Thanks :) Your paper is great! But more sophisticated than what I'll show in part 2. I'll show how to manipulate embeddings vectors as they pass through the residual stream to normalize the bias. So it's not actually fixing the bias, it's correcting one example. And learning about mech interp.

And thank you, Paolo!

ps, if you liked that joke, then you will absolutely love my recordings of the two-day workshop I taught on EEG synchronization in time, frequency, and space.
Here's part 1: mikexcohen.substack.com/p/eeg-synchr...

Awesome.

Interesting perspective. I think peer-review is imperfect but does more good than harm. I've published >100 papers and reviewed... dunno, many 100s of manuscripts. My thinking is that the incentive structure is the root of the problem, although I could be wrong. I agree with your suggestions though.

Fair point. I suppose the question is whether the entire system needs to be revamped (I'd find that unlikely to happen, given that too many people are too vested in it) or whether relatively small changes can have meaningful impact.

I'd love to see the evaluations in academia based on one paper every five years. Not a perfect solution, but it would increase scientific publishing signal-to-noise-ratio by encouraging researchers to put more effort into a smaller number of papers.

Hehe, yep, although I'm not so active here. I'm mostly building up content (tutorial posts on data science and LLM mechanisms) on mikexcohen.substack.com.

Nice to hear from you, Luiz!

That was definitely part of my decision, although I wouldn't call it a "problem." More that I was (and still am) in an extremely fortunate situation, and decided to spend my time in a career that had higher positive impact for more people, and that was also more personally satisfying.

Thank you for taking the time to share you lucid and honest opinion.

Thank you kindly, Thomas. My feelings, reactions, and motivations are definitely not unique, but ex-academics have an unfortunate habit of reticence.

You earned them.

That is a fantastic title and topic. Congrats!

Thank you, Felipe :)