Superalignment 🤖

OpenAI’s alignment approach to keeping AI safe

Author

Anita Kirkovska
July 07, 2023

Hi folks!👋🏻 This is The Prompt! We're your go-to source for all things AI.

I really appreciate all the feedback you gave about The Prompt ❤️. I'll make sure to focus on creating content that's educational and actionable, instead of just sharing the latest news.

And today, we break down OpenAI’s “alignment” game plan 👇🏻

FEATURED

OpenAI’s alignment approach to keeping AI safe

Right now, we're kinda stuck on how to control really advanced AI and stop it from misbehaving.

Our current techniques for aligning AI, such as reinforcement learning from human feedback, rely on humans’ ability to supervise AI.

But, if AI gets much smarter than us — we might not be able to keep up.

So, OpenAI has committed 20% of its computer to data to solve alignment of superintelligence within the next 4 years.

The Game plan

Simply put, they want to make a super smart human-like AI model to evaluate other AI models at scale, for tasks that are hard for humans to assess.

This is very tricky.

Why?

Because we don’t understand current AI models - So how can we build one that we’ll understand and trust to control all these other “misaligned” ones?

Nevertheless, their research priorities for this “superior” model are to achieve:

  • Scalable oversight: Ensure other AI models apply safety guidelines in situations we as humans can't directly supervise (like generalization on unseen data).

  • Automated interpretability: Being able to 'interpret' or understand what the AI is doing, and why it's doing it.

  • Adversarial testing: This is testing intentionally misaligned models and detecting the worst kinds of misalignments automatically.

WHAT ELSE IS GOING ON

🦙 OpenAI is rolling out Code Interpreter to all Plus users. You can opt-in to use the interpreter from your settings, and with it, you can ask ChatGPT to analyze data, create charts, edit files, perform math, etc. Plus, GPT4 is available to all paying users starting today.

👀 Playground AI raised $40M for their text-to-image platform. I’ve been following Suhail from his early AI days on Twitter. Back in June 2022, he went “all-in” into AI, and was documenting his learnings in this Twitter thread. Definitely a must-read thread with so many valuable nuggets.

RESOURCES

The best resources we came across lately that will help you become better at writing prompts & building AI apps.

📚 Personal lessons from LLMs [a must-read ]

👋🏻 Building AI products with OpenAI [ free course ]

🎥 What are transformers models [ great educational article]

TOOLBOX

The latest AI tools to use or get inspiration from.

PROMPT OF THE DAY

TOOL

Midjourney

PROMPT

Mark Zuckerberg threading with a threading machine --v 5.2

RESULT