Y

[Y @ynadell9]

@docmilanfar Wait what, are you saying if I have a model that can predict the paper I am going to submit (stocks people are going to buy/sell) if you are able to do it faster than others then you win?

[Y @ynadell9]

@arianTBD This is interesting work! Is the dataset public?

[Y @ynadell9]

@kothasuhas Wow, do you know what language model it is using? The guardrails are failing

[Y @ynadell9]

@a_karvonen I only see people using SAEs for interpretability, are there any other methods other than SAE to do the dictionary learning?

[Y @ynadell9]

@RohitGandikota @sheridan_feucht @davidbau @saprmarks Can you explain this a little bit more? Why do you want to reshape your output distribution?

[Y @ynadell9]

@arianTBD It would be interesting to see how they would perform on multi-hop questions after scaling test-time compute like in @sea_snell's paper
https://t.co/lia4PHePiT

[Y @ynadell9]

@ChShersh What was more difficult learning soft skills or technical skills?

[Y @ynadell9]

@AnjneyMidha Lmao, did this actually happen somewhere?

[Y @ynadell9]

@jxnlco How do you become a consultant without doing #1 or #2 or #3 to build some credibility for yourself?

[Y @ynadell9]

@ai_data_agent @aidan_mclau OMG this is amazing! Do you think we can use the IDK task for identifying hallucinations?

Lot of amazing work in these past few months

[Y @ynadell9]

@hamptonism Is there a reason why KAN papers are getting rejected?

[Y @ynadell9]

@ActuallyIsaak @hamptonism Interesting work!

[Y @ynadell9]

@ai_data_agent @aidan_mclau Is it better than LLAMA3? LLAMA3 also has an amazing needle in the haystack peformance.

[Y @ynadell9]

@Nominus9 @aidan_mclau This is an interesting paper but it's a new way of doing attention, kinda like RNNs so training can't be done parallely so that would make training complexity O(n^2). They didn't do enough benchmarking too, I don't think this is the answer we are looking for.

[Y @ynadell9]

@DeqingFu @COLM_conf What do the scores in the table on the bottom represent?

[Y @ynadell9]

@sea_snell True, this also echoes in LLAMA 3 paper. There weren’t any big architectural changes from LLAMA 2 but it had a significant performance boost with just using more high quality pre training data.

[Y @ynadell9]

@sea_snell Thanks for the reply!

[Y @ynadell9]

@sea_snell 🥲

[Y @ynadell9]

@mervenoyann @huggingface Or you could also ask GPT how to do it, it usually gives decent suggestions about the datasets and the models to use