Mike Krieger, cofounder of Instagram and now head of Anthropic Labs, shares his early experiences with Claude Fable 5. He describes how getting access to the advanced AI model months before its public release was both exciting and disorienting, fundamentally shifting his understanding of productivity and workflow.
Krieger, speaking with Dan Shipper, delves into the transformative power of building with Fable 5. They examine how this highly capable model reshaped established rhythms, presented new challenges, and unlocked possibilities in AI-native product development. Topics range from practical applications, like a media tracker built over a weekend, to the evolving roles within Anthropic's engineering teams.
This discussion offers unique insights from a leader at the forefront of AI innovation. It highlights how Fable 5 is not just another tool but a catalyst for re-evaluating core aspects of software development and human-AI collaboration. Understanding these changes is crucial for anyone looking to navigate the rapidly evolving landscape of artificial intelligence and its impact on how we build and work.
Key takeaways
- Understanding the full power of advanced AI models like Fable 5 requires prolonged engagement over weeks, not just initial day-one use.
- Fable 5 enables users to delegate entire complex tasks by expressing intent, fundamentally changing how workflows are designed and executed.
- The model exhibits problem-solving capabilities, independently handling unexpected obstacles during task execution, such as generating a temporary backend for a downed remote service.
- Engaging AI in architectural planning helps ensure project scalability and facilitates the creation of shareable documentation for team alignment and early design agreement.
- Efficient AI utilization often requires managing multiple concurrent sessions to break down and execute complex tasks in parallel, maximizing output from powerful models.
- Fable excels at complex queries requiring deep thought, while Sonnet is optimized for quick, less intensive questions.
- The cost and time commitment for building new software have dramatically decreased, exemplified by the contrast between Instagram V1's multi-day all-nighters and modern AI-assisted development.
- AI models bridge the gap between an idea and a functional product, empowering non-developers to create and iterate on tools efficiently.
- The role of a software engineer is evolving, with many previous engineering responsibilities now converging with product management, emphasizing high-level problem-solving.
- Human engineers at Anthropic maintain direct responsibility for product intent and overall direction, acting as Directly Responsible Individuals (DRIs) even while leveraging numerous AI "Clods."
- New workflows involve "meta-maintenance" where engineers actively manage and track tasks delegated to AI assistants, often through personalized dashboards, to ensure oversight and timely human intervention.
- A critical area where current AI models fall short is understanding the complexities of production systems, including real-world incidents and unexpected failure modes that emerge post-deployment.
- Product managers are now using engineering prototypes to quickly test and demonstrate product concepts, effectively resolving internal debates on product direction.
- Fable 5's higher cost is justified by its efficiency in completing complex tasks, reducing the overall effort and time users spend.
- The industry is moving towards evaluating AI cost based on task completion and user satisfaction, rather than just per-turn usage.
- Fable's AI can understand system-wide context and production requirements, proactively reminding users about dependencies like feature flags.
- AI interfaces should decouple work execution from conversation, enabling effective mobile interaction for delegated tasks.
- Implement tight development loops and visual verification, including video captures, for AI-generated code to quickly identify and resolve UI/UX issues like animation jank.
- Ensure comprehensive testing by having AI agents interact with real data in staging environments and utilize robust mock backends for complex systems that are difficult to spin up locally.
- AI-powered dynamic workflows can autonomously orchestrate complex, multi-step projects over extended periods, demonstrated by a full codebase conversion from Python to TypeScript.
Fable 5 Redefines Workflow by Enabling Delegation of Complex Tasks
Mike Krieger, head of Anthropic Labs, shares his experience using Fable 5, emphasizing that its true capabilities reveal themselves after extended use rather than initial impressions. He describes feeling like a "newbie" because his previous methods of prompting and task decomposition became outdated with the model's advanced functions.
The new model signifies a shift in user interaction, moving from step-by-step instructions to expressing broader intent. Fable 5 can understand global context and is expected to complete entire workflows, demanding users rethink their approach to AI interaction and the time horizons for task completion.
Krieger highlights Fable 5's capacity for autonomous problem-solving. He recounts setting complex tasks for the model overnight, which it would complete, even resolving unforeseen issues like a remote service outage by independently generating a scaffolded backend and documenting its actions. This demonstrates a new level of self-sufficiency and delegation previously not possible.
I will, you know, wish Claude a good night, set it up on like a pretty complex task or something of this Done by like two in the morning...really impressive ability to like complete the swing, get itself out of the situation where it's like, okay, alright, well, Mike asked me to do this complex task overnight, I got stuck because this remote service went down. I'm gonna write a like scaffolded like backend for it, for now, so I'll, you know, I'll document that, I'll, you know, go all the way through.
Optimizing AI Workflows: Architectural Planning and Concurrent Session Management
Mike Krieger observes a significant shift in how he interacts with AI models, evolving from companions to fully capable teammates to whom he can delegate substantial work. This redefines his understanding of productivity with these advanced tools.
He highlights the necessity of upfront architectural planning with AI, especially for projects intended for shipping. This involves a collaborative back-and-forth, where the AI can help design scalable solutions and even generate shareable documentation, like an HTML page with diagrams, to align human teams effectively.
This strategic planning enables the execution of tasks in chunks rather than monolithic requests. Krieger's current workflow involves managing numerous concurrent AI sessions, sometimes up to five or six tabs, to handle different long-running tasks simultaneously, optimizing the use of powerful but costly models.
Can you just like make an HTML page like that represents what we just talked about so I can share it with the team?
Optimizing AI Model Selection Based on Task Complexity
Mike Krieger distinguishes between AI models like Fable and Sonnet by their processing depth. Fable is designed for detailed, deep thinking, suitable for complex tasks, while Sonnet provides faster responses for simpler queries. Mike notes Fable offers effort levels that he adjusts for tasks like tweaking UI components.
Initially, Mike admits to using Fable for random questions, but realized this was inefficient when he needed quick answers for things like NBA finals scores. He likens using Fable for simple queries to using "a rocket launcher to kill a mosquito," highlighting that Sonnet is better suited for speed and efficiency in such cases.
The core insight is that not all answers require extensive thought. This presents a product challenge: guiding users to select the appropriate model without overthinking. Potential solutions include defining clear use cases for each model or implementing context-dependent model selections, such as a sticky model for an iOS app.
Mike realized he needed to use the faster Sonnet model more often for quick questions, acknowledging that an answer sometimes doesn't need to be fully thought through by a more powerful model like Fable.
Oh, yeah, I should use this all the time for fast questions.
Building a Self-Modifying Media Tracker with Agent-Native AI
Mike Krieger built a personal media tracker over a weekend using Fable 5, driven by a need to easily track games, TV shows, and recommendations. The primary goal was simple addition of new items and proactive research by the AI for new seasons or sequels. The initial UI was largely AI-generated, and the core interaction allowed users to add media by simply providing a URL to Claude.
The tracker exemplifies an "agent-native" architecture, where Claude not only processes requests but can also modify the application's code. This moves beyond standard tool calls, allowing the AI to integrate editing capabilities. For instance, if a user requested a show about radioactive events in Goiania without remembering its name, Claude could identify it and add it to the tracker.
The system leverages managed agents to take on edit requests, offering features like live previews and code diffs. Krieger described a scenario where he asked the AI to adjust the position of a floating action button for iOS, and it successfully implemented the change. This approach allows for rapid, iterative development and modification of the application directly through natural language prompts to the AI.
This project highlights a significant shift in software development, where the AI becomes an integral part of the development team, capable of understanding user needs and directly implementing code changes to meet those needs. It demonstrates the potential for AI to dramatically simplify and accelerate the process of building and evolving software, even for complex feature requests.
AI Significantly Reduces Software Development Costs
Mike Krieger reflects on how AI has drastically lowered the cost and effort of building software. He contrasts a recent weekend project, which felt easily manageable and iterative, with the intense five days of all-nighters he and Kevin Systrom put into developing Instagram V1. Even with Mike's high productivity as a mobile developer at the time, getting from an idea to a complete product required substantial, focused effort over multiple days.
The initial version of Instagram, though having more features, wasn't an order of magnitude larger in scope than what can be quickly prototyped today. Beyond the initial build, the challenge in Instagram's early days was the slow pace of iteration; even adding a basic feature like hashtags could take a week, diverting resources from other ideas.
Today, AI models are closing the gap between a user's intent and execution, making software creation accessible to a much broader audience. This empowers people who aren't professional software engineers or startup founders to build their own tools without needing consultants or significant funding.
Krieger shares an anecdote about an internal tool built with Fable and internal systems for a recruiter. This person expressed that it was the first time in her life she felt able to build something, highlighting AI's potential to democratize tool creation across various roles.
closing that gap between intent and execution is what I've seen it do to people's ability to build or not, like builders
Software Engineering Transforms, Shifting Focus from Coding Craft to Product Vision with AI
AI is fundamentally changing software engineering, making it more accessible for individuals to build tools that previously required specialized engineers. This expansion enables more people to translate their ideas into real, functional software, fostering a broader range of creators.
The definition of software engineering has dramatically evolved from what it was just a few years ago. Traditionally, it involved deep dives into architecture, intricate coding, understanding specific ORM layers, and fixing post-deployment bugs. These detailed, hands-on tasks are now increasingly handled by AI.
Much of the traditional engineering work is collapsing into product management, blurring the lines between these roles. The emphasis shifts from manual coding and debugging to higher-level conceptualization, vision, and ensuring the product effectively meets user needs, making the PM and engineering split more diffuse.
While some may lament the diminishing 'craft' of coding—the satisfaction of elegantly solving complex technical problems—the overall field of software production remains vital. The focus now is on what needs exist and whether the developed solution is genuinely good, highlighting a more human-centric, collaborative endeavor.
I think software engineering is different. It has like dramatically changed.
Human ownership and meta-maintenance in AI-driven development at Anthropic
At Anthropic, human engineers still hold primary ownership of product intent and serve as Directly Responsible Individuals (DRIs), even with many "Clods" (AI assistants) working alongside them. This human oversight ensures that the actual goals and context of a product are maintained, especially when considering how different products might integrate in the future. This model emphasizes the value of focused human responsibility rather than a fully distributed approach.
Engineers are evolving their workflows to manage asynchronous delegation to AI. Many now create personalized dashboards to monitor the tasks their "Clods" are performing, identify pending requests, and track items needing their human attention after AI or human code reviews. This "meta-maintenance" of AI-driven work is a developing skill, with some aspects likely to standardize while others remain tailored to individual preferences.
A significant challenge for current AI models is truly understanding how code operates once deployed in production environments. Real-world incidents, such as unexpected network link cuts that aren't typical failure modes, present complex scenarios. Mastering these production realities was a major focus for companies like Instagram during its scaling years from 2012 to 2016, highlighting the depth of understanding required beyond just writing functional code.
Understanding what happens to code after it gets deployed, you know, because there's incidents, there's, you know, this was all working well, but like this network link got cut, which isn't in your usual failure mode.
Product Managers Use Engineering Prototypes to Resolve Product Debates
Prototypes are increasingly used to resolve disagreements in product development, a shift from the past where 'code wins arguments' often meant engineers held undue influence. Now, product managers are leveraging this power to demonstrate ideas.
Mike Krieger observes that when debates arise about product direction, PMs can now quickly develop and demonstrate how a concept could work. This empowers them to visually present ideas after trying them in various ways.
This approach fosters a new level of parallelism within product teams, allowing for more concurrent exploration of ideas. It also drives the need for higher-order abstractions, streamlining the development process.
The ability for non-engineers to rapidly prototype transforms how product ideas are validated, moving beyond theoretical discussions to concrete, demonstrable solutions that advance product decisions.
It's been really cool now where sometimes we will have some disagreement or some sort of debate about where to take a product, and often it's the PM that'll say, 'All right, I just tried it ... but look, it actually shows like how this could work.'
Fable 5's Pricing Focuses on Task Completion and User Satisfaction
Mike Krieger discusses Fable 5's cost, acknowledging that while it can be expensive, it offers significant value for the work it delivers. He recounts building a personal app over a weekend with Fable 5, noting that the extra usage cost was minimal, demonstrating its efficiency even for individual projects.
The approach to measuring AI cost is evolving beyond simple "per-turn costs." The industry is shifting focus to understanding the comprehensive expense of completing a task to a user's satisfaction, rather than just the raw usage metrics.
Fable 5 differentiates itself by its ability to execute tasks correctly on the first attempt. This capability significantly reduces the need for multiple follow-up interactions, as the model effectively "thinks through all the little details," saving users time and effort in clarifying or correcting outputs.
I think measuring cost has gotten so, multifaceted now because there's the per-turn costs, and then there's like, what did it cost you not to just do the task, but like complete the task to your satisfaction?
Fable's AI Learns System-Wide Understanding and Discerning Judgment
Fable's AI model has evolved beyond just executing individual tasks; it now possesses a comprehensive "sense of the system." This means it understands broader production needs, like required feature flags, and will actively remind users if these conditions aren't met, even across sessions spanning multiple days.
A key progression for the model is its ability to engage thoughtfully with code review feedback. Instead of a simple "yes, I'll fix it" response, it can exercise judgment. It might accept certain risks for the current fidelity level of a build or even push back against a reviewer's suggestion if it deems it incorrect.
This developed judgment, which allows the model to form its own opinions and rationale, is considered crucial. It transforms the product into a "living, breathing" entity, reflecting a deeper understanding and proactive engagement with the development process.
I see what you mean, but, like, I'm actually gonna push back. I don't, I think that that's actually not right.
Rethinking AI Interfaces for Fable: Mobile, Visual Context, and Collaborative Workflows
The traditional chat interface might not be the optimal way to interact with models like Fable, especially when delegating complex tasks rather than engaging in turn-by-turn conversation. One crucial area for evolution is decoupling the conversation from where the actual work happens. The speaker highlights the utility of mobile interfaces for this, referencing Boris's work with Cloud Code and the speaker's own experience with DevBoxes at Anthropic, where mobile access proved highly effective for keeping up with tasks.
Another limitation of current chat UIs is making the extensive context generated by AI models comprehensible. When models provide a lot of text, it can be overwhelming for users to fully grasp. There's a need to evolve the interface to offer progressive disclosure of complexity, potentially through visual aids like diagrams, to help users understand and digest the AI's proposals and decisions more effectively.
Finally, an underexplored frontier is the development of multiplayer workflows for AI. While sharing chats offers a basic level of collaboration, the future could involve more robust systems where multiple humans and AI agents (or 'clods') can independently contribute and keep up with shared work on a team. This could be particularly useful for incident response or projects with conjoining areas, abstracting the models to act as true teammates rather than just individual conversational partners.
I mean, number one is like decoupling the, the where the work is happening from where I'm talking to about the work.
Implementing Robust Verification Strategies for AI-Driven Development
Mike Krieger emphasizes the importance of tight development loops for AI-driven projects, drawing a parallel to his past experience at Instagram where dedicated build targets enabled rapid iteration on specific screens. He advocates for attaching visual evidence, such as photos or videos, to every pull request generated by AI models like Claude. This approach builds confidence and helps human reviewers quickly identify subtle issues that might otherwise be missed.
Effective verification extends to testing with real data and flows, not just static injections. This involves getting AI agents to log into staging environments with real accounts, simulating genuine user interactions. For complex systems, a robust strategy includes building comprehensive mock backends or in-memory implementations, similar to successful pre-LLM testing practices. Video is highlighted as an underexplored tool, enabling AI to detect subtle visual issues like animation jank that screenshots might miss.
AI agents like Claude are increasingly capable of not only identifying and fixing bugs but also engaging in follow-up conversations with engineers to explain trade-offs. A significant development is the AI's evolving "taste and discernment," allowing it to differentiate between immediate, temporary fixes (like a server restart) and more long-term, architectural solutions. This enables the AI to act akin to a seasoned SRE or engineer in prioritizing and managing production issues, even communicating fixes directly to customers.
Video is a very underexplored tool to give Claude as well.
Fable empowers non-technical people to build and iterate on complex bespoke software.
Fable enables individuals to express the unique complexity of their specialized knowledge through creative software applications. For example, environmental engineers can now build sophisticated PyTorch simulations and visualizations, turning their deep domain expertise into interactive games or detailed models that were previously inaccessible.
A key advantage of Fable is its ability to compose software and facilitate continuous iteration. Unlike older systems where increasing complexity often led to fragile code, Fable allows users to refine and expand their projects over months without fear of breaking the underlying structure. This drastically lowers the barrier to entry for developing and maintaining complex applications.
Fable significantly raises the 'ceiling of complexity' for non-technical users, allowing them to build solutions previously only achievable by experienced engineers. Its code generation capabilities are so advanced that they score highly on benchmarks, approaching the level of a senior engineer and allowing for indefinite project development.
The ceiling of complexity that a person that doesn't start out as technical can now build for solving problems within their domain is unprecedented.
Dynamic Workflows Enable Autonomous Codebase Conversion with AI
Mike Krieger champions dynamic workflows for deep, meaningful work, showcasing AI's ability to orchestrate complex, multi-step tasks over long horizons. He provided an example of this capability at Fable.
He described a scenario where a complex Python codebase was autonomously converted to TypeScript using Bun over a weekend. The dynamic workflow first performed a deep understanding of the code, then generated a specification, and subsequently orchestrated a series of conversion steps.
The outcome was a fully ported and documented codebase that was even better in certain aspects. Krieger emphasized that achieving this would have been impossible with previous models without the specific scaffolding provided by Orkplates.
Looking to the future, Krieger suggests that dynamic workflows will evolve to tune subtasks based on their complexity, potentially utilizing lower-thinking AI or smaller models for less demanding sections to optimize performance and resource use.
Yeah, this thing is a like Typescript and Bun port of that thing, and it's actually better in these ways.
Follow the shows you care about.
Podbrew watches new episodes and turns them into concise briefs you can read in minutes.