Mastodawn

Victor Zhong Jun 1, 2023

Before I start at Waterloo, I will be in NYC with Microsoft Research as a postdoc, where I will work on RL + NLP. If you are in NYC or Toronto and want to get in touch, please reach out!

Show thread

Victor Zhong Jun 1, 2023

I have had a wonder 5 years at the University of Washington. I am tremendously grateful to my advisor Luke Zettlemoyer for his help in my journey - I could not have imagined a better advisor. I also want to thank UW NLP and my office mates in CSE318. I hope our paths cross often in the future!

Victor Zhong Jun 1, 2023

Some personal update: I will join the University of Waterloo as Assistant Professor and Vector Institute as Faculty Member in 2024! I am *very* excited to be back in Canada to help grow the Canadian AI ecosystem! Please apply if you are interested in a PhD at the intersection of NLP and ML!

Victor Zhong Feb 1, 2023

Our work on Reading to Learn, along with terrific work from
@anima_anandkumar and
Karthik Narasimhan was recent featured in @QuantaMagazine !

https://quantamagazine.org/machines-learn-better-if-we-teach-them-the-basics-20230201/

By far the most professional interactions I've had with a news org - a lot of work put into fact checking+editing.

Machines Learn Better if We Teach Them the Basics | Quanta Magazine

A wave of research improves reinforcement learning algorithms by pre-training them as if they were human.

Quanta Magazine

Victor Zhong Dec 20, 2022

Akari Asai Dec 20, 2022

New paper 🚨

Can we solely rely on LLMs’ memories (eg replace search w ChatGPT)? Probably not.
Is retrieval a silver bullet? Probably not either.

Our analysis reveals that LLMs' memorizations are still limited and scaling won't help much in long-tail distributions.
We show that adaptively incorporating non-parametric memories (eg retrieved chunks) can improve performance as well as efficiency.

📜 http://tinyurl.com/2sdeuupn 💻 http://github.com/AlexTMallen/adaptive-retrieval

#PaperThread #newpaper
[1/N]

Victor Zhong Dec 20, 2022

Mark Abraham Dec 19, 2022

"Using #Apple's CarPlay system slowed drivers' reaction times nearly five times as much as driving with a blood-alcohol level of 0.08 — but CarPlay is legal on U.S. vehicles, even as U.S. regulators spend millions on anti-distracted driving campaigns to politely request drivers not use it."

https://usa.streetsblog.org/2022/12/19/opinion-what-if-were-thinking-about-impaired-driving-all-wrong/

ADDED: Link to original study https://iamwebsite.blob.core.windows.net/media/docs/default-source/default-document-library/iam-roadsmart-trl-simulator-study_infotainment.pdf

#roadsafety #CDoH #publichealth #drunkdriving

Opinion: What If We’re Thinking About Impaired Driving All Wrong?

Let’s pull back the cover on some seemingly shocking stats from the National Highway Traffic Safety Administration.

Streetsblog USA

Victor Zhong Dec 20, 2022

Yoav Artzi Dec 20, 2022

Users engaged with natural language systems can provide feedback in realtime, and this feedback is a super duper learning signal! So: deploy, train, repeat!

https://arxiv.org/abs/2212.09710

Last PhD paper w/@alsuhr/[email protected] ... 🧵

Continual Learning for Instruction Following from Realtime Feedback

We propose and deploy an approach to continually train an instruction-following agent from feedback provided by users during collaborative interactions. During interaction, human users instruct an agent using natural language, and provide realtime binary feedback as they observe the agent following their instructions. We design a contextual bandit learning approach, converting user feedback to immediate reward. We evaluate through thousands of human-agent interactions, demonstrating 15.4% absolute improvement in instruction execution accuracy over time. We also show our approach is robust to several design variations, and that the feedback signal is roughly equivalent to the learning signal of supervised demonstration data.

arXiv.org

Victor Zhong Dec 18, 2022

Andrey Kurenkov Dec 18, 2022

Kind of frustrated with the state of Mastodon as an open source project... it seems pretty ad hoc and none transparent process wise. Please comment on / upvote this discussion if you care about this:

https://github.com/mastodon/mastodon/discussions/22428

What's the process for addressing Github issues? · Discussion #22428 · mastodon/mastodon

There are currently 3.2k open issues for this repo and 185 open pull requests. Obviously it's not feasible to address them all, so I am curious how it is decided which issues to address/priorit...

GitHub

Show thread

Victor Zhong Dec 17, 2022

In fact I think perhaps this is a viable business model? Company A serves as a platform for social media and user content. It sells access to that content to other companies B that curate content personalized to end users. A is paid by B while B makes money off of ads or subscription fees.

Victor Zhong Dec 17, 2022

I do think there needs to be some sort of recommendation algorithm to make Mastodon feeds scalable. For example this Jack Dorsey thread describing a platform for content (eg Mastodon) upon which users can choose to run their own ranking algorithms: https://www.getrevue.co/profile/jackjack/issues/a-native-internet-protocol-for-social-media-1503112. I’m already starting to find my feed a bit unwieldy, and I don’t even have that many followed accounts..

a native internet protocol for social media

There’s a lot of conversation around the #TwitterFiles. Here’s my take, and thoughts on how to fix the issues identified. I’ll start with the principles I’ve come to believe…based on everything I’ve learned and experienced through my past actions as a Twitter co-founder and lead:Social media must be resilient to corporate and government control.Only the original author may remove content they produce.Moderation is best implemented by algorithmic choice.The Twitter when I led it and the Twitter …