TE
TechEcho
Home24h TopNewestBestAskShowJobs
GitHubTwitter
Home

TechEcho

A tech news platform built with Next.js, providing global tech news and discussions.

GitHubTwitter

Home

HomeNewestBestAskShowJobs

Resources

HackerNews APIOriginal HackerNewsNext.js

© 2025 TechEcho. All rights reserved.

Time is encoded in the weights of finetuned language models

124 pointsby convexstrictlyover 1 year ago

10 comments

alephnanover 1 year ago
By time, they’re talking about the writing style of a specific time period.<p>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.
评论 #38759422 未加载
评论 #38761403 未加载
评论 #38759236 未加载
评论 #38761544 未加载
评论 #38764355 未加载
convexstrictlyover 1 year ago
Twitter summary: <a href="https:&#x2F;&#x2F;twitter.com&#x2F;ssgrn&#x2F;status&#x2F;1738256456250470853" rel="nofollow noreferrer">https:&#x2F;&#x2F;twitter.com&#x2F;ssgrn&#x2F;status&#x2F;1738256456250470853</a><p>Github: <a href="https:&#x2F;&#x2F;github.com&#x2F;KaiNylund&#x2F;lm-weights-encode-time">https:&#x2F;&#x2F;github.com&#x2F;KaiNylund&#x2F;lm-weights-encode-time</a>
评论 #38758874 未加载
cwmooreover 1 year ago
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?
评论 #38760246 未加载
评论 #38760585 未加载
评论 #38761928 未加载
评论 #38769157 未加载
jiggawattsover 1 year ago
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.<p>Obviously this isn’t time travel, but more of a zeitgeist extrapolation.<p>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.”<p>This could have actual utility: predicting trends, fads, new market opportunities, etc…
评论 #38759380 未加载
评论 #38759094 未加载
评论 #38759282 未加载
评论 #38760210 未加载
bkfhover 1 year ago
Can someone ELI5 this?
评论 #38767747 未加载
simneover 1 year ago
Well, I think this could become one of most underestimated idea in LLM development.<p>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.
airockerover 1 year ago
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.
mjvmrozover 1 year ago
I think the general case is far more interesting than time specifically. There are cool functor&#x2F;analogy ideas here.
lprovenover 1 year ago
I thought it was encoded as a helix of semi-precious stones, but perhaps I am misremembering.
throwaway81523over 1 year ago
What about helixes of semi-precious stones?
评论 #38767239 未加载