Two years ago Quoc Le of Google Brain conceived of “AutoML,” where he used AI to design the neural networks of other AIs. The results were impressive, with AIs beating the hand-built networks of top research teams. DeepMind entered the fray this past week, applying sophisticated techniques that mimic evolution to build AIs for Waymo’s self driving cars. Their approach reminds me a bit of the 2002 NEAT algorithm popularized by YouTuber Code Bullet for toy problems. (link)

AI practitioners spend a lot of time gathering, cleaning, preparing and labeling data. A freshman at MIT was looking for a side hustle with his friend in 2015. After working his first summer as a quant on wall street, he settled on the idea to outsource data labeling via an API. His service combines machines and humans to return clean, labeled datasets. His top customers are the litany of self-driving car companies that need fused, labeled data from multiple sensors. Alex is now sitting on an AI unicorn! (link)

Some hackers have built an AI to keep telemarketers on the phone, preventing them from scamming others by literally chewing up the hours they work. This particular AI stitches together hilarious snippets from prank callers. (link)

Researchers showcased AIs that can generate short, 3 second video thumbnails completely from scratch. These GANs generate all 48 frames simultaneously vs. one after another in 256x256 resolution. The GAN was trained on the kinetics video dataset from Deep Mind. (link)

Feather color matters. The darker colors of a feather save energy, increasing the distance and speed birds can fly. As we study biological systems to influence computer designs, let’s not forget that somewhat inconsequential details may play an important role in our intelligence. (link)

We’ve seen that roughly 4 out of 5 AI projects never launch. A recent report claims something similar, showing 87% of the sampled data science projects remain in the labs. Root causes seem to be silo’d data, poor collaboration across business functions, and trying to do too much. Sometimes its as simple as the lack of a sponsor in IT operations who owns an app that will call your model, take the prediction, and do something with it. (link)

Parking garages of the (near) future will be designed for the inhabitants – semi-autonomous cars – and less for human operators. Daimler and Bosch teamed up on a design that lets humans drop off a car to self park and returns. This mirrors what we see in cloud data centers that are designed primarily for machines, and not for the humans that service them. (link)

Your next pair programmer may be a bot. Someone built an AI to autocomplete lines of computer code, much like GMail finishes your sentences. These are largely generic models tuned and applied to custom datasets. Here the engineer trained a variant OpenAI’s language model (GPT-2) with 2 million lines of code from GitHub. Coders have called its ability to predict code “astonishing.” (link)

Remember all those electric scooters littering the sidewalks in California? Two entrepreneurs started a repo business, “charging” the scooter owner $30 to return a repossessed scooter, plus $2/day up to $60. Shop owners were delighted, calling the startup to clear pathways and sidewalks. Bird thought this was cute and paid $40k to return a slew of scooters, even snapping smiling photos for social media. Repo continued. Now lawyers are involved. (link)

NASA to Elon Musk - wanna bet? This feels like a prize fight! I’m going to get some popcorn and enjoy this show. (link)

Meanwhile, NASA is “gassing up” with nuclear fuel for Mars rover. (link)