JT: Our focus today is really what does AI mean for everyday life. I’m hearing a lot about AI. What is your assessment about where we are and how it is making a difference?
WT: we’re in an unprecedented, interesting era. From a consumer perspective, negative connotation.
It’s an interesting era we are in; these technologies are going to do a tremendous amount in terms of consumers selecting what they buy, Helping patient-centric care.
Combination of data set & availability of resources is fueling AI.
You might hear 90% of the world’s data has been created in the past two years. AI will help us deal with that kind of information overload.
The big difference with programming systems is that AI knows how to understand, reason, learn, interact.
AG: There is a set of techniques through which we can more accurately predict fraud, insurance plans, credit scoring.
This is a jump in the past 15 years.
5 years ago, the ability to do very exciting things with unstructured data, i.e. automating radiology. Then digital networks came along and then we had use case after use case.
AI has lots of programmatic uses.
WT: Algorithms are contributing to oncology.
You take a look at things we’ve done in oncology as an example: the ability to train a system;
Effect is based on training sets that AI is being fed. Are we using the right humans to train these systems?
We need to hold the systems to the same standards we hold humans.
Design principle: It’s always assisted. Not replacement.
JT: Do you feel that the future is assisted?
AG: I don’t necessarily agree that future is assisted. Repetitive tasks can be automated, e.g. radiology. End result is that algorithms are out for anything that involves repetition or mundane tasks.
Humans spend a lot of time doing repetitive things. I think algorithms will be our future radiologists. Any job that demands a lot of repetition, for instance, auditing, mundane legal tasks.
There is probably an element of combination between routine algorithm and more challenging cases for humans
Eventually the algorithm will do the routine, simple tasks, and the more challenging tasks will be given to humans.
WT: I agree. “We’re getting humans to raise their game”. The idea is to get rid of commodity tasks. When I call Comcast, I go through the same questions all over again. So how do you reduce the time from 1 1/2 days to five minutes? How do we get humans to find solutions to more complex problems?
Even in a scenario in which AI is playing a game, AI takes care of the commodity moves, and the final win from the human.
Creativity comes from humans vs. commodity
JT: Great leap forward is to process unstructured data and give insights; Give insights to someone who may only see a small sliver.
WT: Especially important in life-sciences;
They are unstructured data: handwritten reports, etc.
10k new articles on clinical trials
AG: Lets say radiologists can look at 3,000 images a year; AI can look at 3,000 a second.
As long as task is suitable, AI can achieve objective.
Machines have an unfair advantage.
JT: what should we be doing as a community to realize potential benefits of AI in everyday life?
AG: I’m at Google and I think Google does this effectively: when we use the voice assistant or photos we search for our photos by “search my name” and it finds pictures of me. That’s the Google brain team going out and infusing products with digital awareness.
There is a shift that companies should make: being willing to shift to need a handful of outstanding machine learners rather than many; think the right way about talent — small, extremely talented teams.
WT: The developer is this era’s doctor/engineer.
Dominantly, data and application team now work collaboratively. More need for data science.
The data team and application team used to work separately; now there’s more need for data scientists and the collaboration of those teams. Data is the fuel for AI; so that’s an important dynamic to think about.
Those roles didn’t exist until recently. As we go into next phase, new roles and division of labor will come up.
Building a team with tremendous expertise is necessary.
JT: Looking forward, how far should we and can we be taking AI?
AG: “The future is already here; it’s just not widely distributed” Challenges are mainly organizational.
Google brain is an example of how you productionize machine learning in a company.
Also, reinforcement learning: There is a technique now that automates trial and error; it plays enough games that it learns what gets it a good score.
We’re starting to see AI as input to stock trading, ad targeting, etc.
Generative models are a new area of machine learning: they’ll take an image and be able to write a caption for that image. Visually impaired people can now describe scenes using AI.
Digital networks are starting to make their way into existing use cases.
WT: There is no reason to believe that something like tax codes can’t be replaced by AI. In the future, systems will have embedded AI; Internet can provide access to these systems at a commodity level.
JT: You’ve talked about replacing commodity activities, used more broadly. I think about the development of trust that it will take for these technologies to become widespread. I’m skeptical.
JT: how can we develop trust ? How should we think about building trust for broader adoption?
AG: There’s the issue of the market being ready. Do i trust my cancer diagnosis from a machine? Building trust is case specific.
Let’s use radiology: You can start by having a machine operate alongside a human; look at the agreement rates. With medical diagnosis you eventually know for sure. So over time, articles are published, the machine as a track record. If it’s high-performing, maybe it takes over. There is no general answer; very use-case specific.
WT: Agree. I think you have to build these on some principles, ie transparency. As a consumer you want to know if you are dealing with a human or a system. If a system, who taught it? And when it generates a recommendation, you want to know the data set that recommendation was generated from. Human-assisted is important; will yield the type of system people can trust.
JT: Also, no human is perfect; so what are our expectations of the system vs. the human?
AG: “And no algorithm is perfect.” The Tesla will have an accident, just as humans do. The question is: Does the Tesla have accidents at a lower rate than humans? You can be sure that when a non-human has an accident, we’ll view it differently.
Q: I’m a physician at Stanford. Doctors spend only 30% of our time taking care of patients, 70% on data. Are we developing AI at the expense of human intelligence?
WT: not at the expense — -how is AI giving you more time to actually make you more efficient and give you data to better make decisions? Data-entry will be taken
In some cases, we’re giving you more leverage in terms of your data set and efficiency. You won’t have to key in data; it will be learned / read / listened to by a machine.
AG: Future lives will be more interesting when you take mundane, repetitive tasks away.
Let’s say a our mundane roles go away; does that mean fewer of us are needed? Historically we’ve gone through waves of automation and more professions are created. It’s hard to know in this case; is the disruption happening too quickly to adapt? It’s a little scary. If the structure does change, I think all our lives will be more interesting minus the mundane tasks.
Q: I have a 20-year old daughter in college. With so many jobs potentially being replaced, what career advice do you have for her?
AG: Computer programming and machine learning are good bets.
If the job involves creativity, and connecting dots in disparate ways, no machine learning technique can replace that even remotely. I don’t know any machine learning techniques capable of doing that.
WT: In the major revolutions, there’s always been the fear that occupations will be replaced.
We’re in the same era.
Q: Use-cases have been scientific and medical; what about social and political limitations of AI implementations? E.g. Law that involves rules that are like algorithms: would you be willing to replace jury trials with AI, why or why not?
AG: At low levels of the legal system, yes; it’s only once you get to the Supreme Court that those should still be conducted by humans. But there are so many rote cases that come to court again and again; so it seems feasible.
Rote cases could very well be replaced by AI.
WT: Reasoning is still important. You may have data sets that assist the jury to help them make a better decision; but you can’t replace the human factor on the judgment call.
Originally published at blog.cloudflare.com on September 14, 2017.