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.

Ask HN: Does the AI training lag time always put it at a disadvantage?

2 pointsby notanotherericalmost 2 years ago
I noticed that although doing queries to the AI based chatbot like bing or bard, they have a fundamental handicap of being out of date with current events … is this going to be the Achilles heel preventing AI from totally displacing existing search?

1 comment

PaulHoulealmost 2 years ago
Good observation.<p>The Bing chatbot can answer questions about a sports game that happened yesterday because it runs a conventional query and feeds that to the chatbot.<p>I was interested in things like <a href="https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Cyc" rel="nofollow noreferrer">https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Cyc</a> ten years ago and made some progress in using Freebase as a generic database for information, particularly I discovered the first and (I believe) only correct conversion of Freebase to RDF, a process I could run on the weekly data dump from Freebase.<p>One thing I struggled with was that obviously people would need an updated database sooner or later but most users could live with an out-of-date database and I was quite frustrated that companies I contacted who were interested in the data weren&#x27;t interested in a repeatable solution but would instead go with whatever half-baked system they had.<p>It&#x27;s a tough problem for a model like ChatGPT because (1) the model training process is so resource intensive and (2) if you retrain the model the &quot;activation code&quot; used by the neural network will change so systems like LangChain that depend on storing activation vectors will break.<p>Right now the standard procedure for updating the model is to redo the training process completely with added documents, there is a research area called &quot;continuous learning&quot; where you could add a few documents here and there but the direct approaches suffer from &quot;catastrophic forgetting&quot; where the network loses what it already learned if you don&#x27;t blend your documents into a uniform sample.<p>I have a simple model that I retrain every day (or every few days depending on the mode I am running it in) and it takes just one minute so it is OK to build from scratch. I&#x27;ve got another model that does the same task that takes more like 45 minutes to train, for my particular task it doesn&#x27;t really beat my simple model, but for some similar tasks I&#x27;d expect it to better. That would already have me thinking about tradeoffs between training time and the results I get, people who are training foundation models are facing this problem times a billion.
评论 #36744461 未加载