Superpower Daily
Posts
In the future, humans will need to supervise AI systems much smarter than them

In the future, humans will need to supervise AI systems much smarter than them

Google DeepMind used a large language model to solve an unsolvable math problem

Saeed Ezzati
December 15, 2023

In today’s email:

👀 Sam Altman on OpenAI, Future Risks and Rewards, and Artificial General Intelligence
🧸 Grimes is working on an interactive AI toy for kids. Meet Grok.
🧠 A Portable, non-invasive, mind-reading AI turns thoughts into text
🧰 14 new AI-powered tools and resources. Make sure to check the online version for the full list of tools.

In the future, humans will need to supervise AI systems much smarter than them.

The new research by OpenAI in the field of superhuman AI alignment discusses "superalignment." The primary challenge in this area is supervising AI systems that are significantly smarter than humans. The team behind this research proposes an analogy to study this problem: using smaller, less capable models to supervise larger, more capable ones.

The research shows that a GPT-2-level model can be used to supervise GPT-4, eliciting performance close to GPT-3.5, even on complex problems where the small model fails. This demonstrates that strong models can generalize according to the weak supervisor's intent, leveraging their capabilities beyond the supervisor's limitations.

The article addresses the limitations of current alignment methods like reinforcement learning from human feedback (RLHF), which may not scale well to superhuman models. The proposed method, however, shows promise in improving weak-to-strong generalization, suggesting a potential pathway for aligning future superhuman AI models.

To foster further research in this area, the team is releasing open-source code for weak-to-strong generalization experiments and launching a $10 million grants program for research in superhuman AI alignment. The goal is to ensure that advanced AI systems remain safe and beneficial to humanity, and this new research direction offers a promising avenue for empirical progress in AI alignment.

News. Without Motives. That’s 1440. Join Today!

Frustrated by biased news? Tired of trying to sift through 100s of news sources to find the facts? Leave behind the opinions and receive a daily digest of pure facts. Say hello to 1440! Delve into politics, business, science, sports, culture, and more. Join the other 3 million Americans who read 1440 each morning.

Outfit Anyone: Ultra-high quality virtual try-on for Any Clothing and Any Person

Virtual try-on has become a transformative technology, empowering users to experiment with fashion without ever having to physically try on clothing. However, existing methods often struggle with generating high-fidelity and detail-consistent results. Diffusion models have demonstrated their ability to generate high-quality and photorealistic images, but when it comes to conditional generation scenarios like virtual try-ons, they still face challenges in achieving control and consistency. Outfit Anyone addresses these limitations by leveraging a two-stream conditional diffusion model, enabling it to adeptly handle garment deformation for more lifelike results. It distinguishes itself with scalability—modulating factors such as pose and body shape—and broad applicability, extending from anime to in-the-wild images. Outfit Anyone's performance in diverse scenarios underscores its utility and readiness for real-world deployment.

Instagram introduces GenAI-powered background editing tool

Instagram introduced a new AI-powered background editing tool for U.S.-based users, allowing them to change image backgrounds in Stories using prompts. Meta's generative AI lead, Ahmad Al-Dahle, announced this feature, which includes ready prompts like “On a red carpet” or “Surrounded by puppies,” and also allows users to write their own. The tool adds a “Try it” sticker to Stories with these backgrounds, encouraging others to use the feature.

Snapchat also released a tool for paying users to create and send AI-generated images. They previously launched a feature called Dreams for creating themed images, and allowed Snapchat+ users to use prompt-based backgrounds.

Additionally, Meta expanded the availability of its 28 AI-powered characters to all its apps in the U.S., including WhatsApp, Messenger, and Instagram. These characters support Bing search and have an improved context window. Meta also launched a standalone AI-image generator, Imagine with Meta, powered by their Emu model.