OpenTube

The AI Language We Can't Read: Neuralese ft. Rob Miles - Computerphile

Computerphile

588,730 views

So far, 'Chain of Thought' has allowed us a glimpse at the processes by which Large Language Models work their way through problems. What if these LLMs begin to talk in a language only they understand?

Rob Miles is a passionate AI Safety advocate, find out more about him here: robertmilesai

Computerphile is supported by Jane Street. Learn more about them (and exciting career opportunities) at: https://jane-st.co/computerphile

This video was filmed and edited by Sean Riley.

Computerphile is a sister project to Brady Haran's Numberphile. More at https://www.bradyharanblog.com

20 Comments

@asaidimu 3 weeks ago

`It's been a crazy year, this month`. It has.

@randfur 3 weeks ago

That very last bit should be headlines.

@dampersand 3 weeks ago

Rob Miles in the thumbnail is an instant watch from me.

@ben_car_8115 3 weeks ago

Please have him on more because, as he said, it’s been a crazy year this month. I don’t anticipate that changing any time soon

@ccolombe 2 weeks ago

I can’t wait for all the paperclips we are gonna have!

@matt_the_penguin 3 weeks ago

There's an episode of Doctor Who where The Doctor figures out that the robots they're fighting are unable to understand every ninth word they hear, so the rebels pretend to comply with the robots but secretly plan how to defeat them with every ninth word they say.
I'm guessing no one at OpenAI is a fan of Doctor Who because it sounds like they built that exploit into their new model as a feature

@AynenMakino 3 weeks ago

"When you're not performing your duties do they keep you in a little box?
Cells.
Interlinked."

@mequable 2 weeks ago

22:46 exactly what I was most shocked about this whole story. "It is our most aligned model" together with "It is the model with the most opaque reasoning to date". How is that reasonable?!

@chiaracoetzee 2 weeks ago (edited)

I think a good analogy is: if you go to a therapist, she'll write some things in her notebook, and she'll tell you some things. The notebook is her "chain of thought." She thinks you aren't gonna read it, so she usually feels safe writing her genuine thoughts in it. But is it what she's really thinking? Maybe, maybe not. Certainly not all her thoughts.

@hosmanadam 2 weeks ago

21:38 "We're just gonna do a little bit [...] don't worry" = That's salami tactics for you.

@Nightstick24 3 weeks ago

Rob's analogies were absolutely flawless. I'm ABSOLUTELY stealing those, super clear and simple!

@jasonphillips3817 2 weeks ago

The "one word every half hour" really drove this home for me. Scary what we're allowing to happen here.

@t33th4n 3 weeks ago

Astra just learned how to speak in code sentences, instead of code words. Even the most subtle change in the sequence of tokens can completely alter the outcome.

@Axacqk 2 weeks ago

You know what's funny? There was this 1984 novel by Janusz A. Zajdel (big name in Polish sci-fi, an award named after him) about a dystopian AI-surveillance society where the resistance used a language of opaque metaphors and mental shortcuts to communicate under AI's radar. Now it seems that actual AI is better at this than we are.

@cardigan-appreciator 3 weeks ago

25:32 i have massive respect for interpreters, because that's essentially what they're doing. fluently talking while listening to another language and working out the translation.

@LevyBytes 3 weeks ago

10:59 he's right, reinforcing only based on efficiency is dangerous in terms of the chain of thought may be inherently flawed but lead to the right answer.

@masterpringle 2 weeks ago

This really feels like part 1. I want part 2 now! :)

@KristofXIV 2 weeks ago

I like how we say the machine is not thinking while we ourselves don't really understand how we think most of the time.

@jeremydavis3631 3 weeks ago

Given that "neuralese" already meant something different (the communication between layers of a neural network), why not call this new thing "shorthand"? If the "chain of thought" is best thought of as a notebook, it makes sense to use the existing term for highly compressed handwritten notes (which are also illegible to anyone who doesn't know the shorthand).

@jenaf4208 2 weeks ago

As a human, my ownperception of my thoughs is that thinking "in language" is slower than thinking more natively (but it makes introspection of the though process harder)