Ryan Fishman

[Ryan Fishman @RyanFishman]

@sama Remember what it did in my code base 5 days ago.

[Ryan Fishman @RyanFishman]

@sama Long term memory, continuous learning.

[Ryan Fishman @RyanFishman]

@telegram

[Ryan Fishman @RyanFishman]

@sama Allow users to login with their ChatGPT account and allow builders to build their own UI interface ontop of the API but use the users rate limits so you don't have to think of every good UI feature and instead outsource that aspect so you can focus on AI.

[Ryan Fishman @RyanFishman]

@liambolling The surgeon, who is the boy's father says, "I cannot operate on this boy, he's my son!" Who is the surgeon to the boy?

[Ryan Fishman @RyanFishman]

@rand_longevity When will agents be able to implement a PR all on its own in a large code base from just a description 99% of the time.

[Ryan Fishman @RyanFishman]

@SoroushG_ @RewindAI https://t.co/8JWORECkQW

[Ryan Fishman @RyanFishman]

One of my jobs at work is to create reports(sql statements) based on a description of data the customers want. We have 100s of tables. Some of the data they want need to be computed in the Sql statement. Queries end up requiring many joins and computations and inner queries. Simply give tables with descriptions. Some examples of other queries and ChatGPT can spit out a working query. Saves lots of time and can re-use the template.

[Ryan Fishman @RyanFishman]

@tentacionfn Thanks.

[Ryan Fishman @RyanFishman]

@Rizzy1c Cool

[Ryan Fishman @RyanFishman]

@MgmtArham Thanks

[Ryan Fishman @RyanFishman]

@mickey8rangers

[Ryan Fishman @RyanFishman]

@far__el I can explain how A* can be used in inference a Q values used in training if invited to chat

[Ryan Fishman @RyanFishman]

@Douglas_Schon @far__el By the way how did you get to be a speaker?

[Ryan Fishman @RyanFishman]

@Douglas_Schon @far__el Yes, that explains it. But the states are the tokens, and instead of the output of the transformer being a probability of next token, they are the projected final quality of future output of final phrase. A* can be used in inference to try multiple strategies working thr proble

[Ryan Fishman @RyanFishman]

@Douglas_Schon @far__el For inference, you use the A* algorithm and start outputting various possible sentences as you follow the best Q path. Hard to explain over twitter.

[Ryan Fishman @RyanFishman]

@Douglas_Schon @far__el The path is not infinite, it ends with the end character.

The Q values are the the value of the the best future path of the final sequence.

For training the main difference instead of having a softmax for final layer of the you have a regression layer .

[Ryan Fishman @RyanFishman]

@Douglas_Schon @far__el I think the reward is the quality of the total final output, the state is the current path of your output sequence. The actions is the possible next tokens. Currently transformers once they pick a token they are stuck, with this they can backtrack and and do A* on best path.

[Ryan Fishman @RyanFishman]

@fanged_desire @far__el You can evaluate the quality of a final answer on a scale. Lets say 10 for high quality and 0 for bad quality. RL is good at path finding. The path can be from first token to last token. Picking a token is a action. So RL could be find the optimal path for picking token.