@kchonyc
a mediocre combination of a mediocre scientist and a mediocre advisor at @nyuniversity (@CILVRatNYU)
wow young @W4ngatang from the pre-pandemic era! i personally enjoyed this paper arxiv.org/abs/2106.02736 by @kartik_goyal_ , @redpony and @BergKirkpatrick that gave a better motivated approach to sample from BERT.
Sebastian Gehrmann@sebgehr·Been thinking about this one all week... Alex and @kchonyc were way ahead of their time. x.com/d_aumiller/sta…
🤦♂️
Djamé..@zehavoc·@kchonyc @srchvrs Don’t forget the training data collectors that are flooding conferences (at least for ARR) with their AI papers to get reviews
globally there are roughly 30k phd students focusing on AI and adjacent areas. assuming half of 60k abstracts at iclr are placeholders due to neurips, it's roughly one a phd student. from this perspective .. not sure if we should be surprised.
ah ... this is nice: hermes desktop by @NousResearch with privately hosted @deepseek_ai v4 flash.
this was a pretty crazy movie i couldn't take my eyes off the screen even for a second. didn't even notice that 2h40m passed by .. wow imdb.com/title/tt273690…
unsolicited plan for AI slowdown: 1. anthropic locks all employees' accounts/access and closes down API access to their models. anthropic continues to pay the employees. 2. anthropic deletes all code, model weights and data sets of all type. 3. anthropic uses remaining money to sue every departing employee who joins another firm not for non-compete but for violation of confidentiality and IP. 4. when anthropic runs out of money, anthropic starts selling their IP piece by piece over decades. 5. prof ... nope. guaranteed slowdown of AI development.
it's not "out-of-control" at all. openai had all the control in the world (i.e. they could always literally turn off all their machines across all data centers.) they decided not to based on carefully calculated trade offs between finance-economy-reputation-customers-etc. just like any other co'a e.g. anthropic
ACL 2027@aclmeeting·ACL Sustainable Reviewing Policy: We are introducing changes in the @ReviewAcl reviewing and submissions. The changes will cap authors and introduce changes in the reviewing to keep our community sustainable. #NLProc
waiting for Nike's justdoit AI cloud
PSA:
Kyunghyun Cho@kchonyc·if you are the only one who can build this amazing tech and believe it is super dangerous, it’s very easy: you quit for the sake of humanity. if you are not the only one who can do it, you are definitely over-valued. you can’t claim to be both.
finland and @AaltoUniversity always have a special place in my heart
Eric Malmi@ericmalmi·I had the privilege of sitting down with @kchonyc for our new AI podcast "And humanity created intelligence". we discussed the backstory behind the attention paper as well as... 🧵
why is @GeminiApp suddenly answering in the first person pretending to be me ... ? @orf_bnw @_angie_chen ?
still a hot tip
Kyunghyun Cho@kchonyc·a hot tip for @GoogleDeepMind : open source gemini 3.5 pro and become the most beloved tech co overnight.
genius
Minsik Oh@minsik_nlp·@kchonyc We can ask Ox Alpha to hack opencode itself and report what itself is
i don't get it. can someone just ask fable, codex, muse, etc. (maybe not gemini) to hack opencode and figure out where this model is served and who trained this model?
OpenCode@opencode·Ox Alpha (stealth model) is free for the next week - 1M Context - Multi-modal - Zero Data Retention Generous rate limits, near unlimited usage We have capacity for 100T tokens per day, lets see what you can do
rl-based ai cannot implement rsi, because we have to design a reward function and we recursively suck at it.
after finishing <crypto> by levy, just started <Inside Digital Advertising> by mackenzie & caliskan: i wonder what would've happened with strong digital cryptography was already there at the very beginning of the internet. would we have avoided ad-driven internet economy? t.co/FEV6X5Nwof
mission, vision and persistence ❤️
Chris Dyer@redpony·I'm so happy to be able to announce the first general purpose sign-language-to-text translation (SL2T) model from my team that's out today and powering a new ASL input feature on @Android phones (and don't worry, we're not stopping at ASL).
i gave my first public talk in front of korean teenagers earlier this summer. it was most stressful but also the most learning experience i had ever had. it was in korean, which made it even more challenging but satisfying. thanks to hoam foundation for this amazing opportunity and to all the amazing students who came to hear my dull story! t.co/c54iHhVugw
sometimes our job is to use our expertise to build and contribute something valuable back to the community, and @bingyan4science just did it with t.co/9Wj315v7tt. AI for chemistry but well grounded and supported by hard evidence, built by a chemistry PhD & computer science PhD student @bingyan4science ! give it a try
... umm ...
after all, it is truly all about "loop" isn't it? (yeah ... our original name truly sucked ... manifoldify 🤦)
whoever vibe coded the backend of EU border control, please ask claude to make it faster RIGHT NOW ... ugh x.com/kchonyc/status…
oh this border control for leaving finland is one of the slowest i've experienced ever. i mean .... the slowest per agent. wow
if you know, you know. i smell the change and future in the air. denis zorin will be the inaugural deal of the new Courant School! cims.nyu.edu/dynamic/news/1…
this is why it is better to use BoTied by @tagasovska and @jiwoncpark (arxiv.org/abs/2306.00344) than hypervolume which is sensitive to scaling. x.com/openai/status/…
since they all love biosecurity and use it to scare people of AI, i hope they know if it happened in a BSL, the lab would be shut down pretty much right away.
ok gemini seems to work better with google calendar (duh! but finally!) another issue: i can't use @GeminiApp to analyze any legal doc with redlines. it's simply blind to various markups ... 🤦 @orf_bnw @_angie_chen pleaseeeeeee
hold on .. did they change the swedish name of tikkurila? or is my mind playing a trick on me?
google map didn't tell me thisriute was this awesome. just realized it thanks to @PetarV_93 ! 🤣
in introducing @sedielem , razvan mentions two historical artifacts, echo state networks and theano. no one in the audience but old wise ones laughed.
an actual rigorous test for so called RSI is to use it to arrive at the current sota LLM from scratch. if it is exponential, the initial point really shouldn't matter. otherwise, it is just another occasion of computer scientists re-discovering for loop.
i am learning that a lot of people secretly have always wanted to be G-men.
reading it now, and i already see some stark parallel between NSA and some of the so-called frontier labs; both hiw (seemingly) competent they are and how blind they are. x.com/0x00b1/status/…
umm nyt, if you thought susan zhang was worth hearing from, add asian fears in the title together as well. you can't really brush off asian americans aside in this political discourse because one of the victims didn't fit the discourse.
the era of computational math and stat is here: a proof within computational budget and constraints by search rather than a proof under strong assumptions by derivation. along this line of progress, @luciusbynum, rajesh ranganath and yours truly put forward computational identifiability. the link to arxiv below.
preprint at arxiv.org/abs/2606.19361
looking at july 1 x.com/andrewe_dunn/s…
perhaps the most important lesson for all of us who are not dillusional from @huggingface's incident report: ``` ... When we started the log analysis, we first used frontier models behind commercial APIs. This did not work: ... these requests were blocked by the providers' safety guardrails, ... We ran the forensic analysis instead on GLM 5.2, an open-weight model, on our own infrastructure. This had a second benefit: no attacker data, and none of the credentials it referenced, left our environment. ... ... The practical lesson ... have a capable model you can run on your own infrastructure vetted and ready before an incident, both to avoid guardrail lockout and to keep ... credentials from leaving your environment. ... ```
this game is so unfair. spain has 11 players while argentina has only 1 player on the field.
not entierly sure about @OpenAI 's choice of the authoritative source
another hot tip: call it gemini-4-flash and release right now
@sharongoldman well .. okay not only those in bay area, but the concentration there can't be denied.
so sad a lot if us (inc me and my friends) in AI have lost the ability nor desire to hear ourselves talk and to listen to others. a good podcast interview today of jasmin sun
tick tock tick tock x.com/kchonyc/status…
in estimation and inference, i want lower cariance. in simulation and data generation, i want higher variance. how about bias? hmm
i've tried. i've really tried to love gemini. but ... just subscribed to chatgpt again after nearly two years.
half way in on this amazing episode on designing a cancer vaccine by @owl_posting with @iskander and ben vincent. link below
spotify: open.spotify.com/episode/5xk4Oj…
it's unfortunately not even surprising to see this from @GeminiApp . i'm done with google gemini integration with google workspace. cc @orf_bnw @_angie_chen
the final mission at #icml2026: escape coex mall.
a hot tip for @GoogleDeepMind : open source gemini 3.5 pro and become the most beloved tech co overnight.
you're welcome
ahhh this tingly feeling . love it
[네이버지도] 이조복집 서울 강남구 봉은사로112길 28 102호 naver.me/GqfT9GZT
my first poster presentation at icml was exactly 15 years ago. looking forward to another in 2031 x.com/asalam_91/stat…
beer tasting in B1 at coex. if you feel thirsty
if you want to experience where i used to hang out demolishing real kbbq and emptying soju bottles when i was growing up, check out [네이버지도] 돼지굴뚝 서울 동작구 동작대로25길 51 naver.me/56XqGKXV
at icml this week in seoul, please reach out for coffee, beer, etc. especially if you're a student (or a group of stduents) or postdoc (or a group of postdocs).
a few things i have so far in seoul this time: - 강된장 - 우설 - 회덮밥 - 새우회 - 모둠회 - 돼지국밥 - 보리굴비 - 쭈꾸미 - 짜장면
it's great to hear from leslie kaebling and @polynoamial at the global ai frontier symposium in seoul
the most confusing soup ever
full house at the @nyuniversity 's Global AI Frontier Lab's second annual workshop in seoul!
although it's sad (emotionally and financially,) i totally understand this decision. i've immensely benefited from these monthly free credits on @LightningAI over the past few years. thanks, @LightningAI !
i raise you x.com/roydanroy/stat…
kiosks in seoul are completely fucked up for all
traffic lights in seoul are completely fucked up for pedestrians
seriously? come on, @orf_bnw and @douwekiela . x.com/kchonyc/status…
come on .. it's actually become a bit embarrassing for me to even use @GeminiApp
watching congo vs uzabec: korea really is the worst team this world cup
i want to meet and chat with students (groups), startups and reaearch teams in big co's in seoul in the afternoons on monday, tuesday and wednesday. what would be the most effective way to do so? asking for me
arxiv: arxiv.org/abs/2102.00863
