Superpower Daily
Posts
Google's AI Chief Predicts 50% Chance of AGI Breakthrough in 5 Years! 🚨

Google's AI Chief Predicts 50% Chance of AGI Breakthrough in 5 Years! 🚨

👀 AI Cameras Took Over One Small American Town

Saeed Ezzati
November 02, 2023

In today’s email:

🤦🏻‍♂️ How Microsoft is making a mess of the news after replacing staff with AI
🌠 DALL-E 3 Double Exposure Images Using ChatGPT's Custom Instructions.
📚 Get Started Building with Generative AI reading these 12 lessons from Microsoft
🧰 15 new AI-powered tools and resources. Make sure to check the online version for the full list of tools.

Google AI Chief Says There’s a 50% Chance We’ll Hit AGI in Just 5 Years

Shane Legg, co-founder of Google's DeepMind, stands by his prediction from 2011 that there is a 50-50 chance artificial general intelligence (AGI) will be achieved by 2028. This forecast is based on his belief in the exponential growth of computational power and data, inspired by Ray Kurzweil's "The Age of Spiritual Machines." Legg contends that AGI would require an AI system to perform a wide array of tasks without significant gaps in capability, rather than just one specific task. He also highlights the need for scalable algorithms that can handle vast amounts of data beyond human experience. While acknowledging the computational power is likely sufficient for AGI, Legg maintains his prediction is still uncertain, with a balanced chance of AGI being realized within the indicated timeframe. DeepMind has not yet provided additional comments on this forecast.

The Morning Paper for Tech

Want a byte-sized version of Hacker News that takes just a few minutes to read?

Try TLDR's free daily newsletter.

TLDR covers the most interesting tech, startup, and programming stories in under 5 minutes.

No sports. No politics. No weather.

It's read by 1,250,000 software engineers, founders, and tech workers.

Subscribe for free.

12 Lessons, Get Started Building with Generative AI - By Microsoft

Learn the fundamentals of building Generative AI applications with our 12-lesson comprehensive course by Microsoft Cloud Advocates. Each lesson covers a key aspect of Generative AI principles and application development. Throughout this course, we will be building our own Generative AI startup so you can get an understanding of what it takes to launch your ideas.

Instead of just telling robots “what to do”, can we also guide robots by telling them “how to do” tasks?

Robot learning methods usually provide policies with task conditioning using text instructions, goal images, or object-centric representations… but these are all quite removed from low-level control actions that robots need to use! What about robot end-effector trajectories?

These trajectories are close to the low-level robot actions that policies must predict. Better yet, we can recover hindsight trajectories for free during training -- similar to hindsight goal relabeling, but instead applied to the entire executed robot trajectory.

We introduce RT-Trajectory, an extension of an RT-1 backbone that conditions on trajectories instead of language. We train on hindsight trajectories, but at inference, we utilize trajectories drawn by humans, extracted from videos, or even generated by foundation models.

One of the most exciting contributions of RT-Trajectory is showing the ability to apply prompt engineering to robotics. We find that for a fixed policy and fixed initial conditions, we can get improved performance by "just asking the robot better"! 🤯