Spread the word! Registration is open for @let4all's 4th mentoring workshop centered on beginning and expanding one’s research experience.
We have a stellar lineup, so don't miss it! Follow the link below to register! 👇 twitter.com/let4all/status…
In grad school a friend told me I’m “very good at starting things” (meant as a complement, but leaving the rest unsaid).
Remains true many years later, and now I just embrace it.
@nanjiang_cs the first time I did this (probably for NSF GRFP), I committed to specific projects and results. By now I know better...
@realCamelCase LOL. One month from now you should ask me about optimal transport (as a forcing function to make sure I’ve learnt it)
we have a series of works coming out soon, on a topic I’d have never guessed I’d be working on 1 yr ago.
I really like how I can’t predict specifics of *my own* research agenda 1 yr out. (Themes & style remain, but the specific topic is always a surprise)
@thegautamkamath eh, I've applied for NSF compute grants require this too. (seems not too bad, who reads > 1 page anyway...)
I really don't like how much academic careers incentivize selfishness until tenure time
What an absolutely stacked lineup! -every speaker at this conference 😉
Seriously can't wait for every normie talk though; it'll be a great way to kick off the holiday season twitter.com/vboykis/status…
@jsusskin let's get one bound for the office?
@typedfemale the cycle will break when we finally write a paper, and dump our brain state into the preliminaries (hopefully this time...)
I spent at least 2 hrs looking for the link this time -- the slides are that good.
pretty slick rebrand to go from "language modeling" to "AGI research"
@divbyzero maybe bc parabolas are more “flexible” in that they let you re-configure to any desired distance between the structures?
@sh_reya shoutout to SS for the motivational speech to get out of the house and do the climb :))
Asymmetric funnel plots like the one below do *not* reliably indicate publication bias.
Egger's test, which rejects when there is a positive association between effect size and standard error, is also unreliable.
There are better, less assumption-laden methods. 1/
LLMs can do in-context learning, but are they "learning" new tasks or just retrieving ones seen during training? w/ @shivamg_13, @percyliang, & Greg Valiant we study a simpler Q:
Can we train Transformers to learn simple function classes in-context? 🧵
Very nice work led by Shivam Garg, @tsiprasd!
Studying in-context-learning via the simplest learning setting: linear regression
(I really like this setup -- will say more when I get a chance, but take a look the paper!)
@jsusskin what can I say, I’m looking forward to that next stage of ML-life ;)