Mike Krieger: Product Building Lessons from Instagram and Anthropic

Welcome to Generative Now with Michael Mignano, partner at Lightseed, featuring a conversation with Mike Krieger, Anthropic's Chief Product Officer and former co-founder and CTO of Instagram. This episode revisits their discussion about Mike's impressive journey in tech and Silicon Valley.

Mike and Michael explore his transition to Anthropic, the valuable lessons he's carried forward from Instagram, and his strategic approach to differentiating Claude from competitors in the AI landscape.

Timestamp: [0:02-0:42]

Mike's transition from Instagram to Anthropic represents a fascinating shift from consumer app building to AI research and development. After only two months of semi-retirement following his previous company Artifact, Mike found himself eager to return to building products and teams.

Unlike his post-Instagram break where he needed rest after an intense eight-year journey, Mike felt energized and ready for another swing at building something significant. He discovered he missed working with larger teams on multiple parallel initiatives - something he experienced at Instagram but not at Artifact, which peaked at just 13 people.

"I'm a builder. I love building. I love building both products and teams." - Mike Krieger

The opportunity at Anthropic presented a unique confluence of factors: a world-class research team, early-stage product development, and the chance to work on zero-to-one product building within an existing company with momentum and aligned culture. For Mike, it felt like it was "meant to be."

Timestamp: [0:55-3:21]

Building product strategy at Anthropic presents unique challenges compared to traditional consumer apps like Instagram or Artifact. Mike works closely with researcher-founders, developing strategy around constant model and research innovations rather than inventing entirely new social experiences or UX paradigms.

The closest parallel Mike draws from his Instagram days is the yearly drops at Google I/O or WWDC, where new capabilities would be delivered from Cupertino. At Anthropic, this happens on a monthly basis, creating dynamic reprioritization challenges.

"I like to joke with the product team that if we froze our researchers - we love them and they shouldn't be cryogenically frozen - but if we did, we should still have like a year of roadmap ahead." - Mike Krieger

Timeline unpredictability works in both directions. Computer use, for example, accelerated from "probably early next year" to launch-ready in just 3 weeks once it passed safety testing. This requires a portfolio approach with three parallel tracks: labs paired with early research, model-dependent features requiring research collaboration, and traditional product work that uses models without requiring breakthroughs.

Timestamp: [3:22-6:27]

The relationship between research breakthroughs and product innovations at Anthropic follows a dynamic loop rather than a single direction. Model improvements unlock new capability levels, which then enable tight collaboration loops between product vision and model fine-tuning.

Mike uses agentic capabilities as an example: Claude 3.5 Sonnet's refresh unlocked coding startups to build interesting agentic coding tools, while internally enabling improvements to Artifacts through research collaboration. Step changes typically happen at model release time, followed by incremental improvements through prompting and model customization.

The research team maintains a 3-6-9 month roadmap for major deliveries, while the product team can execute tight loops on top of each milestone. This creates a balanced approach where both research discoveries and product vision drive innovation forward.

Timestamp: [6:27-7:42]

Building products at Anthropic requires balancing rapid iteration with safety considerations - a significant departure from Meta's famous "ship fast, move fast, break things" culture. Mike identifies two key constraints that reshape experimentation approaches.

For model releases, safety testing and responsible scaling evaluations create natural deliberation periods. However, since model releases don't happen at breakneck speed anyway, this doesn't drastically impact iteration cycles.

More interesting is how users develop deep relationships with AI models. People become attached to how models react and communicate with them, often detecting changes that don't even exist. This creates a higher bar for experimentation compared to Instagram, where most usage was entertainment-focused rather than work-critical.

"People become very attached and attuned to how the model reacts to them and like how it talks to them." - Mike Krieger

Since Claude users rely on the platform for actual work completion, changes that break established workflows cause real disruption. Mike is still finding the balance between Instagram's 50+ concurrent A/B tests and the extreme of annual software releases, likely favoring UI experimentation over frequent model changes.

Timestamp: [7:43-10:52]

• Mike Krieger's transition to Anthropic was driven by his desire to build products and teams at scale while working on cutting-edge AI technology
• Product strategy at AI companies requires managing unpredictable research timelines through portfolio approaches with multiple parallel tracks
• Innovation at Anthropic flows in both directions - research breakthroughs enable product possibilities, while product vision drives model customization
• Safety considerations and user attachment to AI models create unique constraints for rapid iteration compared to traditional consumer apps
• Users develop deep relationships with AI models, making them sensitive to even minor changes in behavior or personality
• The challenge is finding the right balance between rapid iteration and maintaining stable user workflows for work-critical applications

Timestamp: [0:02-10:52]

Companies/Products:

• Instagram - Mike's previous company where he was co-founder and CTO for eight years
• Artifact - Mike's second company that peaked at 13 people but didn't achieve product market fit
• Meta - Parent company of Instagram, known for "ship fast, move fast, break things" culture
• Anthropic - AI research company where Mike now serves as Chief Product Officer
• Claude - Anthropic's AI assistant product
• Google I/O - Annual developer conference mentioned as parallel for capability drops
• WWDC - Apple's Worldwide Developers Conference referenced for yearly capability announcements

Technologies/Features:

• Computer Use - Anthropic's agentic capability that accelerated from early next year to 3-week launch
• Claude 3.5 Sonnet - Model that unlocked agentic coding capabilities for startups
• Artifacts - Claude feature that required tight research-product collaboration loops

Platforms:

• Reddit - Platform where Claude users discuss and detect model changes
• Cupertino - Reference to Apple's headquarters and capability announcements

Timestamp: [0:02-10:52]

The AI landscape is evolving rapidly with models from different labs leapfrogging each other at breakneck speed. Mike draws parallels to Instagram's competition with Snapchat, using his "non-mathematical formula" for what makes social networks distinct: formats, audience, and vibe.

While formats can be easily copied between platforms, vibe creates lasting differentiation. Instagram started from a curated, polished aesthetic, while Snapchat had a completely different feel - even when both platforms adopted similar features like Stories, they maintained distinct personalities.

"The same people would actually use both, but the vibes could not be more different between the two." - Mike Krieger

Mike sees similar dynamics emerging in AI models. People describe Claude as their therapist, trusted friend, or decision-making advisor. Claude's personality shines through in moments like urgently telling someone "go to the hospital now" in all caps when their friend had a medical emergency.

Timestamp: [10:59-13:18]

Product differentiation in AI will likely follow two key paths: vibe and capability specialization. While short-term product features get borrowed quickly across platforms - a healthy dynamic for the ecosystem - deeper differentiation emerges over time.

Vibe differentiation occurs as users gravitate toward models that match their preferences, naturally pushing products in distinct directions. Each model develops its own personality and communication style that resonates with different user bases.

Capability differentiation means different models will spike on different strengths. Currently, Claude 3.5 Sonnet excels at coding and creative writing, while other models may dominate in different areas. These differences will likely become more pronounced rather than converging to similarity.

"I think what is different over time is probably two things in this case: one is actually vibe, which is like what does it feel like to use a Claude versus a Gemini versus a ChatGPT versus a DeepSeek." - Mike Krieger

The disconnect between standardized evaluations and real user value will likely widen over time. While evals remain crucial for training, they don't capture the full picture of what makes models valuable to individuals and companies.

Timestamp: [13:18-15:30]

Anthropic's focus on presentation layer innovations like Artifacts, Projects, and Computer Use represents a strategic approach to unlocking novel user behaviors. Mike emphasizes that the AI product journey is only "1% finished" - possibly even less - making product primitives crucial for enabling new capabilities.

Projects exemplifies this philosophy. While users could achieve similar results through regular chats or external tools, having a dedicated container unlocks specific behaviors. Users create dedicated spaces for health records, tracking medical information, or organizing complex workflows.

"Hey I have a project for all my health records. It's like one place where I know to go put every single thing." - Mike Krieger

The next phase involves transforming expert use cases into accessible features for broader audiences. This means taking sophisticated prompting techniques that work well for power users and building product surfaces that make those capabilities available to everyone.

Current suggestion chips represent early attempts at lowering barriers, but they often lack the context needed for truly satisfying answers. The focus must shift from entertainment use cases like jokes and poems to tools that connect to relevant data sources and produce actionable work output.

Timestamp: [15:41-17:56]

• AI model differentiation will increasingly depend on "vibe" (personality and feel) and specialized capabilities rather than just feature parity
• Like social media platforms, AI models can have similar formats but completely different user experiences and emotional connections
• Product differentiation in AI is currently short-lived as features get borrowed quickly, but deeper personality differences will emerge over time
• The AI product journey is only 1% complete, making product primitives crucial for unlocking novel user behaviors
• The next challenge is democratizing expert use cases by building accessible interfaces that bring advanced capabilities to broader audiences
• Current evaluation metrics don't capture the full picture of what makes AI models valuable to real users and companies
• Success requires connecting AI to relevant data sources and producing actionable work output, not just entertainment features

Timestamp: [10:59-17:56]

Companies/Products:

• DeepSeek - AI model mentioned as having different approaches to showing reasoning
• Snapchat - Instagram's primary competitor, used as example for platform differentiation
• Gemini - Google's AI model mentioned for comparison and UI evolution
• ChatGPT - OpenAI's model referenced for capability and vibe comparisons

Technologies/Features:

• Artifacts - Claude's product feature for code and content creation
• Canvas - Product interface evolution mentioned alongside Artifacts
• Projects - Claude's organizational feature for containing related conversations and documents
• Computer Use - Anthropic's agentic capability for computer interaction
• Stories - Social media format that moved between Instagram and Snapchat

Concepts:

• Evals - Model evaluation metrics important for training but limited for measuring user value
• Suggestion chips - UI elements that prompt users with model interaction ideas

Timestamp: [10:59-17:56]

The rise of AI agents is fundamentally shifting how the internet and advertising-based business models operate. Mike explores two potential scenarios: AI tools could save people significant time, potentially increasing entertainment consumption and making ads more valuable. Alternatively, if agents become primary web users instead of humans, the entire content ecosystem faces disruption.

Mike draws from his experience with Artifact, an AI-powered news recommendation app, which struggled partly due to poor content quality behind clicks. Local news organizations, hollowed out economically, resort to aggressive ad strategies not by choice but for survival.

"I don't think anybody at these local news wakes up and is like how do I put more ads in here. They're like we got to do it to stay afloat." - Mike Krieger

The emergence of web-browsing agents and summarization tools like Perplexity means users increasingly bypass original content entirely. This necessitates fundamental shifts in business models and the relationship between writers, publishers, and readers.

Timestamp: [18:02-20:14]

AI can create more symbiotic relationships with content creators rather than purely extractive ones. Mike highlights Kora, an email product by the folks at Every, which summarizes newsletters in his inbox. Rather than reducing engagement, the summaries help him identify which content is worth his time, actually increasing his click-through rates.

"I'm finding I'm reading more because I'm like oh that actually sounds great. I actually am going to click through and I'm confident that it's gonna probably be worth my time." - Mike Krieger

This represents a model where AI serves as a better conduit to quality content rather than replacing it entirely. The challenge is avoiding a world where we become completely detached from the writers behind the content.

Mike also discusses evolving publisher relationships, citing his subscription to The Verge after engaging with their newsletter, and reading Defector for sports news. These direct relationships may become increasingly important as AI intermediates more content consumption.

Timestamp: [20:31-22:27]

Reflecting on Artifact's journey, Mike acknowledges how dramatically the landscape changed during its brief existence. The company started in 2021 and shut down recently, but the final version had incorporated LLM APIs for summarization, clickbait rewriting, and other AI enhancements - capabilities that weren't available at launch.

"If you had started Artifact in 2024 rather than in 2021 like we did... you'd have a lot more around like AI aggregation and you'd have even the ability for AI to read you like the top news of the day." - Mike Krieger

The rapid pace of AI advancement raises fascinating questions about product strategy. Capabilities that previously required years to build or weren't achievable without massive research teams are now available and getting cheaper monthly. This forces fundamental reconsideration of product assumptions.

Mike mentions experimenting with Particle News, another AI-powered news approach, and describes an intriguing Artifact prototype that focused on conversing with their three-year corpus of news data rather than traditional feed browsing.

Timestamp: [22:27-24:35]

• AI agents accessing the web will fundamentally disrupt advertising-based business models, forcing shifts in how content creators monetize
• Local news organizations use aggressive ad strategies out of economic necessity, not choice, highlighting systemic industry challenges
• AI can create symbiotic relationships with content by serving as better conduits rather than pure replacements for human consumption
• The pace of AI advancement means product assumptions must be constantly questioned - capabilities impossible in 2021 became trivial by 2024
• Direct relationships between readers and publishers may become increasingly important as AI intermediates more content discovery
• Email has emerged as a significant content consumption channel where AI summarization can enhance rather than replace engagement
• Future product strategies must account for rapidly evolving AI capabilities that get cheaper and more accessible monthly

Timestamp: [18:02-24:35]

Companies/Products:

• Artifact - Mike's AI-powered news recommendation app for iOS and Android that was eventually sold to Yahoo
• Perplexity - AI search and summarization tool mentioned as example of content bypass
• Kora - Email product built by the folks at Every that summarizes newsletters
• Every - Company behind the Kora email product
• The Verge - Technology publication Mike subscribes to and reads regularly
• Defector - Sports news publication Mike reads
• Yahoo - Company that acquired Artifact
• Particle News - AI-powered news approach Mike has been observing

Technologies/Concepts:

• LLM APIs - Technology Artifact incorporated for summarization and content enhancement
• CDN layer - Network infrastructure level where resistance to AI agents might occur
• Clickbait rewriting - AI capability for improving content quality
• AI aggregation - Approach to news consumption using artificial intelligence

Timestamp: [18:02-24:35]

Building for both enterprise and consumer customers simultaneously creates unique dynamics compared to Mike's Instagram experience. While many people sign up for Claude's free and pro subscriptions individually, enterprise customers provide different challenges and opportunities.

In enterprise settings, the buyer (often CIO or CTO) differs from the end user, but this creates richer feedback opportunities. Unlike Instagram's aggregate data analysis across the entire United States, Anthropic can send teams directly to enterprise deployments for troubleshooting and learning.

"What's really interesting is actually the opportunity for feedback and engagement is much greater because you're really mutually invested in the success of that partnership." - Mike Krieger

The enterprise model starts from a different foundation - companies pay for employee access, requiring clear value demonstration to prevent scenarios where subscriptions exist but see minimal usage. This shifts focus toward user education and "enablement" - enterprise terminology Mike had to learn quickly.

Timestamp: [24:40-26:52]

Product development at Anthropic flows in both directions between enterprise and consumer use cases. Knowledge integrations like Google Drive, Google Suite, and GitHub started as enterprise-first features but benefit individual users equally.

Some features like Projects and Artifacts began as general-purpose tools and later received enterprise customization. However, knowledge integrations represent the key to success - connecting Claude to company data to enable real work completion benefits both enterprise and individual users.

"How we succeed is like can we connect Claude to the knowledge at your company and actually help you do real work. That's a company problem, that's an enterprise problem, but individuals are trying to do that as well." - Mike Krieger

An active strategy question involves whether enterprise and consumer versions will diverge over time or remain similar. Mike expects them to stay fairly similar for the next six months, with potential divergence around the types of work users delegate to Claude based on their specific roles and contexts.

Timestamp: [26:52-28:55]

Mike discusses the untapped potential for AI integration in personal life management. Despite having significant productivity applications in work settings, AI hasn't yet meaningfully penetrated personal and home life organization.

A founder of a successful productivity company shared a workflow where he ingests dense school emails from his children's schools and uses Claude to produce actionable task lists. This inspired Mike to consider more AI applications for family life management.

"I should use AI for more things in my personal life... around family life management as well." - Mike Krieger

The conversation reveals a gap in current AI product development - while work-oriented delegation feels natural (productivity tasks, computer use for business), consumer applications haven't reached the same level of integration. Personal life organization presents significant opportunity, particularly for parents managing multiple schools, activities, and family logistics.

Both Mike and the interviewer share similar situations with two young children attending different schools, highlighting how common these organizational challenges are among working parents. The expectation is that Google and other companies will soon provide better AI-powered personal productivity tools.

Timestamp: [28:55-30:10]

• Enterprise customers provide richer feedback opportunities than consumer products due to mutual investment in partnership success
• The buyer-user disconnect in enterprise sales requires different approaches to proving value and ensuring adoption
• Product development flows bidirectionally between enterprise and consumer use cases, with knowledge integrations benefiting both
• Success depends on connecting AI to relevant data sources to enable real work completion across all user types
• Enterprise and consumer versions may diverge based on delegation patterns rather than core functionality differences
• AI integration in personal life management remains underdeveloped compared to work productivity applications
• Parents managing multiple schools and family logistics represent a significant untapped market for AI productivity tools
• Work delegation feels natural for AI while personal life integration requires different product approaches

Timestamp: [24:40-30:10]

Companies/Products:

• Instagram - Mike's previous company, referenced for comparison to enterprise product building
• Google Drive - Knowledge integration platform built for enterprise customers
• Google Suite - Productivity platform integrated with Claude for enterprise use
• GitHub - Code repository platform with Claude integration
• Gmail - Email platform mentioned for potential personal AI integration
• Google - Company expected to develop personal AI productivity tools

Concepts:

• CIO/CTO - Enterprise decision makers who purchase software for employee use
• Enablement - Enterprise term for user education and adoption strategies
• Knowledge integrations - Connections between Claude and company data sources
• Virtual collaborator - Concept for how Claude might automate workflow parts
• Computer use - Feature mentioned for productivity and work delegation

Timestamp: [24:40-30:10]

Anthropic open-sourced the Model Context Protocol (MCP) in October/November as their approach to bringing data into LLMs and extracting data from them. The decision to make MCP open source was strategic - fostering an ecosystem where developers could build on top of it across multiple platforms, not just Claude.

The protocol has gained adoption beyond Claude, with several editors integrating MCP support. Block recently open-sourced their agentic coding tool called Goose, which includes MCP components, demonstrating the protocol's broader utility.

"We made it open source because we wanted to sort of foster an open ecosystem of people building on top of it." - Mike Krieger

The most interesting development is seeing what MCP servers people create. Claude for desktop allows users to write and integrate custom MCP servers, and the community has built connections to personal productivity tools like Apple Notes, Google Calendar, and Apple Reminders - applications that are inherently personal rather than purely professional.

Timestamp: [30:16-31:04]

AI coding capabilities will democratize software creation far beyond traditional software engineering. While people initially think of AI helping engineers write code, Mike envisions AI helping everyone solve problems that can uniquely be solved through coding.

Today, this might mean writing MCP servers or connecting to existing tools like Apple Reminders. In the future, it could involve AI writing code for data analysis tasks and delivering results directly to users, even if they're not producing traditional software.

"It's going to help everybody solve problems that can be uniquely solved by writing code." - Mike Krieger

Mike shares an inspiring example from Anthropic's go-to-market team: non-engineers with no coding background used tools like v0 by Vercel during holiday break to prototype ideas for Claude. They returned saying they'd built fully functioning web applications despite never writing code before.

This represents a fascinating intersection where consumer behavior meets software creation. The challenge involves helping people translate ideas in their heads into visual expressions, potentially through AI assistance and media creation tools that spark further human creativity.

Timestamp: [31:04-33:08]

Mike extensively uses Claude in his CPO role, particularly as a critic and sounding board. Rather than expecting flawless strategy generation, he leverages Claude's ability to identify gaps in reasoning and provide external perspective on product requirements documents and strategy papers.

"What am I not thinking about? What did I forget? What are the holes in this argument?" - Mike Krieger

A concrete example involved Anthropic's leadership team OKR process, where Mike fed their objectives to Claude asking what they were missing. Claude provided three valuable insights that improved their strategic planning.

Claude also serves as an accelerant for routine tasks. Using Google Docs integration, Mike started filling out a table about how Claude could accelerate product development, completed half the cells manually, then had Claude finish the rest. Not only did Claude complete the task efficiently, but it often generated better ideas than Mike's original concepts.

The newest knowledge integrations help with daily preparation, allowing Mike to ask Claude to review briefing documents, prepare for meetings, and get into the right mindset for upcoming conversations.

Timestamp: [33:09-35:31]

Despite building AI products, Anthropic discovered their internal Claude usage was "very unevenly distributed" across the organization. This realization prompted Mike to focus more attention on internal products and productivity to learn from successfully implementing AI in workplace scenarios.

A November hackathon produced compelling results when a go-to-market team member partnered with an engineer to automate their prospect research workflow. The process involved researching meeting attendees, gathering company context, and reading S1 filings - tasks that Claude handles well through summarization.

"We need to do those kinds of things to take the seat of what is a good use of LLMs in the workplace and then make it so that you don't have to think about it." - Mike Krieger

This internal focus represents a strategic shift for Mike, who believes they can learn valuable lessons by making their own sales and engineering teams more productive with cutting-edge Claude capabilities, then externalize those insights into customer-facing products.

The approach treats internal productivity as a testing ground for identifying successful AI workplace applications before building them into formal product offerings.

Timestamp: [35:31-37:10]

• Open-sourcing MCP has successfully created an ecosystem beyond Claude, with community-built integrations focusing heavily on personal productivity tools
• AI coding capabilities will democratize software creation, enabling non-technical users to solve problems through code without traditional programming knowledge
• AI serves best as a critic and accelerant rather than a complete replacement for human strategic thinking in product development
• Internal AI adoption at AI companies can be surprisingly uneven, requiring intentional focus on workplace productivity optimization
• The most successful AI workplace applications involve automating repetitive research and preparation tasks rather than core creative work
• Testing AI productivity improvements internally provides valuable insights for external product development
• Non-engineers are already successfully building functional web applications using AI-assisted tools, indicating rapid democratization of software creation

Timestamp: [30:16-37:10]

Technologies/Protocols:

• Model Context Protocol (MCP) - Anthropic's open-source protocol for bringing data into LLMs and extracting data from them
• Claude for desktop - Platform that allows users to write and integrate custom MCP servers
• Goose - Block's open-source agentic coding tool that includes MCP components
• v0 by Vercel - Tool used by non-engineers to prototype web applications

Companies/Products:

• Block - Company that open-sourced the Goose coding tool
• Vercel - Company behind the v0 prototyping tool
• Apple Notes - Personal productivity app with MCP server integration
• Google Calendar - Calendar application connected through MCP servers
• Apple Reminders - Task management app mentioned for MCP integration
• Google Docs - Document platform with Claude integration used for collaborative work

Concepts:

• OKR process - Objectives and Key Results planning methodology used by Anthropic's leadership
• S1 filings - SEC documents used for company research in sales processes
• MCP servers - Custom integrations built by users to connect Claude with various applications

Timestamp: [30:16-37:10]

Scaling teams at Anthropic presents unique challenges compared to Instagram's hyper-growth period. While Instagram doubled year-over-year, most growth happened in Menlo Park post-acquisition, allowing visual confirmation of team expansion. Anthropic's distributed team makes growth less tangible when it occurs through Google Meet and Slack participants.

The remote nature changes team dynamics significantly. At Instagram, Mike and Kevin held product reviews that became presentations rather than nuanced discussions when too many people attended. At Anthropic, with many people joining from New York or remotely, rooms can actually feel smaller despite company growth.

"It's much harder to see that when it's primarily growing in Google Meet Hollywood Square style or Slack participants." - Mike Krieger

Another key difference is Anthropic's organizational structure. Despite significant product team growth, it represents a small portion compared to research teams, trust and safety, threat modeling, and societal impact groups. This creates a different feeling of scale and responsibility within the broader company mission.

Timestamp: [37:16-39:16]

Mike applies Instagram lessons about maintaining focus on fundamental problems, though it's significantly harder in AI. At Instagram, people generated fantastic ideas that weren't aligned with serving the most people possible. In AI, the temptation multiplies because there's no shortage of cool prototypes possible with Claude.

"If we staff teams on all of them, we'd have a pretty complex product, but we wouldn't make enough progress because we'd be peanut buttering ourselves over a lot of things." - Mike Krieger

The discipline of saying no becomes even more critical but twice as hard in AI. Mike acknowledges he won't have all the best ideas - they'll come from the team through bottom-up excitement. However, maintaining user focus and adherence to fundamental problem-solving remains essential.

The challenge involves harnessing team creativity while avoiding the "peanut butter" effect of spreading resources too thin across interesting but unfocused initiatives. This requires building organizational discipline that Anthropic hasn't perfected yet.

Timestamp: [39:22-40:43]

Mike learned at Instagram that every time a company doubles, processes and culture break in some way. Instagram doubled annually in engineering, requiring constant evolution and "unbreaking" of systems. Anthropic has grown even faster since Mike joined, despite starting from a smaller product base.

"Every time your company doubles, your processes will break and your culture will break in some way." - Mike Krieger

Currently, Anthropic feels one organizational and process refactor away from "things really humming." Mike experiences constant variability where some aspects feel good while others need attention.

The most important lesson from Instagram: no amount of processes and organizational structure beats getting on the ground to do work yourself or maintaining high-bandwidth communication with individual contributors. At Instagram, Mike remained an IC engineer until the year he left, providing crucial tactical understanding.

Maintaining ground-level contact is more challenging as a non-co-founder. At Instagram, Mike knew everyone because he started there and hired much of the team. At Anthropic, he's still building relationships and doesn't have the same tactile feeling for the codebase and daily building experience.

Timestamp: [40:43-42:11]

• Distributed team scaling presents unique challenges in perceiving growth and maintaining team dynamics compared to co-located expansion
• AI product development amplifies the difficulty of maintaining focus due to endless possibilities for cool but unfocused prototypes
• The discipline of saying no becomes more critical but harder to execute when building AI products with broad creative potential
• Organizational processes and culture inevitably break as teams double in size, requiring constant evolution and "unbreaking"
• Ground-level contact through IC work or high-bandwidth communication with individual contributors remains essential for effective leadership
• Non-founder executives face additional challenges in building relationships and understanding tactical realities compared to company founders
• Product teams represent a smaller portion of AI companies due to extensive research, safety, and impact assessment requirements
• Bottom-up innovation must be balanced with user focus and adherence to fundamental problem-solving objectives

Timestamp: [37:16-42:11]

Companies/Locations:

• Instagram - Mike's previous company, used for scaling comparisons
• Facebook - Instagram's acquirer, mentioned for team collaboration context
• Menlo Park - Location where Instagram team growth was concentrated post-acquisition
• New York - Location mentioned for Anthropic's distributed team members

Concepts/People:

• Kevin - Kevin Systrom, Instagram co-founder who conducted product reviews with Mike
• Product reviews - Meeting format used at Instagram for team discussions
• Trust and safety - Anthropic team focused on AI safety considerations
• Threat modeling - Process for identifying AI-related risks and societal impacts
• IC (Individual Contributor) - Technical role without management responsibilities
• Peanut buttering - Business term for spreading resources too thin across initiatives

Technologies/Platforms:

• Google Meet - Video conferencing platform used for distributed meetings
• Slack - Communication platform mentioned for team coordination
• Hollywood Square style - Reference to grid-based video meeting format

Timestamp: [37:16-42:11]

The DeepSeek moment represented a significant cultural crossover for AI, with Mike's wife texting about it and his parents asking questions - indicating broader societal awareness beyond tech circles. Mike identifies three key takeaways from this event.

First, it opened people's eyes to alternatives beyond ChatGPT. While Claude has strong business penetration and API usage, there's still work to do in getting broader awareness and letting people experience and choose for themselves.

Second, Mike draws parallels to market overreactions he's observed over the past decade, referencing Bloomberg writer John Authers' perspective on market swoons. The short-term ability to overcorrect and snowball is real, but stepping back reveals less dramatic actual changes.

"If you had fallen asleep two weeks ago and woken up today... the market moved like half a percent, not that much." - Mike Krieger (referencing John Authers)

The fundamental need for compute remains stronger than ever, especially with reinforcement learning scaling opportunities. As Dario Amodei noted, having more valuable RL scaling should incentivize training for even longer periods - potentially five times longer for five times the intelligence.

Third, DeepSeek brought geopolitical implications to the forefront in an unignorable way, sparking necessary and timely conversations about AI development and global competition.

Timestamp: [42:18-44:58]

Mike emphasizes that enterprise AI adoption isn't about buying models or exchanging tokens - it's about finding AI partners for internal transformation and product development. Speaking at a CEO summit, he found enterprise conversations focused on comprehensive partnerships rather than technical specifications.

"They're like 'we want an AI partner that will help us co-design our big bet on our internal transformation.'" - Mike Krieger

Enterprise customers seek partners who will help design transformation strategies, assist with AI-focused product development, and provide access to communities of excellence. This relationship dynamic means the list of viable AI partners remains quite small, regardless of pure model capabilities.

The past year at Anthropic has focused on building capabilities beyond "tokens in, tokens out" - developing infrastructure, support systems, and partnership frameworks that matter as much as model quality for enterprise success.

Mike notes the reductionist thinking of wiping trillion-dollar valuations based solely on new models, when companies represent much more than their core technology. The application layer and partnership capabilities become increasingly critical differentiators.

Timestamp: [44:58-47:33]

The spotlight on applications and products represents a pivotal moment for AI development. As models become more intelligent and capable of extended reasoning, they'll excel at tasks increasingly disconnected from typical user needs - creating interesting product challenges.

The unsolved problem involves finding real applications for models that use extensive test-time compute. Most people can't articulate how they use reasoning models in daily life beyond asking them to "think about something for a long time."

"Nobody has an answer that's really solved yet... ask it to think about something for a long time and it... that's an unsolved problem." - Mike Krieger

The second frontier involves models playing longer roles in people's lives through agentic action, reflection, and autonomous operation. Computer use and similar capabilities represent entirely new fronts for consumer and business products that will emerge over the coming months.

This application layer focus mirrors the social media era's product innovation cycle, where breakthrough capabilities needed thoughtful product design to reach mainstream adoption.

Timestamp: [47:33-49:00]

Mike sees clear parallels between the social media explosion and current AI development. Both eras feature Cambrian explosions of companies - whether building their own models or AI-focused software. Incubators and accelerators now focus heavily on AI, mirroring the social media wave.

The pattern includes breakthrough companies becoming self-sustaining long-term players, while others get absorbed, consolidated, or see their ideas integrated elsewhere. The question "who's going to be the Instagram of video" found answers in both Instagram itself and emerging players like Snapchat.

"What's who's going to be the Instagram of video and it turns out Instagram being the Instagram of video but also like Snapchat emerged." - Mike Krieger

Cultural breakthrough moments provide another parallel. Instagram's inclusion in hip-hop songs or movie trailers marked transitions beyond early adopters to mainstream culture. DeepSeek achieved similar name recognition without corresponding usage penetration - people aren't comparing chain-of-thought reasoning between models in everyday conversation.

Mike expects the AI journey to unfold faster than social media's eight-year transformation. Already one year into Anthropic, he's experienced four growth moments, suggesting accelerated timeline for mainstream adoption and cultural integration.

Timestamp: [49:00-51:48]

Looking toward AGI/ASI development over the next decade, Mike envisions Anthropic helping people become their maximal versions of themselves. For creative individuals, this means removing impediments to realizing ideas beyond personal creativity, time, and effort.

Mike references Dario Amodei's essay "Machines of Loving Grace" as both manifesto and roadmap, joking with Dario about its practical implications. The vision encompasses helping with life sciences, civic society, and economic prosperity - goals that extend far beyond current capabilities.

"Have we helped people be their maximal versions of themselves?" - Mike Krieger

The greatest challenge involves translating model capabilities into actual societal impact. Models won't magically create change - they need thoughtful products, user adoption, appropriate context, and proper guardrails to succeed.

Mike learned the concept of societal "viscosity" - the resistance that slows idea penetration and adoption. Overcoming this viscosity to achieve more even distribution of AI benefits requires good design, good products, and good building practices.

"It's not just the future unevenly distributed, but actually evenly distributed or more evenly distributed I think is through good design and good product and good building." - Mike Krieger

Timestamp: [51:48-53:47]

Podcast Information:

• Rate and review Generative Now on Spotify, Apple Podcasts, and YouTube
• Subscribe to support the show's growth and reach
• Follow Lightseed at @LightseedVP on X, YouTube, and LinkedIn for more content

Production Credits:

• Generative Now is produced by Lightseed in partnership with Pod People
• Hosted by Michael Mignano, partner at Lightseed
• Weekly episodes featuring conversations with AI industry leaders

Timestamp: [53:47-54:19]

• Cultural crossover moments for AI technology create broader awareness but don't always translate to increased usage or understanding
• Market overreactions to AI developments often obscure the underlying continuity of fundamental technological and business needs
• Enterprise AI adoption focuses on comprehensive partnerships rather than pure model capabilities or technical specifications
• The application layer represents the next major frontier as reasoning models become more powerful but less aligned with everyday user needs
• Social media and AI eras share similar patterns of company emergence, consolidation, and cultural breakthrough moments
• AI development timelines appear accelerated compared to previous technology waves like social media
• Translating AI capabilities into societal impact requires intentional product design, not just technological advancement
• Societal "viscosity" creates resistance to new ideas that must be overcome through thoughtful building and design practices
• The future success of AI companies depends as much on partnership infrastructure and support systems as on core model quality

Timestamp: [42:18-54:19]

People:

• John Authers - Bloomberg finance writer who wrote about market overreactions and perspective
• Dario Amodei - Anthropic CEO who made points about RL scaling and wrote "Machines of Loving Grace"
• Mike's wife - Referenced for texting about DeepSeek, indicating cultural crossover
• Mike's parents - Mentioned asking about DeepSeek, showing broader awareness

Companies/Platforms:

• DeepSeek - AI model that created significant market discussion and cultural awareness
• ChatGPT - Referenced as the dominant AI model that people were previously aware of
• Nvidia - Chip company mentioned for market valuation impact
• Bloomberg - Financial publication where John Authers writes
• Snapchat - Referenced as parallel to AI company emergence
• CEO Summit - Event where Mike spoke about enterprise AI impact

Concepts:

• Reinforcement Learning (RL) - AI training method mentioned for scaling opportunities
• Test-time compute - Computing resources used during model reasoning
• Machines of Loving Grace - Dario Amodei's essay about AI's potential societal impact
• Viscosity - Societal resistance concept that Mike learned about idea penetration
• AGI/ASI - Artificial General Intelligence/Artificial Super Intelligence
• Model distillation - Process of creating smaller models from larger ones
• Chain-of-thought reasoning - AI reasoning process mentioned in context of user understanding

Timestamp: [42:18-54:19]