Interview with Salesforce’s chief scientist Richard Socher

Chief Scientist Richard Socher announced new voice products for Salesforce’s AI tool Einstein last week.Salesforce

• Business Insider sat down with Salesforce Chief Scientist Richard Socher to hear his thoughts on the future of artificial intelligence at work.
• Datasets are top of mind for Socher, who also moonlights as a professor at Stanford. He thinks the AI community needs to be mindful of biases in data, which could impact whether an AI algorithm is ethical and even lawful.
• And while he doesn’t see AI taking over the human race anytime soon, he does have a vision for the future of work which includes a lot less stress for human users — even sales people.

Richard Socher would like to correct the record on artificial intelligence. There’s no reason to fear that self-conscious AI “will decide that humans are a pesky plague,” he said, or at least there’s no research to suggest this is the case.

But what AI can do is reduce some of the pain points of using a customer relationship management platform like Salesforce, and eventually make the sales process a lot less annoying for salespeople and their potential customers alike. 

Socher, chief scientist at Salesforce, is the top brain behind Einstein, Salesforce’s artificial intelligence tool, which purports to add an intuitive layer to the company’s entire CRM platform. He also teaches computer science at Stanford, where he got his PhD with a focus on AI and deep learning.

On Wednesday, Socher and his team announced Einstein Voice, a new product category at Salesforce which includes tools like a voice assistant and voice bots. Whereas Einstein used to only be able to read, now it can listen.

But the future of AI at work is bigger than mere transcribing. Socher sat down with Business Insider to discuss what he sees on the horizon for enterprise AI, and some of the ethical questions people in his field face in this brave new world.

This interview was edited and condensed for clarity.

Becky Peterson: A lot of companies in enterprise are serious about boxing off data so that AI is trained on a customer by customer basis. How does Salesforce handle that?

Richard Socher: We have a set of core values. Trust is our number one value. That sentence is literally burned into my brain, I’ve said it so many times.

So what that means for AI is that we are in a similar situation where the default is, we will never use a customer’s dataset for anything but improving that customer’s models. AI only gets to see and learn from that customer.

There are some areas where that makes a lot of sense. You don’t want two beverage companies helping each other’s sales. There are other things like sentiment analysis on Twitter, where it’s all public data and you just want to know if it is positive or negative.

Maybe it makes sense for customers to eventually say, “I want to join in on a data partnership,” and then eventually train more global models, so everybody benefits because it’s not part of what differentiates their core businesses from another one. But the default for everything is that it’s just trained on your data.

Peterson: These days, changes in enterprise tech reflect what people use in consumer tech. Einstein Voice follows the popularization of voice assistants like Siri and Alexa. So what comes after voice in AI?

Socher: It depends on how you define after and voice. [Laughs] In terms of technology, voice by itself is only sometimes useful, like if you want to dictate a document, when you’re like “I want to say it and I want to see it exactly written in a document.”

But in our case, what has to come after voice is actually natural language understanding, and an integration into important workflows. So we, for instance, have sequence tagging mechanisms where a sales person can walk out of a meeting and dictate all of his or her meeting notes.

I think the future is a fully immersive conversational CRM experience. Where you can just ask questions, give inputs, get outputs.

Just dictating that is not as interesting because then they have to still update all of the custom fields. So they still have to go into the interface, click a dozen times, do some drop down menus, select some fields, and it’s not that fun. It’s actually one of the most dreaded parts of using a CRM. But if you have natural language understanding on top of the transcribed speech, then you can automate that process too. 

So that’s this beautiful combination of unstructured text with the structured data in the database. So we can say, “Oh, you talked about Acme Corporation? That’s this field, that’s this object in your database.”

I think the future is a fully immersive conversational CRM experience. Where you can just ask questions, give inputs, get outputs.

And in the end, ideally everybody should have their own assistant as well as other things like their own doctors, and so on, because AI can learn from the very best people in the world.

Socher: I think the answer, as with all good, complex questions, is it depends.

Put my professorial hat on —  I think that some things will converge, namely the algorithms. We really have to give credit to the entire AI community in that we’re very public in the advances that we make. They are patented but they are also published in peer-reviewed, academic research papers. And we basically give the blueprints to most of the algorithms that we develop.

Including us. We just announced DecaNLP in a natural language processing decathlon, where we developed a single model to solve 10 NLP tasks. So you can ask one model, what’s the summary of this document or this email? What’s a translation of that sentence into German? What’s the biggest state in the United States in terms of GDP? What’s the sentiment of this tweet? You can ask all of these different questions to the same model.

Normally, there are like 10 different models—there are Q&A models. There are summary models. Translation models.

They’re becoming all neural network-types but they’re all very different architectures, so you can’t transfer. In theory, I have a model that can translate into German, and I have a summarization model, but that doesn’t mean that you can all of a sudden summarize in German.

We’re making a lot of progress on these algorithms and we’re able to share that because that’s just how the community works.

If anybody ever tells you we have a secret sauce algorithm that is light years ahead of everyone else, that’s usually bogus. Innovation happens so much more quickly when everyone collaborates.

If anybody ever tells you they have a secret sauce algorithm that is light years ahead of everyone else, that’s usually bogus. yone collaborates.

So that’s one side of the equation where we will see convergence of the underlying algorithms. But the other side is the dataset side. And the reason why all of these companies can tell you their algorithms is that the algorithms are only as good and are really only shining the most when you have a lot of training data for them. That’s where I think we’ll see more specialization and diverse outputs, is who trains on what datasets.

It’s pretty hard to infuse very specific rules into the AI. It just picks up on the data. If you give it bad training data, it will be a bad AI algorithm. And if you don’t give it enough training data, it will be a bad algorithm.

There are certain companies that are very far ahead when it comes to certain kinds of data sets, and sometimes even countries, but that doesn’t mean they are ahead on everything else.

Peterson: Since we are talking about Salesforce, besides the convenience to the user, what will AI mean for the world of sales?

Socher: I think it will make the whole sales process more efficient and be a better experience for both sides. A sales person doesn’t really want to send a bunch of emails to people who don’t want to buy their products. They don’t really want to talk to people on the phone and try to convince into buying something that they are just not interested in buying.

This is what we have with lead and opportunity scoring, where we can basically sort based on all of the interactions that two companies have had, and say today, these 10 people are the most likely to want to buy your product. So I think we’ll make sales people a lot more efficient and hence also more pleasant to interact with for everybody.

In service, we’ll see a lot of automation too. Customers just want to have a quick answer, and if you can just type that answer into a chatbot, you don’t even have to pick up your phone. Instead of waiting in line for half an hour for your representative, you just ask your question to an AI. I think that’s the future.

And in marketing, I think you’ll see all kinds of new jobs that will come out where you can basically personalize your marketing messages and make it much, much easier to create personalized campaigns and also understand in a much more automated fashion how that impacts your customer sentiment about your brand, on social media for instance, because now you can actually literally go through a million tweets and see every time anyone mentions or shows visually your product, what did they say? Is it positive, negative or neutral?

It’s a pretty exciting time in all of these three areas.

 [Here’s what analysts say Salesforce can do to turn its $10 billion of revenue into a $60 billion money machine]

Peterson: Salesforce announced during its last earnings call that it will create an Office of Ethical and Humane Use of Technology. What are some of the ethical issues that as an AI scientist you are confronting?

Socher: I think with AI and ethics, we need to differentiate between two things: One are the unfounded fears, and one are the more realistic issues.

For a couple of years now, because of some of the excitement and indeed breakthroughs in AI, people extrapolate a little too far and are like, “So Richard, what happens when AI will automate all of the things, and no more human jobs are there, and the AI is self-conscious and will decide that humans are a pesky plague and we need to get rid of them?”

That’s just a really unfounded fear. There’s no research. We don’t really know the missing steps toward self-conscious AI that will set its own goals and objectives and things like that.

So it’s kind of a red herring. It’s kind of a distraction. And frankly, it’s fine to think about cool SciFi things, and maybe have SciFi ethics and so on.

But it’s also quite distracting from real issues that we have right now. And those real issues come from datasets that the AIs are trained on and the biases that are in those datasets.

If you want to build an AI algorithm to define who should get a promotion, and your promotion process was kind of sexist before, then your AI algorithm is going to have that same bias in it.

And if you want to apply positive use cases like AI in medicine, and you train it only on middle aged white guys, because they are the ones that keep getting a certain type of scan, then your AI algorithm isn’t going to be as accurate for young, African American girls because it hasn’t seen enough of them in the training data.

Equality is also one of our big values. We really need to carefully think about the datasets that we’re training our AI on. And we have ethics Trailheads to try to also educate our customers around the kinds of issues that they will observe themselves.

In some places, fortunately, in the US, it’s legally important to do this. If you want to classify whether somebody should get a loan or not, and you don’t have a race column in your CRM, but you may have a zip code and income column. And from those two you can almost in many areas of the United States exactly know the race of somebody. So you need to be careful about what things you can include and should include into your classifiers to make sure they are ethical.

The data set bias is a major issue of AI, because as I said, AI is only as good as the training data that it gets.

Socher: This is not like Tay. We’re not going on Twitter and everybody can try to mess with it. In most cases, our customers also want to have their customers have a good experience so they have the incentive to train AI in a way that everybody enjoys working with it.

Peterson: If I am a salesperson on Salesforce, and I am extremely bad at my job, is that going to be reflected in the way Einstein treats me?

Socher: In fact, you’ll probably — and this is one of the beautiful things — you’ll probably get better because of Einstein. The simple example of lead and opportunity scoring, somebody who’s just starting — let’s not say they’re bad, but they’re maybe novices — they’re just starting, they don’t know all of the background stories on these sometimes large companies. They’ve had so many interactions, so many emails sent back and forth. So you might not know that you shouldn’t be reaching out to this company at this point because they have just had some other issue, they’re doing a restructuring, so they’re not going to buy some new thing because they’re dealing with their own issues.

The lead and opportunity scoring could give you insight like that, like you should reach out to them, they’re very excited, and then like in three weeks or four weeks if they didn’t get back, it’s much more likely to succeed. Maybe they forgot about it for some reason but they’re actually interested they just forgot about it.

So there are a lot of things in AI that can help make people more efficient. And I think especially in sales, that’s just a win/win.