"... a software tool cannot be an author of a submission, in part because it cannot consent to terms of use and the right to distribute content."
Remote
Yoav Artzi
@yoavartzi@sigmoid.social
mastodon 4.7.3#NLProc, ML researcher/teacher @ Cornell CS / Cornell Tech (NYC🚡) // Researcher at ASAPP
404 Followers
93 Following
17 Posts
Joined November 19, 2022
Website:
Cornell NLP:
There's so much happening here, not sure where to start. Putting aside the dubious factuality, any idea what is "Mechaka" and how it translates to "story"?
Replying to
This is absurd. Beyond credit, authorship is responsibility and liability. OpenAI assumes neither, and it is nonsensical to attribute either to ChatGPT or expect it to assume it (whatever that would even mean! 🤯). This practice is actively misleading the public about LLMs.
Learning Machines Seminar @cornell_tech@twitter.com: Yoshua Bengio (Université de Montreal) / Reusable Modular and Causal Knowledge Representation for Lifelong Learning / November 4, 2022 https://youtu.be/K8LNtTUsiMI
We are recruiting PhD research interns @asapp@twitter.com for the coming summer! Working on challenging ML/NLP problems, with amazing data, and SOTA models. Apply here:
https://jobs.lever.co/asapp-2/24e2de68-4a39-4a02-823d-3b010e9e9eb9
(also: hiring for full-time research positions, both scientists and engineers!)
When the optimistic hope is that it's "just a sophisticated phishing attack", and not someone is actually buying letters ....
Learning Machines Seminar @cornell_tech@twitter.com: Petar Durdevic (Aalborg University) / Waste Deep – Using Deep Learning methods to improve Wastewater Treatment / December 2, 2022 https://youtu.be/9VHrun3y6vY
Replying to
The agent's goal is to modify the environment so the grounded truth-value of the given statement is a target boolean. The language in lilGym is semantically-rich and human-written. Including: set reasoning, spatial relations, cardinality, and more.
Replying to
This level of reasoning is missing in RL benchmarks, that mostly use simplified synthetic 🤖 language? No real human language 🗣️ Why? Because computing rewards 🥇 requires resolving language meaning -> It's really a 🐔+🥚 situation!
Replying to
How do we compute reward? We annotate all sentences with Python 🐍 programs -> can test every possible state 🚀. This is a hard semantic parsing annotation problem, so we build an interactive platform with auto validation against hidden examples 🙈
Replying to
lilGym includes two types of environments, varying 🤸♂️🚴♀️⛹️♀️🤺🧘♀️ the difficulty of the task. Plus, flexible simplification control to tune complexity.
Replying to
Strong baselines -> a lot of room for improvement. Even with extra help, they are far from expert rewards. Because of the rich environment: more conventional reward without our annotation approach -> learning is hopeless -> flat (zero!) learning curve!
What is the ActuallyFollowing tab? Anyone else seeing it? All I see there is @tallinzen@twitter.com , which is fine, I guess. I mean, it's a world I _could_ live in. Maybe
Decision boundaries ....
Learning Machines Seminar @cornell_tech@twitter.com: Vivek Farias (MIT) / Markovian Interference and the Differences-in-Qs Estimator / October 28, 2022 https://youtu.be/JYpWtUkuW_Y
Learning Machines Seminar @cornell_tech@twitter.com: Wei Xu @cocoweixu@twitter.com (Georgia Tech) / Importance of Data and Controllability in Neural Language Generation / September 23, 2022 https://youtu.be/OTR94_8vSWQ