The Lost Guide for Creating Software in the Age of AI

Transcript

1. A 45-MINUTE TALK FOR FINAL & PRE-FINAL YEAR CS / SE STUDENTS THE LOST GUIDE FOR CREATING SOFTWARE IN THE AGE OF AI github.com/thecloudranger · CNTXT · AI Tinkerers UAE & Riyadh · Google Cloud Developer Group Riyadh 🗺️ ...IN THE AGE OF AI 1 / 45
2. > WHOAMI EXP 22+ years · Backend engineer → Solutions Architect -> Principal Architect / Managing teams · Led AWS Developer Advocacy in the region COMMUNITY Builder @ASEAN and @MENA · UAE Coding Ambassador @ coders(hq) MENTORSHIP Technical leadership · How to build a career path in tech HUAWEI AWS coders(hq) · نـيجمربملا رقم AI TINKERERS UAE & RIYADH 2 / 45
3. 20% Junior developer employment dropped from its 2022 peak Stanford / ADP · Brynjolfsson et al. · Nov 2025 3 / 45
4. Engineers of 35+ years of age in same AI-exposed roles grew 6-12%. This is not an age story. It is a depth of knowledge story. 4 / 45
5. WHAT DO THE PEOPLE ON THE RIGHT SIDE OF THAT NUMBER KNOW? By the end of this talk: 5 mental models + 3 habits + 1 accountability insight 5 / 45
6. SECTION 01 HERE WE GO AGAIN. A brief history of panic 6 / 45
7. A BRIEF HISTORY OF ABSTRACTIONS Every generation thought the current wave would end software careers. None did. Punch Cards & Machine Code 1950s Programmers spoke the machine Assembly Language 1960s Mnemonics replaced raw binary High-Level Languages 1970s C, COBOL, Fortran abstracted hardware OOP & Frameworks 1990s Java, patterns, reusable components Cloud & Platforms 2010s Managed infra and serverless AI-Assisted Development 2020s Copilots, agents, code generation Every new abstraction made us more productive, but never eliminated the need to understand what's happening underneath. 7 / 45
8. THE RECURRING PANIC CYCLE Each era had its "developers are obsolete" moment. Each was wrong about replacement. 2000S "Drag-and-drop will replace coding." IDEs and code generators arrived. 2010S "Anyone can build an app now." Frameworks and low-code platforms arrived. 2018+ "Developers are the new buggy whip makers." No-code platforms arrived. 2023+ "Why learn to code at all?" AI code assistants arrived. You are here. Yet here we are. 8 / 45
9. "We've been here before. We survived the cloud. We survived containers. We survived Kubernetes, BARELY, but we did. We'll survive this too. Probably." Michael Stahnke · cfgmgmtcamp 2026 9 / 45
10. "You may not be replaced by AI. You will be replaced by somebody using AI better than you are." Practitioner wisdom · 2025 10 / 45
11. SECTION 02 WHAT A SOFTWARE ENGINEER ACTUALLY DOES. 11 / 45
12. …right? Write Code. Go Home. 12 / 45
13. WHAT A SOFTWARE ENGINEER REALLY DOES 📋 Requirements & Planning ⚙️ System Design & Architecture ⌨️ Writing Code 🐛 Testing & Debugging 🔍 Code Review 🚀 Deployment & CI/CD 📊 Monitoring & Performance 💬 Communication & Collaboration 🔒 Security & Compliance 📝 Documentation 🧑‍🏫 Mentoring & Leadership 🧩 Problem Solving Writing code is just one piece of the puzzle 13 / 45
14. BUT THE STAKES MATTER. Vibe coding and engineered systems are worlds apart. Same tools. Completely different responsibility. VIBE CODING Personal budget tracker Side project landing page Internal tool prototype Hobby app · Proof of concept If it breaks, you adjust your budget. CONTEXT ENGINEERING HR payroll system: 10,000 employees Healthcare records platform Financial trading system Anything with compliance or SLAs If it breaks, thousands don't get paid. 14 / 45
15. CODE IS CHEAP. SOFTWARE IS NOT. 15 / 45
16. THE PROCESS CODING WAS ALWAYS A PROCESS. Some phases got faster. Some got more important. None disappeared. 16 / 45
17. THE SDLC BEFORE AND AFTER AGENTIC AI Claude Code doesn't skip phases; it compresses some and surfaces others. UNDERSTAND Read requirements, clarify with stakeholders, model the domain. Slow. Often skipped. Always regretted when skipped. DESIGN Architecture decisions, data modelling, interface contracts. Whiteboard sessions and long async threads in Notion. IMPLEMENT Write code. Most of the calendar. Most of the pull requests. The phase everyone optimises for. TEST Write tests after the fact. Often skipped under deadline pressure. Technical debt accrues silently. → → → → UNDERSTAND Spec first, code second. A precise prompt is the spec. Vague intent produces vague output, at 100× the speed. The thinking phase is now load-bearing. MORE CRITICAL DESIGN Modular, small-context architecture is now a prerequisite for the agent to work well. Good design is no longer optional; it's how you give the AI a fighting chance. MORE CRITICAL IMPLEMENT Claude Code handles boilerplate, scaffolding, repetition, and first-pass logic at speed. The phase that consumed most of the calendar now takes a fraction of it. ACCELERATED TEST AI-generated code has no fear of being wrong; only you can verify correctness. Tests written before implementation are now your primary control mechanism. MORE CRITICAL 17 / 45
18. THE ONE THING THAT CHANGED IMPLEMENT GOT CHEAPER. EVERYTHING ELSE GOT MORE IMPORTANT. The engineer who only knew how to implement is the one being displaced. The engineer who owns understanding, design, testing, and review: that's who the AI works for. 18 / 45
19. SECTION 03 THE BEDROCK. Principles that survived 50 years 19 / 45
20. TOOLS CHANGE. FOUNDATIONS DON'T. These were true for punch cards. They're true for AI agents. 20 / 45
21. PRINCIPLES THAT PREDATE THE INTERNET AI eliminates more accidental complexity with every model release. Essential complexity is still yours. BROOKS '86 Essential vs. Accidental Complexity: AI eliminates syntax, boilerplate, scaffolding. It cannot touch the conceptual difficulty of what to build or why. When AI saves you an hour of boilerplate, use that hour to think harder about the design. DIJKSTRA '74 Separation of Concerns + Modularity: AI generates monolithic globs when not constrained. The engineer who knows SoC spots it instantly. "Encapsulation gives AI (and humans) focus." (Byron Salty) DRY · KISS · YAGNI Simplicity as a discipline: AI loves to duplicate logic, over-engineer, and build for hypothetical futures. "It suggested microservices. For a to-do app." (Michael Stahnke). You are the corrective force. BROOKS '75 Conceptual Integrity: A system must feel designed by one mind. AI-assembled fragments lose this. You are the single mind that holds the whole system together. 21 / 45
22. LEARN SKILLS THAT STICK The Lindy Principle: the longer a technology survives, the longer it will continue to survive. UNIX shell 1971 Still the fastest way to operate a system. Still everywhere. SQL 1974 Every company has a database. Every database speaks SQL. Vim / text editors 1990 The terminal doesn't go away. Neither does the skill. Python 1991 Readable, everywhere, and the default language of AI/ML. Go / Rust 2009/2015 Systems-level clarity. Growing fast. Worth the investment. Kubernetes YAML 2014 The ops standard. Boring. Universal. Lindy. "The longer a technology lives, the longer it can be expected to live." (Nassim N. Taleb, Lindy Principle) 22 / 45
23. SECTION 04 FIVE MENTAL MODELS FOR RIGHT NOW. Mental models are shortcuts for good judgment in situations you haven't seen before. 23 / 45
24. MENTAL MODEL 1: VIBE CODING VS. AUGMENTED CODING Kent Beck / Andrej Karpathy · 2025 🎲 VIBE CODING "I Accept All always. I don't read the diffs anymore. When I get errors I just paste them in. I care about behavior, not code." Legacy code on day one. Cognitive debt compounds. 🔬 AUGMENTED CODING "The value system is the same as hand-coding: tidy code that works. I just don't type much of it. I care about complexity, tests, and coverage." Craftsmanship preserved. You own the output. After AI generates a block of code, ask: Could I explain this in a code review? Could I debug it at 2am when production is down? 24 / 45
25. MENTAL MODEL 2: AI IS AN AMPLIFIER DORA 2025 · Google Cloud · ~5,000 professionals AI amplifies you. If your foundations are solid, AI makes you faster. If they are thin, AI makes the thinness visible, faster. The multiplier depends entirely on what you bring to the tool. HIGH-MATURITY TEAMS AI converts to faster delivery, better quality, higher throughput. Foundations × AI = 10× FRAGMENTED TEAMS AI accelerates technical debt and delivery instability. Throughput gains erode under churn. Thinness × AI = 10× more visible 25 / 45
26. MENTAL MODEL 3: INVERSION Charlie Munger / Faz Fazalullah · "Invert. Always invert." "If I wanted to fail as a software engineer in the AI era, what would I do?" FAILURE PLAYBOOK Accept AI output without reading it Skip debugging because AI "handles it" Use AI to avoid discomfort, not accelerate learning Measure output (tickets closed) over outcome Build things you can't explain YOUR PLAYBOOK Read every diff. Understand the why. Practise debugging manually. Keep that muscle. Use AI to accelerate learning, not bypass it Track outcomes. What moved? Why? Ship things you can defend in any review 26 / 45
27. MENTAL MODEL 4: OUTCOME ≠ OUTPUT AI inflates output. It cannot verify outcome. That is your job. You say Business hears "Fixed the N+1 query" Page load: 4s → 0.8s. Conversion up 12%. "Refactored the auth module" SSO now possible. Enterprise contracts unlocked. "Wrote tests for payment flow" Payment failures reduced. $X in recovered revenue. "Set up monitoring" We catch issues before customers do. SLA met. "Ticket closed" ??? Junior engineers ship features. Senior engineers ship outcomes. Staff engineers shift markets. 27 / 45
28. MENTAL MODEL 5: SPEC FIRST, CODE SECOND Simon Martinelli · Byron Salty · U-Zyn Chua: three independent practitioners, same conclusion THE RULE Don't ask AI to build a feature. Ask AI to document the feature first. Review it. Iterate. Then implement. WHY IT WORKS The spec is the source of truth. Code and tests are derived from it. When requirements change, the spec changes first. THE NEW CORE SKILL Translating ambiguous human needs into precise, machine-actionable requirements. This is context engineering. This is what's irreplaceable. "AI is not a compiler; it's an assistant. You are responsible for the output." (Simon Martinelli) 28 / 45
29. SECTION 05 THREE HABITS THAT WILL SEPARATE YOU. Mental models are how you think. Habits are what you do. These compound. 29 / 45
30. BUILD IN PUBLIC GitHub > grades. A shipped project beats an A. 01 SHIP REAL THINGS One project a stranger can find, run, and use is worth more in any technical conversation than a semester of assignments. CONTRIBUTE TO OSS Find a project you actually use. Read its issues. Fix one small thing. Reading a real codebase (domain + implementation) is the skill universities under-teach and AI atrophies. WRITE 100 WORDS A WEEK Public. A post, a note, a LinkedIn update. Over years, this is a portfolio of thinking, not just code. Start this week, not next semester. 30 / 45
31. KEEP THE UNASSISTED REPS METR 2025 RCT: experienced devs were 19% slower with AI, even though they felt 20% faster. 02 DEBUG WITHOUT AI One debugging session per week, no paste. Observe → hypothesize → test. This is the most compressed path to systems understanding. READ 200 LINES / WEEK Someone else's code: a library, an OSS project, a PR. Understand it before writing anything new. In the AI era, you'll spend more time reading code, not less. HOARD WHAT YOU KNOW "Hoard things you know how to do." (Simon Willison). Skills internalised are permanent leverage. Skills only accessed through AI are borrowed leverage. 31 / 45
32. LEARN THE BUSINESS DOMAIN The most underrated skill in the room. AI is excellent at syntax and patterns. It cannot understand why this business cares about this constraint. 03 LEARN THE METRICS Pick the company you'd want to work at. What metric do they optimise for? Latency? Retention? Cost per transaction? Read their engineering blog. LEARN THE VOCABULARY Fintech: IBAN, SWIFT, settlement, reconciliation. Healthcare: HL7, FHIR, audit trails. This vocabulary is what makes you irreplaceable in AI-assisted work. ASK WHY, A LOT "Embrace your inner 5-year-old." (Gillian Armstrong). For every ticket: what business outcome does this feature move? If you can't answer, find out before writing a line of code. 32 / 45
33. SECTION 06 THE ACCOUNTABILITY ADVANTAGE. 250 years of displacement history, and one consistent pattern 33 / 45
34. THE PATTERN ACROSS 250 YEARS OF AUTOMATION Industrial Revolution → Electrification → Mainframe Era → PC Revolution → AI WHAT GOT DISPLACED Handloom weavers (250,000 → 7,000) Lamplighters · Switchboard operators Keypunch clerks · Ledger clerks → Undifferentiated task-defined roles WHAT SURVIVED AND WAS CREATED Licensed engineers who stamp drawings Factory inspectors · Systems analysts The "responsible person" attached to dangerous machines → Roles with named, enforceable accountability Engineering licensure (Wyoming, 1907) emerged specifically to attach a named, accountable human to consequential technical judgment. The engineer's stamp/seal means: "I accept personal and professional responsibility for this work." 34 / 45
35. THE LEGAL LANDSCAPE IS SHIFTING, FAST AI accountability is being written into law. These roles don't go away. Because AI cannot be the accountable party, the human who reviews and signs off on AI output becomes more legally and economically essential, not less. This is the structural argument for why depth matters. EU AI ACT · AUG 2024 Splits liability between providers (heaviest obligations: risk management, conformity assessment) and deployers (human oversight, monitoring). High-risk provisions: 2 Aug 2026. Someone has to sign off. That's a job. EU PRODUCT LIABILITY DIRECTIVE · DEC 2024 Software and AI systems are now explicitly "products" under strict, no-fault liability. Burden of proof eased for victims. Liability runs to manufacturers and economic operators. The accountable human in the loop is legally load-bearing. 35 / 45
36. SECTION 07 CONNECTING YOUR WORK TO THE BOTTOM LINE. 36 / 45
37. DEVELOPER VALUE PERCEPTION How the business measures your impact has changed, and it keeps shifting toward outcomes. Business Stakeholder View 1990-00 2000-10 2010-20 2020-26 Revenue Impact Digital Transform Innovation Rate Product-Market Fit Engineering Manager View 1990-00 2000-10 2010-20 2020-23 2023-26 Lines of Code Sprint Velocity Deploy Freq. DORA Metrics Business Impact ▶ ▶ The trend is always the same direction: from output metrics toward outcome metrics. Start thinking in outcomes from day one. 37 / 45
38. AI accelerates feature-shipping. Decide which lane you're building your career in. 👶 JUNIOR ENGINEER Ships features 🧑‍💻 SENIOR ENGINEER Ships outcomes 🧭 STAFF ENGINEER Shifts markets "AI does quite frankly reduce the floor and raise the ceiling for all of us." (Satya Nadella) 38 / 45
39. SECTION 08 YOUR COMPASS, NOT YOUR MAP. 39 / 45
40. EVERY DECISION MOVES YOU ALONG THIS SPECTRUM. The left side compounds. The right side borrows. FUNDAMENTALS AI DEPENDENT How you learn, what you practise, when you use AI and how, each choice places you on this spectrum. 🧭 40 / 45
41. YOUR COMPASS 1 Fundamentals first, then AI. The athlete who skips form and goes straight to the highlight reel gets injured. 2 You are the pilot, not the passenger. Co-pilots don't land planes. Own the outcome. 3 Code is cheap. Software is not. Anyone can generate code in 2026. Very few people can build systems that are correct, maintainable, secure, and aligned to real human needs. 4 Ask why, a lot. Why are we building this? Who does it affect? How will we know it worked? What happens if it goes wrong? 5 This is an extraordinary time to be a builder. The cost of experimentation has collapsed. Fundamentals + AI = an engineer who can build anything. The question is what you bring to the tools. 41 / 45
42. "AI makes engineering more essential, not less. The fear among junior engineers is real, but it's backwards. Yes, AI can write working code from natural language, but working code and engineered systems are worlds apart." Nate B. Jones 42 / 45
43. THE FUTURE IS BEING BUILT. BE ONE OF THE BUILDERS. Not a passenger. Not a spectator. A builder. 43 / 45
44. Q&A No wrong questions. Only unasked ones. 44 / 45
45. SCAN TO CONNECT linktr.ee/devs_central github.com/thecloudranger · AI Tinkerers Riyadh · Cloud Native Riyadh 45 / 45