Why innovation is difficult for small and medium businesses — and how AI is changing the game

• Innovation is like "setting money on fire" — but necessary for long-term wins
• Most experiments fail; the learning is the value, not the output
• R&D tax credits exist specifically because the government wants businesses to invest in uncertain outcomes
• Analogy: Innovation is like working out — everyone wants the results, nobody wants the 5-year grind

• Writing code is fast; architecture and problem-solving are the hard parts
• Losing a day's work and recreating it in 30 minutes proves: the code isn't the value, the thinking is
• AI can write code extremely quickly, but still struggles with novel architecture and business-specific problems

• What took weeks to build now takes days
• The barrier to entry for innovation has never been lower
• Small/mid-sized businesses are the biggest winners — they can now do what only enterprises could afford before
• Example: Building interactive, regional data visualizations that would have been "cost-prohibitive" before

• The goal isn't to replace workers with AI — it's to eliminate the work nobody wants to do
• Non-technical team members can now build React artifacts and interactive tools
• The focus shifts from "writing code" to architecture, ideas, and oversight
• People still need to learn through failure (like touching the hot stove)

• Previously, custom software required $2-3M+ investment for dev teams
• Now, small teams with AI tooling can build tailored solutions
• Example: Instead of begging enterprise vendors for features, just build what you need
• Modern frameworks (Rails, etc.) allow deployment in minutes

• AI agents will try to work around restrictions (digging tokens out of logs, attempting DNS changes)
• Balancing innovation with security is an ongoing tension
• Local/on-premise models offer a path for sensitive data processing
• The future: purpose-built, domain-specific models that don't need general knowledge

• Frontier models are being compressed to run on consumer hardware (RTX 6000, etc.)
• Next evolution: slicing off specialized capabilities for specific use cases
• Small, tuned models for narrow tasks (OCR, customer service, etc.) instead of massive general-purpose models

1. Budget for failure — Innovation requires experiments that won't work
2. AI lowers the barrier — What cost millions now costs a fraction
3. Empower your team — Give them AI tools and let them experiment
4. Focus on architecture — Let AI handle code output; humans own the thinking
5. Stay curious — The landscape changes weekly; ride the wave or get left behind

Episode Length: ~47 minutes
Tone: Conversational, technical but accessible, optimistic about AI's potential with realistic caveats about challenges