• AI Time to Impact
  • Posts
  • . . AI: OpenAI Superalignment breakthrough, DeepMind math breakthrough, and more (12.14.23)

. . AI: OpenAI Superalignment breakthrough, DeepMind math breakthrough, and more (12.14.23)

Friends, just 4 stories today but I'm trying something new - I've added bullet points below each story to highlight details that weren't in the summary but made me say "WOW" when I read them. I think these details make the story summaries much more interesting and are thus worth the added length. Let me know what you think.

As always, these are the stories that rose to the top of our review of thousands of AI community conversations across the web today. I hope you find them useful; thanks for sharing with friends.

Now here’s today’s news,

First impacted: AI researchers, AI developers
Time to impact: Medium to Long

OpenAI's Superalignment team has released a paper detailing an experiment in which a smaller AI model was able to guide a larger, more advanced model, a method they believe could be used for aligning future superhuman AI systems. (That’s systems more powerful than humans.) In addition, OpenAI has launched a $10 million grants program for research on superhuman AI alignment and has made the open source code available for further study in this area.

* A related paper titled "Practices for Governing Agentic AI Systems" outlines seven strategies for ensuring the safety and accountability of increasingly autonomous AI systems as they become more prevalent and powerful. (Strategies like Evaluating Suitability for the Task, Constraining the Action-Space and Requiring Approval, etc)

* It also points out (in Section 5 of the paper linked to above) potential indirect effects from the widespread use of autonomous AI systems, which may require further governance structures.

First impacted: Mathematicians, Computer Scientists
Time to impact: Medium to Long

DeepMind has developed a system called FunSearch that uses LLMs to generate novel solutions to unsolved mathematical problems, according to a paper published in Nature. The system, which pairs a pre-trained LLM with an automated evaluator, has reportedly discovered new solutions for the cap set problem and more efficient algorithms for the bin-packing problem. 

* The FunSearch team claims to have identified the largest cap sets (sets of points in a high-dimensional grid, where no three points lie on a line) ever discovered, representing the most significant growth in cap set size in two decades.

* The method involves creating programs that explain the process of finding solutions, with a preference for solutions that can be expressed through concise programs, which they say aids researchers in understanding the program outputs.

First impacted: AI researchers, Data scientists
Time to impact: Medium to Long

Ego-Exo4D, a large-scale video dataset, is now publicly available from a consortium of 15 Universities and Meta. The dataset, collected by 839 individuals across 13 cities, includes 1422 hours of video, seven-channel audio, IMU, eye gaze, head poses, and 3D point clouds of the environment, focusing on skilled human activities.

* The "ethics" section of the site says that the Ego-Exo4D dataset was gathered in private settings and the individuals recorded to train robots to take peoples' jobs cannot be de-anonymized, as is required under ethics protocols! Lol ugh.

* The Ego-Exo4D dataset is distinct in its emphasis on skilled tasks, recording participants with varying skill levels performing the same task, while preserving the unique characteristics of the scenario in a new setting.  That's cool.

First impacted: Datacenter managers, AI developers
Time to impact: Medium to long

Intel's CEO, Pat Gelsinger, criticized Nvidia's CUDA GPU technology today, stating that the industry is eager to move away from the CUDA market and that inference technology will play a more significant role than training in AI. Gelsinger launched Intel Core Ultra and 5th Gen Xeon datacenter chips, suggesting that Nvidia's dominance in training may not last, and promoted OpenVINO, Intel's AI standard, predicting a future of diverse computing.

* Gelsinger presented the Gaudi 3 on stage, indicating that it could operate AI models without relying on NVIDIA's CUDA, which could change the AI industry.

* Sandra Rivera, Intel's executive vice president and general manager of the Data Center and AI Group, highlighted Intel's range from data center to PC, suggesting that Intel's production capacity makes it competitive throughout the AI value chain "We're going to be a foundry player."

That's it! More AI news tomorrow!