Monthly meetup / Schelling meetup

This month's meeting is also the Spring 2023 ACX Schelling Meetup (aka "Meetups Everywhere"). New folks are even more welcome than usual!

-----

This is our recurring monthly meetup. Our format is:

1. Unstructured conversation on rationalist and rationalist-adjacent topics is the default
2. Optional short readings are available for discussion too
3. Engaging in good faith and openness to disagreement and nuance are table stakes

This meeting's topics: let's continue our discussion of recent advances in AI. At the rate things are moving, we could probably just stay on this topic indefinitely....

Here are a few great readings and topics from the past 1-2mos we can discuss:

• OpenAI's intro post on GPT-4 — GPT-4 is probably the most advanced large language model (LLM) that's available to the public. I'm happy to bring my laptop if anyone wants to play with it live at the meeting.
• Illustrating Reinforcement Learning from Human Feedback (RLHF) — RLHF is a technique for modifying a raw language model so that its outputs are more acceptable to human reviewers. It's what was used to turn GPT-3 into ChatGPT.
• AI Alignment: Why It’s Hard, and Where to Start — A MIRI post by Eliezer Yudkowsky that's a good introduction to the AI alignment problem.
• Deceptively Aligned Mesa-Optimizers: It's Not Funny If I Have To Explain It — In which Scott Alexander teaches us more about AI alignment, particular meta- vs. mesa-optimizers, in order to painfully explicate a meme.
• The Waluigi Effect — A deep dive into a particularly interesting and recent proposed failure mode for AI alignment that's relevant to recent LLMs.
• Microsoft Research Paper Claims Sparks of Artificial Intelligence in GPT-4 — I haven't read this one yet, but it looks interesting and roughly about what it says on the box.

Or just bring your own questions, readings, or topics. Or we can just play with GPT-4 until we get throttled by the API :)