Tweets

2023-03-26T19:17:42+00:00 | 🔗

@kmjablonka MIT License (it’s in the README sheet)

2023-03-26T17:41:28+00:00 | 🔗

weekend! Can you distinguish AI-generated compounds (sampled from MOSES) and Patent compounds (from SureChEMBL) https://t.co/bXLmrGXTtb (everyone's -- anonymized -- scores are collected and displayed here: https://t.co/IRgHWy7fMj) https://t.co/Q1WNjc8hLn

2023-03-22T06:45:56+00:00 | 🔗

@TrungTPhan How can you spell covfefe wrong? Everyone knows it

2023-03-19T23:34:42+00:00 | 🔗

@andrewwhite01 Crypto should rebrand to "Global Payment Verification"

2023-03-19T18:41:01+00:00 | 🔗

@natfriedman I thought it was to preserve watermark.

2023-03-19T17:09:34+00:00 | 🔗

@xuanalogue @OpenAI @AnthropicAI @alexanderklew I think that many people would simply sample the most probable token and not the “suggested watermarked” output

2023-03-19T16:18:29+00:00 | 🔗

@xuanalogue @OpenAI @AnthropicAI @alexanderklew Given their description of their AI-generation stenography, it seems like giving the log probs would run counter to that. They need to be able to put slightly less probable words/sequences to be able to detect AI-generated text. Sadly, we can’t have both log props and detection.

2023-03-16T21:39:23+00:00 | 🔗

@olexandr Emphasizing the ability to "Flexibly call APIs" is very different from the actual workflow being endorsed :)

2023-03-16T21:35:50+00:00 | 🔗

@olexandr As always over-hype always clouds these discussions. It's possible to be overly critical though misuse. A bit like cloning a github repo for Deep RL for de novo drug discovery, setting the learning rate to 1e3, giving it a dataset of pesticides, and then reporting bad results.

2023-03-16T21:31:55+00:00 | 🔗

@olexandr I'm pretty sure the example you are referencing was GPT-4 + LangChain + Tooling, which seemed evident from the System Card. No, it isn't going to give you a drug. Yes, it can flexibly call APIs, which when fleshed out could be helpful in an actual discovery setting. https://t.co/SibwHT3Nkk