bradfox2

Paper mentions amorphous state plus an annealing step. Should be glass like at the end. If so, heat rates (up and down) of the last step are important and hardly mentioned.

LLM output is scored by another model that produces a reward for the entire sequence emitted by the LLM. The reward model is trained on human preferences or some other metric usually. It's RL because we train on the reward and not some language modeling objective.

The LLM is trained to increase this reward score (or minimize the inverse), which is what makes it RL.

It's still loss being backproped, but the loss is calculated over a different criteria

Compressed air usually.

On my team, there was not intially acceptance from folks that were used to traditional nlp techniques. The models built with transformers were viewed as over parameterized.

It really wasn't until the original BERT paper came out and topped the superglue leaderboards that the move from lstm-cnn based architectures started. I remember feeling at the time that a ~300M parameter model was absolutely huge.

All partially funded by sweetheart lease deals between land ASU owns and billions of dollars in corporate office buildings and hotels thanks to the tax exempt status of the University. Status that is supposed to be used for educational purposes but instead is turning the ASU campus into a commercial property tax haven.

   https://tucson.com/news/local/arizona-ag-sues-to-block-universities-from-leasing-land-to-commercial-businesses/article_c1210c2f-da3f-5a35-9678-6aa10f2efb9a.html

My company builds AI systems for Nuclear plants. In fact, AI systems based on LLMs are already handling issue triage at several plants in the US. Dynamic Operating procedures are something the industry is interested in and in certain scenarios can add significant safety margin. In these scenarios, there is no documented procedure action for the operating crew to follow and they are under high stress/cognitive load.