4 ms·

Of course they're not going to get worse. That would be absurd. The rate of progress will slow down though.

NewJazz 5mo ago
I don't think it would be absurd for them to worsen. If LLMs cause discourse to worsen, but also grow and change, then the trainers are in a conundrum of ignoring new training data or losing track of the zeitgeist.
LtWorf 5mo ago
They might get worse for 2 reasons: 1. AI free training sets no longer exist. This might degrade quality, although some claim that it will not. 2. Cost. Right now they are burning a lot of money to convince people it's good. But they might not be able to keep it up forever and need to increase prices (which few will want to pay) or degrade the quality to save money.
- mirekrusin 5mo ago
  non synthetic pre-training text got exhausted long time ago, now focus is more on quality and rl/post-training. cost will keep going down and more powerful chips will be available, like it always was. reinforcement learning doesn't have fixed ceiling, advancements in software will keep happening as well. things like distilling smaller models will likely become free – ie. through speculative decoding which speeds up larger models (incentive to run during inference) where you have free access to strong distillation (large model's logits are calculated, ie. it's free for pickup for distillation) etc.
  - LtWorf 5mo ago
    > cost will keep going down and more powerful chips will be available, like it always was. cost is not a monotonous function.
    - mirekrusin 5mo ago
      compute and memory cost is monotonic with temporary hiccups
      - LtWorf 5mo ago
        I figure you aren't aware of why there are other kinds of interpolation other than linear.
lelanthran 5mo ago
> Of course they're not going to get worse. That would be absurd. The rate of progress will slow down though. It's unlikely, but not totally improbable - Model collapse means that the subsequent models would get worse over time, not better.
- mirekrusin 5mo ago
  what's "model collapse"? RL has virtually unlimited ceiling.

I don't think it would be absurd for them to worsen. If LLMs cause discourse to worsen, but also grow and change, then the trainers are in a conundrum of ignoring new training data or losing track of the zeitgeist.

LtWorf 5mo ago

They might get worse for 2 reasons: 1. AI free training sets no longer exist. This might degrade quality, although some claim that it will not. 2. Cost. Right now they are burning a lot of money to convince people it's good. But they might not be able to keep it up forever and need to increase prices (which few will want to pay) or degrade the quality to save money.

mirekrusin 5mo ago
non synthetic pre-training text got exhausted long time ago, now focus is more on quality and rl/post-training. cost will keep going down and more powerful chips will be available, like it always was. reinforcement learning doesn't have fixed ceiling, advancements in software will keep happening as well. things like distilling smaller models will likely become free – ie. through speculative decoding which speeds up larger models (incentive to run during inference) where you have free access to strong distillation (large model's logits are calculated, ie. it's free for pickup for distillation) etc.
- LtWorf 5mo ago
  > cost will keep going down and more powerful chips will be available, like it always was. cost is not a monotonous function.
  - mirekrusin 5mo ago
    compute and memory cost is monotonic with temporary hiccups
    - LtWorf 5mo ago
      I figure you aren't aware of why there are other kinds of interpolation other than linear.

lelanthran 5mo ago

> Of course they're not going to get worse. That would be absurd. The rate of progress will slow down though. It's unlikely, but not totally improbable - Model collapse means that the subsequent models would get worse over time, not better.

mirekrusin 5mo ago
what's "model collapse"? RL has virtually unlimited ceiling.