AI summary by TwinMind
React Is the last framework.
Theo - t3․gg
26m 46s
Introduction and Thesis
- Claims React is "the last framework" - not meaning no new tools exist, but last to reach mass adoption
- AI fundamentally changing how abstractions, frameworks, and libraries evolve
- Software development about to get "very weird" due to AI's irreversible impact
Car Analogy: Infrastructure Lock-in
- Pre-car era: infrastructure built for horses (narrow Boston streets example)
- Once car adoption hit critical threshold, format locked in
- Street width standardized
- Making narrower cars provides no benefit due to existing street infrastructure
- Cannot change fundamental car operation due to existing car/street ecosystem
- Can only innovate inside the car (power source, layout, features)
- Cannot change how car interfaces with external world
React's Dominance and Ecosystem
- React has reached critical adoption threshold
- New tools typically support React
- New projects commonly start with React
- 1/3 of developers start coding journey with React (not HTML/JS/Python)
- Massive amount of React code exists on web
- Strong ecosystem with solutions to most problems
- Quality varies but coverage comprehensive
- Most problems solvable within 2 Google searches
- Ecosystem strength creates both solutions and disasters (Webpack example)
AI Training Data Problem
- AI models trained on massive amounts of React code
- More Stack Overflow posts about React problems than Laravel/Rails
- Laravel/Rails provide built-in solutions (less documentation needed)
- React requires external solutions (more documentation created)
- Data quantity determines AI recommendations, not quality or user satisfaction
- AI defaults to React TypeScript syntax even in non-React contexts
- Happens in Solid, Elixir codebases
- AI cannot distinguish between similar syntaxes
Competition Impossibility
- Even superior frameworks like Solid face AI bias
- Solid syntax example: looks nearly identical to React
count, setCount = createSignal()onCleanup(() => clearInterval(...))
- AI sees 100 Solid examples vs 100,000 React examples
- Makes incremental improvements insufficient
- Need to compete with "the street" (data ecosystem), not just other cars (frameworks)
React Team's Own Constraints
- Facebook cannot make major React syntax changes
- Too much internal React code to justify changes
- Stumbled onto AI problem independently through scale
- Adding new React features makes development harder
- AI models lack knowledge of recent features
- Example: Stripe subscription prevention feature
- Recent API addition to prevent double subscriptions
- AI models gave outdated recommendations
- Massive old data trumps new solutions
React Compiler Solution
- React Compiler as innovation layer strategy
- Functions as automemoizer
- Statically analyzes files for memoization opportunities
- Eliminates need for manual
useMemo/useCallback - Lets developers write simpler code, compiler optimizes
- Compiler output still slower than Solid but enables different innovation approach
Abstraction Layers Analysis
- Bottom to top: Assembly → C/C++ → JS Runtime → React/JSX/TypeScript
- Developers and AI both live in top abstraction layer
- AI performance degrades going down layers
- Poor at debugging build/bundle problems
- Better at higher-level abstractions
- General rule: lower layers = higher pay + worse AI performance
React Compiler as Innovation Layer
- React syntax effectively frozen at current state
- All future innovation happens in compiler layer
- Benefits:
- New React features don't require syntax changes
- Can introduce signals, context improvements, performance wins
- Treats React syntax as complete, immutable language
- Downsides:
- Never got JSX 2.0
- Still use
classNameinstead ofclass(Solid provesclassworks)
Python Parallel
- Python as example of "bad but dominant" technology
- Python ecosystem described as "atrocious"
- Virtual environments problematic
- Difficult code bundling/distribution
- Package installation doesn't make sense
- Python continues growing despite quality issues
- Python 2 to Python 3 transition took 10 years
- Key insight: "being good doesn't matter anymore"
The End of Major Transitions
- Python 2→3 style transitions cannot happen anymore
- AI prevents major syntax changes
- Would hallucinate old syntax
- Easier to write old version
- Autocomplete bias toward established patterns
- Projects like Mojo take existing syntax, improve runtime
- Angular "red wedding" style changes impossible now
Innovation Constraints
- Code today won't look meaningfully different in 10 years
- Contrasts with rapid changes in past 3 years
- React architected to continue improving without language changes
- 10x better syntactic/ergonomic solutions "effectively can't win"
Internal Innovation Still Possible
- Electric car analogy: can change internals without changing street interface
- React runtime/reconciler can improve without affecting written code
- React Compiler enables this internal innovation approach
- "Good enough" threshold reached - incremental improvements insufficient
Future Outlook
- No more TRPC, Solid, or React Hooks level innovations expected
- Innovation shifts to:
- How we command systems
- How we orchestrate code
- AI tools generating code (higher abstraction layer)
- Framework innovation era potentially ending
- Hope to be proven wrong on this prediction
