In my Actuator newsletter, generative AI is a frequently discussed subject. I’ll admit that a few months ago, I was a little apprehensive to devote more time to the topic. Everyone who has covered technology for as long as I have has experienced numerous hype cycles and financial burns. A fair dosage of skepticism is necessary when reporting on technology, perhaps balanced by some excitement about what is possible.
This time, it appeared that generative AI was patiently waiting for the ultimate collapse of cryptocurrency. Projects like ChatGPT and DALL-E were waiting to be the subject of frenzied reporting, optimism, criticism, doomerism, and all the rest as the blood drained out of that category.
People who read my writing regularly know that I never had a strong opinion on cryptocurrency. With generative AI, though, things are different. To begin with, there is almost unanimous consensus that in the future, artificial intelligence and machine learning in general will play more central roles in our lives.
In this case, smartphones are very insightful. I write on the topic of computational photography occasionally. There have been significant developments in that area recently, and I believe that many manufacturers have now found a solid mix between hardware and software that both enhances the final product and lowers the entry barrier. For instance, Google performs some genuinely astounding feats with editing tools like Magic Eraser and Best Take.
These are certainly cool techniques, but they also serve a purpose rather than just being features for features’ sake. The true challenge going forward, though, will be to incorporate them into the experience naturally. Most users will be unaware of what is happening behind the scenes in ideal future processes. They’ll merely be pleased that it functions. This is standard Apple strategy.
Another way that generative AI differs from the hype cycle’s forerunner is that it provides a comparable “wow” effect right out of the gate. There isn’t much conceptualizing needed when your least tech-savvy relative can sit at a computer, input a few words into a dialogue field, and then watch as the black box spits forth paintings and short stories. That’s a significant reason why everything took off as rapidly as it did; often, when common people are presented cutting-edge technologies, they need to see what it might look like in five or ten years.
You can currently experience it personally using ChatGPT, DALL-E, etc. Of course, this also has the negative effect of making it more difficult to manage expectations. As much as people are prone to do, it’s simple to assume intentionality in this situation without having a solid understanding of AI. But that is simply the way things are now. We hope readers will stick around long enough to read the clever headline we used to draw their attention.
Warning: This section contains a spoiler. Nine times out of ten, they won’t, and we find ourselves spending months or even years trying to bring things back to reality.
The opportunity to discuss these issues with people who are far wiser than me is one of the great benefits of my employment. They spend time explaining things, and I hope I do a decent job of expressing that for readers (some attempts are more successful than others).
I’ve been figuring out ways to insert questions into talks ever since it became evident that generative AI had a significant part to play in the future of robotics. I find that the majority of experts in the field concur with the prior statement, and it’s intriguing to see the range of effects they anticipate it will have.
For instance, Gill Pratt clarified the role that generative AI is playing during our recent chat with Marc Raibert.
In order to effectively train a robot from just a few examples, we have figured out how to leverage contemporary generative AI techniques that allow human demonstration of both position and force. The code is not altered in any way. This is based on a concept known as diffusion policy. It’s work that we completed in association with MIT and Columbia. Thus far, we have taught 60 different talents.
When I questioned Deepu Talla, VP and GM of Embedded and Edge Computing at Nvidia, last week on the company’s view of generative AI, he responded as follows:
I believe that is evident in the outcomes. The increase in productivity is already apparent. It’s capable of writing emails for me. Although it’s not quite correct, I don’t have to start from scratch. I’m getting 70% of it. You can already tell that there are some items that function significantly better than they did previously. Summarizing is not always the best approach. I refuse to allow it to read and summarize for me. Hence, there are already some indications of increased productivity.
The director of the MIT CSAIL revealed how academics are utilizing generative AI to actually design the robots during our previous talk with Daniela Rus:
It turns out that even motion planning issues can be effectively solved by generative AI. Compared to model predictive solutions, you can receive far faster, more flexible, and human-like solutions for control. That has a lot of power, in my opinion, considering how roboticized future robots will be. Their movements will be far more fluid and human-like.
For design, we have also used generative AI. This has a lot of power. Also, it’s quite intriguing because it goes beyond pattern generation for robots. You must take other action. There can’t just be a pattern generated from data. The devices must make sense in relation to physics and the physical universe. This is why we relate them to
This week, a Northwestern University team revealed the results of its own investigation into robot design created by AI. The scientists demonstrated their ability to create a “successfully walking robot in under seconds.” As these things go, it’s not much to look at, but it’s simple enough to see how the method may be applied to develop more sophisticated systems with further study.
According to research leader Sam Kriegman, “We found a very quick AI-driven design method that avoids the evolutionary bottlenecks without reverting to the bias of human designers.” “We informed the AI that we desired a robot with the ability to traverse land. Then, all we had to do was hit a button, and voilà ! In the blink of an eye, it produced the blueprint for a robot that has no appearance.
The AI algorithm decided to give the little, squishy robot legs. Kriegman continued, “It’s intriguing since we didn’t teach the AI that a robot should have legs. It rediscovered that moving around on land is best done using your legs. In actuality, the most effective mode of terrestrial mobility is on two legs.
According to Formant creator and CEO Jeff Linnell, “generative AI plus physical automation/robotics are what’s going to transform everything we know about life on Earth.” “I believe that we are all aware of the existence of AI and anticipate that all of our jobs, businesses, and students will be affected. That works well with robotics, in my opinion. A robot won’t require programming from you. You will use English to communicate with the robot and ask for a specific action after which it will be determined. There will be a minute for it.
Linnell started Bot & Dolly and was its CEO before Formant. In 2013, Google acquired the San Francisco-based company, well known for its work on Gravity, as the software behemoth set its sights on advancing the sector (the best-laid plans, etc.). It’s all about the software, the CEO tells me, and I’m inclined to think Google agrees given the introduction of Intrinsic and Everyday Robots’ incorporation into DeepMind.