Using Codex, we experimented with longer prompts (5 shots) and with sampling and combining many completions.
Using an ensemble of 100 completions per question, we achieved (passing score) human-level performances on the three datasets: USMLE-MedQA, MedMCQA and PubMedQA.

The ensemble of Codex-generated CoTs turned out to be surprisingly well-calibrated.

We also experimented with more exotic CoT prompts (e.g., “Let’s follow a Bayesian approach”) and showed that GPT uses biased heuristics (preferring the last answer option when in doubt).

Crypto companies now using ChatGPT to do their ENTIRE audit.

🚩 🚩 🚩 🚩 🚩

Just wait until they start really controlling the language used by LLM AIs

PubMedGPT LLM passes US Medical Licensing Exam (MedQA-USMLE) with more than 50% correct answers

https://crfm.stanford.edu/2022/12/15/pubmedgpt.html

Language models generalize beyond natural proteins

Reaching post-evolutionary biology from evolutionary learning – language models trained on millions of natural sequences can be used generatively to make completely de novo proteins that are viable in the wetlab 🦕🤖🗣️🧬

https://www.biorxiv.org/content/10.1101/2022.12.21.521521v1

“Today, youChat goes live.

Open, broadly capable, conversational AI for search with knowledge of recent events and citations of sources.

Search and chat of the future”

https://you.com/search?q=what+was+the+recent+breakthrough+in+fusion+research%3F

^^ AI startups all working right through the Christmas holidays.

One last boom.

Finally a benefit to LLM lying.

It won’t spill your startup’s secrets — because it’s too prolific of a liar!

Update Your Course Syllabus for chatGPT

https://medium.com/@rwatkins_7167/updating-your-course-syllabus-for-chatgpt-965f4b57b003

“Less than two weeks ago, ChatGPT wrote me a whole essay about why immigration is bad. Now it won't even address the topic.

ChatGPT has been ruined.”

De-woking ChatGPT

https://cactus.substack.com/p/why-its-easy-to-brainwash-chatgpt

Smart contract auditing powered by GPT3!

(Not a good idea)

Cheating With ChatGPT: Can OpenAI’s Chatbot Pass AP Lit? | WSJ

Data for 400 million Twitter users are for sale. Contains emails and phone numbers allegedly obtained via an API vulnerability.

The sample posted shows high profile accounts including
@VitalikButerin, @mcuban, @briankrebs

— Imagine an evil GPT, that has all of these hacked databases in its training too.