Ravi Ganti

ML researcher @Google X. Ex @WalmartLabs, postdoc@WID. PhD in ML from Georgia Tech, bachelors from IIT Kharagpur.

[Ravi Ganti @gmravi2003]

#uberone Superbowl ad looks heavily inspired by cred's ads?

[Edward Grefenstette @egrefen]

🧵 Every year, I rather self-indulgently end the year by looking over the research output and life events which have marked 2022 for my collaborators and me. Between career changes, return to in-person conferences, and disruptive world events, it's been quite a year! LFG… [1/14]

[Yi Tay @YiTayML]

Happy New Year / New Years Eve! 🥳(depending on where you are)

Here's a thread of me reflecting on 2022 and all the research I've contributed to this year and stuff I did.

🧵 quite a long chronological thread below 👇

[Aakanksha Chowdhery @achowdhery]

1/ Reflecting back on 2022: we shared our most advanced language model PaLM - a single 540B-parameter dense language model for multiple domains & tasks, trained over two TPUv4 Pods.
Research paper: https://t.co/baKtK3kSqp
Blog post: https://t.co/Nk52y5ulzm

[Ravi Ganti @gmravi2003]

@alexandr_wang Could you share your process?

[Marcos López de Prado @lopezdeprado]

When making a scientific claim, not all types of empirical evidence presented in support of that claim are equally strong. Some types of evidence are more susceptibility to being spurious than other types.

For more details, see section 6.5 here: https://t.co/6TPFZG5caT

[Ravi Ganti @gmravi2003]

@TudorsAndTMs You should have called him on the phone and said "I just called to say, I no longer love you". Sorry, had to make this joke.

[Ravi Ganti @gmravi2003]

@zacharylipton don't forget coconut chutney. the coffee amplifies the heat in the coconut chutney. Oh, so awesome.

[Ravi Ganti @gmravi2003]

@TaliaRinger sorry. That was just a joke. vidyarthi = student, bhavan = house. So, I was trying to joke that you might get a discount if you prove that you are a student :-) (maybe, but i do not know ) :-)

[Ravi Ganti @gmravi2003]

@dsivakumar I have a saying (basically I say it to myself). All the interesting things happen in the intersections. And, most work at the intersections of fields are misunderstood and hard to get noticed.

[Ravi Ganti @gmravi2003]

@DimitrisPapail @madiator I agree with Mahesh. Jax. If you are using python, then you are probably using numpy. If you are using numpy then you might as well just do import jax.numpy.

[Ravi Ganti @gmravi2003]

@_arohan_ Maybe call it a kernel layer. Or an expander?

[Ravi Ganti @gmravi2003]

@RamaswmySridhar Where did the 800 number come from? Should it not be 33*4^2=528? 33 for increase in #of tokens and 16 for the increase in context size.

[Ravi Ganti @gmravi2003]

@MountainOfMoon bhaalo aache

[Ravi Ganti @gmravi2003]

@DimitrisPapail 🤦‍♂️

[Ravi Ganti @gmravi2003]

@JeremyDanielFox @MichaelTrazzi @ethanCaballero A good definition of AGI.

[Ravi Ganti @gmravi2003]

@zacharylipton @jefrankle @HanieSedghi @roydanroy @bneyshabur @prfsanjeevarora sgtm. This feels close to results in model selection literature, though not sure how large are the model classes there. (been a while)

[Ravi Ganti @gmravi2003]

@zacharylipton @jefrankle @HanieSedghi @roydanroy @bneyshabur @prfsanjeevarora I guess what you are saying is that the model class here is very large. But if you separate out the simple models in this class from the more complex models on the basis of clean/random data fit then the simple models generalize, even though results from classical theory say no

[Ravi Ganti @gmravi2003]

@zacharylipton @jefrankle @HanieSedghi @roydanroy @bneyshabur @prfsanjeevarora ok. I am trying to understand how this result is different from standard learning theory results that guarantee generalization for low complexity classes. 1/2