What Historical AI Breakthroughs Have Been Unappreciated by The Mainsteam Media?
Recently had the chance to interview Joshua Browder, the Founder/CEO of DoNotPay. The interview is published in full on HackerNoon, and below is an excerpt that I think aptly summarizes the metoric rise of AI from an entrepeneur in the the thick of it.
David Smooke: Since the launch of ChatGPT, the AI boom’s been in mainstream media’s crosshairs. PyTorch and TensorFlow were monumental achievements that maybe weren’t fully appreciated until something more user friendly was built with/atop/beyond it. What future AI breakthroughs have you excited? And what historical AI breakthroughs do you think have been under-appreciated by the media?
Joshua Browder: It feels like we are making years worth of progress every month in A.I. and things that were not possible even last Fall are possible today.
The first major breakthrough was when GPT 3 was coherent enough to hold a conversation. At that point, we built an A.I. that can cancel subscriptions. As you may know, some companies, such as The New York Times, make you chat with an agent, just to cancel a subscription. It felt like magic the first time we canceled a magazine subscription with A.I.
Then came GPT-4. The reasoning functionality for what we were trying to accomplish seemed like an order of magnitude improvement, so it allowed for more sophisticated products. Recently, we launched A.I. bill negotiation, where our robots log in to your utility account (such as Comcast) and start chatting with an agent to get you a discount. In some cases, the big companies are using A.I. (and we are using A.I.), so the two A.I.s are battling it out. With GPT 3, this use case would not have been possible.
Multimodal, where A.I. can accept different types of inputs (such as images), is probably the most unappreciated breakthrough by the media. I don’t think many consumers realize that ChatGPT can “see.” At DoNotPay, we are using GPT-4 vision to assess parking signage, such as when our system prompts GPT-4 to determine: “is a tree covering the sign?”
Latency is still the thing that needs to improve the most. 6 months ago, both the large language models (and the voice models) would take too long to hold a conversation on the phone. For our purposes, a lot of consumer rights disputes get handled over there. It seems we are finally at the point where we can build phone bots to complete tasks on peoples’ behalf, though we still have some minor technical improvements that need to happen.
Read the Full Interview: 'Multimodal is the most unappreciated AI breakthrough' says DoNotPay CEO Joshua Browder, AI Blogs on HackerNoon, and Get Interviewed by HackerNoon.