Introduction: Amid the continuous leap in large model capabilities and the rapid proliferation of AI programming tools, industry discourse is shifting from "Can the model complete the task?" to "How can model capabilities be organized into products, workflows, and commercial systems?"

Over the past year, products like Claude Code, Codex, and Co-work have entered developer and knowledge worker environments. AI is no longer just a chat interface answering questions—it has evolved into a production-grade interface capable of invoking tools, executing tasks, and validating outcomes. But as the consensus grows that "agents will become the next generation of software," a more critical question emerges: Who can first transform model capabilities into reusable, distributable, and scalable work systems?

This article is adapted from an interview with Mike Krieger on ACCESS Podcast. Mike Krieger is co-founder of Instagram and currently serves as Chief Product Officer at Anthropic, leading Anthropic Labs—responsible for guiding the team in exploring the next frontier of products following Claude Code.

In this conversation, Mike Krieger doesn't merely discuss what Anthropic's next product will be; instead, he dissects AI product competition into a set of deeper structural questions: How do model capabilities enter real-world workflows? How should AI companies organize innovation internally? How should platform companies manage boundaries with ecosystem clients? And as AI execution power grows stronger, where does human judgment fit within the production chain?

First, product form is shifting from "chat" to "task." Previously, large models primarily existed as dialogue interfaces—users input prompts, models generate responses. Now, Claude Code, Co-work, and Claude Design represent a different product logic: enabling AI to persistently work toward a goal, calling tools, generating results, and verifying outputs along the way. This means the key to AI products is no longer just answer quality, but task decomposition, contextual continuity, tool invocation, and result validation. Whoever can encapsulate these abilities into seamless workflows will be closer to the next productivity gateway.

Second, organizational methods are shifting from "large-team planning" to "small-team experimentation." The operation of Anthropic Labs resembles an embedded startup unit within a larger corporation: two or three people starting out, with biweekly reviews to assess whether projects should continue. In the past, corporate innovation labs often fell into long cycles, ambiguous responsibilities, and delays caused by "we could still improve it." Now, models have reduced build costs, making judgment, taste, and decision speed the true scarce resources. Thus, organizational efficiency in the AI era depends not only on engineering headcount but on the ability to validate directions faster with smaller teams.

Third, the boundary between platforms and applications is being redrawn. Claude Code’s success has transformed Anthropic from a mere model supplier into a direct shaper of application form. The controversy between Claude Design and Figma illustrates how model companies entering app development inevitably collide with customer and ecosystem partner interests. Previously, foundational model companies provided underlying capabilities, while vertical applications like Cursor and Figma handled UI and scenario encapsulation. Now, model companies must also showcase agent-first future forms through their own products. This implies that AI platform competition is no longer just about API battles—it’s a competition over product paradigms.

Fourth, the stronger AI becomes, the rarer human judgment grows. Mike repeatedly emphasizes that while Claude can write code faster, prototype quicker, and execute tasks more efficiently, it cannot replace the most difficult part of going from 0 to 1: asking the right questions, understanding real users, defining product north stars, and determining what is truly "right." In the past, execution was the primary bottleneck in knowledge work. Now, execution is accelerating via models, so human value shifts toward pre-execution judgment, creativity, relationship networks, and organizational capacity. AI won’t automatically eliminate tough decisions—it will amplify wrong directions even faster.

If we compress this conversation into a single insight: After Claude Code, Anthropic is not seeking a single blockbuster product, but rather a systematic method to transform model capabilities into production systems. In this sense, the subject of discussion is not merely Anthropic’s next product roadmap, but the structural turning point across the entire AI industry—from "model competition" to "system competition."

Below is the original content (edited for readability):

· AI product competition has shifted from "stronger models" to "how capabilities land"—essentially, large model companies are now vying for workflow entry points.

· The significance of Claude Code goes beyond coding; it proves agents can sustain task execution under clear goals, pushing AI from chat tools to production systems.

· The core value of Anthropic Labs lies not in how many products it releases, but in rapidly validating what model capabilities should evolve into next.

· Co-work represents Anthropic’s effort to extend Claude Code’s methodology beyond programmers—abstracting "coding capability" into general work automation for non-programmers.

· OpenAI Codex’s pursuit forces Claude’s advantage not just on technical lead, but on Anthropic’s ability to integrate Claude Code, Co-work, and Claude.ai into a unified experience.

· Model companies entering application development intensify boundary conflicts with customers—but this is an inevitable path to defining the next generation of AI product forms.

· The faster AI executes, the more human value concentrates on pre-judgment, product taste, and problem definition, because wrong directions get amplified faster by AI.

· AI’s impact on employment isn’t solvable by any single company—it fundamentally forces society to reconsider skill retraining, distribution mechanisms, and irreplaceable human capabilities.

Alex Heath (Host): What’s Anthropic’s next big product after Claude Code? This week, we’re joined by Mike Krieger, co-founder of Instagram and now heading the “moonshot” team inside Anthropic.

Mike Krieger (Chief Product Officer, Anthropic): One of the darkest days I’ve had at Anthropic was naming it 3.5 v2. I can explain why we ended up with that name.

Alex Heath: We recorded this conversation in person during Anthropic’s recent Claude Code event in San Francisco. At that event, Anthropic announced a new major compute partnership with Elon Musk. So, you're now basically going to space with Elon?

Mike Krieger: Absolutely. Yes, we’re actively seeking new, even unexpected sources of compute.

Alex Heath: We discussed what Mike is currently working on, the intense competition between Anthropic and OpenAI, and Mike’s view that even as AI grows stronger, certain aspects of human work will remain crucial.

This is Access.

Mike, great to see you at the Claude Code conference in San Francisco. I was just reminiscing about our last conversation. You’d only just taken over Labs then, but it’s been months since—about four, right?

Mike Krieger: Yes, nearly four months.

Labs in Action: Biweekly Eliminations, Small Teams Validating Big Products

Alex Heath: Nearly four months. For those unfamiliar with Labs, let’s start here—it’s a particularly unique organizational structure. When I visited your office months ago, we also talked about it. What exactly is Labs? What’s its mission inside Anthropic?

Mike Krieger: Simply put, my understanding of Labs today is what I’d call Labs v2. We can explore Labs v1 later—what it did and what Labs v2 aims to do.

But I believe Labs focuses on two things.

First, narrowing the gap between Claude’s theoretical capabilities and everyday user experiences. That is, while Claude can theoretically do many things, how do these capabilities actually enter people’s daily work and lives? What products, prototypes, or projects do we need to build to demonstrate how to unlock more of this potential and shrink that gap as much as possible?

Second, we function more like a “frontier reconnaissance unit,” assessing which directions models should evolve toward to meet diverse user needs.

So, a successful Labs project doesn’t necessarily need to become a final product. It could be a prototype. We might build something and realize: the model isn’t good enough yet to complete this task. Then we park it temporarily, revisit it when the next model version launches, or turn it into an evaluation metric for future model development and keep iterating.

Thus, unlike pure product company labs—where success is measured by “did you ship a product?”—at Anthropic, Labs’ value extends beyond shipping. It can influence Anthropic’s future direction.

Alex Heath: Labs has indeed produced some hits, right? Claude Code and MCP are among them. Any others?

Mike Krieger: Agent Skills is another significant thing Labs has created. Also, I can mention one project we never released but was extremely helpful for research: computer use—enabling Claude to operate a computer.

I joined Anthropic in May 2024. Next week marks my second anniversary—we internally call it “antiversary.”

Alex Heath: Anniversary?

Mike Krieger: Antiversary. Everything at Anthropic must relate to ants. I initially resisted it. We don’t say dogfood—we say antfood.

After joining, we began building Labs. One of the earliest ideas proposed was: Why not try having Claude use a computer?

Alex Heath: Computer use.

Mike Krieger: Exactly.

Alex Heath: What model era was that?

Mike Krieger: That was Claude Sonnet 3.5. It was also the first model I helped launch. I started working on that release in my third week. We joke that Anthropic has no onboarding projects—instead, they throw you straight into a hard one. I was thrown into launching in my third week.

Sonnet 3.5 was an interesting model because it was among the first to genuinely unlock partial coding scenarios. Not quite full agentic coding yet, but early signs were visible.

So, we integrated Sonnet 3.5 and built a computer use product around it. But it had many issues—too slow, low accuracy, poor visual perception. It would see the screen and say, “I need to click that button,” but end up clicking somewhere else.

Yet, building this “not fully usable” test framework itself was immensely helpful. Later, when we reached Sonnet 3.5 v2—this naming story can wait—we could plug the new model into that same framework for testing.

We tried 3.6 next—still not good enough, but showing slight improvement. Then 3.7—my memory of that day is vivid. I was in New York on a business trip meeting the local team. Suddenly, someone messaged me: We think the old Labs project—the computer use project we’d shelved nine months ago—is finally showing life on Sonnet 3.7. We believe it’s time to open it up as a capability and publicly discuss it.

That spanned about nine months. Every few months, we’d test a new model in the same framework. Even though Labs had temporarily paused the project, it remained highly useful—it became a benchmark set for evaluating how computer use capabilities evolved in models.

Alex Heath: When you first joined Anthropic, you were CPO. I remember thinking: Mike Krieger, co-founder of Instagram, who’s known for consumer products—how did he end up at an enterprise AI company?

Mike Krieger: Yes.

Alex Heath: We probably discussed this back then. I thought it was an interesting choice. Looking back, it was correct. Timing was also perfect.

I’m curious—when you first joined, you were CPO overseeing the entire product line. Yet, the concept of “AI product” itself is fuzzy and evolving fast. How did you transition to Labs about four or five months ago? From what I understand, you’re now more of an IC—individual contributor? Do you still manage people?

Mike Krieger: I don’t manage anyone now. We’re just entering the performance review cycle.

Alex Heath: So this is what you wanted, right? You’re escaping writing performance reviews?

Mike Krieger: Exactly. I opened the system, saw what I needed to evaluate, and realized: you only need to write your self-assessment and your manager’s feedback.

Alex Heath: Just that?

Mike Krieger: Just that.

Alex Heath: Now it’s all done by Claude.

Mike Krieger: Claude does help write some evaluations—it’s quite useful. It doesn’t do the whole thing, but at least it helps recall: What did I actually do over the past six months?

I’ve noticed that throughout different stages of the company, the alignment between what I’m passionate about and what the company needs has varied.

When I joined, the product and engineering team was maybe around 30 people—roughly half and half. Of course, there were engineering teams working on research infrastructure and scalability, but if we look only at those directly building products, it was mainly Claude.ai and what we then called API—back then, even before it was called Claude Platform—probably around 30 to 35 people, incredibly small.

At that time, it still felt very much like a startup in early stages—many things were still undefined. For example, “What is this product?” hadn’t even taken shape. Claude.ai at the time had no Projects, no Artifacts—basically just a list of conversations between you and Claude, with almost no extra features.

So joining Anthropic back then felt like joining a company still searching for its product identity. Of course, it already had momentum.

Alex Heath: When you joined, the Claude 3 series had already launched, including Opus, Sonnet, and Haiku.

Mike Krieger: Yes. That was Anthropic’s first attempt at a model series close to cutting-edge levels. There were still so many product-level challenges: What should this product become?

Although my background leans more toward consumer products, I was excited because during the period between Instagram and Anthropic, I had invested with Kevin, co-founder of Instagram. We had a whole investment thesis, one of which was “the future of work”—how work will be accomplished in the future.

Anthropic seemed highly likely to unlock that theme: What happens when you have a very intelligent assistant helping you work? I didn’t anticipate just how disruptive it would become.

Alex Heath: Back then, I probably thought: This is a pretty interesting small AI company—maybe it can help me understand some investment themes.

Mike Krieger: Yes, possibly help us understand topics we were already considering. But in reality, it changed far more than I imagined.

That was Phase One: a tiny team, only a handful of projects underway. Fast forward to late last year, the product team had grown to hundreds of people. We had a full portfolio, much work focused on deployment, understanding customer needs, client-facing roles, management layers, and all the usual growth-related complexities.

I gradually realized that some people genuinely love this kind of work and excel at it. I deeply respect them. But for me, I had a great mentor who described this state as the “zone of competence”—something you’re good at, can handle, but isn’t what truly ignites you or drives you deep down.

This is actually a dangerous position. You can stay here for years, perform well, but it’s not where your passion and energy truly reside.

So, late last quarter, I began discussing this with Daniela. I said, the company has grown. We’ve compressed what usually takes five years into roughly two years.

Alex Heath: Yes, I think your growth has been solid.

Mike Krieger: Yes, growth has been strong. Team size and product portfolios have expanded quickly. So I said, I think I want to start a new company.

Daniela asked: Is this because you want to leave Anthropic, or because you want to shift what you do within the company? I said, I love this company. The people are amazing, and I love the technology and mission.

Just then, we were restarting Labs. Because Labs v1 was so successful—every project graduated, no one remained. So Labs was effectively put on hold.

So we decided to restart Labs, and I returned to the builder role. Everyone who sees me now says: “Mike, you look so happy.”

Alex Heath: Some of your colleagues told me earlier today—they said Mike is in great spirits, really enjoying himself.

Mike Krieger: Yes. Of course, I’m still my own harshest critic. So every day I ask: How can I do better? What can we do? What can we build? What are we actually validating?

It’s not easy. But it aligns much more closely with what truly drives me.

Alex Heath: We don’t need to dwell on this too long, but I’m genuinely fascinated by internal “moonshot” or “zero-to-one” labs within tech companies. Alphabet is perhaps the most famous example, but similar attempts have existed throughout tech history—some succeeded, some didn’t.

Anthropic’s Labs v1, at least from a product outcome perspective, was clearly successful.

Mike Krieger: Yes.

Objective: Find the Next Claude Code

Alex Heath: This must bring you significant pressure. You’ll think: Okay, Claude Code is already here—I must create something comparable.

Mike Krieger: Yes. Interestingly, we have an internal saying akin to a mission statement: We must find the next Claude Code.

By the time Labs v2 launched, this standard was already high. Since then, Claude Code has continued growing, raising the bar even higher.

I believe several things matter. Having entrepreneurial experience helps because you can never fully replicate that feeling: two or three people against the world, with limited funds. If we can’t deliver a viable product, we must shut down, return capital to investors.

Alex Heath: Just like you and Kevin did back then.

Mike Krieger: Exactly. I remember the existential questions pressing on me daily: What if this fails? Can I still maintain independence? Can I still pursue what truly matters to me?

This sense is hard to artificially manufacture within a large company unless you design complex structures. I’ve seen such approaches. Before launching Labs v2, I studied many internal lab cases. Some companies give teams equity in what they create, or design other similar incentive mechanisms.

Alex Heath: Right—various “patchwork” arrangements.

Two-Week Experimentation Cycles

Mike Krieger: Yes, they’re all trying to patch that feeling.

But we found that the real solution isn’t primarily incentives. Because Anthropic already attracts many highly proactive, ambitious, and mission-driven individuals. Our bigger concern isn’t lack of motivation to make projects succeed—it’s avoiding turning this into a comfortable place where “good enough” ideas drag on for months.

So our approach is to shorten cycles. Labs v1 operated on four-to-six-week reviews, with each project getting support and about six weeks to prove something.

Now we use two-week sprints. Every two weeks, each project presents to the entire Labs leadership team. It’s not a “kill committee” meeting—usually rational discussions, unless a project is seriously off track.

But we rigorously ask each project one question: What did you learn in the past two weeks? And: Have we learned enough?

Sometimes, a project has completed its purpose. You say: This project is great—it’s proven what it needed to. We don’t need another two weeks.

Things are changing so fast now, and with these models, building is incredibly fast. So letting a project drag on for four more weeks carries a huge opportunity cost.

I believe this is one of the best things Labs v2 has done: We’ve accelerated the flywheel. At least every two weeks, we measure our learning velocity—even if we can’t yet measure external impact.

Alex Heath: Are most people in Labs former founders?

Mike Krieger: Many are.

Alex Heath: Is that your screening criterion?

Mike Krieger: In Labs, we have two main roles. We call all projects “bets” because they’re inherently high-risk, high-volatility bets. Then there are “bet leads”—project leads.

If you’re the de facto responsible individual (DRI), these people are almost always founder-background.

But Labs also includes other members—not founders, but highly hands-on builders. For example, someone who was an early employee at a startup and experienced the zero-to-one journey; or simply someone who loves building things.

We have a colleague at Labs who joined very early at her previous startup and then joined Labs early too. She has that “I can cover the entire tech stack—wherever needed, I go” capability.

I believe this kind of ability is equally important as founder background. You can’t have only pure founders—you also need complementary “founder-team chemistry.”

The Future Company Form: Small Teams + AI

Alex Heath: My co-host Ellis couldn’t make it today, but he sent me some questions to ask you. One was: Is the “founder super-team” you’ve assembled in Labs a new organizational model for companies, especially in the AI era? Or is it just a special case for Anthropic?

Mike Krieger: A great question.

I believe we’ll see more such teams in the future: small teams working alongside models. Models aren’t perfect, so you still need founder-type talent with judgment, taste, and direction.

For example, yesterday I spent about two hours with the team discussing a Labs project I’m involved in—specifically, multiplayer in the product, i.e., how collaborative interaction should look.

Those two hours were incredibly valuable—and profoundly human. We were just bouncing ideas in a room, refining concepts. Then for the next ~12 hours, I let Claude asynchronously process those ideas.

But the key is needing someone with judgment: Which questions must be thoroughly discussed upfront? Which tasks should be implemented immediately? Because if you don’t specify, the model will make many decisions for you.

So I believe founder-type talent remains critically important. But once you get the structure right, you can achieve a lot.

I recently did a small retrospective on what Labs has done over the past few months—just four months, but some projects were conceived in January, tested in February, wrapped up in March; others conceived in February are now becoming Claude Design.

So a lot has happened—very, very fast.

I believe the future will see more companies adopt this model: smaller teams, greater autonomy, one truly accountable person driving the project, and starting with fewer people.

This is one of the most important lessons we’ve learned in Labs v2: Don’t assign five people to a project upfront. Start with two or three—more like a startup.

Alex Heath: Next comes my favorite part—I’ll try to extract more info about these projects from you.

When we met in March, you mentioned researching Claude running long-horizon tasks—long-duration problems, handling longer-term assignments. You were leading that project at the time. Are you still leading it?

Mike Krieger: No. This project is typical: we put it aside, but its insights later fed into the Outcomes feature of today’s Managed Agent release.

The core idea was having Claude execute tasks aimed at final outcomes or goals based on rubrics—evaluation criteria—not just responding to isolated prompts. This aligns spiritually with what we explored in Labs previously.

I was also tracing how this project evolved. Two themes in today’s morning release directly stem from this Labs project.

One is Outcomes. The other is in Boris’s demo—you saw Claude Code verify its work via screenshots, testing its own output.

This was a key direction we pushed. It’s one thing for Claude to look at its own code and say, “This seems fine.” It’s another entirely to fully explore and validate it.

In my project, we even explored having Claude record its entire work process, then watch the video back and judge: “Oh, this animation is broken.” Some issues can’t be caught via screenshots alone.

So yes, the project we discussed in March was still in development then. Now we’ve put it aside. It still runs internally—we sometimes use it for demos. But its main value lies in serving as upstream inspiration and capability validation source.

Alex Heath: Another project you mentioned was having Claude choose its own form—like deciding, “Next conversation, instead of command-line interface, I’ll become a website.”

Mike Krieger: Yes.

Alex Heath: Is that Claude Design—or something else?

Mike Krieger: It shares spiritual alignment with Claude Design but isn’t the same thing.

We’re now asking: Fundamentally, Claude Design is a “agent + canvas” form. You can imagine many such combinations.

Even within Claude Design, I’ve used it for many things—for example, writing technical specification documents. My favorite way to write technical docs now is using Claude Design. Because you can visualize information flow: How does it move? How else could it move? Then you can directly watch it evolve.

This might be my second favorite use case. First is slides—I now frequently use Claude Design for presentations. But you can imagine the canvas hosting other formats. So this is a direction Labs is exploring: like Claude Design, but for broader types and use cases.

I find this incredibly exciting.

Alex Heath: So is it essentially more generalized productivity software? Is that the core idea?

Mike Krieger: I think this is a topic worth digging deeper into: a productivity tool—ideally highly personalized, tailored to your own productivity stack. I find this a very interesting trend.

Alex Heath: What blank spaces in AI are you currently thinking about?

Mike Krieger: As models keep improving, I find their potential to become useful partners in life sciences particularly fascinating.

I’ve seen more examples recently. For instance, on X, someone did a full genome sequencing at home. Later, he professionalized it—now he offers doorstep services.

I’m interested in this because I enjoy “self-knowledge”—wanting to understand how these things actually work.

Some experts told me there’s already a significant gap between models from months ago and current ones like Opus 4.7. Today’s models are genuinely useful in parsing genetic data, deriving insights, or reading lab results.

Previously, it was just “oh, cute—what it says is something a doctor would say,” or repeating rules-of-thumb we already know. Now it’s genuinely valuable.

So personalized medicine is a blank space I’m very interested in. I feel we’re at a tipping point. There’s still a lot of “overhang”—technologies already possible, but products and applications haven’t fully unlocked their potential.

Alex Heath: Indeed, many startups are emerging—Superhuman, Superpower, Ro, Function Health—all doing similar things.

Mike Krieger: Exactly. In January, when we did early exploration of Claude for Healthcare, one of our partners was Function. You can import your lab results into Claude for further analysis.

This week, I also started using Subco, a supplement service. I input all my supplements, and it knows how they interact, then gives advice—like: You don’t need eight; you can reduce to four.

They haven’t done the next step yet, but I can imagine someone combining your latest lab results or genetic data to determine if you’re hypersensitive to certain supplements.

So the entire field is fascinating. This is just the “optimization” side. Imagine the underserved areas—like many people lacking access to quality local healthcare. Perhaps AI can help bridge those gaps.

So there are many intriguing blank spaces here.

Mike Krieger: Another area is also interesting. About a year ago, I attended a consumer AI conference hosted by Forerunner. On stage, I discussed with several founders: What will be the breakout case for consumer AI—truly viral applications—beyond chat and endless assistants?

To date, I haven’t truly seen it. Maybe health is one, but it doesn’t quite fit the consumer app category I described earlier. AI-driven dating products might have a chance, but they naturally carry an uncomfortable undertone.

Alex Heath: So you’re not doing a Claude Dating Service?

Mike Krieger: I don’t think we’ll do a Claude Dating Service. But perhaps a customer using the Claude API might build something similar. I don’t foresee us doing it internally.

Still, I believe there’s an interesting blank question: Can AI genuinely help us better understand ourselves, the world, communities, and foster connection—not isolation?

For example, I’ve recently become interested in civic engagement: How to assemble representative groups for public issue debates. I’m not an expert, but I’m intrigued by this problem.

In such contexts, AI’s role isn’t to decide for people, but to help identify representative groups. Keep people involved, but ensure we hear the right voices.

Alex Heath: I wonder if you still have that “social media product” instinct?

Mike Krieger: Probably less so now. But if there’s anything interesting in this direction, I think Sora once explored a compelling idea: You describe the algorithm you want, and the system generates it for you.

Alex Heath: This direction is everywhere now—Threads, X to some extent. I think this will become a standard capability.

Mike Krieger: Spotify has it too. I’ve used it myself. I create some “weird” playlists. For example, my daughter and I couldn’t agree on music, so we let Spotify generate a playlist combining our tastes. The result? Pavement mixed with Frozen soundtrack. Amazing.

Alex Heath: I love that. Very interesting.

Mike Krieger: Spotify’s AI DJ is indeed clever. But I think it’s a way AI is helping us personalize ourselves.

I’ve also been thinking about another point—not sure if it’s suitable for Labs, but someone should explore it: Can AI become a useful filter, filtering out noise from the outside world?

I’ve started using Dispatch and Co-work partly for this reason. Otherwise, I’d be a hopeless news addict—constantly checking every news site, reading everything. It has value, but sometimes I wonder: Am I just re-reading the same stories?

Now I have something like a daily digest in my workflow. It aggregates sources I normally read. I still often click through to originals, but it at least helps me spot trends, so I don’t need to open twelve websites every morning.

Alex Heath: As long as you keep reading The Verge’s Sources newsletter, Mike, you can keep doing the synthesis.

Mike Krieger: Of course, that’s going straight into my inbox.

Alex Heath: I vaguely saw a leak on X—don’t know if true—suggesting you’re considering more proactive features. Maybe something called Orbit? Possibly bringing what you mentioned earlier into Co-work or another interface layer, targeting more consumer-facing users.

Do you see this as an opportunity?

Mike Krieger: Yes. Whenever you see heavy users creating such usage patterns organically, it naturally raises the question: What would it look like if we made this capability native?

For example, scheduled daily digests, or proactive monitoring of certain things. We rely heavily on Slack internally, so it can proactively monitor Slack; it can also monitor email.

I’ve started doing this myself. My morning Claude routine includes a few things. One is the news digest I mentioned. The other is having it scan Superhuman—the email client I use personally. Since MCP has launched, it can now scan my inbox.

This is very useful because it understands email categories and tells me: This one you really should read; these can wait. So now, before opening my inbox, I check Claude’s summary.

There are also shopping-related things. For example, waiting for a product launch—instead of compulsively checking daily, I let Claude do it.

Alex Heath: Is this for sneakers?

Mike Krieger: I should use it for sneakers. But it might not be fast enough—sneaker drops require faster systems.

And there are so many internet rabbit holes—those information trenches you accidentally fall into. Of course, there’s a balance: You don’t want to remove the fun entirely. Occasionally getting lost online can be fun. But maybe you can do it more consciously, rather than mindlessly scrolling in the background.

From Claude Code to Co-work: AI Products Begin Serving Non-Programmers

Alex Heath: I’m a Co-work user—I use it daily. I even used it to prepare for this interview and for writing—pretty much everything. It’s already changed my workflow. I’ve recommended it to my wife and friends. But I still feel it’s early-stage. Most people probably aren’t using it like I am. From a product perspective, Co-work still feels somewhat disconnected from traditional chat and Claude Code. I don’t think they should remain three separate things long-term.

Mike Krieger: I agree.

Alex Heath: Months ago when I visited your office, some of your colleagues said Co-work might eventually become the front-end interface for everything.

Co-work was already a Labs project before you joined, right? But you were still CPO at launch. So I’m curious—what was the insight behind Co-work? Did you suddenly realize: We need to abstract coding capabilities so ordinary people can use them? Was this an opportunity—or something else?

Mike Krieger: I’d love to credit Dario.

Anthropic simultaneously considers many different things at any given time: research, product, business, compute, policy, social impact, etc. It’s not a typical company—product and go-to-market aren’t the only two priorities, and the CEO’s attention isn’t solely focused there.

Dario shifts his focus between domains based on the most urgent issues. Shortly before Co-work’s launch, he told us: I’ve noticed a fascinating trend—people using Claude Code in personal contexts. But for most, it’s hard because you need to open a terminal, which is already a barrier.

So he posed a question: What if we built a “Claude Code for everything else”? That was a very useful insight. It fits his working style perfectly—he doesn’t hand you a product sketch or define the product fully.

Alex Heath: That’s your job.

Mike Krieger: Exactly. He just said: This is a product problem—go solve it. The real shaping of Co-work came from a fusion of two personalities. Of course, many contributed, but I specifically think of two people.

One is Felix, one of the main maintainers of Electron, deeply familiar with desktop software—or at least desktop web software. He’s constantly thinking: How can people accomplish work on their own computers? The other is Boris, with deep expertise in Claude Code.

When we brought these two together with their respective teams, we had Co-work ready in just a few weeks—not that all work happened in those weeks, but because substantial prior thinking already existed on both the Claude Code and desktop sides: How to make this capability accessible to non-programmers?

Because these models are fundamentally powerful agentic engines—intelligent engines with agent-like capabilities. The challenge is how to hand this engine to more people.

But I fully agree with your assessment: Currently, it feels like we’re not launching an organizational chart, but rather our harness strategy—or testing and runtime framework strategy. This is not intuitive for most users.

Alex Heath: And it’s a poor user experience. For example, threads in Co-work don’t sync to the Claude mobile app. So I can’t take the interview document I’ve been preparing in Co-work directly to my phone.

Mike Krieger: Yes. Some products naturally point toward their next evolution.

For example, I’m a heavy user of Dispatch. It allows remote access to Co-work, but your computer must be on. So the next natural question is: What if I don’t have to keep my computer on all the time? How do we evolve in that direction?

Claude Code has already gone down this path—half a year to a year ahead. For example, Claude Code Remote—something I use regularly. I can initiate a coding task while away, and often by the time I return, it’s already submitted a pull request.

You can imagine Co-work heading in a similar direction.

But I agree—we’re currently in a phase where, for innovation, we let many things grow organically. That’s great. But from the perspective of context and continuity, whether you’re conversing here or there shouldn’t be a question users need to ponder.

My wife often asks: I don’t remember—was that in coding? In chat? In Co-work? This shows the abstraction layer is flawed—we need to fix it.

Alex Heath: That’s actually reasonable. When I talk with the OpenAI team, their super app strategy and their push around Codex seem to be the kind of ultimate form that you’ve already started building with Co-work.

I’m curious: First, do you think this “super app” path is the correct direction for AI products? Second, does this increase pressure on you? Because I feel OpenAI has recognized this as a massive opportunity—to bring coding capabilities to those who can’t code.

Mike Krieger: Yes, and they’re moving fast.

Alex Heath: Codex is also excellent—new models are strong too. Codex is doing well for them.

Claude had significant growth in coding, and the market still generally views Claude as leading in code ability. But Codex is definitely catching up, right?

Mike Krieger: Yes, it has become very strong.

One of the most interesting periods in my Instagram days coincided with facing competitors like Snap. They were in the same space, but approached things differently. In a fast-moving market, you realize some ideas come from the other side, and some come from you.

Alex Heath: So in this analogy, is Anthropic the Snap?

Mike Krieger: I don’t know who’s who. Market dynamics differ, valuations aren’t the same.

Alex Heath: Yes, it’s different.

Mike Krieger: But company cultures are very different.

At least in my understanding, not all competition evolves this way, but the Instagram vs. Snap rivalry was, to some extent, like that: Each had strengths, and each viewed the future in slightly different but aligned ways.

The question became: How does each company evolve to reach those different positions?

Take Instagram: We were known for posting one very polished photo per week. We wanted to move toward more freedom in sharing. Introducing Stories was crucial for us.

For Snap, they also wanted more influencer and celebrity content, as that demand was growing. So both sides had to manage their own evolution paths.

This analogy might be stretching a bit. But returning to us, having OpenAI in this space is interesting. I think for us, two things are most important.

First, as we discussed: Making our product ecosystem more unified and logical. Whether on web or desktop, it doesn’t have to be a single app or super app, but at least the various building blocks should interconnect meaningfully.

Second, continuing to narrow the gap between capability and usability. Because even today, with Co-work or any of our current products, this gap persists.

I just spoke with a recruiting colleague who’s one of the most extreme Co-work power users and the highest non-programmer Co-work user internally. His workflow looks astonishing.

But even at Anthropic, most recruiting team members don’t work that way.

So regardless of whether the final form is a super app, the goal is the same: Can we create product forms that allow this recruiting colleague’s experience to spread quickly, so others can easily adopt similar workflows? Or enable everyone to reach the same level of confidence and fluency using Claude?

Alex Heath: I was actually hoping to hear more about Co-work today. Though I know this is a developer conference. I’m curious—how is Co-work progressing compared to Claude Code? Is it growing faster? Any data?

Mike Krieger: I’m not sure how it compares relative to the overall coding product landscape, since coding itself is growing extremely fast. But Co-work’s growth trajectory has been similar to, if not faster than, Claude Code’s initial trajectory. This is incredibly exciting.

I’m personally thrilled because I’ve always hoped Claude’s impact could extend beyond pure code scenarios.

Almost every few days, someone in our internal productivity-focused Slack channel shares a new Co-work milestone. This growth is truly exhilarating.

When we speak with enterprise clients, they say: Great, we adopted Co-work, and we’ve seen this phenomenon—some people truly grasp it, others still need a lot of hand-holding, and different departments use it differently.

This is a good question because it shows that, under the right conditions—right people, right context, right product form—this product can do the right things. But the gap still exists.

I believe that once we can get most people to reach the “Oh, AI is really helping me work” state without much onboarding cost, growth will accelerate further.

Alex Heath: Speaking of competition, we discussed another topic about a year and a half ago: competing with customers. At that time, you hadn’t launched Claude Code yet.

Mike Krieger: Yes.

Model Companies Enter Application Development

Alex Heath: We discussed then: What happens if a major customer like Cursor builds similar products? At the time, as CPO, you had to make that judgment. Now you’re no longer CPO—you don’t need to make such decisions anymore.

But how is it now? We’ve already circled around this topic. For example, the controversy between Claude Design and Figma. You were on Figma’s board, so you stepped down. Dylan later said some friendly things—seems it wasn’t personal.

But this incident is a clear example. Many in the market see the trend: Wow, Anthropic seems to be entering every key vertical. The Figma situation feels especially nuanced.

Mike Krieger: This is definitely more complex than before. When I first joined Anthropic, we only had Claude.ai. I felt it wasn’t really competing with anyone—it was more a brand-new product form.

Now we still carefully and seriously consider what products to build.

I believe it only makes sense when we can offer a unique value through a product. Because there are many excellent customers on the Claude Platform, and many internal people deeply committed to supporting their success.

So the question is: Are we showcasing a direction that reveals possibilities to the industry? Ideally, this should create a win-win effect—many companies start adopting this mindset.

Claude Code is a great example. Before, much attention focused on editors, not terminals. After Claude Code launched—especially with the way we launched it and the attention we gave it—the entire industry largely shifted toward that direction. Of course, some had already been thinking about terminal scenarios.

This is what I hope for these products—not that they become the sole product in their domain. That would be bad. I hope for diversified products.

More ideally: First, they should make sense within our product portfolio. For example, if you’ve already connected all MCPs in Claude.ai, this should enhance other products too.

Second, they should demonstrate a forward direction. For example, Claude Design is very agent-first—emphasizing intelligent agents and giving them significant control over output. This creates a very specific product experience.

For me, it’s very useful for tasks like creating slides. I can think alongside the agent and produce content together.

Alex Heath: Don’t you think it will cannibalize Figma?

Mike Krieger: I believe they serve different use cases.

Figma excels in production, refinement, and collaboration—formal production, fine-tuning, and teamwork. These are Figma’s strengths. I really admire the Figma team—they’ve refined their product perfectly for this context.

My use of Claude Design is different. I’m not a full-time designer. I’m more generating visual communication materials or doing interactive explorations. The goal isn’t precision or final production, but capturing early mockup feel.

I once used Claude Design to prototype our iOS app. The pixel quality differed greatly from what we’d eventually release, but it pointed in a direction: I believe we should be able to achieve this in that space. So both will keep evolving. I used Figma and Claude Design as examples, but any similar product works the same way—everyone keeps moving forward.

I hope our good ideas enter other products, and vice versa. I believe we’re collectively exploring: What product forms maximize the use of agents?

This week, I had a great conversation with a researcher. He said: The more you restrict an agent’s behavior, the more you try to impose overly specific constraints, the less likely the emergent magic will happen.

I think Claude Code exemplifies this well, and Co-work reflects it to some degree. In contrast, Claude.ai feels more like chatting and executing very specific tasks.

We’ll continue building such products. Their most important role is helping more people understand this paradigm and embed excellent experiences into their own products.

If these products are powered by Claude, that’s great. We also hope Claude remains the best—or at least one of the best agentic underpinnings—i.e., the intelligent foundation behind next-generation AI products.

In summary, this situation is indeed far more complex than before. But we still adhere to the same principles: carefully consider which domains to enter, and ensure platform-building blocks remain open for everyone. For example, Managed Agents. Now, what you can build based on Managed Agents is just as powerful as what we build internally.

We will never withhold a model release due to safety concerns. We won’t do it just because “keeping model capabilities gives us an edge.” This is a principle we’ve consistently upheld.

We only release a full product when we believe it truly expresses something new.

Alex Heath: So you and Dylan are still on good terms? He’s been on our show before, so I need to confirm.

Mike Krieger: I really like Dylan and deeply respect what they’re doing.

Alex Heath: This is another question from Ellis. Nowadays, especially in early-stage startup circles, people are debating whether consumer startups are still viable. They wonder: What will Anthropic do next? What’s the next Claude Design? What will OpenAI do next?

Compared to that, the enterprise market seems safer. Dario mentioned today he thinks a company will emerge this year achieving $1 billion with just one person. I suspect it’ll be an enterprise company.

But how do you see the current state of consumer startups? If you were the younger Mike who co-founded Instagram with Kevin, facing today’s environment, what would you do differently?

Because the world has changed significantly. Starting consumer startups now seems harder.

Mike Krieger: If I simplify the breakout of consumer products—viral growth—into two essentials:

First, some new capability or new form. For us, it was the smartphone camera. Now no one calls it “camera phone” anymore—having a camera is default. Or perhaps richer media formats, like TikTok, Reels—because streaming capabilities were mature by then.

Second, a mechanism for distribution that enables virality.

For us, it was still a bit of a “wild west” era. You could share Instagram photos to Facebook, Twitter, Tumblr, Posterous—showing how early Instagram was. All these platforms allowed free linking back to your product, creating a growth loop.

But today, it’s completely different.

I’ve chatted with founders like Matt from Locket. They’ve found ways—like entering TikTok’s conversational spaces—to drive some interesting growth moments.

But I believe today’s distribution ecology is harder to predict and more uncontrollable.

So perhaps the biggest trend or challenge today isn’t any AI player—it’s this: When facing such a distribution ecology, how do you achieve breakout?

Previously, you might have competed for user attention against Facebook. I’m just throwing an example—Facebook’s average daily usage might have been 15 minutes. Now you’re competing against TikTok. I suspect TikTok’s average daily usage exceeds an hour. It’s quite crazy, but that’s reality—you’re trying to steal attention from there.

So if you need both: new capability or form, and a new distribution mechanism—then on the former, it might be a new type of interaction with AI-driven entities; or conversely, AI helping you go outside, “touch grass,” i.e., return to the real world, and generate more authentic interactions.

Either way, you still need to solve the second issue: distribution.

Many attention flows are now shifting toward chat intelligences. But I don’t think anyone has truly figured out: What does “viral” mean in a chat context? Today, this question might even be invalid.

Perhaps something new will emerge. For example, we launched MCP apps as an open standard.

Alex Heath: Could this become a new app store-style distribution model?

Mike Krieger: Possibly. You could integrate apps. Maybe the first truly viral MCP app will be something consumer-oriented. That would be fascinating—because it could be a new distribution mechanism.

I once spoke with someone from a nonprofit explorer project. We wrote an article explaining how MCP is used in certain scenarios. He said: It’s interesting to see what “viral growth” means for a nonprofit explorer project.

The number wasn’t huge, but he said: When checking connector lists, they were ranked in the top five for a few days—right next to Gmail and Slack. This brought them a lot of attention.

So now I’m thinking out loud with you. That’s also part of the beauty of these conversations—we might not get a perfect answer immediately. But perhaps we’ll see something like: Can you provide strong practical value within frameworks like Claude, ChatGPT, or Gemini, and find an interesting way to spread?

I’m eager to see that happen.

The Stronger AI Gets, the Rarer Judgment Becomes

Alex Heath: You recently recorded a podcast with Dan Shipper from Every. I listened—I loved how you used Claude to rebuild Burbn. For those unfamiliar, Burbn was the app before Instagram, which you later pivoted into Instagram. Claude recreated it in minutes.

I’m curious: First, emotionally speaking, watching AI recreate something you and Kevin spent months building in minutes—how did that feel? Second, if 2010’s Mike had access to today’s Claude, would Instagram exist?

Mike Krieger: A great question.

Alex Heath: Because I guess AI might not find that direction on its own.

Mike Krieger: Yes. After completing it, my immediate reaction was: First, I knew exactly what I wanted it to do. And I told Dan a fascinating point: It actually overbuilt. It added filters to Burbn, but Burbn didn’t have filters at the time. We added them later when launching Instagram.

Because Burbn was purely web-based—there was no WebGL or similar tech for filters back then. So seeing it self-add this feature was interesting.

But I think returning to our earlier discussion about Anthropic’s internal founder team, a large part of product work remains asking hard questions and showing the product to real users.

I both love and hate the moment when you first expose a product to users and get hit head-on. Someone says: What is this? I have no idea how to use it. Or: I’m so confused.

My first job after graduation was a UX researcher—weekly lab tests with participants. I still love that. You bring people in; whether the product was written by Claude or a human, it doesn’t matter. They care only: Does it work? Can it be used? Does it create a delightful moment?

These things remain hard. Claude won’t solve them for you. So if we’d had Claude in 2010, we’d have delivered faster in many areas—especially when the goal was clear.

For example, after launch, we quickly knew we needed @mentions. That would take about a week—from UI to text layout engine to server-side persistence, handling many details. With Claude, we’d have completed it faster and delivered value sooner.

But the journey from Burbn to Instagram—I don’t think Claude would change much. Except during certain coding sprints, it would speed things up. At the time, I didn’t feel the actual coding from 0 to 1 was the main bottleneck. The real bottleneck was the thinking and exploration process.

So I’d say: Hard things remain hard.

I have a concern: If LLMs automatically do too many decisions for you, won’t they actually hinder you from finding a more orthogonal, more unexpected product form?

Certainly, you can reverse it. You can say: Okay, help me generate three alternatives—then I’ll pick which feels right. I’ve done that myself.

But it will never relieve you of tough product decisions. In fact, you need to make them more.

Alex Heath: I feel this is becoming even more important.

I feel this in my own work. My work isn’t writing software, but media. I can delegate more to Claude—let skill files remind it of my preferred style, etc.

But when I actually read what it produces, I still think: Hmm, not quite right. Then I communicate with it like I would with a journalist in an editorial room.

Mike Krieger: Yes.

Alex Heath: I feel that as its capabilities grow, this role becomes even more crucial. I don’t want to lose my intuition in this process. I sense many people, overwhelmed by the rapid rise of AI programming, may be over-delegating. I hope we can later return to a more balanced state.

Mike Krieger: I believe human intuition remains vital. I really like your “journalist and editor” metaphor. Unless the journalist is completely off track, most of the time, the editor sees a draft that’s already 90% close to what they want.

Alex Heath: I’ve definitely encountered completely off-track cases before.

Mike Krieger: But most of the time, the editor thinks: This draft is already 90% close to what I want—just tweak a bit to reach 95%; or it’s 80 to 85 points.

But if you’re involved from the start, you might get closer to 100%. Of course, these things can’t be precisely quantified.

I’ve had this experience building things inside Labs. Now I’ve learned: Best to have thorough discussions with Claude before it writes a single line of code—clarifying things deeply. Then I say: Okay, this direction feels right—we’ve effectively co-created a spec. Now go implement it.

If I just give it a high-level functional description, it will indeed build the feature. Our entire validation mechanism ensures it runs functionally.

But then when I review it, I think: If I were doing it, I’d do it slightly differently. The problem is, it’s already done. That “not quite right” feeling—I don’t like it.

So we’re shifting toward a method: Clearly articulate the north star first, then help Claude converge efficiently and effectively toward it.

Someone in our internal team put it well: When programming with Claude, the core of the work is clearly expressing the north star—final direction—and helping Claude converge toward it.

This is a guide—sort of a manager. Of course, managers in software usually focus more on people development. So maybe more like an architect. But I haven’t found a better word yet.

It’s more like a guide or Sherpa for the project and Claude: You move toward the goal together, not expecting it to auto-complete from the start, nor completely excluding your early involvement in shaping direction.

Alex Heath: There are many interesting parallels between Instagram and Anthropic. I remember early on, Instagram faced a major headache: compute—specifically, the server capacity needed to host the app. You didn’t have enough servers. At one point, you had to figure out a solution.

Mike Krieger: Yes.

Alex Heath: And that deal came very quickly.

Now it’s interesting—more than ten years later, compute is still an issue. Though I know you’re not directly in charge of compute, I assume it’s a constant consideration inside Anthropic. When you joined this “small AI lab,” you probably didn’t anticipate this.

Mike Krieger: Two small anecdotes illustrate this well—one also shows how much I’ve learned in the past two years.

When Claude 3.5 launched, on launch day—though our scale was much smaller than now—users started adopting it rapidly. We were nearly maxing out our allocated chip resources.

I remember asking the infrastructure team: Okay, what if we run out of resources? Can we just add capacity?

Because in my past Instagram world, unless you were using some very specific hardware on AWS, there was always more hardware available out there.

They said: No, no, no. If these resources are gone, they're truly gone. Those GPUs are fully allocated. Of course, we're working hard to get more resources.

That's when I realized: oh, this is a very different environment. You can't just click "launch instance" like you used to and solve the problem.

Alex Heath: So that's why you now look like you're about to go to space with Elon.

Mike Krieger: Absolutely correct. We are actively seeking new, even unexpected sources of compute power. It was a very rapid learning curve for me—I immediately grasped and internalized it.

We're now deeply thinking about compute constraints. Even within a single product, you don’t want to sacrifice intelligence. The question becomes: how do you deliver as much intelligent capability as possible without wasting compute? For example, what tasks can be handled asynchronously? When must you use the largest model? When can you safely use a smaller one?

So compute is indeed a critical consideration. I think from many angles, this is healthy. It also brings us closer to our customers’ reality. You mentioned customer concerns earlier—they live in a world where they’re essentially buying tokens and reselling them in some product form.

So in many ways, we align with them, all wanting a healthy ecosystem. Over the past year, you’ve seen the market clearly shift toward token-based pricing or usage-based pricing—billed per token or per unit of usage.

I remember looking at this ecosystem a year ago and thinking: our customers want to charge higher prices or deliver more value, but they’re constrained by the current pricing models.

Now, more and more customers can genuinely deliver value around intelligent capabilities—or at least allow users to customize the trade-off between cost and capability. For example, you could say: I want to optimize cost, so I’ll use a less cutting-edge model, knowing exactly what value I’m getting; or if I really want to nail this, I’ll use fast mode, consume more tokens, because I have this adjustable knob.

So yes, at Anthropic, compute is far more critical than it was in the Instagram era. But in a way, there’s a similarity: Instagram’s growth was also exponential, at least if you look at the numbers.

Alex Heath: Dario said your Claude Code grew roughly 80x in the first quarter. I don’t know if Instagram ever saw an 80x growth.

Mike Krieger: No. I think apart from that initial spike—from 100 to 100,000 in the first week—we’ve had steady, solid percentage growth, but nothing quite like that.

Alex Heath: We’ll wrap up here. You’re still quite early at Anthropic. Although “early” sounds like two years already, in the AI world, that’s still very early.

The culture at Anthropic strikes me as fascinating. The more I interact with people inside the company, the more I sense it. Many outsiders see headlines, hear about Mythos, hear Dario talk about perhaps 50% of jobs disappearing. They say: they’re just creating panic, or there’s a regulatory capture strategy here.

But I find Anthropic’s culture incredibly unique. I’m curious—what do you see as the biggest disconnect between how the outside world perceives Anthropic and what Anthropic actually believes internally about what it’s doing?

Mike Krieger: This is something very hard for others to believe. Because I can keep saying: we’re a highly transparent company. What we say is what we truly mean.

Dario may be the most candid person I’ve ever met. He expresses his thoughts not after careful calculation: “Will this help me raise the next round?” Not at all.

As I got to know him better, seeing how he acts internally and externally, I became clear: this isn’t the driving force behind his communication or product strategy. Of course, saying something is one thing; consistently acting on it is another.

But maybe I can illustrate with an internal example. From a business growth and usage perspective, things are obviously going well this year. But I believe the company remains grounded—not losing focus—because Anthropic’s goal isn’t to build a massive commercial enterprise.

Anthropic’s goal is to do everything in our power to steer the world toward a better AI future. Once you view everything through that lens, it explains so many downstream decisions.

Of course, I can’t demand others stop being skeptical. Ultimately, we must continue proving through actions: we’re not manufacturing panic, nor are we suppressing certain things just because it makes us seem more forward-thinking.

Let me give an example. Internally, our belief is: we should be able to safely release a model at the Mythos level. And so far, we haven’t done it—that’s actually a bad thing. We’re not proud of that.

Because if we did it right, every positive use case would become possible. Earlier, we discussed life sciences—there are genuinely exciting possibilities on models like Mythos. I’ve personally written a lot of software internally using Mythos, and it’s exceptionally strong in that domain.

Alex Heath: Your internal software doesn’t pose cybersecurity risks, does it?

Mike Krieger: No. We’ve taken steps to ensure this is as safe as possible. But that’s how we see the world: we want to empower people with these capabilities, not suppress them. We’re working extremely hard to avoid letting security concerns block the release of these capabilities.

So yes, this requires proof. I think it’s fair for the outside world to demand that proof.

But when I hear people say: “They’re just trying to inflate the next valuation,” “They’re just trying to look impressive,” “They don’t actually believe any of this”—at least within the range where you’re willing to trust me, I can say: that’s not the real driver inside Anthropic.

Still, as I said, our job is to prove it over time.

Alex Heath: One final question. When someone comes to you and says: Mike, I’m really worried about my job. I’m concerned about my family, my children, my grandparents’ financial security. How would you respond?

Mike Krieger: I wouldn’t tell them “don’t worry.” Because I truly believe massive change is happening. We don’t know how fast it will come. People can differ on timelines, but change is already underway.

Lately, I’ve received emails from people I’ve met socially or professionally—many of them parents whose children just graduated college. They ask: what should we do? I usually share something I genuinely believe: this isn’t a problem one company or one government department can solve alone. It requires a societal-level conversation.

If there’s one thing we’ve been trying to do, it’s to catalyze that conversation. Even if it’s sometimes interpreted as creating panic. Whether it’s different tax structures, or reskilling—retraining and skills redefinition—these need to happen together. You’ve probably noticed we’ve started becoming more concrete in recent policy proposals.

Answering this for one individual would be a long response, because the issue itself is complex. But I’d tell them: you’re not alone. This is a shared, complicated challenge.

I also believe some things will remain fundamentally human, intangible, and profoundly important: relationships, curiosity, creativity, and the ability to bring people together toward a common goal. These will remain critically important. I don’t think AI will replace them anytime soon.

If you can develop that capacity—or at least continuously nurture and cultivate it—I believe it’s vital.

Also, don’t treat this moment as a fixed state. Things will keep evolving. Even in this uncertain, challenging period, if a friend’s child doesn’t land their dream job, it doesn’t mean the situation is set in stone. Nothing is permanent.

If people stay curious and proactively explore what lies ahead, they might end up helping create an entirely new career category or stepping into a new role within their own organization.

I believe the landscape will keep shifting. So don’t see this current uncertain, difficult phase as a static condition that locks everyone in place.

Alex Heath: Alright, Mike, I’ll let you get back to the lab. Thanks for coming to talk.

Mike Krieger: Great to meet you.

Alex Heath: Thank you.

Mike Krieger: Thank you.