How to run exploratory bets and kill them

1 The bottom line

Staff each exploratory bet with a convinced team and one named lead, set the evidence for success and stopping in advance, and let the team call the kill. Most ideas are wrong, so cheap, honest endings protect your best people and your money.

2 Key takeaways

Trust your instincts, but expect most ideas to fail

Instincts right (Pincus)95%

Ideas wrong (Pincus)75%

Ideas bad (Cherny)80%

• Give each bet a team that wants it and one lead who decides. Many people can explore, but someone must own the call to double down or wind down.
  • Mike Krieger, Anthropic12:00 says the bet lead role matters even more now because the team is moving so quickly.
• Plan for most ideas to fail. If failure is the base rate, ending a bet is routine, not a scandal.
  • Mark Pincus, Zynga4:05 says your instincts are right 95% of the time, but your ideas are wrong 75% of the time.
  • Boris Cherny, Claude Code15:25 says his first internal post got two likes, and his first month of prototypes mostly never shipped.
• Write down what evidence proves you right and what makes you stop, before you build. Cheap building removes the old filter that killed weak ideas.
  • Claire Vo, ChatPRD12:44 says that under engineering scarcity the bottom of the list never shipped, and cheap execution removes that filter.
• Let the team call the kill, and thank them for it. The team usually knows before the leader does.
  • Josh Woodward, Google Labs18:20 replied to all to thank a PM who recommended not launching a Gemini feature after early data fell flat.
• Scale what catches fire, and cut something else to pay for it. A big bet needs a named trade-off, not extra work stacked on the team.
  • Jimmy Liu, DoorDash24:45 grew the DashPass team from four or five engineers to over 100.
  • Sameer Ahuja, GameChanger27:00 paused basketball scorekeeping to fund livestreaming, and basketball now has dedicated teams.

The team usually knows before the leader knows.

Josh Woodward, Google Labs

2.1 Set up the lab

Give exploration its own home, its own goal and a bridge back to the core business.

• Give the lab independence and a clear purpose. Josh Woodward, Google Labs48:08
  • Do not nestle it under an existing business unit. In some companies it reports to the CEO.
  • Decide whether it exists to graduate features into existing products or to invent new product lines. Define winning and losing.
  • Large-company labs usually fizzle after about three to four years, or fail to commercialize their output.
• Label each project as zero to one, one to ten, or ten to a hundred. Josh Woodward, Google Labs47:20
  • Each stage needs different people and different metrics. At zero to one you look at eyeballs and reactions.
  • At ten to a hundred you look at waterfalls, CPAs and converting people to magic moments. Larger graduated products can crowd out zero-to-one work, so name the stages.
• Build a bridge from the frontier team to the main business. Josh Woodward, Google Labs44:34
  • Staff it with people who test every new model and tool. Give them an environment to build in.
  • Without the bridge, the rest of the company dismisses them as an R&D group making toys.
• Fund shared foundations separately. Ami Vora and Mike Krieger, Anthropic20:35
  • Put memory, MCP and file storage in a foundations team, not in the frontier product teams.
  • Anthropic found chat and co-work had different memory, MCP and file storage. A third experiment would have felt disconnected from the start.
• Use the lab to pilot new ways of working. Josh Woodward, Google Labs49:52
  • Labs saw job roles blurring and piloted a builder job ladder before it went across Google.

2.2 Staff the bets

Put convinced people in small teams, with one lead who holds the pen.

• Staff a bet only with a team that has real conviction. Ami Vora and Mike Krieger, Anthropic20:15
  • Do not assign a team to an area they are not excited about just because leadership likes it.
  • Let teams with exciting directions explore, even if two or three overlap.
  • Krieger says a manufactured B team bet tends to end with a bad product. He has never succeeded at tasking a team with an area they did not want.
• Name one bet lead for each exploration. Ami Vora and Mike Krieger, Anthropic12:00
  • Many people explore. One person decides whether to double down or wind down, and where to add or remove people.
  • Krieger's phrase: "we're all lost together, but somebody's got the pen".
• Keep pods small. Dianne Penn, Anthropic24:50
  • Large teams on ambiguous ideas slow things down, so a pod is sometimes a single engineer.
  • Select for people who love zero-to-one work.
  • Labs incubated Claude Code, skills, MCP and Claude Design this way.
• Do not schedule idea generation. Nurture side projects instead. Josh Woodward, Google Labs6:50
  • Good ideas rarely come from design sprints or blocked-off ideation time. They come from curious people hacking on weekends and over breaks.
  • Create space for side projects and watch for the one that catches your eye.
  • NotebookLM, Flow, AI Studio and Beam all started with very small, obsessed groups.
• Ask whether you would invest your own money. Sameer Ahuja, GameChanger51:10
  • Read the deck, but mainly look for conviction in the person's eyes.
  • He never asks the parent company to back an initiative unless he has absolute conviction himself.

2.3 Fund and plan

Plan in short windows, fund convictions instead of feature lists, and pay for big bets with a visible trade-off.

• Plan on a rolling six months and push for 50 to 100 day milestones. Josh Woodward, Google Labs43:10
  • Match the window in which labs release new models. Keep a multi-year vision, not a multi-year roadmap.
  • Remove blockers so people can build.
  • A project that began 48 hours before his interview could not have been planned.
• Replace the feature roadmap with convictions and the evidence that would prove them. Claire Vo, ChatPRD13:41
  • Write the direction you believe in over one to two years. Define what evidence shows you are right and what you would need to see to stop.
  • Allocate tokens, effort and investment based on what the experiments teach you.
  • She argues a roadmap of features and dates becomes ammunition for a slop cannon once building is cheap.
• Make the plan ambitious, and stop ranking by effort. Claire Vo, ChatPRD20:00
  • Describe the future you believe should exist, what would prove you wrong, and what earns more time and attention. Skip small onesie-twosie items.
  • Ask what is worth proving, not what can be built.
  • She says RICE becomes pretending when everything is buildable12:44. The last 12 to 18 months were a velocity game, and next year is the ambition game.
• Incubate cheaply, then invest hard once a bet catches fire. Jimmy Liu, DoorDash24:45
  • In annual planning, every team was asked what DashPass benefit it could build, which made it a companywide effort.
  • The signal was higher order frequency, retention and LTV among members.
• Back a contrarian bet with customer evidence and numbers. Jimmy Liu, DoorDash29:00
  • He combined conviction with customer anecdotes and financial modeling. He had talked to more customers about affordability than anyone at the company.
  • The skeptical COO flipped within a few months of launch. Liu also suggests building credibility for about six months at a new company before betting on yourself.
• Pay for a big bet by naming what you stop. Sameer Ahuja, GameChanger27:00
  • GameChanger stopped building basketball scorekeeping to fast track livestreaming. They stayed multisport and scaled basketball later.
  • When game activity fell 99.9% in March 202022:46, leaders met on a Friday and got a September deadline. Product and engineering committed within hours, with no detailed plan.
  • They launched the MVP around Labor Day 2020, and livestreaming became the rocket fuel for growth.

2.4 Run the bet

Say what kind of bet it is, protect customer trust, and stay with it long enough to learn.

• Label each shipped feature as a probe, an experiment or a promise. Claire Vo, ChatPRD18:18
  • A probe is a small bet to explore. An experiment is a durable test against a conviction. A promise is something customers can build on.
  • Be honest about the label inside the company and with customers. Track conviction strength and durability next to effort and impact.
• Keep a feature behind a flag until you believe it is right. Claire Vo, ChatPRD19:26
  • Code is abundant but customer trust is not. Customers will build on what they see, so the wrong ship burns trust.
  • Her product graph, a working insights engine and auto wiki, stayed flagged off while she rebuilt it until she had conviction.
• Queue several ideas, and treat the current one as probably wrong. Mark Pincus, Zynga4:05
  • Isolate the instinct, then test many ideas around it so you fail for the right reason.
  • Start with the premise that the new idea will likely fail and have four other new ideas ready.
• Be stubborn about the problem, not the solution. Claire Vo, ChatPRD17:04
  • Hold durable convictions about where the market is going, with zero ego about specific features.
  • Bad stubborn is using cheap tokens to move the goalposts and keep shipping the same idea.
• Avoid the churn trap. Claire Vo, ChatPRD11:34
  • Because shipping is cheap, teams see noise and ship something else. They never learn or compound on what they built.
  • Decide in advance what evidence you need, so a quiet launch is not read as failure or success by default.

2.5 Kill it or scale it

Make ending a bet safe, fast and informative, then fold winners into the product.

• Let the team call the kill. Josh Woodward, Google Labs18:20
  • Warning signs are fading passion and having tried everything you can think of.
  • Build a culture where people can say it is not working, and thank them publicly when they do.
• Give people safety to try ideas, and hold them to cutting losses. Boris Cherny, Claude Code15:25
  • Innovation cannot be forced with a roadmap. Expect that 80% of ideas may be bad.
  • Accountability means cutting losses on bad ideas rather than investing more.
  • Cherny spent a month on throwaway prototypes before Claude Code emerged.
• Treat a failed launch as input, and fix forward. Ian Silber, OpenAI59:40
  • What matters is how you react and keep listening to users, not what you shipped once.
  • IGTV flopped at Instagram because it baked in wrong assumptions and kept some constraints. The team changed them and redid it as Reels, which succeeded.
• Set an end date, because a doomed bet can drag on. Adam Mosseri, Instagram1:00:20
  • Running an idea fully can be the way to learn whether it had market fit.
  • Facebook Home, a fork of Android with HTC hardware, was a spectacular failure. Mosseri says he learned more in that year and a half than in any other, and he ended it when it had dragged on.
• When a bet works, consolidate it and protect the main experience. Ami Vora and Mike Krieger, Anthropic26:48
  • Expect a power law in usage. Make the main experience great and use it as an entry point to the others.
  • Half the job is saying no to another sidebar item or tab, and rolling things in once proven.
  • At Instagram the main feed was about 80% of use, and improving explore might move it from 10 to 15%.

2.6 Where they disagree

The sources split on how many bets to run, how long to hold them, and who gets the final word.

• Many overlapping bets versus only launching believed hits.
  • Ami Vora, Anthropic18:20 would rather have five overlapping products that work in different ways than one overly constrained one.
  • Mark Pincus, Zynga33:15 says to kill hope, because hope is confidence without basis. Separate products you ship to learn from products you ship because you believe they are hits.
  • Pincus says Brian Chesky already knows he has a hit before launching, and that the word viable should be killed from minimum viable product.
• Kill quickly versus hold on long enough to learn.
  • Woodward and Cherny stress cutting bad ideas early. Claire Vo11:34 warns that quick abandonment means you never compound on what you built.
  • Her fix is to set the evidence bar in advance. Dianne Penn, Anthropic24:50 adds that a turned-off bet can be revisited in one to two model generations.
• Leader's call versus a quiet side project.
  • The Anthropic sources want one bet lead who decides. Tony Fadell29:30 says that when the decision maker says no to something you see coming, keep a quiet skunk works version and bring it out later.
  • Jobs said Windows support for the iPod was over his dead body, yet the Windows version drove its takeoff. He also resisted a stylus, which later became a major feature for professionals and artists.
• Decide in a day versus build the evidence first.
  • Sameer Ahuja, GameChanger22:46 made a bet-the-company call on a strong gut feeling, with no 30 to 45 day presentation, because the world had shifted. Jimmy Liu, DoorDash29:00 won his skeptics with customer data and financial modeling.
• Chasing revenue versus being early.
  • For new device platforms like Vision Pro or Android XR, Yoav Sharon, Simply34:20 says the aim is to be early in a future platform, not immediate return. He suggests a fast pod working closely with the platform to ship at release.
  • Design partnerships are a side benefit. His "not about an ROI right now" stance differs from DoorDash's rule of scaling only once the data shows a bet catching fire.