LLMS

–

07 December 2022 17h15

Anecdotally, its math reasoning (excl. arithmetic calculation) improves substantially when you request LaTeX notation.

→ View original post on X — @_karenhao

Large Language Models in Drug Industry: Measurable Benefits and Regulation

By

@_karenhao

–

07 December 2022 17h15

This is what I like about this particular application of large language models. The drug industry is highly regulated, and there are clear ways to measure whether the output of a protein-language model is doing more harm than good.

→ View original post on X — @_karenhao

Protein-Language Models Accelerate Drug Discovery for Difficult Diseases

By

@_karenhao

–

07 December 2022 17h02

Protein-language models can be used to find them faster: either by enhancing existing molecules with targeted edits or by generating entirely new ones that could provide new hope for diseases like pancreatic cancer and ALS, for which more effective drugs have been elusive.

→ View original post on X — @aibreakfast

GPT Explains Nonexistent Idioms Accurately

By

@aibreakfast

–

07 December 2022 14h39

GPT can accurately explain idioms that don't exist

AI behavior with l33tsp34k, hyphens, Zalgo text, and Base64

By

–

07 December 2022 6h16

Less protocol-complaint but likely also dumber. You see a similar phenomenon if you ask it to speak in l33t5p34k, or with hyphens between all letters, or especially with Zalgo text. Base64 is interesting though because it probably sees a *lot* of translation pairs in pre-train.

text-davinci-002 lacks RLHF; ChatGPT belongs to 002 line

By

–

07 December 2022 5h51

We now know that text-davinci-002 doesn’t use RLHF either, and based on the model names used in API calls ChatGPT appears to be part of the 002 line, rather than 003 which uses RLHF.

Prompt injection shatters ChatGPT’s coherent personality illusion

By

–

07 December 2022 5h50

You can prompt-inject ChatGPT into behaving however you like, but this shatters the illusion of a coherent personality. ChatGPT forces you to type your own disclaimer to prove you understand it’s not real. Otherwise you talk to the protocol droid.

→ View original post on X — @jiqizhixin

DeepMind UCL Fine-tune 70B LM for Human-aligned Statements

By

@jiqizhixin

–

07 December 2022 5h48

DeepMind & UCL Fine-tune a 70B Parameter LM to Generate Statements Agreeable to Humans with Diverse Opinions

ChatGPT correctly answers Super Bowl question about Justin Bieber’s birth year

By

–

07 December 2022 3h17

Here's a related tweet, from the pre-ChatGPT era. Nowadays, ChatGPT can correctly answer the question, "What NFL team won the Super Bowl in the year Justin Bieber was born?" without any CoT prompting, unlike prior GPT-3 versions.