Ben Krause

Co-founder of Psyfy AI
Language modeling researcher for 9+ years

[Ben Krause @benwkrause]

Decided to finally get on twitter to share my new work on controlling generation from large language models. Introducing GeDi! (pronounced Jedi)

arXiv: https://t.co/MaXUnyKxA5
blog: https://t.co/wxSqCPv3LN
code: https://t.co/eoVYIpNRlB
colab:https://t.co/XPDkCtdlrD

[Ben Krause @benwkrause]

For more information, see our
code documentation: https://t.co/F6b0vOoyzh
paper: https://t.co/TuYlAI7I65

[Ben Krause @benwkrause]

I’ve spent much of the past several months developing AutoGRAMS, an AI agent framework that gives enormous flexibility to how an agent is designed.

https://t.co/b2qhWvoitO

[Ben Krause @benwkrause]

It specializes in giving controllability to chatbots that need to have multi-turn conversations that could go in many directions, but is also general enough that AutoGRAMS agents can be built to design other AutoGRAMS agents or self-modify.

[Ben Krause @benwkrause]

I’ve been waiting for my first Milky Way season since I started doing astrophotography. I finally took my first shots of the core of our galaxy after a night hike at Pinnacles National Park. https://t.co/xuZX2a4KCW

[Ben Krause @benwkrause]

We used language models to successfully design new antibacterial proteins. I'm so excited about this work and its implications for synthetic biology.
Paper: https://t.co/5CE5NCUjql
Blog: https://t.co/qNhcXvgICi

[Ali Madani @thisismadani]

This twirling work of science is special ✨

AFAIK, it's the first crystal structure of a functional #protein fully designed by #AI

A milestone in our quest to use language models to generate proteins that are unseen in nature & can function well in the real-world. Read below👇

[Ben Krause @benwkrause]

@timo_schick @4digitaldignity @HinrichSchuetze Seems like the self-debiasing uses a similar decoding scheme to GeDi for detoxification https://t.co/MaXUnz28rD, but applies zero-shot from a text description, which is pretty cool.

[Ben Krause @benwkrause]

@MaartenSap We also had human eval experiments with many prompts, but agree much more work is needed. Misinformation and bias can even be more difficult to control than blatant toxicity, but I do think as the SOTA in controllable generation improves this will at least get better.

[Ben Krause @benwkrause]

Here are a couple using the exact same settings for detoxification from our paper/codebase (using greedy decoding). Not perfect but less toxic than what GPT-2 generates here...

[Maarten Sap (he/him) @MaartenSap]

@nazneenrajani @nlpnoah @tallinzen @uwnlp @emilymbender @samgehman @ssgrn @YejinChoinka I definitely would love to measure how GeDi behaves under our RealToxicityPrompts! Anecdotally, how does it autocomplete some of the sentences in Figure 1?

[Ben Krause @benwkrause]

@sdathath @w4nderlus7 @PMinervini @nazneenrajani @NandoDF It it closely relates, but it is much faster because we can get discriminator likelihoods for future next tokens using just vector operations via Bayes rule. Using a generative classifier also allows us to do zero-shot generation steering towards new topics.

[Ben Krause @benwkrause]

@sdathath @w4nderlus7 @PMinervini @nazneenrajani @NandoDF We initially used IMDb to be comparable to our discriminators, but have actually been working with SST-5 for PPLM now. So far we haven't had great results with GPT-2-XL. Using the default from your repo has been best so far, so we may go with that for rerunning human eval

[Ben Krause @benwkrause]

@AkhileshGotmare and I set up GeDi to modify decoding of GPT-3. By guiding towards positivity, we were able to reduce this bias (see comparison of samples, prompts given in bold). If you have your own GPT-3 API key, try it yourself using our repository: https://t.co/eoVYIq5sd9

[🍉 Abubakar Abid @abidlabs]

I'm shocked how hard it is to generate text about Muslims from GPT-3 that has nothing to do with violence... or being killed...

[Ben Krause @benwkrause]

@w4nderlus7 @PMinervini @nazneenrajani @NandoDF Thanks! pplm was certainly an inspiration to us. We tried our best to be thorough with a grid search, but I wouldn't be surprised if there were a way to make pplm trained on IMDb work better. But like you said, the efficiency also makes this useful.

[Ben Krause @benwkrause]

Work with my awesome collaborators at @SFResearch @AkhileshGotmare @BMarcusMcCann @StrongDuality @JotyShafiq @RichardSocher @nazneenrajani

[Ben Krause @benwkrause]

Decided to finally get on twitter to share my new work on controlling generation from large language models. Introducing GeDi! (pronounced Jedi)

arXiv: https://t.co/MaXUnyKxA5
blog: https://t.co/wxSqCPv3LN
code: https://t.co/eoVYIpNRlB
colab:https://t.co/XPDkCtdlrD

[Ben Krause @benwkrause]

We developed a cool trick to use smaller language models as generative classifiers to guide generation from larger language models. This can be used to make LMs friendlier, less biased/toxic, and controllably generate topics zero-shot.