Regular readers will notice: This is the natural next chapter after Pod Shots #130: Dark Factories. Where Simon Willison gave us the engineering reality check, this week three operator-CEOs (Lütke, Foroughi, Jolly) plus two of the most thoughtful dev-tool builders on the planet (Zechner and Ronacher) tell us what they've actually changed inside their companies. If you followed Andrew Wilkinson on $40/day engineering teams (#122) or Dave Killeen's AI chief of staff (#129), you'll recognise the patterns - now applied at $160B+ scale.

Want something completely different? Try these from the archives:

Today's Pod Shot

Six episodes: Tobi Lütke says Shopify's best engineers haven't written code this year. Adam Foroughi runs $5 billion in cash flow on roughly 400 core people. Xero's CPO Diya Jolly says the standard planning rules of thumb no longer work. Scott Galloway warns that the AI brand has collapsed for everyone earning under $200k. Dev-tool builders Mario Zechner and Armin Ronacher describe self-modifying software emerging in the wild. And a State Department under secretary explains why "AI with a Western soul" is now a diplomatic priority. The common thread: when the cost of building collapses, the things that don't scale - taste, judgement, relationships, regulation - become the actual bottleneck.

Top tip - Read the key insights, then jump to the Lütke and Foroughi sections back-to-back. They're the same playbook (radical leanness + AI-native operating model) seen from two completely different vantage points.

Tobi Lütke, Founder & CEO of Shopify, interviewed by Harry Stebbings on 20VC
Adam Foroughi, Co-founder & CEO of AppLovin, interviewed by David Senra on Founders Podcast
Diya Jolly, CPO & CTO of Xero, interviewed on First Round Review's In Depth
Scott Galloway, Professor at NYU Stern, interviewed by Steven Bartlett on The Diary of a CEO
Mario Zechner & Armin Ronacher, creators of Pi and Flask respectively, interviewed by Gergely Orosz on The Pragmatic Engineer
Sarah Rogers, Under Secretary of State for Public Diplomacy, interviewed by Katherine Boyle on the a16z Podcast

Estimated Reading Time: 14 mins. Time saved: ~7 hours! 🔥

Key insights:

Tobi Lütke is unusually candid for a public-company CEO running a $160B business. Asked how many people Shopify will have in five years, his answer is striking: "My real hope is 7 and a half to 8,000. Right. Like at 100x productivity level." Roughly the same headcount as today, but doing two orders of magnitude more.

The change is already visible internally. "Many of our best engineers have not written code this year," Lütke says. "Like December changed everything. Opus changed everything." Shopify has built an internal agent called River that "lives in Slack" and now does "a ludicrous amount of Shopify's engineering" - people prompt it in public channels so the whole company learns from each other's prompts.

Lütke has also changed his mind about who wins. He used to believe new graduates with no priors would have the biggest advantage as AI natives. The data inside Shopify said otherwise: "Senior engineers just steer these things in such a way that like they can accomplish incredible feats in very very little time." The accumulated reps of senior engineers translate directly into better agent steering - which is just another form of programming, at a higher abstraction level.

The contrarian view: Lütke is bullish on layoffs being misattributed: "AI will be blamed for absolutely everything... It's the perfect Girardian scapegoat." His read is that current layoffs are still primarily over-hiring from the Covid era unwinding, not AI. That may be true at the aggregate level - but at the individual job level it's cold comfort.

Key Takeaways:

If Lütke describes the destination, Adam Foroughi has already arrived. AppLovin generates roughly $5 billion in annual free cash flow on a core team of about 400 people on the advertising side. The company peaked at around $250B market cap.

The story behind that ratio is brutal. AppLovin's stock collapsed 92% in 2022, from $115 to $9. Most CEOs would have battened down. Foroughi did the opposite - he bought back roughly $6 billion of stock, partly with debt, at the bottom. That buyback eventually generated something like $50-60 billion in proceeds. But the more interesting decision came later.

In 2024, while the business was nearly doubling year-over-year, AppLovin cut about 40% of its team. Foroughi's logic: "A players don't like to work with B's, C's, D's. I mean D's you're gonna just burn your A's out." The new CTO, Giovanni, joined in November 2022 and kept asking Foroughi uncomfortable questions: "Why do we have this person? Why do we have this team? Why do we have like these processes?" Either Foroughi addressed those questions, or he'd lose the talent asking them.

The hiring veto is the operational mechanism. Every single hire requires Foroughi's personal approval. "If they come to me, I know that they're desperate. It's everything's going to break down around me. People are working so much that they really need this extra hire. Okay, go make it." Job postings used to backfill automatically when someone left - he killed that, because "in order to hire and train and get value out of a new employee, you're talking about 6 to 12 month process. What happens in the period of time when that person quits and you don't have that person? Why do you need to back fill if you can survive without that person?"

How this connects: This is the operationalised version of what Lütke describes aspirationally. Both CEOs have reached the same conclusion: with AI, you don't want B-players in core roles - you want a small core of A-players with the support structure to be 100x effective.

Key Takeaways:

Xero's CPO and CTO Diya Jolly delivers the planning-shaped hole in everyone's roadmap. Where previously a chief product officer might delegate a problem and trust the team to figure it out, the rules of thumb that governed resource allocation - the classic Google "30% existing, 30% new, 20% moonshot" splits - don't really hold anymore.

"Right now for most people you better be taking a lot of risk in your road map," Jolly says. "These rules of thumb don't really work right now with everything that's happening with technology change." A concrete example: Xero built automatic bank reconciliation despite customers insisting AI couldn't do it. They hit 97% accuracy and saved each small business 22 hours per month. But they had to push through the customer "no" to discover it.

The shift in altitude is the most interesting observation. Jolly used to spend her time as a CPO operating at a higher level - now she's flying lower. "You're trying to change people because you as the senior product and engineering leader are the catalyst for change right now... Now it's like should we do this so we have more flexibility, should we be model agnostic, how quickly should we be able to change models, what can be probabilistic, what cannot be probabilistic, do we need a UI here, will the UI be chat based or will it be a sidecar?"

Right now, Jolly says she spends 50% of her time on big ambiguous problems that can't be delegated. To protect that time she carves out one full meeting-free day every week and another two hours daily. "If you don't ask to if you don't tell people that that is the situation, people are going to trample over the calendar."

The contrarian view: Jolly is explicit that being demanding without support is destructive. "If you demand without giving people the support to succeed and letting them know that taking risk is okay and failure is okay, you're going to have a culture where nobody takes risks." Her batting-average target for OKRs is 0.7 - if you're consistently scoring 1.0, you haven't stretched.

Key Takeaways:

If the first three guests are operators inside the wave, Galloway is the social commentator watching it crash on the shore. His single most striking data point: "Your view of AI is directly correlated to your wealth. The only cohort that has a positive rating of AI is people making over $200,000."

The framing is uncomfortable but useful. Wealthy people see AI as portfolio fuel and a productivity multiplier. The middle class sees higher electricity bills and Sam Altman saying things like "stop complaining about energy costs - think about the amount of energy it takes to raise a child." The brand of AI, Galloway argues, has fallen further faster in the last 18 months than almost anything except the US brand abroad.

His thesis on job impact splits the difference between the doomers and the deniers. He doesn't buy mass apocalypse - the unemployment rate in the US is 4.5%, youth unemployment 8.8%, both around historical averages. New business permits per capita have doubled in the last decade. But he is worried about velocity: "There is a scenario you don't need 100% unemployment like Musk is predicting. At 20% unemployment the French had a revolution. In Weimar Germany turned very ugly."

The catastrophising, Galloway argues, is partly self-interested marketing: "I'm Dr. Frankenstein and I've created this monster, but I don't know how to deal with it. So I'm going to go peace out to Sanrope. That's just not very helpful." The CEOs benefit when their technology sounds seminal and dangerous - it justifies the valuations.

The genuinely useful skill prediction: Galloway argues the most-underweighted skill for young people is now the ability to endure rejection. "Because of AI and because of these frictionless relationships that people are engaging in online, I think a lot of young people are losing the ability to endure rejection." His mentoring practice with young men explicitly targets this - the goal is to get a "no," because resilience to rejection turns out to be the secret of every self-made entrepreneur he knows.

Key Takeaways:

Mario Zechner built Pi - a deliberately minimalist coding agent - because the existing options frustrated him. Claude Code "would inject stuff behind your back" and modify system prompts. Open Code would ping LSP servers after every edit, confusing models with errors that weren't really errors yet. Pi strips it back to the essentials: read, write, edit, bash, plus an extensive hook system so users can extend it themselves.

The unintended consequence is the interesting part. "Pi doesn't have MCP," Zechner notes. "People just ask Pi to build MCP support into Pi. Pi doesn't have a plan mode. Armin goes and 'my plan mode must be fantastic, bespoke and super.'" Users with non-technical backgrounds modify the entire TUI to fit their workflow - they just ask Pi to modify itself.

Zechner's deeper thesis: "I think where we are going is software that modifies itself on behalf of the user's wishes and needs and the agents can do that now if you give them enough rope to modify themselves."

Armin Ronacher (creator of Flask, ex-Sentry) adds the warning side. After interviewing ~30 engineering teams about their AI adoption, the patterns are consistent: Christmas 2024 was the inflection. Quality dropped because keeping it up "takes some effort." PRs are getting longer and more psychological to review. "There's so much stuff that you need to do that if you actually do the right solution... it is the kind of complexity that kills you at scale."

The MCP-versus-CLI debate is also worth listening to. Ronacher's view: MCP is essentially "input in, do some stuff, maybe some state transition" - it fills your context quickly and composability is hard. CLIs let you pipe outputs - "the model only sees the end result." Both predict code execution (not protocol calls) will dominate.

The contrarian view: Ronacher pushes back on the doom narrative around open source. The volume has changed, but "the amount of actually useful and maintained projects has probably not changed a lot." Zechner has built tooling that auto-closes any PR from a contributor he hasn't manually approved - a friction mechanism specifically to filter agent-generated noise.

How this connects: Where Willison (Pod Shot #130) named the "vibe coding vs agentic engineering" distinction, Zechner and Ronacher are showing what it looks like inside actual maintainer workflows. The friction they're adding back (manual contributor approvals, refactor mercilessly, no rubber-stamping PRs) is the muscle memory engineering culture is having to redevelop.

Key Takeaways:

Pulling Lütke, Foroughi, and Jolly together, the same operating model emerges three times:

Smaller core teams of A-players. Lütke targets flat headcount at higher productivity. Foroughi runs $5B in cash flow on ~400 core people. Jolly is restructuring Xero around a leaner model where every role gets re-examined.

Active culture maintenance. Lütke deliberately recruits "eights" on the Enneagram - the disagreeable problem-namers most companies push out. Foroughi removes B-players to protect A-players. Jolly says her chief product officer job is "essentially setting the direction" while building "very strong lieutenants" who can execute.

Personal protection of thinking time. Jolly carves out a meeting-free day every week. Lütke runs his calendar around the idea that great leaders "must be exothermic." Foroughi acts as head of recruitment for his entire company because hiring is the highest-leverage decision.

Speed of organisational change. Foroughi moves fast: "We have an idea and we just go." He bought studios on a hypothesis (own data to train Axon) and sold them when the hypothesis was validated. AppLovin's Axon 2 model launched in April 2023 and the company's market cap went from under $4B to a peak of $250B in about 18 months.

The PM implication is sharper than it looks. If your team can be 100x more productive but only with A-players and only with senior judgement on top, the implicit message is uncomfortable: the median PM and the median engineer in most organisations are now actively in the way. The companies winning this transition aren't the ones with the best tools - they're the ones willing to make the people decisions Lütke, Foroughi and Jolly have made.

Key Takeaways:

Sarah Rogers is the State Department's under secretary for public diplomacy - and her brief now includes "digital freedom." Her framing is direct: she's pursuing "transparency, truth, and reconciliation on prior censorship" while making freedom of expression a primary prong of US public diplomacy.

The geopolitical AI argument is sharper than most. Rogers borrows Tyler Cowen's phrase: "AI with a Western soul... AI that reasons in an individualistic way, a rules-based way that prioritises user consent, for example, those are all Western principles. And that is going to be the underlying reasoning model on which so much of the world's communication and commerce runs."

Her concern with EU regulation isn't abstract. She cites former European Commission official Thierry Breton sending Elon Musk a letter in August 2024 threatening regulatory penalties if Musk aired an upcoming interview on X with then-presidential-candidate Trump. "The interview hadn't even happened yet," she notes. The recent €20 million fine against X eventually came through ostensibly-content-neutral enforcement. Rogers' point: viewpoint-skewed enforcement of nominally neutral regulations is "both insidious and inevitable when you have something that's this politically pitched."

For PMs building AI products, three regulatory pressure points matter:

The under-discussed PM angle: Rogers is explicit that the regulatory environment should "favour viewpoint neutrality" while still allowing platforms to moderate spam, pornographic content, and content with foreign provenance. That distinction - viewpoint-neutral content moderation versus viewpoint-based suppression - is going to matter increasingly for any product team building user-facing AI systems.

Key Takeaways:

On team composition: The Foroughi and Lütke models suggest the next 24 months are about A-player concentration, not headcount growth. If you're a PM building or running a team, the question is no longer "how do we hire faster" - it's "how do we identify, retain, and protect the small core of people who actually compound with these tools."

On planning: Take Jolly's warning seriously - "these rules of thumb don't really work right now with everything that's happening with technology change." Roadmap commitments based on legacy planning ratios are less reliable than ever. Shift to outcome-based commitments, shorter cycles, and explicit acknowledgement that the variance between AI-accelerated tasks and integration-heavy tasks has widened dramatically.

On code quality: Take Ronacher's vibe-coding hangover seriously. If your team is shipping more code faster, the code review process probably hasn't scaled with it. Build in friction - manual approvals, refactor cycles, explicit ownership - or you'll inherit a code base nobody can maintain.

On thinking time: Take Jolly's calendar militancy seriously. If you're not protecting at least one full day per week for ambiguous problems, you're optimising for someone else's priorities.

On the political backdrop: Take Galloway's "AI brand has collapsed" seriously. The political environment around AI is going to get worse before it gets better. PMs building user-facing AI products should expect more regulatory scrutiny, more user scepticism, and more pressure on pricing models that look extractive.

On founder psychology: Lütke's "people who build companies are fundamentally crazy people" is worth remembering. The CEO behaviours that look unreasonable from outside (no calendar, no meetings, calling in sick to get thinking time) are often the rational response to a job that has no other defenders against bloat.

The AI-changes-engineering progression - each episode adds a new layer:

#135 adds the operator-CEO layer: Where Willison gave us the engineering reality, Lütke, Foroughi and Jolly show what running a real business through this transition actually looks like. Zechner and Ronacher provide the dev-tool builder's view of what's emerging in the wild. And Rogers reminds us that the political and regulatory backdrop is moving as fast as the technology.

That’s a wrap.

As always, the journey doesn't end here!

Please share and let us know what you liked or want changing! 🚀👋

Alastair 🍽️.