Peter Zukerman

NLP researcher @ Infoscitex, former CLMS @UW, @UofIllinois.
日本語, ru

[Peter Zukerman @peter_zukerman]

Are you a beginner in natural language processing (#nlp #nlproc)? Here are my top 3 recommendations for resources to get you started:

[Kyle Chan @kyleichan]

The paper looks at 328 prefecture cities in China from 2010 to 2023. The paper uses a classic staggered DID approach to show that EVs as a share of new and used car sales increased more in cities that got a high-speed rail line compared to cities that didn’t.
2/

[Sasha Luccioni, PhD 🦋🌎✨🤗 @SashaMTL]

Maybe if you shared some numbers, we wouldn't have to make shit up 🤷🏼‍♀️

[Peter Zukerman @peter_zukerman]

@hyusapx Cool idea. None of these properties are “emergent” and it definitely doesn’t “understand” chemistry, any more than it “understands” etymology

[Peter Zukerman @peter_zukerman]

@oledoteth Wasn’t this book heavily scrutinized and less than 10% of the experiments could be replicated?

[Stefania Druga @Stefania_druga]

🚨1/7: Our new paper from #FAccT2024 explores how anthropomorphization in descriptions of “AI" systems influences users' trust. Does calling products "intelligent" make people trust them more? #FAccT2024
@NannaInie @Stefania_druga @peter_zukerman and @emilymbender

[Peter Zukerman @peter_zukerman]

@ScriptingJapan Where can I find info about your textbook!

[BURKOV @burkov]

Don't let them fool you: AGI today is no nearer to us than it was two years ago. While ChatGPT might appear to be a step closer to AGI, from a scientific standpoint, it's not: training a neural network to predict the next word is not groundbreaking science. Achieving AGI would necessitate multiple significant scientific breakthroughs. I'm talking about genuine breakthroughs, where real scientists engage in real science, not just enlarging the scale of the autocomplete.

[Nanna Inie @NannaInie]

What motivates people to trust 'AI' Systems?

✨BRAND NEW✨and definitely shiny paper still dripping from that fresh wet arXiv DOI.

Qualitative study with more than 450 participants across the world.

@VILLUMscience @CCER_itu @ITUkbh

https://t.co/ZdavmdICpE

[Crémieux @cremieuxrecueil]

Top-tier AI researchers are from all over, but most of them reside in America.

[Peter Zukerman @peter_zukerman]

Last day of Apollo, trying Narwhal. Really like that it loads links before comments and lets you view in same screen as comments, incentivizing actually reading the article before engaging

Other than that some QoL missing but anything is better than official

[Peter Zukerman @peter_zukerman]

@RobbieLemos I feel like you’re calling me out 😅

[Peter Zukerman @peter_zukerman]

Detecting AI-Generated Text, huge issue right now. Paper from late April says that "detecting AI-generated text should be almost always possible but one
would need to collect more samples depending on the hardness of the problem".

What do you think? Is it a dataset issue?

[Peter Zukerman @peter_zukerman]

Here's the paper:
https://t.co/GdqPgVNqTX (2023 Apr 10)

[Peter Zukerman @peter_zukerman]

I'm not the biggest fan of #duolingo but I've decided to give it a shot for German for 30 days and see what happens. Usually, I track known words with Anki so I'll have to come up with some workaround, but I'm curious to see how I feel after a month.

#languagelearning

[Peter Zukerman @peter_zukerman]

Here's the paper with those findings
https://t.co/xXnUK3vfLi

[Peter Zukerman @peter_zukerman]

Researchers found that Bing Chat, NeevaAI, Perplexity AI and YouChat responses only back up 51.5% of sentences with citations, and of those, only 74.5% support the sentence...

Who thought this was remotely OK? Even if it's 99% that's still not enough... what are your thoughts?

[Peter Zukerman @peter_zukerman]

Recent paper exploring this - distillation mechanism to curate task-specific models, outperforming LLMs: https://t.co/obh5ARZZOK (2023 May 3)

[Peter Zukerman @peter_zukerman]

LLM hype is still very real, but in terms of industry, I think shrinking model sizes, using less training data, and curating models to be task specific is the way to go.

For example, supervised prompting towards smaller task-specific models.

What do you think?

[Peter Zukerman @peter_zukerman]

scroll through “for you” on Twitter without falling behind ChatGPT/gpt4/autogpt/#aihype impossible challenge

[Peter Zukerman @peter_zukerman]

@gregisenberg I’m not sure if speaking 20 of 700~ languages constitutes “every” language