Hacker News new | past | comments | ask | show | jobs | submit login
Time is encoded in the weights of finetuned language models (arxiv.org)
124 points by convexstrictly 11 months ago | hide | past | favorite | 55 comments



By time, they’re talking about the writing style of a specific time period.

Feels like a click bait title. Of course language model weights encode different writing styles. The fact that you can lift out a vector to stylize writing is also more interesting, but that’s also nothing newly discovered here. It should be obvious that this is possible given that you can prompt ChatGPT to change its writing style.


Besides what the sibling comment said, what's most interesting (imo) is that you can manipulate the vectors like that. The fact that you can average the vectors for January and March, and get better results for February, is pretty surprising to me.


This also generalises: https://arxiv.org/abs/2302.04863


Generalizing vectors in generative models seems like an incredibly useful thing to know about, if you want to use them more effectively. Blew my mind when I saw someone demonstrate doing vector math on a GAN a couple years back to move an "input image" around the space of outputs.

Maybe this could be useful for singling out post-LLM text and generating output that excludes it.


Why would it pertain only to writing style?


Interesting that writing style works, but other reflective actions don't.

Like, "only use the the 2000 most common words of the English language" or "the response should be 500 words long".


It does work on other reflective actions, parent is just wrong; in the paper, they specifically run the experiment on a dataset of political affiliation over time


From the title, I was thinking "of course the neural network of the LLM is a [cause-effect] sequence of words" thus time is encoded in each connection.




Thanks!


the X version worked fine for me. I don't know what you want to achieve by posting a link to a third party website.


I don’t have a twitter account, and many don’t. Twitter is slow, nitter is fast. Twitter has never reliably displayed threads for me. Nitter does.


>Twitter is slow

It’s faster than it’s ever been, and seemingly without 85% of its staff. Says a lot


And yet nitter is still faster with only 47 contributors total. Says more.


... because it is essentially a caching proxy to twitter?


So is twitter.com. Unless they aren't using a CDN lol


I think the advertiser boycott is a major contributor. Advertisements are slow...


X's reduced user base might make X faster than ever.


Some data for me:

Nitter: I get all 8 posts in the thread in 18 requests, 207 kB, 169 kB transferred.

X: I only get the first post, 128 requests, 11.45 MB, 2.11 MB transferred.


No, no it is not.


Twitter requires JS to work though.


Twitter doesn’t show replies if you are not logged in. As others have said, I also don’t have an account. So this link provides the full context. The twitter link only shows the post and no replies.


Twitter doesn't even show most recent tweets from profiles unless you are logged in now. They show a summary of the profile's activity. Nitter is great if you don't have a Twitter account.


I’m not on Twitter, and found it valuable to see the replies!


Allowing people without twitter accounts to view it.

Allowing those who would otherwise avoid twitter to view the content.


x.com links requires being logged in to even read the thread.


Twitter's not very usable these days.


He said he would sink the company...


All of the links were to a third party website.


I think I like time. Though spectral, indeterminate, presently a fixture, essential moments last forever but occur daily. Why would any network encode time if it were all just a crystal vase?


what are you on?


Crystal vase


Don't worry about the crystal vase.


I gotta get me some of that


Because people have to publish papers, that's why.


Beautiful. Thoughtful. Clever. Wise. In brightness like the face of Odin, in hearing like Moo, in spring and morning most goodly delight. Doing poetic justice to itself. Bringing up crystal vases! Per-bloody-fect.


Sooo… if I’m reading this right, it’s possible to force an AI into extrapolating into the future. As in, it’ll answer as-if its training was based on data from future years.

Obviously this isn’t time travel, but more of a zeitgeist extrapolation.

I would expect that if an AI was made to answer like it’s from December 2024 it would talk a lot about the US election but it wouldn’t know who won — just that a “race is on.”

This could have actual utility: predicting trends, fads, new market opportunities, etc…


Kind of. You still need some data from the "future" to extrapolate it: In the paper, they take an LLM finetuned on 2015 political affiliation data, and add to it the difference between 2020 and 2015 Twitter data, and show that the performance is better when the new model is asked about 2020 political affiliation.

So, the LLM still needs to know about 2020 from somewhere. In a way, you teach it about the task, then separately you teach it about 2020, and this method can combine that to make it solve the task for year 2020.


nah, this is not what they're talking about.


I don’t think it’d be nearly as accurate as purpose built future predictors.

LLMs aren’t a silver bullet for everything.


Ah, the bitter lesson teams it’s ugly head


> LLMs aren’t a silver bullet for everything.

Please explain this to my Product org.


lolol


Maybe less zeitgeist, but it would be really interesting to see what extrapolating future writing styles are like.


Here ya go: lorizzle.nl


Can someone ELI5 this?


A vector is a position in a dimensional space. In 2D space a vector is a point (x, y) like (1, 3) or (-2.5, 7.39). We can also do simple math on vectors like addition: (1, 3) + (2, -1) = (3, 2).

LLMs treat language as combinations of vectors of a very high dimension -- (x, y, z, a, b, c, d, ...). The neat thing is that we can combine these just like the 2D vectors and get meaningful results. If we have the vectors for the concepts "King" and "Woman", adding them gives a vector close to the one for "Queen"!

Once you know this, you can extrapolate and look for ways to categorize groups of vectors and combine them in new ways. As I read it, this research is about finding the vector weights for text from specific time periods -- i.e. January of 2021 -- and comparing them to the vectors for text from a different period -- i.e. March of 2021. It seems that all the operations are still meaningful, you can even do something like averaging vectors in January and March and getting ones that look like vectors in February!


Well, I think this could become one of most underestimated idea in LLM development.

To be honest, it is relatively obvious idea, to make vectors from timestamps and feed them to LLMs, but for some strange reason, nobody made this before and looks like, this is mostly unnoticed in NN community.


I think a more general way to think about it would be to add any data and reduce weight. For eg, if we want to create geography vectors, we would add all geography data to fine tune and then take a difference. Now add this to any other model with same architecture, and you have a geography capable llm.


I think the general case is far more interesting than time specifically. There are cool functor/analogy ideas here.


I thought it was encoded as a helix of semi-precious stones, but perhaps I am misremembering.


What about helixes of semi-precious stones?


Why is short story from 1968 mentioned here? Was it popular? Is it good? Were there some recent adaptations or homages?


Similarity of titles is all.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: