Has anybody tried binaural beats for focus? I can hear them and it drives me crazy.

ChatGPT Plus is the best investment ever. Did I write that already?

I’m always jumping around concepts (((((

https://twitter.com/hubermanlab/status/1677744568962514945?s=46&t=tJA3XyJ6UXEb-9Si1PfeJQ

Fed part of our code for a matrix compression project to the Code Interpreter in ChatGPT Plus, and it got several things wrong. See, I don’t believe in the AI conquering the world any time soon because you have to always check and recheck and recheck. It also has problems with einsum ahaha, I am not the only one.

Some house spirit destroyed my NVIDIA driver installation on the laptop David gave me. It cannot reach CUDA out of the blue.

I haven’t done any calculations there in a while but I didn’t do any kernel update or reinstalled anything. Wtf

He’s not wrong (in a certain sense). I don’t know about any unsupervised system that would be close to human performance on some task without any fine-tuning on supervised data.

However, the definition of AI can be disputed.

Pretty cool!

this paper's nuts. for sentence classification on out-of-domain datasets, all neural (Transformer or not) approaches lose to good old kNN on representations generated by.... gzip (c)

It's super elegant. If x1 and x2 are similar, then their joint compression (lzw/whatever) is more efficient. (c)

https://twitter.com/giffmana/status/1679399205570326528?s=46&t=tJA3XyJ6UXEb-9Si1PfeJQ

Impossible to scale (as with most KNN distances, I guess?) 🤷🏻‍♀️ but the idea is super pretty.

I dreamt about our project, David said that it is called nightmare. Understood some stuff about my own patterns of thinking which I didn’t like at all.

Discovered that I kinda like to write papers.

I find the posts like this one ridiculous.

https://www.tg-me.com/kovinagorelik/298

I just cannot believe that people really try to fit language learning in a square box.

For me the only way to learn a language is to understand in what form you like to consume the information, what sources you prefer to get it from and in what activities. And you have to learn grammar of course, this is crucial. And then you just do it.

It has to be extremely customized. Absolutely tailored. Otherwise it won’t work.

That’s why I think that this whole language learning industry that they sell to people is a fraud.

Well, I agree that learning grammar is a bit tricky.

Since I remember myself I always was a bit of a linguist. (Not a bit, very much so, even if a self-taught one).

I admit that without this base it can be complicated to do it fast. But I am sure there is a way.

Woke up with the thought that I’ve written 2 papers this summer and I have to write one more by 14.08.

Look how good my make up looks next to a pytorch profiler.

The more papers I have opened trying to write the state of the art, the slower I write. I have 30 papers open rn and I should close 27, read the three that are left and finally do it.

This summer I wrote …ok, in the last one I contributed but it was a big big journal article and there were a lot of things to write… three papers. And I haven’t even started my PhD.

If I didn’t love the team I am working with I would never do anything similar. It’s a big deal, believe me, even when you have help. But yeah, it seems I found taste for academics.

This is what my UPC master thesis supervisor, Marta R. Costa-jussà told me on various occasions: Ksenia, you are born to work in academic field. Go and do your PhD, don’t waste your time.

Thank you, Marta, looks like you are right.

Well… not to be calculative in order to ask for merits but in terms of fairness… whether I indeed wrote the last fresh paper or not.

The last paper is a reusing of the first paper which I wrote from scratch almost by myself.
I didn’t write the state of the art, Z&DG did, except the small part of LLMs that became the beginning of the introduction they also wrote (because me and GPT-4, we crafted it so well).

DP reused the architecture of the system part which I had to literally rewrite (the structure of the section) during 10 min at around 9pm tonight, and then Z and him finished it off with the meat of what he wrote grafted in the right places.

I did all the evaluation and results. We decided on the metrics we were going to use on Friday afternoon, on Saturday I had the results, I spent all Sunday on transferring them to the google sheets and the effing appendices with the millions of effing examples translated from Spanish to English (thank you GPT-4, you are a present from our technological gods).

DP wrote the conslusions and stuff.

Ok, I wrote a pretty big and really important chunk, I admit.
Also edited and proofread. I was a minion of Z this time, I’ve never seen such an academically efficient person in my life.

Ok, so yeah, I guess I also wrote it.