3 ms·

The cost to compute these language models should eventually lower. Will OpenAI then release more frequently, or release larger LLMs? They'll probably to try ach

by jasfi 3y ago

The cost to compute these language models should eventually lower. Will OpenAI then release more frequently, or release larger LLMs? They'll probably to try achieve both goals in some proportion.

ben_w 3y ago
Training on web pages created after ChatGPT (and equivalently Stable Diffusion) was released, has a strong risk of trying to create the next AI (of whatever type) on the barely-even-proofread output of the current models.
rtkwe 3y ago
A big issue with constant retraining is going to be the self referential consumption of it's own generated material as training material. There's already been the studies that these models quickly break down when they're fed their own generated data as training data. It won't immediately degrade them as it's a small percentage but there's already a lot of people using these models to generated spam junk out there and it'll only get worse with time.
- visarga 3y ago
  > A big issue with constant retraining is going to be the self referential consumption of it's own generated material as training material That's just a hypothetical scenario. Usually LLMs don't generate text used indiscriminately to train new models. In creative tasks, the generated texts are the result of multiple explorations and corrections, filtered by the user, then posted online and commented on by other people. So there is plenty of correction and feedback mixed into the process. A child model could learn something from this feedback. Then there is code and math, where it is possible to run tests or double check by a different method. These tasks could be iteratively improved by LLMs by incorporating the feedback. Another field ripe with feedback signals is using LLMs in games and simulations, where you can specify a goal easily but finding a solution requires search and learning. And a related application was RLHF, where a preference model was trained on human judgments and used to fine-tune a LLM. That preference model can be used to filter bad training data as well. To generalize, LLMs alone can't do it, but LLMs with external systems can get feedback and improve. The garbage-in-garbage-out scenario can be avoided with a bit of external feedback.
- YeGoblynQueenne 3y ago
  >> There's already been the studies that these models quickly break down when they're fed their own generated data as training data Not "these models" if you're referring to large language models pretrained and fine-tuned with Transformers. The only study I am aware of that explored the effect you discuss [1] had to make do with very different models as proxies: Variational Autoencoders, Mixture of Gaussians and a smaller LLM with 125 million parameters that hardly qualifies as one. It is not safe to extrapolate those results to really Large-LMs trained with Transformers, as it is not safe to extrapolate any results on one neural net architecture on another. Note the Mixtures of Gaussians aren't even neural nets at all, but a completely different statistical model. The question of what will happen once a few generations of LLMs have trained on their ancestors' outputs is an interesting one that has not yet been clearly answered. In the long run, future developments may even mean it's not really a valid question at all because such feedback retraining can't happen in practice- who knows? I for one remain curious to see how this goes. ______________________ [1] The Curse of Recursion: Training on Generated Data Makes Models Forget https://arxiv.org/abs/2305.17493 https://arxiv.org/abs/2305.17493
- jasfi 3y ago
  Ranking algorithms should help with this, working on human feedback. Pages that spew badly generated text likely won't rank well on the top search engines. The big LLM names, such as OpenAI, would probably partner (e.g. with Microsoft), to tap into this if they needed to. They may have their own ranking algorithms anyway.
  - rtkwe 3y ago
    That's really just a temporary bandaid as the whole goal of these spam sites is to rank on search and the goal of LLMs is to appear to be human generated. You're stuck in an arms race against you're own improving performance if you're counting on filtering to keep your own output out of your inputs as an LLM company/trainer. Also there's not a singular search rank for a site it's rank depends on the search query so I'm not even sure how you'd apply that metric.

by jasfi 3y ago

The cost to compute these language models should eventually lower. Will OpenAI then release more frequently, or release larger LLMs? They'll probably to try achieve both goals in some proportion.

ben_w 3y ago

Training on web pages created after ChatGPT (and equivalently Stable Diffusion) was released, has a strong risk of trying to create the next AI (of whatever type) on the barely-even-proofread output of the current models.

rtkwe 3y ago

A big issue with constant retraining is going to be the self referential consumption of it's own generated material as training material. There's already been the studies that these models quickly break down when they're fed their own generated data as training data. It won't immediately degrade them as it's a small percentage but there's already a lot of people using these models to generated spam junk out there and it'll only get worse with time.

visarga 3y ago
> A big issue with constant retraining is going to be the self referential consumption of it's own generated material as training material That's just a hypothetical scenario. Usually LLMs don't generate text used indiscriminately to train new models. In creative tasks, the generated texts are the result of multiple explorations and corrections, filtered by the user, then posted online and commented on by other people. So there is plenty of correction and feedback mixed into the process. A child model could learn something from this feedback. Then there is code and math, where it is possible to run tests or double check by a different method. These tasks could be iteratively improved by LLMs by incorporating the feedback. Another field ripe with feedback signals is using LLMs in games and simulations, where you can specify a goal easily but finding a solution requires search and learning. And a related application was RLHF, where a preference model was trained on human judgments and used to fine-tune a LLM. That preference model can be used to filter bad training data as well. To generalize, LLMs alone can't do it, but LLMs with external systems can get feedback and improve. The garbage-in-garbage-out scenario can be avoided with a bit of external feedback.
YeGoblynQueenne 3y ago
>> There's already been the studies that these models quickly break down when they're fed their own generated data as training data Not "these models" if you're referring to large language models pretrained and fine-tuned with Transformers. The only study I am aware of that explored the effect you discuss [1] had to make do with very different models as proxies: Variational Autoencoders, Mixture of Gaussians and a smaller LLM with 125 million parameters that hardly qualifies as one. It is not safe to extrapolate those results to really Large-LMs trained with Transformers, as it is not safe to extrapolate any results on one neural net architecture on another. Note the Mixtures of Gaussians aren't even neural nets at all, but a completely different statistical model. The question of what will happen once a few generations of LLMs have trained on their ancestors' outputs is an interesting one that has not yet been clearly answered. In the long run, future developments may even mean it's not really a valid question at all because such feedback retraining can't happen in practice- who knows? I for one remain curious to see how this goes. ______________________ [1] The Curse of Recursion: Training on Generated Data Makes Models Forget https://arxiv.org/abs/2305.17493 https://arxiv.org/abs/2305.17493
jasfi 3y ago
Ranking algorithms should help with this, working on human feedback. Pages that spew badly generated text likely won't rank well on the top search engines. The big LLM names, such as OpenAI, would probably partner (e.g. with Microsoft), to tap into this if they needed to. They may have their own ranking algorithms anyway.
- rtkwe 3y ago
  That's really just a temporary bandaid as the whole goal of these spam sites is to rank on search and the goal of LLMs is to appear to be human generated. You're stuck in an arms race against you're own improving performance if you're counting on filtering to keep your own output out of your inputs as an LLM company/trainer. Also there's not a singular search rank for a site it's rank depends on the search query so I'm not even sure how you'd apply that metric.