This is an AMA episode of The Pragmatic Engineer podcast where Gergely Orosz answers subscriber questions read by Vladimir Gignak (CTO at Wordsmith), covering his career transition from Uber engineering manager to full-time tech writer, AI’s impact on software development and hiring, big tech dynamics, career advice, and the business of The Pragmatic Engineer newsletter.
From Uber to writing
Gergely left Uber in 2021 after four years (first as IC, then engineering manager) when COVID layoffs hit and his team’s mission dissolved.
He had ~$400k post-tax from Uber stock (originally $500k grant) giving him 2-3 years runway.
Original plan: finish “Software Engineer’s Guidebook” in 6 months, then start or join a startup — he was tired of middle management politics, especially handling layoffs across European regulations.
Writing the book took far longer than expected (similar to unknown software migrations); he wrote three shorter books but the main one stalled.
His brother (second-time founder) advised: only start a startup if ready to commit 10 years; Gergely wasn’t excited enough about his RFC-tool idea.
Two motivations for startup: potential “fuck-you money” (exit → $25M post-tax) and desire for small “us vs world” teams — but realized if successful he’d just end up writing anyway.
Saw Lenny Rachitsky had 2,000 paid subscribers for product management newsletter; reasoned 10x more software engineers exist, no paid newsletter served them → tried Substack for 6 months, it worked immediately.
AI-native SDLC
“SDLC” means different things: traditional waterfall (year-long planning), agile/scrum/SAFe (rigid ceremonies at non-tech corps), or modern tech-company flow (light planning → code → deploy → iterate).
Closest real “AI-native SDLC” at scale: Anthropic (Claude Code team) — hundreds of engineers, research-lab culture, minimal design docs, constant prototyping, heavy dogfooding, rapid iteration from user feedback.
Questions remain: how much strategy/planning happens? Pricing changes frequently. Reliability concerns (e.g., Spotify outages despite “responsible AI” claims).
Most big companies aren’t retrofitting SDLC; they’re building internal AI infrastructure (custom agents plugged into internal services) — Google, Ramp, Uber all doing this.
Business pace hasn’t changed: Uber still needs weeks of driver outreach for events; going too fast risks forgetting basics.
AI and hiring
Hiring is becoming messier, more friction, more subjective — no clear rules yet.
Pre-AI two worlds: Big Tech (LeetCode for raw intelligence, CS basics, pressure tolerance, scalability, willingness to endure BS processes) and startups (practical take-homes, open-source contributors).
AI breaks both: algorithmic interviews solvable remotely; take-homes completed well by AI.
Emerging model: filter via take-home (AI allowed), then in-person/whiteboard for Big Tech; startups do deep discussion of candidate’s homework — checking reasoning, research, course-correction ability, not just AI output.
Linear scales “work together” trial weeks but loses candidates who can’t take time off.
As a candidate: more time investment, feels unfair, more subjective evaluation.
Engineers currently thriving
Engineers in high demand: at startups/known tech cos, product-minded (don’t stop at borders), early adopters of AI — many moved into AI infra roles (RAG, fine-tuning, model selection, inference costs, Groq/Cerebras decisions).
5 years ago: hired cloud-savvy engineers; now: hire inference/AI-infra-savvy engineers.
Compensation asks high (from Google/Meta/well-funded startups) but demand exceeds supply.
Struggling: engineers with no AI-building experience at current job, no “modern company” pedigree — hard to jump tiers (consulting → product → venture-funded → AI labs).
Junior roles: saturated for web product; low-level/embedded/hardware less saturated, different world (less AI assistance ~30% vs ~100% for high-level).
Advice for juniors: get pedigree (top school/internship), build impressive side projects, contribute to OSS (AI contributions often rejected), accept stepping-stone jobs, excel wherever you are.
Meta’s war mode
Leadership (Zuckerberg) likely sees existential threat in AI ownership — not that they don’t see morale damage.
Meta unique among big tech (Google, Microsoft, Uber) in laying off 10% + reassigning 10% without consent after record year.
Pattern: Meta went “wartime mode” for Google+ (2010s), then metaverse, now AI — but revenue is high, ads business strong, products growing; employees don’t see the enemy.
Best engineers (10+ year tenure, good WLB) reassigned to unwanted work (data labeling in “AI ADO” org) — some making best of it, but signal: leadership no longer cares about engineering as a whole.
Speculation: Zuckerberg wants to break out of application-layer dependency (no platform ownership), possibly reactive.
AI at Big Tech vs. startups
Google: trying hardest — free reign for internal AI tools (chaotic but productive), only big lab with competitive model (Gemini eating ChatGPT share), e.g., editor prefers Gemini for queries.
Meta: bogged down training own models, morale dropping, people don’t see the point.
Microsoft: political — Copilot vs Core AI orgs, GitHub under Core AI, reliability issues, Azure capacity fights, more focused on politics than AI.
Apple: secretive, engineering culture “absolute trash” (duct tape everywhere), not doing much visibly — hope: local AI on hardware leveraging their strength.
Amazon: trying hard (Q internal tool, own models) but all subpar; engineers drag feet, prefer Claude Code — example of how hard retrofitting innovation is at scale.
“Little tech” (public but smaller: Uber, Ramp, Intercom, Block) doing better — no identity crisis, don’t need to own full stack, integrate best tools (Claude/Codex), optimize for business.
Anthropic / Claude Code as model
Anthropic (Claude Code) and Codex team are best examples of AI-native dev at scale — but hard to copy because Anthropic is an AI lab: product is the model (Claude), Claude Code is byproduct/revenue generator until model is perfect.
Everything revolves around model training (pre/post-training every few months); tooling is beehive around that one thing.
To copy: become an AI lab where tooling is byproduct — not replicable for normal companies.
Interest: how startups change workflows — but many stuck at traction phase; AI-native doesn’t matter without customers/market.
Coinbase example: very AI-native, but crypto market dictates results — AI-native may reduce headcount but not change business fundamentals.
Better strategy: use AI where it naturally solves a problem (e.g., incident response first pass), not “AI-native” as ideology.
Tech debt vs speed with AI
False dichotomy: not speed OR quality — segment by time (stage) or codebase area.
Uber 2016 example: terrible architecture (polling every 5s, huge blobs) unblocked mobile/web teams for years — tech debt sped them up early.
Ken Beck’s 3X: Explore (prototype, quality irrelevant), Expand (product-market fit, hacks OK to scale), Extend (mature, pay down debt).
AI also enables faster refactoring — no excuse not to do it periodically.
Infrastructure: more quality attention; product: more speed attention.
Standards in AI tooling
Too early for standards; AI changes fast.
MCP (Model Context Protocol) from Anthropic emerged accidentally: small non-threatening lab → big cos adopted (politics: not Google/OpenAI).
If Anthropic tried MCP today, cos would resist lock-in.
Standards will emerge organically, not by design.
Types of engineering managers
No single right model; tradeoffs:
Non-coding EMs: more people focus, fix org/personal frustrations, HR/policy changes, cross-team systems.
Coding EMs: more technical guidance, better technical discussions, less bandwidth for people/org work.
Industry swinging hard toward “managers must be technical” — people-management-focused EMs will feel underappreciated.
Pendulum likely swings back eventually.
Elon Musk at Twitter/X mandated coding EMs with 20+ reports — “pretty insane.”
Preventing AI adoption theater
Theater phase (leaderboards, mandatory usage, token volume targets) largely over — was relevant during autocomplete/GPT-4.0/Cursor-tab era.
Shopify had token leaderboards then deprecated them.
Now: strong models (Opus 4.5, GPT-5, Claude Code) → everyone uses it naturally; measuring usage like lines of code — meaningless.
Measuring AI productivity
Business productivity = incremental revenue (new money from AI-enabled products) OR cost savings.
Crypto exchange example: more volume = market, not AI.
AI labs (Anthropic, OpenAI) and AI-product startups (AI incident review) show clear AI revenue.
Otherwise “iffy, finicky” — AI may be more like cloud (ubiquitous infrastructure, cost flexibility) than mobile (new market creator).
Gergely’s Uber lens: always knew how his team made money; could answer “what if I hire/lose 2 people?” in revenue terms.
Misconceptions about AI
“AI makes things easier” is wrong — if your life gets easier, you’re not trying hard enough or delegating too much.
Gergely finds work harder with AI (thinking just as much or more).
The value of CS degrees
CS degrees becoming prestige gatekeepers like law/architecture — not due to curriculum but market.
2015-2020: bootcamps (3-12 mo) got well-paying jobs due to shortage; ended.
Top unis (MIT, Caltech, Waterloo, Imperial) still hunted but fewer competing offers; mid-tier harder.
Self-taught with 5 years exp + SR/infra work couldn’t find role for 1 year.
Degrees matter for: employer filtering (reduces resume volume), visas (critical for moving West).
Easier than ever to do own thing, but companies pickier.
AI at Pragmatic Engineer
Gergely spends most time researching/writing; increasingly builds backend for newsletter (group subscriptions, support) — CRUD on Render.
Uses Codex (GPT-5), Claude Code, Cursor, Factory — rotates tools.
No AI for writing: experiments produced artificial output; loves writing as thinking process — best social posts are byproducts of revisiting topics.
Analogy: favorite photography YouTubers are working photographers with side channels, not pro YouTubers.
Uses deep research for topics (e.g., “Ramp engineering culture”) — frees some time but didn’t reduce work; now over-relies on deep research (web-finding skill may atrophy, but doesn’t trust internet anyway).
Coding fluency degrading — okay with that tradeoff.
AI tempts building more software (less intimidating) — building self-service signup flow now.
Future-proofing your career
Best future-proofing: work at company doing relevant modern product work with AI experimentation.
Banking/rigid places may block this — but many leaders want AI adoption; propose internal AI project (part-time, internal tool) — win/win.
Google example: encourages AI experiments on product teams.
Hands-on beats full-time CS degree (curriculum lags industry); part-time degree possible but harder.
Side projects only if genuinely motivated (e.g., health app you need); otherwise do at work.
Finding motivated peers
If classmates/colleagues unmotivated: find different friends (online communities, other classes, change teams/companies).
Ali Spittel (high school) joined online communities, contributed to OSS.
Gergely: banking colleagues nice but not tech-obsessed; Skype — everyone obsessed, huge difference.
Advice for aspiring game developer (high schooler)
Construction analogy: DIY tools/materials/YouTube exist, but pros hired because most don’t want to do the work.
Game dev in 10 years: studios (startup or AAA) will hire grads from top unis with side-game portfolios.
Reference: Jonas Tyroller episode — built games 10-20 years on side, one hit 1M sales with 2 people.
Start building games now.
EU job market
Tier model: Tier 1 = local (supermarkets), Tier 2 = regional, Tier 3 = global (Big Tech).
Volatile market → staying put can be good, but don’t stop looking.
2023: brutal, no hiring; now: layoffs but many companies hiring — opportunity to jump tier (startup, product, autonomy, AI tools).
Engineers with hands-on AI experience in high demand; zero years experience = stuck.
Be opportunistic: check openings, talk network, don’t ignore recruiters — can always decline offers.
Using AI for learning
AI helps only when you want to learn — start with goal, use as tool.
Don’t discard books, videos, tutorials, building things.
Biggest misconception: AI doesn’t make learning easier when unmotivated.
One fewer excuse to learn; if you don’t want to, don’t.
Making money as a creator
Last public numbers: Year 1 — 2,700 paid subscribers; now >10,000 + podcast sponsors.
Avoids specific revenue to avoid “how do I do this?” coaching requests.
Best Uber year (Netherlands): €288k ($320-330k) — €120k base, ~€30k bonus, rest equity.
Stopped tracking money; focused on one great article for 1.5-2 years.
Realized: own business can exceed Big Tech pay; loves daily autonomy (“I’m sitting here because I want to”).
15 years as developer built skills, structure, connections — guests often from network.
Not attached: if business fails, okay as long as helped people; leaves important content free (e.g., Meta article) even if hurts revenue.
What’s next for The Pragmatic Engineer
No VC funding → no forced expansion.
Goal: make Pragmatic Summit regular — SF annually (Feb), add Europe (London) annually.
Growing team slowly; wants more ambitious research (e.g., utilities companies — boring but critical).
Bunq and Pollen
Two near-legal incidents:
Bunq (Netherlands neobank): Early article on hiring practices (intelligence tests pre-tech interview). Got juicy insider stories, wrote damning piece (hit piece). Slept on it: realized article had zero positives, company employs/grows people, one Egyptian engineer said Bunq gave him visa chance when no one else would → now at Meta. Decided not to publish; deleted it. Journalists later asked for details.
Pollen (events co): Covered layoffs; COO dismissed on all-hands (“not BBC, small pub with agenda”). Gergely investigated: unpaid salaries, canceled health insurance, deliberate double-charge at outage. Sent article → Pollen cried “libel” → self-censored (stress, effort). BBC later did documentary (Gergely helped). Realized investigative journalism not for him.
Spotting trends
Ideas fester; long collection list.
Trends pop when multiple people mention simultaneously.
Example: Jan 2024 — used o1/Claude heavily over holidays, impressed; researched, saw same sentiment → wrote “coding by hand is over” early; got flak (“AI shill”) but conviction from experience + evidence.
Book updates
Guidebook surprisingly durable for AI — light on coding, heavy on non-technical (business understanding, architecture) which is more relevant.
Will update lower levels when industry settles on actual working practices (best practices for AI-era coding) — may take a while.
Favorite books & tech products
Philosophy of Software Design (John Ousterhout): only book comparing architecture approaches via student groups; deep vs shallow modules concept.
Tidy First (Kent Beck): thin, crisp ideas; thinking behind it valuable even if writing less code.
Perplexity Deep Research: fastest deep research; wishes Google did this for search; pays for it, no affiliation; dislikes their new “computer” push.
What won’t change in engineering
In 5 years: just as big (or bigger) demand for professionals who care about the craft.
True professionals: know industry state, know tools, used most, understand tradeoffs, no ego, choose right tool for job.
Involves: coding tools, testing, deployment, verification — care about things average person wouldn’t (like architect seeing structural details vs pedestrian seeing glass).
High-confidence changes with right tooling (scaffolding when needed).
AI may scare away those who never cared about software — just wanted quick buck.
