Parent: pp-rewrite
Honest framing: anyone giving confident AI capability predictions over 1–2 year horizons is bluffing. The following is directional, not prophetic.
• Spec extraction ~80–90% useful with iteration
• Spec→Python ~70–80% useful given a good spec
• Productivity multiplier: 2–3× over a manual port
• Verification (humans reading code) is the dominant cost, not generation
• Context windows of 5–10M tokens with significantly better long-context utilisation. Whole-codebase reasoning becomes practical, not aspirational.
• Agentic coding tools mature — multi-day autonomous work on well-scoped tasks becomes routine
• Migration projects of this kind: 5–10× multipliers, not 2–3×
• Spec-driven workflows become standard tooling rather than bespoke prompting. Expect "port this codebase" to be a button somewhere.
• Verification gap widens: AI writes faster than humans can review, and the bottleneck shifts decisively to validation
Genuinely speculative. If current trajectories hold:
• End-to-end migration with AI-driven verification (test generation, behavioural equivalence checking, formal-ish methods on hot paths) becomes plausible for codebases of this size
• A senior engineer plus AI could plausibly do in weeks what takes a team months today
• Claude vs Gemini distinction probably matters less. Frontier models converge in capability for mainstream tasks; differentiation moves to tooling, integration, ecosystem
• Progress is uneven. The hard parts (subtle semantic equivalence, async behaviour preservation, "why does this break in production but not staging") may not improve at the same rate as the easy parts.
• Verification remains the binding constraint. Even with perfect generation, accountability requires human-readable evidence of correctness.
• Enterprise adoption lags capability by 12–24 months. What's possible at the frontier in mid-2027 may not be available within sanctioned tooling until mid-2028.
• Compliance, audit, and regulatory requirements add friction independent of AI capability. WebLogic admin platforms typically sit in tightly governed environments.
Don't wait. The spec produced today remains valuable in 1–2 years; re-running generation against a better model is cheap. The long pole is comprehension and verification of the existing Perl system, and that work needs to start now regardless of which model does the generation.