Welcome to the weekend edition of InsideAI. I'm Rob May, CEO of Talla, the AI powered CX Automation platform, and host of the AI at Work podcast. Be sure to check out our latest episode with Matthew Mattina, head of the ML research lab for ARM.
I'm also an active early stage AI investor with almost 70 investments, so if you are working on something cool please reach out.
A new machine learning algorithm can predict how well people understand concepts. Researchers at Dartmouth College connected engineering students to an fMRI scanner and had them takes STEM-related tests and answer picture-related questions. The algorithm produced "neural scores" that could predict a student's performance based on how certain parts of the brain lit up. According to Engadget, the process could help teachers improve their courses by showing what teaching techniques best resonate with students. - ENGADGET
An app that used GANs to "undress" women in photos has been shut down. The app, called DeepNude, generated a lot of controversy over its use of AI to create "deepfake" images of nude women without their consent. Motherboard first reported on the app on Thursday. By Friday morning, its creators stopped selling it, saying that they “greatly underestimated” interest in the app and that “the probability that people will misuse it is too high.” - MIT TECHNOLOGY REVIEW
Apple published a new journal article about machine learning called “Bridging the Domain Gap for Neural Models.” Apple developed techniques of “unsupervised domain adaptation” that could improve the performance of deep learning models in specific scenarios. - PATENTLY APPLE
Facebook, Google, and 38 other tech companies released new benchmarks for evaluating AI tools. The benchmarks, which are called MLPerf Inference v0.5, focus on image classification, object detection, and machine translation. There are separate AI benchmarks across different platforms, including chips, cloud computing platforms, and smartphones. The idea is to help companies compare their AI tools to see which work best for them, according to Peter Mattson, general chairman of the consortium of companies known as MLPerf (machine learning performance). - WSJ
Some of AI's top thinkers authored a paper about how machine learning can be used to fight climate change. The paper suggests 13 ways that the technology could help prevent human destruction, including building better electricity systems, monitoring agricultural emissions and deforestation, and creating new low-carbon materials. - THE VERGE
Robotics has been a difficult place to play as an investor this past decade. For every Kiva, which was a huge win, there were dozens of failures that seemed like good ideas. In other words, robotics has been just as risky as startups in general. Except, robotics companies were slower to get off the ground and required more capital so that made them even worse. I know a few investors who won't touch a robotics deal any more. But I think the changes in the marketplace are starting to make it a good time to get back in the robotics game.
One big change is standardized parts. For example, if you are building something that needs a robotic arm, there is a 90% chance you can use something already in market, and not build your own. I recently made an investment in a company that is leasing a robotic arm, adding a few novelties to the end of it, training the arm + novelties to do certain tasks, and then renting that full system to the customer (who pays up front each month). Basically, this "robotics" company is a software and ML training company, not a hardware company, and the business model is relatively capital efficient because of how they charge the customer up front, then pay the monthly lease on the robotic arm. Robot companies weren't this way in the recent past.
We are a long way from a full housekeeping assistant robot, and there are tons of innovations required to get us there, but the market is in a nice phase where we have moved beyond robot point solutions to robots that are adaptable to non-precision workflows, within certain constraints.
The way I think the robotics market will be built out is that no one will build a huge all-in-one robot-as-a-platform just yet, because the requirements for what that should be are still murky. Instead, every time so many use cases innovate the (roughly) same new piece of a robotics system, someone will realize there is the opportunity to standardize and platformize it. Each time this happens, the market opportunities will jump as a new group of robotics entrepreneurs no longer needs to hire someone for, or retain deep expertise in that piece of the system.
What's next? Probably standardized grippers of certain types and more types of standardized robotic locomotion. It's a good place to be as an investor, and I expect the use cases where robots make sense to explode in coming years.
Thanks for reading, and see you next week.