Interview #104 Andrew Moore, Co-founder & CEO, Lovelace

SPEAKER_1 (0:00)

AT&T Business Wireless helps businesses stay connected. I own a landscaping business in San Fran. We get booked for jobs all over the Bay Area. I need a reliable connection so I can check on my crews working in Hayes Valley and over in Fruitvale. With AT&T Business Wireless, I always know if I need to dispatch a team or some celebratory donuts to a job site.

Donuts it is. Keep your business connected with AT&T Business Wireless. Visit att.com/5gnetwork.

Tony Hoang (0:26)

When the connection matters, it has to be AT&T.

Welcome to The Artificial Intelligence Podcast, ranked in the top 2% of over 7 million podcasts on Spotify. My name is Dr. Tony Huang. Today we're with Andrew. Andrew, welcome to the podcast.

Andrew Moore (0:46)

Thanks Tony. Really glad to be here.

Tony Hoang (0:48)

So for the viewers listening at home, Andrew, can you give us like a brief description of your background?

Andrew Moore (0:53)

Sure. Yes, I started out as a professor at Carnegie Mellon University. Then I've done in my career stints at Google, where I've been involved in Google Ads, and I was responsible for the Artificial Intelligence division of Google Cloud.

Tony Hoang (1:09)

So right now, chatbots are great at giving answers that sound right. So but in like a bank, at like a hospital, or like in like a military setting, sounding right isn't enough.

What does like an AI need to understand about like a situation before it answers, before like the answers can be trusted?

Andrew Moore (1:30)

Exactly right. So the thing is sometimes you ask a question, especially if you're doing something like investigating potential financial fraud, or trying to figure out the due diligence about a particular company. The questions you're asking there need you to really touch hundreds of thousands of different pieces of information to formulate your answer. So when I, when I as a consumer, ask Google about my banana plants, it just has to look up like three Reddit articles on banana plants, combine them together, give me a good answer. I'm delighted. On the other hand, I have to say, is that shipment going through the Straits of Homers right now legit?

That is something which touches so many bits of information, you've got to have a different approach for answering the question.

Tony Hoang (2:18)

So in your opinion, like, what does reading the room mean for like a machine?

Andrew Moore (2:25)

It's a very difficult for a machine to just look at a small question and really figure out what the right way of answering it is. So I do expect that at the point where a machine is being asked to make a serious decision, like something involved in planning a military strike or like something involved sending to the Treasury a warning that someone might be involved in money laundering.

Those kind of things don't look for snap answers. They look for well-sighted, thorough answers.

Tony Hoang (3:00)

Can you tell me about a time you saw like an AI give an answer that sounded completely reasonable but missed what was actually going on? And then like, how did someone catch that?

Andrew Moore (3:12)

Yes, a perfect example there is one day, one of our AIs saw a huge number of ships heading out of port from a Chinese port towards Taiwan. And so the first thing a quick AI would have done is say, up, this looks like an invasion.

What our AI did was it read the local Chinese newspapers to understand what might be causing this. And it realized, up, this is the start of fishing season. And this is something which, oh, look, let me check my records. Yep, this happens every year at this time, or a whole bunch of fishing boats are all headed out to fish at the same time.

Tony Hoang (3:52)

Now, I've noticed that many big AI projects at orgs never make it past the pilot. In your opinion, can you tell me about the ones that you've watched fail and what actually kills it?

Andrew Moore (4:06)

There are a couple of things which are really bad for these kind of projects. And one of them is if you're trying to solve with technology something which is actually an organizational problem. So common scenario is there's something which involves a whole bunch of people laboring away to do something. Like, for instance, if you are responsible for figuring out if a transaction is dangerous, if you're looking for large financial transactions, there's a large room full of people who get alerts from automated systems and have to do a bit of an investigation, sometimes involving ChatGPT or Google. Then the ones that they're alarmed about, they send up to another huge room of people who do secondary screening on those.