Artem Lukanin

Wake up babe, a new draft of the Jurafsky&Martin NLP book appeared web.stanford.edu/~jurafsky/slp3/

# On the "hallucination problem" I always struggle a bit with I'm asked about the "hallucination problem" in LLMs. Because, in some sense, hallucination is all LLMs do. They are dream machines. We direct their dreams with prompts. The prompts start the dream, and based on the LLM's hazy recollection of its training documents, most of the time the result goes someplace useful. It's only when the dreams go into deemed factually incorrect territory that we label it a "hallucination". It looks like a bug, but it's just the LLM doing what it always does. At the other end of the extreme consider a search engine. It takes the prompt and just returns one of the most similar "training documents" it has in its database, verbatim. You could say that this search engine has a "creativity problem" - it will never respond with something new. An LLM is 100% dreaming and has the hallucination problem. A search engine is 0% dreaming and has the creativity problem. All that said, I realize that what people *actually* mean is they don't want an LLM Assistant (a product like ChatGPT etc.) to hallucinate. An LLM Assistant is a lot more complex system than just the LLM itself, even if one is at the heart of it. There are many ways to mitigate hallcuinations in these systems - using Retrieval Augmented Generation (RAG) to more strongly anchor the dreams in real data through in-context learning is maybe the most common one. Disagreements between multiple samples, reflection, verification chains. Decoding uncertainty from activations. Tool use. All an active and very interesting areas of research. TLDR I know I'm being super pedantic but the LLM has no "hallucination problem". Hallucination is not a bug, it is LLM's greatest feature. The LLM Assistant has a hallucination problem, and we should fix it. Okay I feel much better now :)

The energy and carbon costs of deploying AI models have largely been unknown.. until now! 🌏🚀 With @strubell and @YJernite, We tested 88 models on 30 datasets from 10 different tasks from different modalities and found some pretty cool stuff! A thread 🧵:

Finally our Lightning Talks featuring: Max Nigri - Search workload acceleration, @avlukanin - Is LLM the right tool? , @wrigley_dan - Decompounding with Querqy, Maximillian Werk of @JinaAI_ - Vector Search - Why top-k is harder than top-1 ... [5/7]

@NiestrojRobert var is an anti-pattern. Only B can be useful for a short period of time, but if you extract new BankAccount() into a method, the type is lost, so it's a delayed burden on the reader

#Java question: what is your opinion on using var? Tomorrow i got a meeting on how our team want's to use var or not. I'm a big fan. There are also people how never used it and are against it. I guess the use of var depends. Here are some examples. What are your thougts?

The new search metric in town !! ;) #haystackconf

So it's your last chance to grab a ticket to Haystack Europe on Wed/Thur next week - we have an awesome lineup including talks on neural search, vector search & other #AI-powered search - come and join us online or in Berlin! haystackconf.com

We know it takes the world to map the world. Together with our partners at @OvertureMaps, we've launched the first worldwide open map dataset that will power the next generation of innovations in mapmaking. What will you make with it? Read more: bit.ly/44HEuFw

@JnBrymn And then your email will be suggested, when I write my email

I think language models will make it possible to write a beautiful email by prompting it with only a bulleted list of things to say. AND I think language models will make it possible to read emails faster by taking the original content and condensing it into bulleted lists.

TIL you can run SQL queries directly against CSV files as a one-liner using the default sqlite3 command line utility til.simonwillison.net/sqlite/one-lin…

Thanks to those who attended our tutorial on Keyphrasification at @ecir2022 . Our tutorial videos are available - youtube.com/channel/UCOPZb… All tutorial materials - keyphrasification.github.io #NLP #AI #MachineLearning Hope this helps the community.

Learning about adapters, replaceable language layers in deep neutral networks for multilanguage tasks #ecir2022

Crowd in front of the demo on “DicTAG: A Customizable Annotation Tool for Ground Truth Creation” by Fabio Giachelle, Ornella Irrera and @giansilv #ecir2022 @ecir2022 @examode

Types of false information, and how sometimes the difference is only in intent, which is very difficult to detect automatically @IAugenstein during #ECIR2022 Keynoteaddress @ecir2022

#ECIR2022 (Springer respectively) publishes its proceedings #openaccess for four weeks. The proceedings include a few nice #recsys papers, too.

‼️ There are some npm packages that are now deleting your ENTIRE hard drive's content or causing infinite loops in the name of peace. If you’re using Vuejs you MIGHT BE affected. These packages are: - node-ipc - colors This article discusses the details👇

@DmitryKan @danielwarna @FlaxSearch What is K (how many clusters) in your K-means? 10 for top 10 results?

Recording of our Haystack talk with @danielwarna is now available (powered by Aditya Jitta who helped with the demo and slides). lnkd.in/dv3VuudR Thanks @FlaxSearch for releasing it so quickly and for having us as the concluding presentation of the year!

This writeup about how the #log4j gets exploited is quite good. Particularly about how the strings can linger in the system for days until they hit a vulnerable program. blog.cloudflare.com/inside-the-log…

@JnBrymn But my most contribution was this tool, if you want to study Persian 😉 artyom.ice-lc.com/pvc/

@JnBrymn 😯 You can find some Hāfez's poetry with translation and Persian audio on a site I helped to create: pit.farsi.rocks/menu06.html

I found this beautiful and mysterious poem written by Persian mystic poet Hafez 600 years ago mooji.org/music/karuna/9 And in related, strange, and disappointing news I found that much of Hafez's poems have been faked aljazeera.com/amp/opinions/2…