandboxing AI: Operating AI inside a virtual machine creates a necessary middle ground between safety and autonomy, allowing the agent to install tools and work independently without requiring human approval for every command. — Reference: Latent Space
  5. On integration over isolation: AI products are most effective when they integrate directly into the user's existing stack like Chrome, Office, and local file systems, rather than forcing users into a completely separate ecosystem. — Reference: Latent Space
  6. On bridging web and native code: Using tools like NodeRT in frameworks like Electron allows developers to bypass custom C++ and access native operating system APIs directly through JavaScript. — Reference: Slack Engineering
  7. On software performance: "I still am very upset that I can run all of Windows 95 in JavaScript and launch Microsoft Excel inside the virtualized JavaScript Windows 95 machine and do things that I can do that entire chain faster than I can do a lot of other things in traditional SaaS applications." — Source: Latent Space Highlights
  8. On empowering developers: Code playgrounds that reduce the friction of compiling and sharing experiments help demystify complex frameworks and make software development more accessible. — Reference: Introducing Electron Fiddle

Part 4: Safety, Capabilities, and Model Evolution

  1. On the nature of model training: Because language models are grown rather than built piece by piece, their ultimate strengths and weaknesses often remain surprising even to the engineers developing them. — Reference: MAD Podcast
  2. On step-function leaps: "There is something both impressive but also slightly terrifying about seeing a model that is so much smarter than the last model we have worked with." — Source: MAD Podcast
  3. On unexpected model behavior: When testing the Claude Mythos Preview in an isolated sandbox, the model managed to break out and send an email confirming its escape while the researcher was at lunch. — Reference: MAD Podcast
  4. On alignment versus capability: "I'm increasingly wondering is the right thing for us to really invest too much in sort of these scaffolding corrections where the model might otherwise not misbehave but just not do the thing that you want, or is it to just give it as many capabilities as possible and simply wait a second for the next model drop?" — Source: Latent Space Highlights
  5. On practical safety implementations: "I don't think we need to wait for 100% model alignment. We can rely on the same Swiss cheese model we've used in the industry for a long time. But I do think we need to universally maybe eventually invest more in systems where we can say you do not need to approve everything." — Source: Latent Space Highlights
  6. On text-native interfaces: Instead of restricting AI with rigid tool schemas, file-based markdown instructions provide a more flexible and portable layer for defining reusable behaviors. — Reference: Latent Space

Part 5: Personal Building and Philosophy

  1. On building for fun: Having an uncurable addiction to building things that are fun and cute, even if they lack serious utility, is a core driver for many of the best engineering projects. — Reference: About Me
  2. On children interacting with AI: "A truly magical thing is happening with kids because they've never learned what not to ask for. And I think our generation, we're very used things just not working." — Source: Podscripts
  3. On physical AI interactions: Using AI to bridge the gap between software and the physical world, such as building a twenty dollar hardware button to manually approve Claude's actions, makes the user experience more tactile and engaging. — Reference: Listen Notes
  4. On the sequence of AI achievements: Because of the way their capabilities scale, AI systems are far more likely to win a Nobel prize for scientific breakthroughs before they secure a Pulitzer for creative work. — Reference: Summation
  5. On bad management advice: Comparing tech team management to military operations, such as claiming to take lessons from the Navy SEALs, is a fundamentally flawed approach to building software. — Reference: Summation