April 7, 2022– Advisor: Simon Erickson
I recently attended the MIT EmTech Digital Conference, where several industry leaders discussed the most important technologies that are impacting the field of artificial intelligence.
I could probably write a novel about all of the things that I learned. But rather than subject you to that, I’ve put together a shorter Cliff’s Notes below that captures several of the key takeaways in a simple bullet format. (And if anyone interested in that longer novel, here’s a link to my unedited conference notes. Warning in advance that they’re messy.)
This was the key theme of the conference. Companies are putting much more of a focus on labeling and curating their data upfront to better train their AI models. This is vastly different than the approach to AI a decade ago, which was unleashing the AI on tons of unstructured data, like millions of images from ImageNet databases and having it figure out which ones were cats.
Taking a closer look at the data itself and then involving subject-matter experts to help label it is an effective way to unleash AI on harder applications. It is similar in many ways to the approach Pandora used in the Music Genome Project. But the stakes are much higher now.
Labeling data is incredible time-intensive and expensive. There’s an opportunity to use automation as a way of enhancing the speed of labeling while still preserving its quality.
Subjective humans are also being replaced by automated filters, in order to standardize the data and its reproducibility. It’s also hard for AI models to work cross-industry, and sometimes even within the same industry, due to different machines collecting the data. One example is interpreting and labeling data from X-rays, which are different in type and in the year they were produced based upon different hospitals. There needs to be a way to create fit-for-purpose AI systems, who can tackle smaller projects like this.
C3.AI (NYSE: AI) — a company that my colleague Steve Symington is quite fond of — is one who’s up for this challenge. There are “a lot of $1 million to $5 million enterprise-grade AI projects floating around” that no one seems to be addressing. And only around 20% of America’s largest enterprises are today recognizing a profitable ROI from deploying AI. Last-mile AI is the most difficult, which requires operational excellence and continuous feedback loops to be successful.
Ethics is an important and often controversial topic in AI. It’s important to ensure that everyone is counted and treated fairly, including the outliers.
Often times, organizations are training AI models with the wrong data. As one example, hospitals were incorrectly correlating the seriousness of diseases to the total cost of the treatment. This induced a bias in the results, as it was actually just hospitals serving richer communities with better insurance plans that were being labeled as having sicker patients.
80% of the AI models today are running on CPUs, mostly from Intel (Nasdaq: INTC). 20% of the models are running on GPUs, mostly from NVIDIA (Nasdaq: NVDA). Custom silicon is developed by large tech companies for specific applications, like Amazon’s (Nasdaq: AMZN) Inferentia for its Alexa application. The Amazon Echo is now in 50 million households.
Now that the world has AI models, it also needs models to monitor those models. Explainable AI refers to trying to explain whether processes are in or out of control, based upon a continual feedback loop.
This matters a lot for self-driving cars. Autonomous vehicles look to keep processes in control, and have failsafes that kick in when things that aren’t previously recognized appear. For example, if a purple elephant steps in front of your car (something the AI has never seen before), the car still needs to recognize it as an outlier and know to stop before hitting it.
More broadly, the developing field of metacognition aims to understand the decision-making process that AI goes through. It’s similar to why our human brain makes decisions in the way that we do.
Synthetic data has been around for decades, but it’s becoming much more important now as the enterprise world is embracing simulations. Unity Software (NYSE: U) is one company who is ahead of the curve here, since it has already created the physics required for designing games in the Metaverse. This translates nicely, to companies who are trying to model physical concepts such as gravity or acceleration and bring them into the digital world.
Here were several other topics that were interesting. If you’d like, I’d be glad to stir up a conversation about any of them on our 7investing Community Forum.