Gideon Mann

@jordiae @sebgehr @mdredze @drosen Fantastic to have you this summer! Thank you!

Last day of my internship at Bloomberg NYC. I have to say it has been an incredible experience. Many thanks to my awesome mentors and advisors @sebgehr @mdredze @drosen @gideonmann!

@srush_nlp 💯 . are you in the "end of programming" camp??

@gideonmann somedays I think we're not freaking out enough about this stuff.

Llama2.rs (github.com/srush/llama2.rs) - A one-file Rust Llama2.c. Pretty safe, and fast.

@srush_nlp kinda unreal that this works.

I don't actually know Rust, but it seems good. I had gpt translate the code, and then I fixed the compiler errors until it ran fast. Would love a code-review if anyone has time.

@yuntiandeng @UWCheritonCS @VectorInst @allen_ai @YejinChoinka @srush_nlp @pmphlt Fantastic!

Excited to share that I'm joining @UWCheritonCS as an Assistant Professor and @VectorInst as a Faculty Affiliate in Fall '24. Before that, I'm doing a postdoc at @allen_ai with @YejinChoinka. Immensely grateful to my PhD advisors @srush_nlp and @pmphlt. This journey wouldn't have been possible without them. Looking for #NLProc students eager to explore this exciting field with me!

Just heard that Eugene Charniak @BrownCSDept passed away. May he live on in blessed memory.

@tszzl Postscarcity == manufactured scarcity

the assets that matter in postscarcity are where rents accrue: credentials, land, status, and cartels

@fishnets88 Love 'em ! How do you handle the text? Post-processing?

Work in progress.

@moyix @AnthropicAI Does ChatGPT get these right? Do you have a code questions benchmark???

The @AnthropicAI model Claude is not great at bit-level code. I tried using it to solve the Matasano Mersenne Twister challenge and it a) couldn't count bits correctly b) assumed it could invert a function by just doing the operations in reverse order

@goodside 😳 (1) do You need more of a prompt ? (2) how complex can you get the assertions? Nested?

Prompting GPT-3 to reliably generate text and JSON data in a precise format using Python assertions, f‑strings, and variables declared only in our imaginations.

@moyix Congrats!!

Really lovely news to start the new year with - I’ve been promoted to Associate Professor with tenure! I’m so grateful to my friends, colleagues, students and collaborators at NYU and beyond for helping me get here! <3

Just heard the news about @ScottMinerd . I met him briefly, randomly, and he was surprisingly humble, kind, and generous with him time. Really a loss. My condolences.

@srush_nlp @ihnorton Would you have it just fill in the related work? If it was good enough to do that... would it even be meaningful to require a related work section? ... Does that imply that the paper of the future is just the code & data?

@ihnorton It's interesting, but the details are complex. Like do you prompt with the main result and outline? Seems at that point you might as well write it.

I'm honestly kind of embarrassed trying to write essays knowing ChatGPT exists.

Congrats to the fifth cohort of @Bloomberg's #DataScience Ph.D. Fellowship! Learn more about each of our four 2022-2023 Fellows and what their research is focusing on: bloom.bg/3v8TBI0 #AI #ML #NLProc

Introducing task vectors! A new way to steer models by doing arithmetic with model weights. Subtract to make models forget, add to make them learn 📜: arxiv.org/abs/2212.04089 🖥️: github.com/mlfoundations/…

@spolu 💯a phenomenal approach.

"Memorizing transformer style" or similar approach become predominant as they are always up-to-date and infinite in size. They enable a new wave of products that can operate on entire companies data or entire codebases, or entire chat history.

A list of predictions for 2023 for the field of LLMs🧵

Don’t write code often? Been a while? Worry no more! Introducing the “Fuzzy Compiler” 🧮 Just make up your own code, no worrying about syntax. Translates to working code w comments. See example 1:

Our beautiful, maximalist urban future awaits.

@moyix @yoavgo who has feelings about this

One thing I'm curious about: does human feedback in the form of InstructGPT-style RLHF constitute the kind of agreement and engagement with other language-users that is deemed necessary for "belief", "understanding", etc?

@brianluidog The model wants what the model wants.

It hates my prompt so much that it 1) made up a horrible fund name 2) misattributed both quotes 3) made the Munger quote directly contradict the fund strategy 4) wrote the absolute banger "you don't have to worry about not knowing what you're doing - we take care of that for you"

Last night, I wrote a tool which allows you to scrape data out of GPT models using SQL queries. Just make up a relevant database, and use your imagination of what kind of query you'd like, and it will return the data within GPT's parameters. Results are structured. Example: