Hacker News Reader

They need to be retrained often is my experience for general usage.

I have SO many cases, for example when shopping, when they make assumptions based on their cutoff date, and they need to redo everything based on lots of searches.

Sure, for many use cases old models will be useful but for general usage a model that's 5 years old will probably be garbage for most people, unless they prompt it extensively, which the general public won't do.

It’s helpful for a model to be somewhat static when it’s involved in business processes.

Relying on an LLM's baked-in syntax knowledge is suboptimal, compared to leveraging reasonable context affordances (skills, examples, tool calls, MCP servers, search) to ensure adherence to precise APIs and language specs. Don't get me wrong; a model specifically trained on your desired framework is obviously preferable for more consistent generation of idiomatic code -- but relying on that in a large general-purpose model is a questionable proposition IMHO.

> But I'm also interested is retraining needed? how relevant will a model be in N years if it's just "static", and if you do need to retrain and fine-tune over time, is it still profitable? how often is that needed?

Just wait until what Ilya Sutskever has for Safe Super Intelligence (SSI) to self update the weights without retraining them to bring training costs down close to $0.