Gergely Orosz – Podcast Stage | Craft 2026 (Conversation Outline)

A podcast-stage conversation at Craft 2026, recorded by the ABK Podcast (“A Bátrak Klubja” / Hungarian online leadership club — abkpodcast.hu, “leadership anonymous,” covering tech and leadership), the last of two days of interviews. Hosts include the main interviewer and Fanni; the guest is Gergely Orosz — ~10 years a software engineer, ~5 years an engineering manager (at Uber), author of the Pragmatic Engineer blog/newsletter. The framing question: what is the current state of engineering management/leadership in the AI era? The conversation ranges over writing authenticity, the decline of management, team sizes, token economics, an apparent AI plateau, and career advice. (A narrative interview — rendered as topic threads at full fidelity.)

1. The Origin of the Pragmatic Engineer Blog

1.1 “Nobody reads your blog” — until they do

• Started ~10 years ago writing his software-engineering learnings for himself; the first many months had only a few hundred visitors.
• Published every 2 weeks for ~3–4 months, then stopped.
• Realized people were reading when his shared hosting crashed — a Hacker News submission drove a few thousand visitors it couldn’t handle.
• That post argued “if you have comments in your code, it’s just bad code — it invites a refactoring.” Hacker News split: half “finally someone said it,” half “what an idiot” — and seeing experienced engineers debate his writing motivated him to keep going.
• Lesson: “you never know who might come across your writing” — the host draws the parallel to their podcast (keep going even when no one seems to be listening); the “Cheetah engineer” post was the one that broke through on Slack channels.

1.2 AI-written content kills trust

• “With AI I can personally tell when something is AI written — I just stop reading.”
• The Michael Novati dispute: Novati (the “coding machine” archetype at Meta was created after him — hugely valuable, wrote enormous amounts of code, but wasn’t in the “right archetype” for promotion) wrote a post about problems inside Meta with good ideas but reading like “AI slop.”
  • Novati said he’d spent many hours brainstorming with AI and that Gergely “doesn’t appreciate his effort.”
  • Gergely’s reply: ”If it feels like AI, I will assume there is no effort.” Over DMs, Novati was clearly hand-writing — “your writing is good”; he only used AI to “make it more polished,” but that polish is exactly why Gergely stopped reading (“I’m interested in the Michael, the person I know”).
• There’s value in putting out ideas in any form — not just public blogs, but internal emails and sharing with colleagues; “don’t underestimate how saying something obvious might mean someone says ‘oh yeah, me too.‘”

2. Human Connection vs. the Ease of AI (Fanni’s thread)

• Fanni: the conference’s recurring theme is the re-imagination of human connection — people long for connection yet find it easier to turn to AI.
• Why AI feels easier: it gives the illusion that hard work can be replaced easily — coding used to induce “time blindness” (an hour feels like 5 minutes) but felt good when done; AI does it in seconds — but the output turns more generic, less creative.
• Product managers create much of the tech content online (good at explaining, writing docs, interviewing); over the last 6–12 months many he looks up to turned to AI to churn out more content, feeling more efficient / more family time — “and all I see is generic; I started to mute and even block some of them.”
• “A lighter form of AI psychosis”: you believe the AI output is the same quality as before; maybe the first time it is, but you don’t notice it slowly degrading.

2.1 Trusting AI like a human is a mistake

• The Meta security vulnerability: engineers didn’t think they were introducing regressions by trusting AI and AI reviews, but shipped “the most embarrassing security-inflicted vulnerability” to production in a big monorepo — thousands of people, no one noticed, because they’d all started to just trust AI. (Covered in his keynote.)
• Pairing analogy: with a human (“Jane”) you build graduated trust (small → medium → large tasks) but still review the truly critical stuff. With AI we start treating it like a human, but it ”will randomly hallucinate,” and the model can silently degrade (fewer resources behind it under traffic, or a new model). Humans aren’t used to trust that degrades.
• Better mental model (from Magda Bernecker, ThoughtWorks, and Simon Willison): think of AI not as a human but as a weird alien — human-like abilities but strange in many ways.
• Bottom line: hard work is never wasted. If you become “an AI prompting machine,” you’re easily replaceable; those putting in the hard work will get ahead.

2.2 The host’s pet peeve — AI-crafted messages

• 360-degree feedback now finishes on time but reads like “an answer from Claude”; Slack messages visibly Claude-crafted. “You don’t need to craft it — just tell me what you want. English isn’t my first language, but trust me, I’ll understand it.”

3. The Decline — and Renewed Importance — of Engineering Management

3.1 Management was already under threat before AI (end of ZIRP)

• In US/Western-Europe VC-funded companies, engineering management is “almost dying” — starting with the end of zero-interest-rate policy, not AI.
• Sourcegraph (code-search tool; raised ~$200M, grew to ~100 people): around 2022–2023, before AI mattered, they laid off all middle managers, with engineer-leaders reporting to founders — a cost exercise.
• Reason: companies realized they wouldn’t grow as fast; a big management layer is needed to absorb doubling head count, not a flat/shrinking one.
• Twitter: managers had to code AND have ~20 direct reports — but if you code, you do no people management.
• AI accelerates the trend: the thinking is “every engineer can use AI agents, so they do more, and managers can too” (e.g., AI agents collecting performance stats).
• Now large companies do it: Coinbase laying off middle management; Meta converting managers back to individual contributors in every layoff.

3.2 Why tech had small manager-to-report ratios

• Other engineering fields (mechanical, chemical) run ~40 reports per manager; tech ran ~6–10 for reasons:
  • Retention — a manager who cares reduces attrition (important when keeping engineers matters).
  • Managing change — technical managers say “stop, let’s create a tiger team (you, you, you) to build better canary rollouts for 2 months,” explain the why to stakeholders, and do a lot of invisible good work.
• Now there’s “no appreciation for that” — CEOs are “busy vibe coding” and think they’re super productive. “We’re going to rediscover in 2–3 years that good management is important” — hands-on, empathetic, big-picture.

3.3 Leadership has never been more important (people are lost)

• The happiest engineers are where leadership is firm on where they stand:
  • Zach Sward (co-founder, opencode, a Claude-Code competitor): “We’re in the AI business but I think we’re overusing AI, we have quality problems, let’s slow down and build fewer things with higher quality” — and “I don’t think competitors are killing us because they use AI better.”
  • Linear’s CEO (jokingly, “with heavy heart”) announced they would not lay off anyone and keep investing in people, reinforcing quality (everything to 300 ms).
  • Kelsey Hightower (recent podcast guest): challenges startups to ”explain what you do without mentioning AI” — forcing the important things (service quality, growth, churn).
• Good leadership can come from anyone on the team — people rally around those with confidence in the right things who won’t just repeat the hype.

4. Team Sizes: Two-Pizza → One-Pizza

• John Deere (a ~200-year-old tractor company) and others: two-pizza teams (~8–10) are turning into “one-pizza teams” (~3–4 people) — smaller, but still teams.
• Host’s observation: a single person who’s product-minded, design-aware, and technical can, with an agent, ship something in days (hackathon-style) — is that good/sustainable, or a bubble?
• The teenagers-in-a-bar analogy: the industry is like teenagers who found the bar unlocked, got drunk, and keep drinking through the hangover; adults would enjoy the fine liquors and call it a night.
• Teams doing really well with AI are doing roughly what they did before:
  • Linear (~50 engineers, taking on Jira): 2–3-person teams (sometimes individuals), all full-stack and product-minded, ask for help, regular demos, ship only quality things — unchanged with AI, which accelerates them “but not that much.”
• The pace of business hasn’t sped up: SaaS customers don’t want a new product every day. Faster bug fixes are the one clear win; new features/products still need months and time to see if they work.

4.1 OpenAI as a cautionary tale of shipping fast

• OpenAI Agent Builder (a visual agent-workflow UI): built in 6 weeks with Codex, released October with fanfare — buggy, and OpenAI never went back to fix it → abandonware.
• Sora (AI-generated social site): launched to mixed reception, shut down ~6 months later.
• But maybe it’s not all stupid: it builds the muscle of building from scratch.
  • Google famously kills projects (killedbygoogle.com — Reader, Google+, Weave…) — seems wasteful; Amazon supports AWS services to end-of-life (~120+ services; recently retired 12 after 5 years) and “does not waste effort.”
  • Yet in the gen-AI race, Google is the only big-tech keeping up with the labs (NotebookLM, Gemini, Vertex), while Apple has no AI strategy, Amazon struggles (Nova “a joke”; its wins are hosting OpenAI/Anthropic on Bedrock), and Microsoft Copilot is “an embarrassment.”
  • Gergely’s theory: Google never forgot how to build, launch, and scale — so the “wasteful” habit built the muscle. “Pick your poison.”

5. Engineering Culture (a hijack)

• Talking to people inside companies: the only places where people are truly happy are OpenAI and Anthropic (and fast-growing startups like Ramp, Supabase).
• Everywhere else people say “our culture sucks” — including inside Google’s Gemini team (interviewing to leave for OpenAI) and the Copilot team (terrible morale, infighting, even when doing well).
• Maybe in a change this big it’s normal for everyone to think it’s better elsewhere — “people change companies and realize it’s the same [mess] there.”
• “We’re in the middle of a bigger change than we’ve ever seen — make notes; in 2–3 years these will be great stories” (“let me tell you what it used to be like”). Fanni: “It’s a paradigm shift.” Gergely: “That’s a fancy way of saying it.”

6. Bugs, New Features, and Resistance to Change

• Host: users hate new features (“why do I have to see this new button?”) and instead want their daily-use bug fixed — which sits low in the backlog.
• Gergely (from Corey at Linear): have a zero-bug policy — fix bugs really quickly; AI is genuinely good here. Linear stopped work for 2–3 weeks, fixed all bugs (with AI), and now maintains an absolute zero-bug policy (hard bugs may need a human).
• Use AI like outsourcing — outsource what you don’t care about; but you do care about not having bugs, so use AI to generate one-shot suggestions you take or reject.
• Developers hate change (from Chris Lattner, creator of Swift and LLVM): before LLVM (~2005) there was GCC; people said a more modular compiler “will never work,” but LLVM won and now powers Objective-C, Swift, and more. 15 years later, Lattner proposed a big change to LLVM and the long-time maintainers said “nah, that’ll never work” — “developers just hate change.” AI is changing how we work with no best practice yet, so it’s a stressful situation we must push through.
• Fanni: it’s human nature — we all fear change and seek security and variety; human connection is the basis, and conferences let us come together and share our fears.

7. Google’s Silent Bundling, and Token-Maxing → Token-Optimizing

7.1 Google quietly killing SaaS businesses

• Host: Google silently bundles features (e.g., a meeting-transcript feature into Google Meet) that kill standalone SaaS products; Gemini capabilities are under-discussed, while other big companies struggle. Even Apple is reportedly using Gemini to level up Siri.

7.2 Token-maxing is already dead

• Token-maxing (“world’s most defunniest trend”) appeared ~a month ago at Amazon/Microsoft/Meta because of leaderboards and no cost attached to AI usage.
• It’s basically over: the labs turned on API pricing everywhere, and GitHub Copilot killed its unlimited plan → very real, measured cost. “The only places still doing it are idiots who’ll get a big bill at the end of June.”

7.3 The new trend: token-optimizing

• Engineers getting good reviews / spot bonuses are the ones finding ways to save a percentage of tokens (heard from friends at Shopify).
• Smart token routing (opencode, Factory AI, Devin) picks the right model for the task.
• Next: companies running DeepSeek or other cheap models on their own infrastructure — “we’re speed-running everything” (cf. cloud-cost optimization, then the Datadog-bill panic ~2023).
• Only the AI labs (and cursor/OpenAI et al.) still don’t care about usage — they spend crazy tokens but are rushing toward IPOs.

8. IPO Signals and the AI Plateau

8.1 IPOs may indicate the top

• Anthropic has filed for an IPO (fall); SpaceX has filed too.
• Historically (~2000s) companies went public because they couldn’t raise more venture funding (Google IPO’d small in ~2004 and rose ever since).
• Now investors give companies money “until forever,” so going public — with its reporting/financial-disclosure burden — is only worth it when that money can’t come from anywhere else, which suggests being closer to the top.
• Aside: SpaceX shares have long been buyable by insiders (ex-Uber alumni networks); IPO opens it to retail investors for the last big raise.
• Suspicious timing: just as Anthropic goes public, it said it “would like to pause all AI research across the world” — “how convenient, just as they’re in the lead.”

8.2 Progress feels like it’s slowing

• ”Opus 4.7 and Opus 4.8 are the first two models where I’m just not impressed” — feels like regression in ways, hard to measure.
• He’s hopeful we may have hit a plateau (“it would be nice to catch our breath”): even if AI froze today, “it will take us a decade to integrate it” — “I can describe what I want in imperfect terms and get code that represents my thought process.”
• Skepticism of the scaling bet: labs bet that doubling the training fleet yields exponentially better capabilities — “I’m not seeing that necessarily.” He also hears many complaints that Anthropic’s safety guardrails now refuse coding tasks (“that’s too dangerous” when touching an HTTPS library). Models have been “good enough for coding since November 2025.”

9. Career Advice for (Aspiring) Engineering Managers

9.1 Why management is valuable (his Uber experience)

• Offered a manager role at Uber while acting as team lead; hesitated (he likes coding and is good with people, which doesn’t require being a manager), but said yes to “see how the sausage is made” — hiring, making the case for head count, firing, performance management, politics.
• At Uber you could go back to IC keeping the same salary; people who spent a year+ as manager and returned became the best right hand to their manager (“I did your job, I have respect for it, let me take stuff off your plate”) and their careers went up.

9.2 Managing people ≠ managing agents

• “I always laugh when people say ‘with agents you’re managing agents.’ No — management is about people problems”: someone calls in sick, burnout, two people fighting over something stupid, misunderstandings.
• His own example: he canceled one-on-ones and team meetings last-minute (looked lazy/like a jerk) because, behind the scenes, someone up the chain wanted to fire a team member over a stupid metric, and it took 2 months with HR to stop it — invisible work he couldn’t explain. “Management is fixing the invisible broken stuff — the job isn’t very visible.”

9.3 The advice: build AI infrastructure, and keep thinking

• Reality: the work is undervalued now (CEOs/CTOs fixated on AI), and the job market is bad — even in a lousy place people won’t leave quickly; pay raises: forget it for a few years (appreciate one if you get it).
• For career stability (if your layer gets cut, you can get a job): get hands-on with how AI works — every company hiring engineering management wants someone who can figure out what AI infrastructure fits the organization.
  • Build AI infra at your company: commit a little, talk with devs about their experience, understand it end-to-end. “If you build that out, you’ll be in so much demand.”
  • Do less of the other stuff: cancel half your one-on-ones (weekly → biweekly, biweekly → monthly), and use the time to think “how can I help my org move better — where are the feedback loops I can make faster?”
• The Kelsey Hightower call-center story (pre-AI): hired as a call-center agent doing repetitive tickets (password resets, ~5 min each) with a huge backlog. One day he stopped taking calls and built self-service automation (e.g., a self-service password-reset flow); his manager was supportive, more people joined, and 2 months later the center processed far more tickets far faster. “Ask forgiveness, not permission.”
• Don’t stop thinking (from Dax Raad, opencode): open code succeeded on positioning — he looked for a gap (“no popular open-source AI coding harness”); his job used to be 95% thinking / 5% coding, and with AI it’s now 96% thinking / 4% coding. “I take that as inspiration.”

People & References Cited

• Gergely Orosz — guest; Pragmatic Engineer; ex-Uber engineer/manager.
• Hosts — ABK Podcast (A Bátrak Klubja), incl. Fanni.
• Michael Novati — ex-Meta “coding machine” archetype; the AI-slop-writing dispute.
• Magda Bernecker (ThoughtWorks) & Simon Willison — “AI as a weird alien” model.
• Zach Sward — co-founder, opencode.
• Dax Raad — opencode; “95%→96% thinking.”
• Corey (Linear) — zero-bug policy.
• Kelsey Hightower — “explain what you do without mentioning AI”; the call-center automation story.
• Chris Lattner — creator of Swift and LLVM; “developers hate change.”
• Companies: Uber, Meta, Sourcegraph, Twitter, Coinbase, Linear, John Deere, OpenAI (Agent Builder, Sora, Codex), Anthropic (Claude, Opus 4.7/4.8), Google (Gemini, NotebookLM, Vertex, killedbygoogle.com, Meet), Apple (Siri), Amazon (AWS, Nova, Bedrock), Microsoft (Copilot), Ramp, Supabase, Shopify, Factory AI, Devin, DeepSeek, SpaceX, Datadog, GCC.
• Concepts: Pragmatic Engineer blog, Hacker News, AI slop, AI psychosis, graduated trust/hallucination, ZIRP, tiger teams, two-pizza/one-pizza teams, zero-bug policy, token-maxing vs. token-optimizing, smart token routing, IPO-as-top signal, AI plateau/scaling bet, build-AI-infra career advice.

Video: https://www.youtube.com/watch?v=Xd__vUlc1F4 — Transcript via yt-transcript.sh; outline generated from the transcript.