I published a new article on Habr: “AI vs Open Source: What Actually Changed and Where the Line Is”

With the emergence of working code models, a more down-to-earth development path has appeared: formulate a requirement, write tests, and get a small, understandable module with no unnecessary dependencies. This isn’t a war against OSS – it’s a shift in the equilibrium point.

Main Takeaways:

What Changed

  • Before: “library first.” Search for a library, accept transitive dependencies, read documentation.
  • Now: “description -> tests -> implementation.” Small, testable modules instead of monolithic “combines.”

Where AI Already Replaces Libraries

  1. Mini-implementations: indicators (EMA/SMA/RSI), statistics, risk rules
  2. Narrow integrations: REST/WebSocket clients with just 2-3 needed methods
  3. Skeleton generation: backtest scaffolds, data schemas
  4. Adapters: mapping between exchanges, code migrations

Where AI Should NOT Replace OSS

  • Cryptography and secure protocols
  • Binary protocols (FIX/ITCH/OUCH/FAST)
  • Database engines, compilers, runtimes
  • Numerical solvers and optimizers

Practical Advice

  • Keep modules small
  • Describe behavior in simple words
  • Do minimal checks for confident merges
  • Generate without external dependencies

In algorithmic trading, this is especially relevant: fewer dependencies means lower risks, more compact artifacts, easier audits, and faster iterations.

Key Takeaway: Choose your tool based on context. A narrow task that’s easy to describe and verify is a candidate for generation. Everything else – go with proven OSS.