kushal thaman
@kushal1t
Followers
604
Following
399
Media
55
Statuses
559
member of technical staff @stanfordnlp
Palo Alto, CA
Joined May 2021
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Cheney
• 502132 Tweets
Michigan
• 179111 Tweets
England
• 140983 Tweets
Ireland
• 120349 Tweets
Lookman
• 40429 Tweets
Grealish
• 30130 Tweets
Arkansas
• 27353 Tweets
Big House
• 24969 Tweets
#MustafaKamalinAskerleriyiz
• 23885 Tweets
Super Eagles
• 20541 Tweets
PLANETARIO FURIOSO
• 19499 Tweets
Lee Carsley
• 18345 Tweets
Penn State
• 15909 Tweets
Ann Arbor
• 15182 Tweets
Ewers
• 10758 Tweets
Pedro Almodóvar
• 10640 Tweets
![Sergio +[@sonjahaxhiaendi(secondo account]](https://pbs.twimg.com/profile_images/1173545156634714112/60U-sH5u.jpg)
Pinned Tweet
Excited to share the first paper of my undergrad: "Incidental Polysemanticity" !
We present a second, "incidental" origin story of polysemanticity in task-optimized DNNs.
Done in collaboration with
@vclecomte
@tmychow
@RylanSchaeffer
@sanmikoyejo
(1/n)
Interested in mech interp of representations that deep networks learn?
If so, check out a new type of polysemanticity we call:
💥💥Incidental Polysemanticity 💥💥
Led by
@vclecomte
@kushal1t
@tmychow
@sanmikoyejo
at
@stai_research
@StanfordAILab
1/N
3
20
104
9
7
35
just wait until someone tries to shift the goalpost by saying “algebra, nt and geo are just <insert symbol manipulation/small search space/pattern matching>, combinatorics requires *real* reasoning”
5
1
52
@debarghya_das
@paulfchristiano
, IMO 2008 silver, founder of ARC
@paraga
, IPhO gold 2001, ex-CEO of Twitter
1
2
44
@PradyuPrasad
i read newton’s principia in high school (the original text was a bit difficult to parse, so i went with ‘Principia for the Common Reader’ by Chandrasekhar). i’ve been told it’s a rather popular choice, and many new textbooks (morin, kleppner) borrow heavily from principia still.
1
0
38
sam altman is speaking at a discussion event at the Stanford AI Club () this Wednesday (April 24)! it'll be a small-group event so we can have a high-context discussion. fill out the form to attend!
Stanford AI Club is inviting
@sama
for a small group discussion this Wednesday 4/24! Stanford affiliates apply at
1
3
16
1
2
33
Applications to Atlas Fellowship 2023 are out! If you are a curious high school student who wants to understand how the world works -- and change it -- apply at .
I did the program last year, and it was an amazing experience!
2
13
27
anon i know you like to be edgy but can you write a better abstract than this one?
1
2
26
had fun asking Sam lots of questions today!
We hosted
@sama
for our first speaker event! In a small group of ~20 people, we talked about everything from progress in scaling, the compute supply chain, agents, safety, takeoff speeds, timelines and more. 🧵 (1/n)
1
3
28
0
0
20
@PradyuPrasad
guy who worries about GPU centralization but can't do math
bro you're worried about the wrong Jensen's inequality
2
1
13
now that 4o is free, many people will want to cancel the $20/month subscription... unless OpenAI releases a new, frontier model that you can only access with Plus?
3
0
15
incredible talk by
@ashVaswani
at Stanford on the Transformer, how it’s evolved over the years, and exciting future research directions!
0
0
13
the latest draft of our paper is now out, with many new results!
Excited to share the first paper of my undergrad: "Incidental Polysemanticity" !
We present a second, "incidental" origin story of polysemanticity in task-optimized DNNs.
Done in collaboration with
@vclecomte
@tmychow
@RylanSchaeffer
@sanmikoyejo
(1/n)
9
7
35
2
3
12
3
0
11
@fionaleng_
contributing to sample size, a majority (and plausibly >80%) of my friends from olympiad camps a few years ago that stated they wanted to be physicists are now interning/ft-ing at js/hrt/2sig etc.
0
0
11
apparently this has become a hot take to say out loud, but the last 45 minutes of "Oppenheimer" (2023) is peak cinema and is absolutely critical to the movie and the life of J. Robert Oppenheimer.
1
0
11
people who might be interested, for visibility:
@aryaman2020
@itsandrewgao
@ChengleiSi
@rm_rafailov
@robert_csordas
@XiangLisaLi2
@RylanSchaeffer
@AaryanSinghal4
@cannonkissane
@kolelee_
@karansdalal
@_Aaditya_Prasad
@ayushh_agrawal
5
1
10
since ARC is back into the spotlight, may I also take this opportunity to remind people of another benchmark LLMs are quite poor at, by
@ylecun
and folks at Meta:
1
1
10
tbc i do believe combinatorics problems (p3/p6) are typically the hardest (followed by NT), and IMO 2024 was unusual in that the hardest problem was an algebra one. simply pointing out that this type of goalpost shifting is the perfect example of nebulous thinking in ai progress.
0
0
10
also, if you aren't around, i'll be asking sam questions, so DM me if you have questions you'd like sam to answer!
2
0
10
@prafdhar
has been an inspiration since high school for me; he won golds at IAO, IPhO and IMO for India back in the day, and went to work on really cool projects on VAEs, GANs, DDPMs, consistency models and much more after undergrad for OpenAI.
GPT-4o (o for “omni”) is the first model to come out of the omni team, OpenAI’s first natively fully multimodal model. This launch was a huge org-wide effort, but I’d like to give a shout out to a few of my awesome team members who made this magical model even possible!
137
344
4K
0
0
10
1
0
10
It's not clear what the implications of this are for mathematical research. Since the method used was very general, there would seem to be no obvious obstacle to adapting it to other mathematical domains, apart perhaps from insufficient data.
3
8
166
1
0
9
@rm_rafailov
@rajammanabrolu
my favorite papers by high schoolers are
@rowankwang
’s and
@AchyutaBot
’s , both excellent works.
0
0
9
correct me if i'm wrong, but it doesn't seem like OpenAI has been doing (publishing) work in deep RL for years now. the last work they published in the subfield dates back to late 2019 (safe exploration strategies in deep RL). why did they abandon it?
2
1
9
@michael_nielsen
+ Purcell on electricity and magnetism
+ Griffiths on quantum mechanics
+ Schroeder on thermal physics
+ Blundell on thermal physics
+ Campbell's biology
+ Chandrasekhar on Newton's Principia for the common reader
+ Barto and Sutton on RL
0
0
7
@itsandrewgao
@sama
glad you had fun! for stanford ppl who want to participate in future events, join the stanford ai club!
@stanfordaiclub
0
0
9
@deedydas
@AravSrinivas
“just” is a bit of a stretch; that’s 4 more points (out of the total possible 7) for every single indian contestant on the hardest problem of IMO day 1…
1
0
8
tfw you're sitting in a bay area cafe and can overhear two people intensely debating the use of tanh v/s leaky ReLU as activation functions...
0
0
6
what could the 'zero to one industry-defining' product be? agents that actually work seems like a plausible story.
I'm hiring at OpenAI.
We're building what (I think) could be an industry-defining zero to one product that leverages the latest and greatest from our upcoming models.
If you like product, deep technical challenges, and writing the future: my DMs are open!
59
132
2K
1
0
7
@__Charlie_G
if you can i’d be interested in seeing how loss curve fares against a gpt-2 that uses SwiGLUs
0
0
6
having a lot of fun attending the Stanford Center for AI Safety Annual Meeting!
0
0
7
@BlancheMinerva
@PradyuPrasad
Yep. So just to clarify, I read it on my own mostly out of interest, and I meant that it's a popular choice amongst quite a few olympiad people I know. I agree with Pradyu's main claim, I just think it's at least not universally true for physics.
0
0
7
@vclecomte
@tmychow
@RylanSchaeffer
@sanmikoyejo
We then proceed to provide a theoretical model of how this happens, analyzing the learning dynamics by studying the interaction between the forces of sparsity (namely feature benefit, interference and regularization), and run various experiments to confirm our results!
(4/n)
1
0
7
@vclecomte
@tmychow
@RylanSchaeffer
@sanmikoyejo
We present a non-mutually exclusive origin story of polysemanticity, showing that it can arise incidentally, even when there are ample neurons to represent all features — due to a single neuron correlating to unrelated features at the start (e.g. by random initialization)!
(3/n)
1
0
7
@vclecomte
@tmychow
@RylanSchaeffer
@sanmikoyejo
Polysemanticity is a property in DNNs that arises when individual neurons represent a mixture of unrelated features, making them hard to interpret.
The classic story is that NNs learn more features than there are neurons, causing superposition.
(2/n)
1
0
7
it’s a usual saturday morning in Wien, but there’s a sense of premonition in the air at Messe Wien, the
@iclr_conf
venue. maybe people are starting to feel the AGI…
0
0
7
@QualyThe
Seems plausible it was dealing with the board about Tigris uncandidly: (e.g. by not communicating about his private discussions with OpenAI investors about Tigris)
0
0
7
10/10 recommend planning a bird's eye view of 2024 (monthly/quarterly/semesterly). it's way more useful than i realized. i actually envision myself accomplishing the things i want, make useful estimates of their timelines & get a better sense of the year overall.
1
0
6
if americans found out about need aware admissions for international students and then did it for everyone there would be riots in the streets
1
0
7
this was dwarkesh’s best episode yet!
Had so much fun chatting with my friends
@TrentonBricken
and
@_sholtodouglas
.
No way to summarize it, except:
This is the best context dump out there on how LLMs are trained, what capabilities they're likely to soon have, and what exactly is going on inside them.
You would be
39
125
1K
1
0
7
@BogdanIonutCir2
@EvanHub
@_robertkirk
@gwern
and others interested: what open questions around path dependence (and mode connectivity etc.) in model training are you most interested in seeing progress on next? why?
(interested in running experiments & hunting useful directions)
3
0
5
E[number of times i hear someone say the phrase "or something" this weekend] = ?
2
0
5
@nickcammarata
have you written more about how one should *get into* meditation at greater length? wondering if there are tweet threads from you lying around, or if you’d be able to recommend a reading.
1
0
4
fun fact: the majority (and maybe as much as 80%!) of FLOPs are spent in the MLP layers.
0
0
6
RIP 😢
So sorry to see Jim pass. He impacted so many lives so positively including my own. I remember marveling at his Chern-Simons theory when studying pure math, marveling at the workings of his hedge fund when visiting it, and benefiting from his vision in funding neuroscience!
2
4
66
1
0
6
mildly annoying that the github API hasn't changed 'account password:' to 'personal access token:' despite changing the functionality
0
0
6
ok but who's working on a truthfulQ*A* benchmark for an AGI eval (me, i just trademarked the name)
0
0
6
this is exactly the sort of acknowledgement i like to see on ML papers
1
0
6
i suppose people went to see a "The Social Network" but on early-to-mid 20th century physicists, and were disappointed they didn't get that. that is a great but separate movie that needs to be made.
0
0
4
munnar is epic and y'all should definitely visit! here are some of my shots there from a few months ago:
I was finally able to see these breathtaking views of Munnar, Kerala towards the end of 2023🤌🏻
41
55
1K
1
0
5
A really simple but nice fact in linear algebra that I'd previously paid little attention to is that the theorem "any two bases for a finite-dimensional vector space have the same cardinality" needn't hold true if your scalars come from a ring instead of a field.
0
0
4
@emilyzsh
idk if elo will work for a v subjective, high dimensional space like books. i can get behind bayesian ratings tho? the best they have rn is
2
0
5
@RylanSchaeffer
@vclecomte
@tmychow
@sanmikoyejo
yes, thanks Rylan! it was a very rewarding experience. 🙂
0
0
5
twitter spaces are full of ppl trying to wrap their heads around the bellman equation bc of a rumor and its weirdly wholesome
0
0
5
sydney's back 👀
0
1
5
unless you want to watch croppenheimer, make sure you get tickets in theatres with the 1570 (15-perf, 70mm) or the DLx (eg DL2) screens.
0
1
5
0
0
5
0
0
5
@vclecomte
@tmychow
@RylanSchaeffer
@sanmikoyejo
Lots of more interesting findings in the paper, indicating the ample scope & need for further research!
If you have questions, contact me or the lead author
@vclecomte
! 🙂
0
0
5
@Jsevillamol
i suppose it’s harder to say if this is still true. 4o mini alone serves 200B/day…
GPT-4o mini launched 4 days ago.
already processing more than 200B tokens per day!
very happy to hear how much people are liking the new model.
570
535
8K
1
0
5
- non-rotationally symmetric noise breaks rotational symmetry and induces sparsity! (e.g. bipolar noise)
- sparsity increases with l1/noise up to a certain point, then goes down
- interplay between interference and the push for sparsity causes a different form of polysemanticity!
1
0
3
rubber duck debugging has been surprisingly useful while running deep learning experiments
0
0
4
have been reading Cade Metz’s ‘Genius Makers’ and it mentions some really cool anecdotes from the first decade of the deep learning revolution!
0
0
5