Portability or ‘don’t marry your mannequin’
Andy Oliver is true: “The newest GPT, Claude, Gemini, and o-series fashions have completely different strengths and weaknesses, so it pays to combine and match.” Not solely that, however the fashions are in fixed flux, as is their pricing and, very doubtless, your enterprise’s danger posture. As such, you don’t wish to be hardwired to any specific mannequin. If swapping a mannequin means rewriting your app, you solely constructed a demo, not a system. You additionally constructed an issue. Therefore, profitable deployments observe these rules:
- Summary behind an inference layer with constant request/response schemas (together with instrument name codecs and security alerts).
- Hold prompts and insurance policies versioned outdoors code so you’ll be able to A/B and roll again with out redeploying.
- Twin run throughout migrations: Ship the identical request to previous and new fashions and evaluate through analysis harness earlier than reducing over.
Portability isn’t simply insurance coverage; it’s the way you negotiate higher with distributors and undertake enhancements with out worry.
Issues that matter lower than you assume
I’ve been speaking about how to make sure success, but certainly some (many!) individuals who have learn up so far are considering, “Certain, however actually it’s about immediate engineering.” Or a greater mannequin. Or no matter. These are AI traps. Don’t get carried away by:
