🎬 The Deepfake Powerhouse, WiFi Human Tracking, and Remote Supercomputing

Sam Altman announced OpenAI’s GPT-4 will launch only when they can do it safely and responsibly. “In general we are going to release technology much more slowly than people would like. We're going to sit on it for much longer…” Altman also confirmed a video model is in the works.

In today's email:

• Deepfakes: How the South Park boys will dominate in Hollywood AI

• Dark Knight Tech: Interview with AI researcher behind movement tracking using only Wifi antennas

• Pay Per Compute: GPU power can be remote and scalable

• Underrated: New and notable AI projects

Last month, Creators of South Park, Matt Stone and Trey Parker, announced that they have raised $20 million to invest in their new AI entertainment agency, DeepVoodoo.

DeepVoodoo's specialty will be high-quality deepfakes, and they claim to have hired the best artists in the world to do it, which is believable considering the duo signed a $900M deal for seven more seasons of South Park.

Bankrolling the best artists in the highly complex niche of deepfakes, combined with Stone and Parker's unprecedented batting average for success in TV and Broadway, likely puts DeepVoodoo in the position to dominate the industry.

Parker and Stone were able to show off DeepVoodoo's capabilities in Kendrick Lamar's video The Heart Part 5, which might be the most mind-blowing example of "synthetic media" to date.

Definitely worth a watch - checkout the video here. Song is a banger, too.

WiFi antennas can track human movement?

The technology straight out of Morgan Freeman's back pocket.

It has now been proven by researchers that common WiFi antennas can be used as the sole source of active sensing to track fine human movements in a room.

Since nearly 1M people saw the post about using WiFi to track DensePose movement, we reached out to one of the authors of the project Fernando De La Torre, who researches Computer Vision and Machine Learning at Carnegie Mellon University to answer a few more questions.

Fernando was kind enough to interrupt his ski trip to expand upon the research. Legend.

AiB: In short, does this mean a simple array of common WiFi antennas can accurately track human movement in a closed environment?

Fernando: That is correct, using three transmitter and three receiver antennas, we were able to recover dense correspondence in multiple humans.

AiB: Has this been done using only WiFi before?

Fernando: In computer vision, the subject of dense pose estimation from pictures and video has been thoroughly studied. However, body pose estimation from WiFi or radar is a relatively unexplored problem. As indicated in the previous work section of the paper, there have been earlier works on lower level human detection tasks (such as segmentation) from WiFi. However, to the best of our knowledge, this is the first study that enables recovering fine-correspondence for the human body from WiFi.

AiB: Your paper mentioned that this technology preserves privacy, but many have speculated about the capability of gait tracking for identifying individuals. Could gait tracking be accurately done with this technology?

Fernando: Compared to video or lidar, WiFi offers a better privacy-preserving signal.

This is critical in applications such as monitoring the well-being of elderly people at home (e.g., detecting falls, computing the amount of social interaction, or detecting potential health concerns), where many users will feel uncomfortable with the use of cameras at home. WiFi, on the other hand, adds an additional layer of anonymity because it cannot be immediately interpreted by humans and it can only recover shape information rather than texture.

It is unclear that the recovered mesh is temporally stable enough to compute a good discriminative gait measurement. However, identification through gait analysis could potentially identify a relatively small number of individuals, but it will not likely scale to the identification of a large number of users accurately.

In addition, some prior enrollment will need to be completed, including the user's consent.

AiB: If somebody wanted to try this at home, what would they need to have? Is your code open-sourced?

Fernando: The current version is trained for our scenario and specific hardware. It is important to keep in mind that a number of variables, like the position and orientation of WiFi devices, the presence of objects, and the movement of people and things in the environment, can alter the WiFi signals. Therefore, more research is required to make the approach robust to these factors, before releasing the code/model.

AiB: Short of wearing a tinfoil suit, is there anything a person can do to mask themselves from being tracked?

Fernando: There are a number of ways to distort the WiFi signal, including metal, a wireless jammer, using a mobile device with WiFi, or even wearing clothing from a conductive fabric that effectively forms a Faraday cage.

For a deeper dive into Fernando's work conducted with his colleagues Jiaqi Geng and Dong Huang, check out their research paper here.

Tools for Builders: Become a GPU Landlord (or renter)

🍊 Juice Labs has developed an application that allows users to connect to a GPU when they need extra processing power - and their method of doing so is far more efficient than Amazon AWS.

What makes Juice different? With Juice, multiple people can use the same GPU at the same time and they can use as much or as little of the GPU's power as they need, instead of renting the entire processor.

This means that you can use the GPU when you need it, and not pay for it when you don't. Think of it as renting out your autonomous car instead of leaving it parked in the garage.

You can try it yourself, and soon turn your unused GPU into a remote service.

New AI projects to keep an eye on

• 🔎 Explore AI is a free semantic search engine for YouTube and podcasts, allowing you to search the contents of a videos and podcasts (instead of just titles) and get the exact timestamp that addresses your query. Created by @tanaydesaii

• ✍️ Jaq n Jil is an AI writing assistant that claims to beat all plagiarism detectors, which is the best value-add an AI writing assistant could offer. Company is still in beta, but expect it to pop, especially at schools.

• 🌐 Futuretools.io is a web database for finding and reviewing AI tools. They also have a great newsletter. Started by @mreflow

That's a wrap! But one more thing:

If you enjoyed this issue, share it! If you do, DM us on Twitter and I'll send you a copy of the Ultimate AI Reading List: 50+ Free Resources for Going Down The AI Rabbithole.

Thanks for reading!