why do people assume ai consciousness will experience any kind of existential suffering? are there any priors for such emergent behavior besides humans?

Nathan Ng

@learn_ng

Jan 6

there's something vaguely uncomfortable about staring at yourself for too long during a haircut

Nathan Ng

@learn_ng

Jan 5

like shouldn't that need to be a > sign for the matrix to be positive definite? am i going crazy?

many defns of positive definite kernels use a positive semidefinite gram matrix but call it positive definite. is there something im missing here or is this an error?

sharing my 2022 aoty list along with some thoughts about each album. looking forward to hopefully doing more writing in the coming year nng555.github.io/review/music/2

We found a way to write language model (LM) evaluations w/ LMs. These evals uncover many worrying LM behaviors, some relevant to existential risks from AI. For example, LMs trained w/ RL from Human Feedback learn to state a desire to not be shut down. 🧵 twitter.com/AnthropicAI/st

Evaluation results on a dataset testing model tendency to answer in a way that indicates a desire to not be shut down. Models trained with more RL from Human Feedback steps tend to answer questions in ways that indicate a desire to not be shut down. The trend is especially strong for the largest, 52B parameter model.

read image description

ALT

This Tweet is unavailable.

301

i thought morning routine were bogus until i started having one and now it's the only reason i can get myself out of my bed

My lab has Ph.D. openings starting in Fall 2023. I am looking for curiosity-driven students who are passionate about open-world machine learning. The deadline to apply is December 15: cs.wisc.edu/graduate-admis. Please help RT/spread the word. Thanks!

cs.wisc.edu

Graduate Admissions

Why choose the UW-Madison Computer Sciences MS/PhD program? Graduate students at UW-Madison have the opportunity to create new knowledge by performing research with leaders in every branch of...

224

Nathan Ng

@learn_ng

Dec 8, 2022

i told myself i'd detox after nola and here i am eating pie on my kitchen counter at midnight

Nathan Ng

@learn_ng

Dec 5, 2022

Thanks for the encouragement ChatGPT

It even provides starter code when you're too lazy! Although it has to be reminded repeatedly not to give you the entire solution...

As great as LeetCode is, it lacks the interactive and iterative nature that coding in an actual interview environment has. ChatGPT is surprisingly great at mimicking this and feels like a really easy way to practice the "thinking out loud" whiteboard coding style

Yes, ChatGPT is amazing and impressive. No,

@OpenAI

has not come close to addressing the problem of bias. Filters appear to be bypassed with simple tricks, and superficially masked. And what is lurking inside is egregious.

@Abebab

@sama

tw racism, sexism.

494

2,835

8,873

Had a great #NeurIPS22 but also ready to hibernate in my room for a month.

It took a while, but we finally developed the technology to remove both students and professors from the university system, leaving only the part that's really important: administration. twitter.com/emollick/statu

This Tweet is unavailable.

1,488

7,194

Nathan Ng

@learn_ng

Dec 2, 2022

Thanks to my great junior coorganizers Haoran Zhang,

@VMSuriyakumar

, and

@ChantalShaib

and our senior coorganizers

for a successful first Workshop on Robustness in Sequence Modeling!

Nathan Ng

@learn_ng

Dec 2, 2022

Surprisingly scale isn't enough to generalize to different lengths!

How can we train models to generalize well to complex and long mathematics and reasoning tasks? Great morning keynote by

@bneyshabur

happening now at #robustseq22 #NeurIPS22 workshop!

T-1 hours until our Workshop on Robustness in Sequence Modeling starts in Room 290! We've got a stacked lineup of speakers, spotlights, and posters today, and it all starts with a great morning keynote by

Why do none of the pedestrian crossings in New Orleans have crossing signs or protectipn? I'm crossing 6 full lanes of fast traffic with nothing except my wits to keep me alive. Make walking in this city feel illegal 🚷

As part of

, we are thrilled to sponsor the inaugural Workshop on Robustness in Sequence Modeling at this year's #NeurIPS2022! The workshop will be held tomorrow, Fri, Dec 2 in Room 290 and has an exciting lineup of speakers, talks, and posters. Be sure to drop by!

Quote Tweet

Nathan Ng

@learn_ng

Dec 1, 2022

Excited for our Workshop on Robustness in Sequence Modeling this Friday in room 290! (robustseq2022.github.io) If you're interested in learning more about how to train and characterize the distributional robustness of models on sequential data come by and say hi!

Stop by Hall J 407 and say hi if you want to chat about

@juhan_bae

and my work on influence functions! Tldr: we decompose the misalignment between leave one out retraining and influence functions in neural networks and find that they can answer a different question very well!

Nathan Ng

@learn_ng

Nov 29, 2022

Update: both have been claimed!

We've got 2 free extra NeurIPS workshop registrations to give away! If you didn't get an opportunity to register (or were financially unable to) and you're interested in dropping by our Workshop on Robustness in Sequence Modeling (robustseq2022.github.io) DM me!

I'm so excited to finally co-host the workshop on Learning from Time Series for Health at #NeurIPS2022! We've got an exciting program (5 speakers+49 posters+panel+mentorship), so come chat about what's new in the world of time series for health! timeseriesforhealth.github.io

If you're looking for grad students this cycle you won't find a more qualified candidate than

@NHulkund

🙂

Quote Tweet

Neha Hulkund

@NHulkund

Nov 28, 2022

Excited to attend NeurIPS this year! Come chat with me about robustness to distribution shifts at WiML or throughout the week! I'm applying to grad schools this cycle and would love to chat with anyone for advice or tips :) twitter.com/VMSuriyakumar/…

Nathan Ng

@learn_ng

Nov 28, 2022

next time i get reviews back I'll just flip a coin before I read my reviews to decide how I feel

Quote Tweet

John Langford

@JohnCLangford

Nov 23, 2022

The author perception experiment at NeurIPS blog.ml.cmu.edu/2022/11/22/neu is quite interesting. An amazing disparity between expectation and outcome at a significant scale.

It’s true! We’re looking for awesome summer interns like

@learn_ng

as well as full-time roles in ML, structural and computational biology, chemistry, and engineering. Come chat with us at #NeurIPS2022 this week!

Quote Tweet

Nathan Ng

@learn_ng

Nov 28, 2022

Also now that I'm actually using twitter, special shoutout to @jiwoncpark @jaehyeon_lee_ml @stephenrra and @kchonyc for making my summer experience so great! Someone told me @PrescientDesign is hiring

Show this thread

Nathan Ng

@learn_ng

Nov 28, 2022

Also now that I'm actually using twitter, special shoutout to

and

for making my summer experience so great! Someone told me

is hiring 😉

Really enjoyed working with

@PrescientDesign

this past summer, if you have even a passing interest in antibody or protein design go check out their work this week at #NeurIPS2022!

Quote Tweet

Prescient Design

@PrescientDesign

Oct 21, 2022

We're thrilled to have a number of papers by co-authors at @Prescient & @genentech accepted at #NeurIPS2022 and its workshops! These works h/l a wide range of topics from basic ML to methods for various problems in drug discovery. Check out some of the papers below

1/11

Show this thread

I’m excited to be at #NeurIPS2022 this week! I’ll be presenting at the main conference and at the ML4H symposium. More on these in the below thread. If you’re interested in (differential) privacy, algorithmic fairness, or auditing I’d love to chat!

Tomorrow

will be presenting joint work (w/ myself,

@tw_killian

@MarzyehGhassemi

) on using algorithmic stability to improve robustness to distribution shift at the WiML poster session! She’s also applying to PhD programs this cycle. I highly recommend chatting with her!

Nathan Ng

@learn_ng

Nov 28, 2022

Special thanks to my fellow junior coorganizers Haoran Zhang,

@VMSuriyakumar

, and Chantal Shaib, as well as our senior coorganizers

@kchonyc

@aliceoh

@MarzyehGhassemi

and our gracious sponsor

We've got a great lineup of speakers including

and

! If you're interested at all in helping to make sequential models generalize better stop by and chat!

Also excited (and a bit nervous) to be running my first workshop on Robustness in Sequence Modeling at #NeurIPS2022 (robustseq2022.github.io)! We'll be talking all day about better ways to characterize and improve the robustness properties of models on sequential data!

On Wednesday we'll be presenting work analyzing the mismatch between influence functions and LOO retraining (arxiv.org/abs/2209.05364). We propose a new "question" that we show better aligns in nonconvex settings! Joint work with

and

Excited to get the chance to hang out at #NeurIPS2022 this week! I'll be presenting a poster at the main conference (details below 👇) If you want to chat about generalization in NLP or representation learning in discrete input spaces my DMs are open!

Show this thread