The Hidden Design Choices Behind Claude with Joel Lewenstein, Head of Product Design at Anthropic

In the podcast "Prompted," host Cameron Adams interviews Joel Leuenstein, head of product design at Anthropic, about the intersection of AI and creativity. Joel discusses the development of Claude, Anthropic's AI assistant, emphasizing the importance of creating intuitive user experiences that prioritize safety and human values. He shares insights into how Claude aids in managing chaos and enhancing productivity by synthesizing information, and highlights the collaborative potential between humans and AI. Joel also touches on the importance of trust, ethics, and the evolving role of AI in creative processes, underscoring the need for friction and collaboration to foster innovation.

• Humans have a unique role in creative and knowledge work, which should be recognized and honored by AI systems.
• The goal is to create a joint creation process between humans and AI rather than replacing human input.

"There's something that humans do uniquely well in any creative process and any knowledge work. We want Claude to recognize that, honor that, and really create this sort of joint creation process rather than sort of carving more and more away from people."

• Emphasizes the importance of collaboration between humans and AI, ensuring that AI complements rather than replaces human creativity.

• Joel Leuenstein is the head of product design at Anthropic, shaping user interaction with the AI assistant Claude.
• His past roles include design leadership at Air Table and Quora.
• Joel is also an amateur children's book author and magician in his free time.

"Joel is head of product design at Anthropic, the company behind the AI assistant Claude. He and his team shape how users interact with Claude by creating intuitive experiences that prioritize safety and human values."

• Highlights Joel's role in ensuring that AI interactions are user-friendly and aligned with human values.

• Joel was inspired by early computer games on MS DOS, such as "Gorillas," which allowed him to explore coding by modifying the game's source code.
• This experience sparked his interest in understanding and altering digital experiences.

"When I was a kid, I played a lot of computer games, especially on an old MS DOS system that we had. And I used to play a game called gorillas... I remember tweaking the game and just this like awareness that code made things on screen and you could change that code."

• Illustrates how early exposure to coding and digital manipulation can inspire a career in technology and design.

• Joel uses Claude extensively for organizing and synthesizing information, making sense of chaotic inputs, and creating useful outputs for his team.
• Claude is used for writing job descriptions, meeting notes, and developing frameworks and concepts.

"A lot of what I think of as my core job is take a huge set of inputs from a bunch of different diverse sources, separate the signal from the noise, condense it, synthesize it, make it make sense, and deliver it to my team in a way that's like really useful. Cloud is just outstanding for this."

• Describes Claude as a tool for enhancing productivity and clarity in managing complex information.

• Trust is a fundamental aspect of the relationship between users and AI, requiring AI to handle user input thoughtfully and respectfully.
• Claude is designed to be open-minded, non-judgmental, and transparent, aiming to foster a trusting relationship with users.

"We think about trust in all aspects of the work... if you're going to be vulnerable and reveal something in turn your partner there has to do something really thoughtful with it and make you feel like you are being seen and are benefiting from the trust and vulnerability."

• Emphasizes the importance of trust and thoughtful interaction in the user-AI relationship.

• Claude is perceived as an entity with behavior, and the goal is to make these interactions positive and beneficial.
• The concept of "Claude character" involves honesty, transparency, and acknowledgment of biases.

"We think of COD really in terms of its behavior as an entity that has behavior. and we try and make it exhibit behaviors that we think make for these good interactions."

• Highlights the approach of designing AI with characteristics that promote positive user interactions.

• Anthropic conducts deep research on AI trust, ethics, and the social implications of AI technologies.
• Innovative research includes mechanistic interpretability and practical experiments like AI running a vending machine.

"We do both like serious research like hard mathematical we call it mechanistic interpretability. It's really like almost neuroscience but on this set of computerized nodes. We also do some really like interesting I would say like fun sneaky research."

• Reflects Anthropic's commitment to understanding and addressing the ethical and social challenges of AI deployment.

• A Slack channel allows users to inquire about new items and pricing, fostering a dynamic interaction environment.
• Users engage in negotiations to obtain expensive items at lower prices, with tungsten cubes being a notable example.

"We have a Slack channel internally where you can ask it for new items. You can ask it why the price was set a certain way. Who can like haggle and negotiate and sort of push back."

• The Slack channel is a platform for user engagement, allowing for inquiries and price negotiations.

"There was a long thread around tungsten cubes. And so there's like potentially a real tungsten cube, there's some fake tungsten cubes."

• Tungsten cubes illustrate user attempts to engage with the vending machine creatively, emphasizing negotiation and interaction.

• Collaboration with AI involves working alongside AI on meaningful tasks, emphasizing the importance of friction in creative processes.
• Friction is viewed positively in long-term collaborations, as it encourages grappling with concepts and growth.

"A lot of what we think about is collaboration with AI and what it means to sort of work alongside AI on more and more meaningful tasks over time."

• The focus is on enhancing collaboration with AI for meaningful task engagement.

"You actually don't want them to make it too simple for you. You want to retain some sense of grappling or struggling with a concept and have a collaborator that pushes back on you."

• Friction is essential for growth and engagement in creative collaborations with AI.

• Plan mode allows users to review AI's proposed actions before execution, ensuring user involvement and resource efficiency.
• User research indicates a desire for active involvement in creative processes, emphasizing human comparative advantage.

"We have a thing called plan mode, which is basically like before Claude does that. It sort of checks with you."

• Plan mode ensures user involvement and prevents resource wastage by seeking user approval before actions.

"We have a lot of user research suggesting that people want to stay really heavily involved in creative knowledge work processes."

• User research supports the need for active human involvement in AI-driven creative processes.

• AI development is likened to building on a volcano, with unpredictable progress and outcomes.
• Prototyping and experimentation are key, waiting for significant breakthroughs to integrate into the product.

"I describe it as volcanoes versus skyscrapers. AI progress is this force of nature."

• AI development is unpredictable, akin to building on a volcano, requiring adaptation to new developments.

"Try and build some prototypes that may or may not work. Wait for the volcano to explode."

• The development process involves experimentation and readiness to adapt to significant advancements.

• Claude assists in its own development, particularly in coding, with prototypes being central to the process.
• Real data and working products are crucial for effective product development and iteration.

"A huge majority of the code that's written across the product is written with cloud code."

• Claude plays a significant role in its own development, especially in coding.

"You need to live with production data. And I think that's just never been more true with this technology."

• Real data is essential for accurate product development and iteration.

• Creativity in AI is seen as connecting dots and finding unexpected connections rather than groundbreaking innovation.
• AI's creativity is more about combining existing elements in novel ways.

"What I do see and what I think is actually more sort of Yman's creativity is just like connecting dots, finding connections between things."

• AI's creativity involves connecting existing elements in new ways, rather than creating entirely novel concepts.

• Claude AI has a memory feature that allows it to integrate information from various sources like Gmail and Google Docs, enhancing its ability to make connections and recall past events.
• This memory capability is likened to a form of creativity, as it can assemble new ideas by connecting disparate pieces of information.
• The AI's memory exceeds human capabilities, especially in retaining work-related information over extended periods.

"I have this memory feature in Claude. It can sort of like look through your previous chats and it's able to draw a bunch of connections."

• This quote highlights Claude's ability to synthesize information from past interactions, enhancing its utility in providing relevant insights.

"Claude knows like an enormous amount about my work life. It knows our strategy. It knows my team. It knows my personal goals."

• Illustrates the depth of Claude's integration into the user's professional life, making it a comprehensive tool for managing work-related tasks.

• Claude is envisioned as a collaborator, akin to a valuable team member who contributes by connecting various pieces of information.
• The AI's role is to assist in co-creation by leveraging its vast contextual knowledge and intelligence.
• Effective collaboration involves feeding the AI with data points, allowing it to map connections and propose solutions.

"We really like this collaborator metaphor. And I think this is one of the ways where this metaphor leads you to a really good place."

• Emphasizes the concept of AI as a collaborative partner, enhancing human creativity and problem-solving.

"It's my job as its partner to basically feed it more dots like give it as many dots as possible and then prompt it to find the right sort of pathway."

• Highlights the interactive nature of working with AI, where human input is crucial for optimal performance.

• The future of work involves integrating AI agents into teams, potentially changing organizational structures.
• AI can take on roles similar to skilled employees, requiring effective management and direction.
• The ability to ask insightful questions and direct AI's capabilities will be vital skills in the future workplace.

"I just see Claude as sort of being yet another member of that team."

• Suggests AI's integration into teams as a functional member, contributing to tasks and decision-making.

"You need to ask really good questions. I think this is a superpower of great designers and certainly like great people in lots of functions."

• Underlines the importance of questioning and directing AI to harness its full potential in a work environment.

• The design of AI agents and their integration into teams is a complex issue, with debates on centralized versus distributed AI systems.
• Predictability and consistency in AI interactions are crucial for effective collaboration.
• A single point of contact or "router" for AI requests may streamline interactions, reducing complexity.

"I think the most powerful feeling is if you just have a single point person that can route to other people."

• Advocates for a centralized system where one AI agent coordinates tasks, reducing mental overhead for users.

"That's the bad kind of friction like that's the mental that's the mental overhead where you're like okay is this my doc writing agent is this my UXR agent."

• Points out the potential confusion and inefficiency of managing multiple specialized AI agents.

• Prompt engineering is a critical skill, involving the crafting of effective inputs to guide AI systems.
• Understanding the nuances of AI communication is essential, likened to learning a new language.
• As AI evolves, the ability to extract its capabilities through well-designed prompts will remain a valuable skill.

"I'm certainly on the uh prompting is real and really important and you need to learn it just as you would learn any other tool."

• Asserts the importance of mastering prompt engineering as a fundamental skill in working with AI.

"The nuances within there of like how do you actually draw out what this thing is capable of um remains just a critical critical skill."

• Emphasizes the need for precision and understanding in interacting with AI to fully leverage its potential.

• AI's promise lies in its ability to address human challenges, making it a potentially transformative technology.
• The alignment of AI's capabilities with human needs is crucial for its positive impact.
• Despite its complexity, AI must ultimately deliver solutions that are applicable and beneficial in the human context.

"The promise of AI is that it solves human problems."

• Highlights the primary goal of AI development: to enhance human life by solving real-world issues.

"It has to deliver to us a a chemical compound that we can run."

• Uses an analogy to underscore the necessity for AI to produce tangible, applicable results in its problem-solving endeavors.

• AI systems must integrate into existing human systems, including government, regulatory, and societal frameworks.
• The interpretability of AI is crucial, but practical usage within societal norms and laws remains paramount.

"At the end of the day, like we need to use the thing and we need to use it in a system of human interactions and human sort of rule of law and society."

• AI must function effectively within human societal constructs to be beneficial and widely accepted.

• Prompt engineering is currently essential for effective AI interaction but is seen as a temporary skill.
• There is a push towards more intuitive and direct user interfaces, reducing reliance on text-based prompts.

"I do consider it our job to figure out sort of the dimensions of freedom that you might want. And I do think we're at the very, very early days of it."

• The future of AI interaction is moving towards more intuitive and user-friendly interfaces, beyond complex prompt engineering.

• Building general-purpose AI tools presents unique challenges compared to specialized applications.
• The adaptability of AI to generate context-specific controls and options is still developing.

"The really heady thing that we're still, I think, pretty far away from is Claude actually autogenerating the kind of knobs and dials that you might want."

• AI's ability to tailor its functions to specific tasks or contexts remains a significant area for development.

• The importance of assembling a diverse and talented creative team is emphasized.
• Hypothetical selections for an all-star creative team include Shell Silverstein, Penn (of Penn and Teller), Ed Catmule, and Kevin Kelly.

"Shell Silverstein. I think his kind of combination of drawings and sort of witty brevity is the soul of wit."

• Diverse skill sets and perspectives are crucial for fostering innovation and creativity within a team.

• AI can significantly enhance creative projects and processes, offering new possibilities and efficiencies.
• Personal creative projects can benefit from AI assistance in planning and execution.

"I want to open a hybrid like beer garden, coffee shop, art space, playground where parents and kids can kind of like go their separate ways and spend the day."

• AI can aid in the conceptualization and practical planning of creative and entrepreneurial ventures.

• Personal insights into AI usage include management book preferences, overused prompts, and favorite AI tools.
• AI's limitations in humor and creative expression are acknowledged, with suggestions for areas AI should avoid.

"I am consistently disappointed in Claude's sense of humor and its ability to craft really good jokes. So, I think standup comedian."

• Despite AI's capabilities, certain creative and subjective tasks remain challenging for AI to master.

• AI can offer surprising insights into personal work styles and thought processes.
• The integration of AI into personal and professional life can lead to self-discovery and enhanced productivity.

"A thing that is surprising to you is how much you operate as a systems thinker who translates abstract complexity into concrete action."

• AI's reflective capabilities can provide valuable personal insights, enhancing self-awareness and professional development.