JavaScript is not available.

Dec 16, 2022

At OpenAI a lot of our work aims to align language models like ChatGPT with human preferences. But this could become much harder once models can act coherently over long timeframes and exploit human fallibility to get more reward. 📜Paper: arxiv.org/abs/2209.00626 🧵Thread:

314

1,322

Jeffrey Ladish

@JeffLadish

Wait snakes probably lost and then *re-evolved* *eyes*??

Quote Tweet

Anasazi Cannibal Groyper

@SattyuzeiYatura

Jan 24

Snakes are descended from burrowing lizards that lost their legs and eyes as an adaptation to the underground environment. When they emerged aboveground after ~50MY, they had to re-evolve eyes from scratch, resulting in an ocular anatomy unlike that of any other animal on Earth.

Show this thread

147

Katja Grace

@KatjaGrace

Jan 26

Alec Stapp

@AlecStapp

begging people to stop romanticizing the pre-industrial era

430

4,422

18.6K

Quote Tweet

Replying to @gbrl_dick

Unfortunately the full argument is not just "we can't rule out danger" but also "we have specific reasons to expect danger via training AGIs" (which don't apply to software in general) as per my recent paper: twitter.com/RichardMCNgo/s

Deep learning has taught us that we don't need to understand a capability in order to train it into a model. So Chollet's skeptical argument should actually alarm us: if we don't understand the models we're training, then we can't guarantee they'll behave in desirable ways.

Quote Tweet

François Chollet

@fchollet

It's hard to take seriously someone who talks about how "AGI" is imminent yet cannot give you a precise definition of what they mean by "intelligence" or tell you how it works in principle

Show this thread

173

Alec Stapp

@AlecStapp

*chatGPT passes United States Medical Licensing Examination (USMLE)* “This just tells us how terrible the exam is” Wild to watch people move the goalposts on AI in real time

467

Rival Voices

@nosilverv

Crazy to think about the pedagogical purchase this image has had

sotonye

@cybersygh

Jan 20

This is nuts

1:03

703.8K views

448

7,006

22K

Rival Voices

@nosilverv

God this is good

145

659

Sia - Elastic Heart feat. Shia LaBeouf & Maddie Ziegler (Official...

This one is... quite something.

youtube.com

Sia's official music video for 'Elastic Heart'. Listen to Sia: https://sia.lnk.to/listenYDAs featured on 1000 Forms Of Fear. Click to buy the track or album ...

I spent so long looping through this one, it's mesmerizing for reasons I can't articulate.

Quote Tweet

Oli

@SatyreContraire

Feb 3, 2021

only one thing is essential no matter how you do it you should never stop never stop Vibbin'

0:35

899 views

Quote Tweet

A well-timed pause is one of the most powerful techniques in basically any type of performance. Here's a masterful demonstration of that, from two of the best dancers in the world: youtube.com/watch?v=UOHG2K

Show this thread

Anderson & Brenda - Hormone soup (Spirit Bird)

The best bachata is slow, simple, and captivating.

Quote Tweet

tired: watching porn wired: watching these two dance youtube.com/watch?v=NziOyR

Show this thread

2:20

2.3K views

Ongoing thread of my favorite dance videos. Let's start with a new one a friend just sent. It's a slow burn, but it's so worth the wait. The section from 3:00 to 5:30 is particularly exhilarating.

youtube.com

Esta foi uma dança que despertou muitos conteúdos durante a execução dela!Foi a primeira dança desse dia entre eu e Brenda, não tínhamos preparado nosso corp...

Neel Nanda

@NeelNanda5

Excited to announce that our work, Progress Measures for Grokking via Mechanistic Interpretability, has been accepted as a spotlight at ICLR 23! (despite being rejected from Arxiv twice!) This was significantly refined from my prior work, thoughts in 🧵 arxiv.org/abs/2301.05217

0:21

35.7K views

Quote Tweet

Neel Nanda

@NeelNanda5

Aug 15, 2022

I've spent the past few months exploring @OpenAI's grokking result through the lens of mechanistic interpretability. I fully reverse engineered the modular addition model, and looked at what it does when training. So what's up with grokking? A

... (1/17) alignmentforum.org/posts/N6WM6hs7

Show this thread

464

Nicholas says, NGDP Targeting Now!

@captgouda24

Jan 19

It is not at all obvious that there should be a strong negative relationship between distance and trade — the marginal cost of distance is very low! This paper argues that business travel is *very* important for trade, using the 1985 USSR airspace liberalization.

It's Bad On Purpose To Make You Click by

@slatestarcodex

ft.

@eigenrobot

237

(((ل()(ل() 'yoav))))

@yoavgo

the real issue w/ "understanding the math behind machine learning" is that, esp with the newer models, this will not really buy you anything. you can have a complete mechanistic understanding of everything in the process, it will not help you understand why/how these things work.

1,143

Robert Wiblin

@robertwiblin

Aug 9, 2021

Kids today don't even know their history.

255

1,314

Jan 19

In my experience there's a particular mental sensation that arises when I'm in the middle of making a mistake. I'll give more details later - but before I anchor you, I'm curious if you relate to this, and if so how you'd describe it.

148

Henrik Karlsson

@phokarlsson

escapingflatland.substack.com

I don’t know anything about dating, have barely ever done it. But I am married to the coolest person I’ve ever met. So some friends asked for advice on finding someone like that, and I said, LOL why not. An essay

Looking for Alice

Not dating

114

701

Eliezer Yudkowsky

@ESYudkowsky

Reminder: *In principle*, minimizing prediction loss on Internet text could potentially grind *far* past human-level intelligence. Proof: Somewhere on the Net is a list of <hash, plaintext> pairs, in that order. A great predictor has to be smarter than the process predicted!

151

Jan Leike

@janleike

Reinforcement learning from human feedback won't scale. It fundamentally assumes that humans can evaluate what the AI system is doing. This will not be true once AI becomes smarter than humans.

800

It also conflicts with the prevalence of imposter syndrome, status anxiety, and low self esteem. I'd guess the published findings focus on very specific tasks + stated beliefs, not broader attitudes + revealed beliefs; if so they're much less practically relevant than they seem.

I'm confused about how to reconcile this type of overconfidence with the fact that most people take way too few risks in so many domains (e.g. moving to new places, asking people out, switching jobs, trying new hobbies, etc).

Quote Tweet

Rob Henderson

@robkhenderson

Jan 14

"All mentally healthy people, especially men, show a systematic bias towards overconfidence in a wide range of domains. In particular, people tend to (1) overestimate their capabilities (2) overestimate their control over events, and (3) underestimate their vulnerability to risk"

340

Andreas Stuhlmüller

@stuhlmueller

Jan 5

1/ Process supervision is safer and more transparent than end-to-end training of language models, but it's not clear that it can remain competitive. In our new paper we share our experience applying it to Elicit, and the workflows and tools we developed: arxiv.org/abs/2301.01751

Stephen Malina

@an1lam

Jan 14

If humanity were wiped from the face of the earth but it was otherwise left as is, which would reach human level intelligence first?

rats
22.5%
octopi
28.5%
some sort of corvid
22.5%
whales or dolphins
26.4%

1,864 votesFinal results

What if another person's worldview *isn't* actually reasonable? Well, it still seems much more reasonable from *their* perspective than yours! So if you can't see why, you're missing something - and as a rule you're missing more of other people's perspectives than you think.

Other related terms are "cognitive decoupling" and

@juliagalef

's "scout mindset". But again I want to emphasize the holistic aspect. Many people can decouple in isolated thought experiments; far fewer can decouple from foundational worldview assumptions.

Quote Tweet

Julia Galef

@juliagalef

Mar 8, 2021

1/ Hey guys, I’m excited to announce my first book is coming out April 13! It’s about something I call “scout mindset” - the motivation to see things as they are, not as you wish they were; to be intellectually honest & curious about what's actually true penguinrandomhouse.com/books/555240/t

Show this thread

Jan 14

This should be seen as a subset of "cognitive empathy" (aka "perspective-taking") focused on overall worldviews rather than everyday parts of people's perspectives (like personal conflicts). Seems worth distinguishing because it often anticorrelates with other types of empathy.

A term I've found useful lately is "intellectual empathy": the ability to see the ways that someone else's worldview is reasonable, even when it disagrees with your own. Emotional empathy is a vital skill, but intellectual empathy seems to be even rarer and more valuable.

245

1,411

Scott Lincicome

@scottlincicome

Jan 13

😥wsj.com/articles/the-d

157

941

6,257

The world if software engineering were as fun as coding interviews.

is being an absolute champ and has opened half a dozen Soylent bottles for me. Now all I need is dictation software and I'll be back in action.

Not painful any more, btw, just inconvenient. Painful part was the blisters the size of walnuts, which have now been excised.

Pictured: me trying to give a thumbs up (but also laughing because it's quite a funny situation to be in).

Update: unsurprisingly in hindsight, routine cryotherapy wasn't meant to hurt this much. Anyway, I now have second degree burns on both hands and will see a plastic surgeon tomorrow. Good reminder: nobody's infallible. If something feels like it might be going wrong, speak up!

Quote Tweet

Jan 9

Just had a routine medical procedure that was nevertheless the most painful experience of my life, leaving me curled up and nauseous. I'm now feeling very visceral empathy and horror for the thousands (millions?) of people who underwent surgeries before anaesthesia.

Show this thread

What We Owe The Future by William MacAskill

Jan 9

The best approach is might just be exposing people to more examples of extreme suffering. I have some disagreements with

@willmacaskill

's recent book but he does an incredible job at conveying the visceral importance of moral progress:

whatweowethefuture.com

A guide for making the future go better.

There's also a lot of work to be done in getting people to internalise the horror of extreme suffering. It's insane that we can't even agree that hell is bad! Increasing our civilisational empathy is one of the most important projects possible.

Quote Tweet