Software Engineering :: Pair Programming :: AI Assistant :: Advocacy :: Software engineering with LLMs in 2025: reality check
⪼ Made with 💜 by Polyglot.
This video is a keynote-style talk that blends a story-driven analysis with research and interviews, aimed at educating and inspiring the audience about the real state of AI coding tools in 2024. The speaker explores the gap between the hype-driven claims from tech executives and the nuanced, ground-level reality experienced by software engineers. Drawing from conversations with engineers at AI startups, big tech companies, independent devs, and even legendary programmers, the talk examines how AI tools are used, how effective they are, and why the actual impact may be subtler and more interesting than media headlines suggest. The speaker concludes that AI is driving a step change in software development, likened to the leap from assembly to high-level programming, but with new, non-deterministic challenges and vast potential.
- CEOs and tech leaders make bold claims (e.g., 90%+ of code will be AI-generated), often for PR or investor appeal
- Ground-level experiences show mixed results, with notable failures and successes, especially in complex or production settings
- Transparency about the real-world limits of AI coding agents is rare but valuable
- AI dev tool startups report high internal adoption, often "dogfooding" their products with claims of 90%+ code written by AI tools (Cloud Code, Windinsurf, Cursor)
- Big tech (Google, Amazon) deeply integrates AI into custom IDEs and workflows, but adoption is pragmatic and gradual, focused on reliability and developer trust
- Amazon’s API-first culture makes it easy to bolt on AI automation, and internal MCP (Model Context Protocol) adoption is high, though not widely publicized
- Some fast-moving startups become “AI-first,” sharing tips and adopting new tools like Cloud Code across teams
- In specialized or cutting-edge fields (e.g., protein design), AI code tools sometimes lag behind the speed and quality of expert engineers
- Veteran independent engineers report a new excitement about programming, describing AI tools as a generational shift (enabling new languages, higher output, rekindling the joy of coding)
- Founders and CEOs are often more excited than the engineers; top-down enthusiasm doesn’t always trickle down to daily workflow
- AI tool usage is significant (about 50-70% weekly adoption in surveys), but the transformative effect is still not universal or 10x for most developers
- AI tools tend to work better for individuals than for coordinated teams or organizations
- The real productivity boost is unclear—hours saved per week are real but not yet revolutionary
- AI coding tools are enabling a step change, like the shift from assembler to high-level languages, but with new complexities (e.g., non-determinism)
- Legendary programmers see parallels with past revolutions: microprocessors, the internet, smartphones
- The biggest opportunity: what was previously expensive or out-of-reach in software is now “ridiculously cheap,” encouraging more experimentation and ambition
- The AI coding revolution is real but messy, uneven, and not quite what the hype suggests
- Experimentation is critical—developers and companies should keep trying new approaches to understand what AI tools can (and can’t) do
- The field is moving fast, and those willing to adapt stand to benefit the most