There is an exhilarating mix of thrill and hazard at the frontiers of tech.

Tuesday, May 9, 2023|

The internet has been influenced by Vint Cerf since he helped create it. He wears many hats, among them being the VP and chief internet evangelist at Google. He is to receive the medal of honor from the institute in Atlanta, and prior to the event he spoke with the website about his work.

When I joined the company in 2005, there were 5,000 people already. My normal attire is three piece suits. I thought I would be raising the quotient of the company by joining. Almost 18 years later, there are over a thousand people and I have failed miserably.

I hope you don't mind if I take my jacket off. Go ahead. We have always had an interest in artificial intelligence, and Sergey has come back to do a little bit more on that side of things. Over the past decade or so, that has increased.

The acquisition of DeepMind was a great choice. You can see some of the outcomes when you play Go and win. The more productive is figuring out how 200 million proteins are folded up. There are large language models.

We are trying to understand what these things can and can't do, and how they go off the rails, and how do you take advantage of them to do useful work. The thrill of discovery and the risk of danger create an exciting mix. I don't want to say the dangers of the large language models, but I did say there are dangers. I told the investment bankers that they shouldn't try to sell stuff to their investors just because it's flashy.

Don't try to apply it too fast without knowing how to put guardrails in place. I wanted people to be more thoughtful about which applications made sense. The Society of automotive engineers have different risk levels for the self driving cars, which could be applied to artificial intelligence and machine learning. If it goes down a dark path, you might want to put some friction into the system to deal with that, especially a younger user.

Suddenly, the risk factors are very high, as you get to the point where you are training these things to do medical diagnosis or make investment advice, or make decisions about whether someone gets out of jail. We should be aware of those risk factors. As we build applications, we can be prepared to detect excursions away from safe territory so that we don't accidentally harm ourselves. We need some kind of protection.

I am not an expert in this area, but I am starting to wonder if we need something like that for the natural language network. We can see when it starts to go off the rails. A second network that is observing both the input and the output might stop the production of the output. It is closer to executive function than a conscience function.

I want to be careful, but only by metaphor. I know that Microsoft is doing something like this. Their version of GPT-4 is called Prometheus. I had the impression that the Prometheus natural language model would detect and intervene if it thought that the interactions were going down with dark path.

I thought that they would implement it in such a way that before you actually say something to the person that is going down the dark path, you intervene and prevent them from doing so. I think that it actually produces the output and then discovers that it has produced it, but then it says, "Oh, I shouldn't have done that." I don't want to talk to you anymore about that. It is similar to the email that you get occasionally from the Microsoft Outlook system that says, "This person would like to withdraw the message.

" It makes me want to read the original message, even if I wouldn't have before. Yes, exactly. It is similar to saying, boy there is something juicy in here. It is an interesting place to work with the models.

Do you get the same kind of flavor that you got from working on protocols and other big shared things over the years? There are new properties showing up in the protocol world. In the online packet switch environment, flow control is a huge problem, and people have been working on it for years. I don't think a lot of us thought about the domain name business. Suddenly, all kinds of properties show up, people with interests that conflict and have to be resolved.

It is an even more weird environment where people buy internet addresses for $50 each. As I watched the auctions for address space, I thought I was stupid. When I was in charge of the Defense Department, I should have allocated the slash eight, which is 16 million addresses, to myself and sold it for 50 years. Simple systems can surprise you.

When there is a large number of them interacting with each other When something gets monetized, you should anticipate there will be emergent properties and possibly unexpected behavior, all driven by greed. I want to know if you are working on any other stuff. When I see cutting-edge tech being applied to people who need it, people with disabilities, people who like just have not been addressed by the current use cases of tech, I am always happy. Are you still involved in the accessibility community? We call them ERGs because we have a number of them.

Some of them are sponsored by me because they have hearing problems. There is a group for people with disabilities and their families who share their stories with each other because they don't know what the solutions are for other people. It is nice to know that you are not alone in some of the challenges. I am the executive sponsor of the Grayglers, a group for people with a little gray in their hair.

The challenges that arise as you get older are the focus of attention. There is a lot of Web 2. 10 years ago, it was completely impossible, broke all the screen readers, and had this kind of stuff. If you don't have this standard, you're leaving out millions of people.

I like to hear about interesting projects or organizations. Make accessibility part of your startup's products and culture from the beginning. In order to make things accessible, you have to have some intuition. I have come to the conclusion that we need to show people examples of something that is not accessible, and let them ingest as many examples as we can give them, because their neural networks will eventually figure out, what is it about this design that You learn a lot more from what doesn't work than from what does.

We did a two-day event looking at research on accessibility and trying to frame what this is going to look like over the next 10 or 20 years. The technology may be able to act as an augmenting capability for people that need assistance. There is great excitement but also great disappointment because we haven't used it as effectively as we could have. Alexander Graham Bell was working on a telephone that couldn't be used by people who are blind, which is why he was working on it in the first place.

It is a contradiction of priorities. Even if you can't see what the models are seeing, they can describe it. One of GPT-4's first applications was for blind people to see the world around them. We are experiencing something close to that right now, because of GPT-4's new capabilities.

Since I have hearing aids, I am using the caption capability. At the moment, there is no setting for closed caption on this one. I use the Chrome browser and it has a capability to detect speech in the incoming sound. packets are coming in and they are known to be sound, it passes through an identification system that produces a caption bar, which you can move around on the screen.

That has been helpful for me. For cases like this, where the application doesn't have captioned video, the caption window automatically pops up. I think we can do this in 100 different languages, although I don't know if we've activated it for more than four or five. As you say, these tools will become more and more normal, and as time goes on, people will expect the system to adapt to their needs.

I want to mention something that I found somewhat unnerving. An example of a conversation between a reporter and a chatbot was recently encountered by me. He chose to use the chat bot's output to speak to the system. He picked the style of a famous British explorer.

Even though the text was well formed, the weight of the assertions was added when they were wrong. Even when the thing doesn't know what it's talking about, the confidence levels are very high. I brought this up because we are allowing these indicators to be used to fool us. They used to mean it was David Attenborough.

It is just his voice. I realized that there was an ancient example of this problem that happened 50 years ago at Xerox PARC. It meant the first draft of anything you type to be printed out beautifully formatted with lovely forms and everything else, when they had a laser printer and the Alto workstation. Normally, you wouldn't see that production quality until after the text has been edited and formatted.

The first draft stuff looked like the final draft. People didn't understand that they were nuts, that they were seeing first-round stuff, and that it wasn't complete. It occurred to me that we have reached a point where technology is fooling us into giving it more weight than it deserves, because of certain indicia that used to be indicative of the investment made in producing it. I don't know what to do about that.

I think we need to make it clear what the provenance is of the thing we are looking at. Don't make any assumptions, like how we needed to say this is first-draft material. In a world where we have the ability to give content with attributes that we would normally interpret in one way, provenance is a very important concept. We should listen to that.

We have to think more critically about them. The attribute is being delivered artificially. It's possible that it's maliciously. Certainly that too.

Critical thinking has become an important skill because of this. It doesn't work very well if you don't know the provenance of the material you're looking at. I think we need to invest more in provenance and identity in order to evaluate the quality of what we are seeing. I wanted to ask you about the internet in the area.

This one was started way back in 1998. When it was possible to think about designing and building a communication system that would span the solar system, I was quite excited because I am a science fiction reader from way back. Many of the space agencies around the world are involved in the team: JAXA, the Korean Space Agency, NASA and so on. There is a growing team of people who are either government funded or volunteers.

The Internet Society started a special interest group called the interplanetary networking Special Interest Group in 1998. 900 people around the world are interested in this stuff. We are running it up in the International Space Station because we have standardized it. It will be available for the return to the moon and Artemis missions.

I will not see the end result, but I will see the first couple of chapters. It is not crazy to actually think about that. It takes a long time like all my other projects. It must have been a real challenge, but also a very familiar one.

Building something like this is what you've been doing your entire career for. This is a different set of capabilities. You put your finger on it. This is a different space than the one that works for the internet.

There are some really interesting problems where you have internet networks on the moon that are connected to other internets on other planets. We have to figure out how the Domain Name System works in the context of internets that aren't on the planet. .