4 ms·
1. Very expensive - why do you care? 2. Opinionated - not necessarily bad, and can be tuned. 3. Politically biased - that's a bullshit criterion. You want a t
by jstx1 3y ago
1. Very expensive - why do you care?
2. Opinionated - not necessarily bad, and can be tuned.
3. Politically biased - that's a bullshit criterion. You want a tool that's reasonably objective - being in the middle between the two current political extremes is different from being objective.
4. Doesn't learn new skills - it does.
5. Controlled by a small group of people - that's why there are other implementations of LLMs.
- roflyear 3y agoHow does it learn new skills?
- jstx1 3y ago1. In-context learning - you can teach the model things without retraining the underlying model parameters. 2. Actual model updates (it's not like we'll be stuck with the 2023 version forever).
- roflyear 3y agoI don't think you can do 1, can you demonstrate that is possible? I don't argue you can't add to its trainingset, of course you can.