Two Roads for AI in Software Engineering — and Neither Is What You Think | Ai++ Ep. 3
AI for Software Engineering: The Big Picture
Two fundamentally different approaches to using AI in software development are emerging — and each comes with its own risks and rewards. In this episode, I break down the bottom-up approach (AI as a coding assistant) and the top-down approach (AI building entire systems), explore why trust is the central challenge, and explain why domain-specific languages and formal blueprints may be the key to making it all work.
Timeline
00:00 – Intro & channel overview
00:30 – Today's topic: AI for software engineering
00:41 – Two emerging use cases: bottom-up vs. top-down
00:50 – Bottom-up: AI as a productivity tool (Cursor, Claude Code, etc.)
01:20 – The trust problem: calculators vs. AI
02:02 – The Munich Taxi-Cab Experiment and risk compensation
02:40 – What happens as AI gets more trustworthy (and reviews get lazier)
02:55 – Mitigations: formal verification, compartmentalized architecture, zero-trust design
03:50 – The skills atrophy problem & Asimov's "Profession" (1957)
04:23 – Top-down: AI creating entire systems (Lovable, Replit, Base44)
05:04 – The real bottleneck: specifying what to build
06:03– Domain-specific languages (DSLs) as a path forward
06:44 – Phased development with formal handoffs and blueprints
07:22 – The construction analogy: why formal blueprints matter
08:17 – AI-built vs. AI-enabled: they're not the same thing
08:46 – What's next
Topics covered in the series
— Ep. 1:The Bedridden Genius: A Mental Model for What AI Can Actually Do → • The Bedridden Genius: A Mental Model for W...
— Ep. 2: Beyond Chatbots: The Case for AI-First Software Architecture → • Beyond Chatbots: The Case for AI-First Sof...
— Ep. 3: Two Roads for AI in Software Engineering — and Neither Is What You Think → watching now
— Ep. 4: 100% AI Code at Anthropic. 19% Slower Everywhere Else. Why? → • 100% AI Code at Anthropic. 19% Slower Ever...
— Ep. 5: One Sentence Can Hijack Your AI. Here's How to Stop It. → • One Sentence Can Hijack Your AI. Here's Ho...
— Ep. 6: One Million Lines of Code. Zero Keystrokes. Welcome to Harness Engineering. → • One Million Lines of Code. Zero Keystrokes...
— Ep. 7: AI Reads Every Word You Say. It Still Gets You Wrong. → • AI Reads Every Word You Say. It Still Gets...
— Ep. 8: The Path to Autonomous Agents Was Mapped Decades Ago → • The Path to Autonomous Agents Was Mapped D...
— Ep. 9: AI Is a Revenue Multiplier → • AI Is a Revenue Multiplier. So Why the Obs...
— Ep. 10: Fluent in AI? That Was Just the Trailer. → • Fluent in AI? Congrats — That Was Just the...
— Ep. 11: Business Automation Was Once DIY. AI Just Revived It. → • Business Automation Was Once DIY. AI Just ...
👉 Subscribe and hit 🔔 — upcoming episodes will provide concrete, implementation-ready recipes for every challenge discussed today: AI phone systems, zero-trust security for LLMs, human-in-the-loop design patterns, and probabilistic architecture in practice.
💬 Share your biggest pains and wins while developing software with AI - drop me a comment.
🔗 If you prefer reading to watching, check out my blog: https://anatoly.com/blog
or chat with my AI Twin: https://anatoly.com/twin
About AI++
Hosted by Anatoly, a software architect with 35 years of Silicon Valley experience and the founder of Rishon, AI++ explores how businesses of all sizes can use AI to achieve tangible results — cost cuts, reduced effort, and lower risk.
🔗 Learn more about Rishon:
• Zero to Production App in 30 min
https://rishon.com
#AI #SoftwareDevelopment #SoftwareEngineering #AIAgents #GenerativeAI