James Cham — Investing in the Intersection of Business and Technology



About this episode

Connect with James

Twitter
Bloomberg Beta

Listen

Timestamps

Links

"Street-Level Algorithms: A Theory at the Gaps Between Policy and Decisions", Ali Alkhatib and Michael Bernstein (2019)

Watch on YouTube

Transcript

Intro

There's still an enormous disconnect between what an executive expects to be able to do and what the software developer or what the machine learning person or data scientist actually understands is doable.

You're listening to Gradient Dissent, a show about machine learning in the real world. I'm your host, Lukas Biewald.

James Cham is a partner at Bloomberg Beta, a fund that invests in machine learning and the future of work. He's invested in many successful companies, including my first company CrowdFlower and my second company Weights & Biases.

I've worked with him for a really long time, and he always has really smart things to say about technology trends. I'm super excited to talk to him today.

How investment in AI has changed and developed

So James, you've invested in AI for a long time. You were the first investor in CrowdFlower, my first company, and you're the first investor in Weights & Biases. I was curious to know your perspective on...how your thinking around investing in AI has changed over the last 15 years.

Clearly the market has changed, but I was curious to understand how your thinking has changed.

You know, when I invested in CrowdFlower, I didn't understand that I was actually investing in AI. I thought that there was a broader collective intelligence problem that you were solving. I was really enamored with both crowdsourcing and flash teams, at the point.

And to be honest, I kind of still am.

I still sort of...in some ways that I think about AI — or machine learning more specifically — kind of as a misnomer, I think that it's actually a collective intelligence thing that's going on.

That's on the broad, theoretical side. The big change on the investment side, I think, is we went from a place where people actively didn't want to invest, or where I actively — there are a couple of folks that you and I both know who I actively encouraged not to use the word "machine learning" because I thought it hurt their chances to raise money — to a world in which now we live in where there's an incredible amount of investment.

What's interesting about the incredible level of investment right now is that we're still sort of at the cusp of getting actual great business results, right? And so we're sort of at that point right now where I think all the pieces are almost all there.

But they're not quite, and everyone feels that...you have that little bit of impatience where everyone kind of wants to get it, and the talent's not quite there, or the executives don't quite understand it.

That's an uncomfortable, but also really exciting point to be in.

Do you think there's some chance that we're set up for disappointment?

We are always set up for disappointment. You know that as well as I do.

That's true.

Lukas and I, I'm lucky enough to...every two weeks we have our little morning chat. I feel like we have recurring themes and one of them is this continued question of, "Where are we in the market?"

And you have to admit that the last few quarters, there's this sense that everything is coming together, right? But at the same time, as you feel like everything's coming together, you're still looking behind you to say, "Oh goodness. In what way are we overselling? In what way are people misunderstanding things?"

At least to me, it feels like there's still base levels of understanding that are missing. And it still feels to me like there are opportunities to define the market in the right way, rather than the buzzy silly way.

When do you think investors kind of flipped from feeling like machine learning was a science project to machine learning was a good business to invest in?

I mean, you've always done early-stage seed-stage investments. That's probably where the change happened the earliest, but when was that and what was going on that caused that change in mindset?

You know, there's this little joke around Google and Facebook where, you know, "What do startups really do? We commercialize things that Google figured out five years ago," right? And then we bring it to the rest of the world.

There's a little bit of that sense that that's not ridiculous. That you saw the kind of changes that people were able to implement and build inside the big FAANGs and then realize that this should be more broadly available.

So you had that on the one side.

And on the other side you had these remarkable...well, okay, how do I think about this? I think on the academic side, you had a few things happen. On the one hand you had great results, just super impressive results.

But also there's a way in which academics sort of figured out how to play the game, in the sense that the machine learning world was well-defined enough now that people could compete on some basis that they understood.

I remember there was this guy who gave this great pitch around how to think about advances in machine learning. He made the point that, actually, maybe it's really about the size of the dataset. Do you remember who that guy was?

Do you think that's still true?

That was Lukas by the way. That was Lukas. Just to be clear, just to be clear.

Do I think that what is still true?

Well, I do think the size of the dataset is incredibly important. And I think maybe 5 or 10 years ago, I thought it was really the only important thing, and the advances in algorithms seemed pointless to me at the time.

But I think in retrospect — maybe I didn't have such a quite extreme view — at that time it wasn't clear that deep learning worked much better than traditional methods. There hadn't been a lot of improvements in algorithms for a really long time, so almost all the advances felt like it was coming from bigger datasets.

But now I look at OpenAI and DeepMind, and it feels like a lot of the advances that are happening there is, on one hand, coming from bigger datasets making more advanced modeling possible, but also advances in compute.

I've got a nuance on the extreme claim you used to make. Which is, I actually think it's that with the availability of large datasets — but also with the understanding that these large datasets were available — it meant that everyone understood how to play the game.

It meant that you have a whole wave of academics and companies and corporations and groups and teams saying, "Oh, we can play with these sets of data in interesting and novel ways."

What that meant is that the thing that was the scarce commodity...or the way that you basically laid that piece out, meant that people were able to work on it. And then that's where you get all these exciting advances. In part because everyone agreed on how to think a little bit about the data.

Creating the first MI landscape infographics

You know, I wanted to ask you too...I think one of the things that you did really well was maybe starting a real trend in content marketing among VCs, when you and Shivon put out the machine intelligence infographic where you laid out all the companies.

I was curious, what caused you to start it? And then I feel like it became wildly successful and you stopped doing it. Many other people have picked up where you left off, but without the same — in my opinion — quality that you had.

Can you tell us the story behind that?

Sure. When the fund started, I think there was a sense that we were at the tail end. Incorrectly, there was a sense that we were at the tail end of a bunch of investment around big data, and that there were a lot of failed, big data projects sitting around.

And so then the question was, "What are you going to do with all that investment and understanding and collecting data?" One of the claims, or one of the guesses, was that you'd use that data for machine learning, right?

There are a bunch of AI applications. And my old colleague, Shivon Zilis, pushed that insight a lot. I think in part because she felt it just intuitively, but also she was surrounded by a set of folks who were playing around different places with it.

I think we were both sitting around thinking, "Wow, this is just so hard to understand," and we couldn't make heads or tails of it.

Basically, what happened was...you know, Shivon being just a really great synthesizer, but also someone who's quite dogged, decided to go work with another friend of hers who figured out ways to cluster different types of businesses. She basically clustered a bunch of different types of businesses that included a number of keywords around AI, and then categorized it, and then stuck it on a map.

I think that was like a two-month process to actually go through all of that and have all these horrible spreadsheets. Because it was super...there are products now that do this, but it was super manual in some ways.

And what was exciting about it was, the moment she put it together — I give her all the credit for actually doing the real work — then suddenly it felt like this world was legible for the first time.

Then I think we kind of assumed that there should be people working on this full-time, rather than having this just be a part-time job, and they would do a better job of it.

For a few years, basically Shivon would take some time off right around the summer to just do the state of what's going on. I think it was really good. The categories were not always right, but at least it gave something for people to agree or disagree on it. And then made a bunch of connections for folks that I think would...it's still valuable to this day.

Why did we stop? I don't know. Like, there are too many companies, right?

Part of it is there are too many companies. Part of it is, I do think there is a new class of journalists who now think that way, right? Who think that mix of computational plus willingness to do the work, plus not sort of subject to the day-to-day grind of reporting the next story. And they should be coming up with those conceptualizations.

But I haven't totally seen...I do think it was a novel contribution at the time.

The impact of ML on organizations and management

One thing that I know you are very interested in — because you talk to me about it all the time — is how organizations function as a collection of humans trying to work together towards a common goal. I feel like you think about that more than most, and you think about machine learning more than most.

I was curious how you think — or maybe how you've seen — organizations adapt to machine learning becoming more mainstream within them. And I'm curious if you have predictions on how organizations might continue to evolve as machine learning becomes a bigger and bigger part of them.

We're not yet at the point right now where machine learning is boring enough that it could be adopted easily. We're still in the part of the market, or part of the phase, where there's plenty of exploration and plenty of definition and ecosystem definition to be had.

And you see some of that in like slightly misguided arguments around augmentation versus automation. I think you only have those sort of theoretical questions when people don't have actual solutions they're dealing with day-to-day, right? But I think that there's definitely...that's the first part.

The second part is...management theorists have thought for a long time — or talked about — the idea of a learning organization. That organizations will actually get better over time because they learn things.

Generally that's just been a metaphor, right? Because of course organizations are not people. They don't have minds, they don't learn anything. Maybe things get codified, and processes or rules.

Part of what's exciting about machine learning — in the next, in the pre-AGI version of machine learning — is that we could actually digitize a bunch of decisions that get made on a day-to-day basis. And we can actually literally learn from them, right?

Something as boring as, "Do I go to this meeting or not go to this meeting?" or something as important as, "Do I invest in this project or not?"

All those things in the world we live in right now have almost no consequences. No one actually follows up on a consistent basis to make sure or understand whether things work or not. Or they do, and it's incredibly expensive and difficult.

Just think about...not you guys, but maybe some other theoretical organization will have to spend all this time just digging down to figure out "What product...", like "What random marketing campaign actually happened or didn't happen?", or how well it worked. And just the amount of automation people need to put in in order to systematize that.

What's exciting about...at least to me, what's exciting about the sort of data-rich ML world we could be living in, is that those decisions we can now find out whether they actually work or not. And then we can actually maybe consistently start making better decisions.

Now, there are also a bunch of...you were going to say something, what were you going to say?

Well, let's take your example of "Should I go to a meeting or not?" How do I ever even know in retrospect if I should have gone to a meeting? How could an organization really learn whether or not it makes sense to go to a meeting?

Okay. I think there's...one of the other angles that I'm very interested in is that intersection around machine learning and the social sciences.

You'll talk to management folks that are rather on the AI side, and there's always this question of "What's the objective function?" The interesting thing is that on the social sciences side, they've learned the lesson.

Which is, "I don't know. We'll have some objective function and it'll be good enough to sort of manage, but it'll never be perfect. That actually will have to change over time because the most interesting systems that are all dynamic, they're dynamic because people are interesting."

That, pnce you decide that one metric is the right way to measure whether a meeting is good or not, people will start to learn that and they'll start to game it. They'll be like, "You know what, whenever Lucas smiles twice...I'm going always make sure to make it, I'll tell some stupid joke." And it'll detract from the actual purpose of the business, right?

I think that the illusion is that you'll come up with some perfect metric. And I think the actual goal is to continually come up with metrics that slightly will change over time, and you'll understand what works or doesn't work, but that'll be okay, right?

In traditional organizational science, there's this great paper called, "On the folly of wanting A and measuring B." And I think that problem is going to be forever. But that's part of the fun of the job, right?

That's part of the fun of creating organizations and social systems.

I totally agree with that, but I feel like...I don't want to harp on this case too much, but I'm curious because I always wonder myself if I should go to a particular meeting or not. How would you even make an imperfect measure of that?

What do you even imagine looking at to-

-so you can certainly imagine it visit. You can imagine it as, "Is the meeting useful to you?" You can also imagine it in terms of, "Is the meeting useful to increase the collective intelligence of the organization?"

And you can certainly do direct measures, which we can just literally ask you, "How good was that meeting?" afterwards. Or we can literally ask the team, "How good was that meeting?" afterwards? Or we can literally look at the number of things you write after that meeting, or we can literally look at the number of times that you nodded or didn't nod.

Which is just to say all those signals are increasingly cheap to gather. And when they get cheap to gather, that's when we actually get interesting innovation.

When it's incredibly expensive — when you need to hire McKinsey to do some study and then hire a bunch of people to build some very expensive bespoke system — then it's not that useful, right?

Because then your ability to move and play with the edges of your social system becomes too difficult. And then you're sort of...your chance to actually design it on the fly and continue to understand it...

That's, I think, the interesting edge around social systems.

Interesting.

Where do you see machine learning making a meaningful difference in organizations today?

In all the normal places, right? We're now finally getting good enough to cluster large-scale bits of information in ways that are meaningful, so that we can provide consistent responses.

I think that that piece of it — which is the big version of machine learning; finding the most critical decisions you need to make, the most digitized pieces, and then finding ways to consistently improve and collect it — I think that that's where most of the energy and opportunity is right now.

But that'll change, right? That'll change. I think that the exciting...does that make sense, first of all?

Yeah, totally.

Demystifying ML for executives

Let me take one slight digression, as we're talking about this. Of course, the real answer is that executives could know how to apply machine learning if only they understood a little bit more than what they learned from reading or watching a movie.

There's still an enormous disconnect between what an executive expects to be able to do and what the software developer or what the machine learning person or data scientist actually understands is doable.

I do have to make the pitch — which I think I've done too many times to you — which is, I do remain convinced that the three- to four-hour class that you used to teach to executives on how to think about machine learning probably is the best...if you were to say, "What's the best way to improve the way people think about machine learning?", you should make your boss's boss take a three-hour course and just sit around and play with a very simple machine learning model.

Because in that process, they will at least have some intuition about how incredibly powerful, unsexy, brittle, finicky, and incredibly scalable some of these models that you build will actually be

Well, it's not the core of our business, but I am passionate about doing it.

And really it's not that we shut down those classes, there wasn't actually much demand for it. Or maybe we didn't pursue it aggressively enough. There was much more demand for the tools that we build.

But I'm curious, when you did the class...

Go ahead.

Maybe I'm actually just soft-balling a pitch to you, but I'm curious. It seems like you really liked that class and really felt like your team got a lot out of it.

But really, what was it that you feel like you took away from those couple hours of building models?

What you did is...it was to a wide, non-technical audience. Well, a few technical folks. What you did is you gave a little overview and then you had them fire up an IDE, open up some things in Python, access some data of...I forget, what were they? Socks? What were the images?

Oh yeah. FashionMNIST, for those in the audience.

That's right.

You gave them a very straightforward framework, but you had them play around with slightly different approaches. You gave them the opportunity to see the results, and you gave them the opportunity to play with different parameters. And then you introduced a few curve balls.

It was actually a very straightforward exercise, but it was curated, and it was accessible to a wide range of folks.

What was interesting about it was that for the first time, rather than thinking about the grand vision of machine learning, you had a wide range of folks thinking about it from a very concrete...sort of the way that a developer would, right?

Where you're actually dealing with data, and you're thinking, "What does this actually do?" And you're thinking, "Oh my goodness, it's totally broken. But, by the way, I could also just apply this to 50,000 images instantly."

Which is an amazing feeling for someone. And is a different feeling that you get from building software, right?

I think that that intuition...I'm kind of convinced that you could teach this to Nancy Pelosi, and she'd learn something, and she'd make better policy decisions as a result of that.

I'm kind of convinced that if you...we've done a slight variation of this with a couple other executives and it worked really well. At least to me, it feels like that shift in mindset — and also just like the little bit of finger-feel — meant that folks just had better intuition. I think it made a huge difference.

And then they also ask better questions.

Why signals of successful startups change over time

One thing that always surprises me about VCs — because so many come from a quantitative background, and I feel like there's so many investments being made — is the lack of rigor in the decision-making processes, as far as I can see.

I'm curious. At Bloomberg Beta, do you use any machine learning or any kind of...is there any kind of feedback loop where something's successful and then you decide to invest more in that?

Only for top of funnel, only for top of funnel.

In our case, we're seed-stage investors, right? So our process for follow on is very different from a bigger fund.

But I will remind you though, part of the fun of venture is that the game is constantly shifting. If it was exactly the same game, if the business models were exactly the same, then it'd be kind of like everything else. It'd be no fun to be routinized.

Part of the excitement of the job — but also part of the opportunity, and the only reason it exists — is that there are chances for new business models to emerge where the old metrics no longer make sense.

Those sorts of windows come around every so often. To be honest, where there's that kind of uncertainty, where there's either a key technical uncertainty or key business model or market uncertainty, that's where the amazing opportunities come from.

You've been doing venture for quite a while now and have seen a lot of big wins and losses.

Is there anything consistent in terms of what you saw at the beginning of a successful company? Or does the venture market sort of adapt to whatever that is, and the opportunity goes away?

I'm sure you reflect on this because it's kind of your main job.

Are there any kind of common threads that you see in the successful businesses that you've backed?

Inevitably there are arbitragers that exist, or there are ways to tell signal from noise. But because the market is clever, and you're dealing with founders who are really smart and care a lot about what they're doing, you're going to end up seeing...they'll end up imitating those signals of success, right?

There's a little bit of this constantly shifting game where you're looking for a new signal to say that, "This thing means that these guys are high quality," or "This insight is really important."

Then they'll figure out, "You know what I should do? I should make sure I game Hacker News, and then I'll get all my buddies to go on Hacker News, and then we'll coordinate." And that'll no longer be a signal, right? Or, "You know, what I really should do is I should make sure that all my friends are consistently starring my open source project on GitHub."

Just meaning that once you figure it out then...this goes back to why I think these sort of dynamic models are so much fun, right? That's the whole point of it. You march on and think, "Okay, what's another signal of success?"

I'm curious. At this moment, if I showed up and I was pitching an ML company and my customers were maybe the less tech-forward enterprises — I feel like I probably shouldn't name names because some of them are Weights & Biases customers — but if my customer base was, you know, Proctor & Gamble and GE, would that be more appealing to you than if my customer base looked like Airbnb and Facebook?

How would you compare those two? Is one obviously better?

I do think it entirely depends on the nature of the product and the nature of the solution.

The way that I think about it is that there's like a gradient of admiration. And in different types of markets, different people are higher up. Imagine that map, right? The higher up in terms of admiration...

In some places, in some markets, in some set of developer tools, then actually it does matter a lot whether or not the early adopters come from the tech-forward, or from Facebook, or whatever.

But in plenty of markets — and increasingly as machine learning gets mainstreamed — the questions will all be around business benefit. And then the question is, "Who are the companies that other people admire, or look up to, or aspire to become in those specific markets?"

I think that's part of the shifting nature of the game.

I see. Is the gradient of admiration always clear to you?

I mean, you could...Okay. The secret fun part of the game is when you figure out what that gradient looks like before everyone else does, and then you play with people who are higher up there.

You figure out, "Yeah, everyone's going to admire the data scientists at Netflix," you know, whenever that was true. And then you play with them and then you come up with much better insights. Or you know, when it was true about whatever organization.

It's not that complicated to think about, right? You just ask people, "Who do you like?" or "Who do you look to?" And I think that constantly shifts.

ML and the emergence of new business models

One of the things that we were talking about, that I thought was intriguing, was you mentioned that businesses focused on ML — even if they're not selling ML, but using ML for applications in different industries — you expect them to have a different business model potentially.

My thought is that the business model would match the market that they're selling into, but you felt differently. I'm curious to hear your thesis on that.

So, I'm a VC. I'm only right occasionally and I believe most things provisionally, right? But I'm pretty sure about this one.

I'm pretty sure that we underestimate the effect of technical architectures on emerging business models.

If you were to go back to like Sabre — which IBM builds for American Airlines, right? When they have a bunch of mainframes — in some ways that business model, which is "We'll charge a bunch of money to do custom development for you," really comes partly out of the technical architecture of the way that mainframes were centralized in some other place.

And the moment that PCs come around — or they start to emerge — there's a way in which we think about maybe the best business model ever, which is to say, the one that Bill Gates creates. You know, which you charge money for the same copy of software over and over again.

It's an incredible business model.

That partly rises because Bill Gates and Microsoft and a bunch of folks were stubborn and clever and pushed through an idea. But part of it was also because there was a shift in the technical architecture, that you ended up with a bunch of PCs.

And so then a different business model...because there are different economic characteristics of how that technical architecture is both rolled out and how it's developed and how you get value, then some different business model might make sense.

Then you see the same thing for the web, right?

When you have a ubiquitous client in 1995, I think everyone realizes that that means something new, and it takes five or six years before people come up with the right way to talk about it. But subscription software really only makes sense, and only works, in a world where you have a ubiquitous client that anyone can access from anywhere.

Which is sort of a shocking idea, right? You compare it to like delivering CDs, before. Or before that, someone getting a printout of some code that they were supposed to re-type in.

In each one of those cases, it's enabled...and there's some new dominant business model that comes about because the technical architecture shifts. Of course, that only enables it. It's really the people who build the thing, and market it, and sell it, and come up with the new dominant business model. They still have to do that.

But it just strikes me that the shift that we're going through right now around machine learning, or data-centric applications, or this change in collective intelligence; however you want to talk about it, the nature of building those applications is different enough, and the technical architecture is different enough that there should be some other business arrangement that ends up becoming the better one for both consumers and for some due-dominant customer.

You think about how on the machine learning, model-building side, like you just think about the amount of data you're trying to own and control and understand and manage. And you think about how that changes what's a scarce resource.

It just strikes me that there's something there. So to be honest, I'm constantly looking.

In my mind, what's my grand dream? My grand dream is to meet that person who's working inside one of the big companies, who's been frustrated because she's understood how the grain of the wood of machine learning lends itself to some new business. And her boss's boss is like, "That's stupid. We need to maintain our margins," or whatever.

Solving it, that's the grand dream. That I'll find that person, and be able to invest, and partner with them for a number of years.

In your imagination, are there ways that that model could look? As opposed to...it's a little bit hard to imagine these new things, but, you know, subscriptions have been around for a while.

Do you imagine a move to more of a usage-based pricing, or maybe companies that are willing to pay for your data and combine the data. I'm trying to picture what this could be.

Let me describe something. I led a little conference chat the other day, a little session about this. Anywhere I go, I try to lead a session on this because I'm kind of obsessed.

Certainly, usage-based is quite good and interesting, but I would just contend that in some ways, usage-based sometimes puts me as a vendor at odds with my client. Because I just kinda want you to do more of the thing, right?

Sometimes it's not really useful because...I don't want to name names, but we are certainly in a world right now where people are wasting a lot of money — either on compute or storage without clear business value — and then they're going to some day actually figure it out. And then cause a lot of trouble, right?

I think that that's the pro and con of usage-based.

There's certainly some notions around data co-ops, where the realization that as these models get better, when our...we share our data, maybe we share upside together. I think there are a bunch of folks who are trying variations of that.

The dream of course, always is to be in perfect alignment with your customer. One way that happens is you have something like a value added tax or a VIG, where you benefit when they benefit.

But right now — in the world that we live in — understanding that benefit is so hard, right? Because it requires an enormous amount of infrastructure, and management layers, and AB testing, and blah blah.

Just think about all the problems, all the reasons why it's never worked. Maybe someone will figure that out. Maybe all the objections that we've had for the last X years around why this sort of benefit-driven business model doesn't work. Maybe it'll work with some twister turn of how we think about machine learning models.

You had me convinced many years ago a competitor would come along to Salesforce, that would aggregate the data, and use it in smart ways. And Salesforce has this inherent disadvantage because they're so careful about keeping everybody's data separate and not building models on top of it.

Do you still believe that's coming or do you think there was some wrong assumption that you're making? Or has it happened quietly and I haven't noticed it?

No, it hasn't happened yet.

I mean, Salesforce is this enduring great business, right? That's going to last for decades and decades. That said, it still does strike me that there's an inherent tension.

You think about all the trouble that they spent convincing me — or convincing people like me — to work with them, because we believed that the data was safe in their cloud. And then just the idea that I might share data with other clients is crazy and terrible, at least from that point of view.

So there's that inherent tension in the traditional, or the now established, SaaS view of the world.

I think it's very hard for the incumbents then to move off of that sort of way of thinking about the world. But harder yet is convincing their clients and their customers who've been trained to think that way, right?

There's a funny — maybe not funny — story, where Microsoft got in a lot of trouble at some point for sending information back to their main servers about how PCs were doing. They would crash, or there'd be some bug report, then they'd automatically send it back.

That was a huge scandal, because "How could Microsoft be looking and stealing all my information?" The hilarious thing...not hilarious to Microsoft, but the hilarious thing about that is, that's right as Google Docs is starting.

In the case of Google Docs, Google literally sees every single thing I type. I mean, literally stored on their servers. And somehow, because it's a different configuration or different expectations around the business, I'm okay about it.

I think something similar will happen with some emerging sets of machine learning-driven businesses.

It's interesting that you say that.

You had a really interesting viral tweet at one point showing how much better Google's transcription was than Apple's. Which I thought was really interesting, and actually made me think about the same point. Apple is so known for being careful with privacy and Google is known for being much more laissez faire with people's data.

But it's not clear to me that Google has used that perspective to create a huge advantage. At least in terms of market cap. Do you think over time Google's point of view will really serve it, or has something changed?

I think that in that case, it's a little bit of a slightly different nuanced thing, right?

I mean, why was that Pixel 6 voice recorder so much better? It was better in part because they had an on-device model, that was one part. And another part of it is that they just collected data with much more thoughtful ways.

What did that mean? That meant you had a very fast, very accurate local experience.

The fact that that's true...that's definitely true, but it's also confounded with the fact that Google is a very large organization right now, and they've got lots of things they worry about and lots of ways that they're unwilling to take risk.

In my ideal world, someone who built the sort of technology that Google did around voice would have decided that, "Oh, you know what? aAtually, this should be part of some SDK or some API, and we should just make this available for everyone." And developers should be building a bunch of products.

That's the other thing that I think we're on the cusp of, because we're just at this point where there's this massive investment in infrastructure, and research, and tooling around machine learning. And we're right at the point where maybe people will build products that are actually good, right?

We're just at the point where the lessons learned around how human-in-the-loop works, the lessons learned around experiences on user interface; all those things, they don't quite take...or value-added to the end user. We're just at the point where there'll be enough variation that some ideas will actually take hold.

So I'm sort of excited about that part too.

Are you starting to see that? Because I feel like maybe I'm too impatient, but I can't believe how much better all aspects of NLP have gotten in the last few years. I feel like transcription is now solid, translation now works.

I mean, it basically works. You can communicate for sure with people that you don't speak the same language with by using a translation system. Hugging Face and OpenAI's GPT-3 have incredible demos.

And yet I don't feel like it's impacting my life that much, except for asking Alexa to play me music.

You're exactly right. We're at the point right now where I'm hoping that your listeners are building products because now it's easier to access it.

You know, there's this talk about democratization of machine learning. We talk about this often, I feel like. But I think it kind of misses the point.

The point is that by making this more broadly available, it also means that the extraordinary person on the edge — who might not have had access to try this before, the person with the crazy idea that will make a huge difference once we actually see it — that they can start working as well.

That's part of the exciting thing that I think everyone misses as they talk about the way that this whole world is shifting. But you're exactly right. That we should be deeply dissatisfied with...

On the one hand, all the progress that's made voice and parts of NLP, we should be super impressed with it. And we should be deeply dissatisfied because the products, and the product minds, and the UI folks, and the business minds have not yet figured out how to take advantage of those advances in ways that actually make sense, and go with the grain of the technology.

What James considers when investing

One thing that I would imagine being hard as an early stage investor investing in machine learning is that it's so easy to demo successful cases of machine learning.

I feel like no other field is it quite as easy to make a compelling demo. And yet it feels like to make a working product, it's often going from like...bringing the error rate down from 1% to 0.1%, or something like that.

Do you have trouble evaluating?

Here's my secret [?]. I'll give you one of my current secrets.

Okay. Tell me.

I just assume it doesn't get better.

If the application requires the thing to go from 95 to 98, or 98 to 99...what if it doesn't get better? Will users still get value out of it? If users still get value out of it — because of the way they configure the problem — then it's an interesting product.

But if you're sitting there thinking, "It'll just be another month before we go from 98 to 99.5," then I'm like, "Well, I don't really know if I believe that."

This goes back to one of our earliest conversations around search quality. This is like many, many years ago.

What's the beauty of search? The beauty of search is that when it's wrong, I'm okay about it. There are whole sets of products in which you can take advantage of the fact that it's super fast, it's consistent, and when it's wrong, I'm okay about it.

You do that over and over again, or you find the products that do that; then those are interesting applications.

For an investor, you're doing an extraordinary job of not bragging about your portfolio, but give me some glimpse of the future. What's the exciting stuff that you're seeing lately?

There are two parts that I want to talk...that I sort of want to highlight.

On the ML infrastructure piece, I still think that there are analogies or lessons to be learned from traditional software development.

I think that you guys have done such a good job of understanding so many pieces out of that, but I still think that...you think about QA, like figuring out how to consistently do QA.

I think there are lots of lessons to be learned from normal software development to be applied to computer vision and structured data, and those sorts of release processes. There's a company called Kolena that's in the middle of figuring out parts of that.

You look at companies like...we talk to Sean every so often. You look at the demo, like the publicly available stuff about Primer. And just imagine what they're actually doing under the hood.

If you go to primer.ai and you look at their ability to synthesize huge amounts of data and lots of articles, and just make sense of the world; and imagine applying that...in their case to a bunch of national security use case. If you look up various things that are happening in the world right now and the word "primer", you'll see these demos. They can't show you what they're actually doing, but you get that sense of, "Oh, this is changing the way that people are actually doing things right now."

That's the sort of thing that I feel like on the application layer, but then also in the development part we're just sitting on right now.

Going back to my secret [?] — which is I just sort of assume it's not necessarily going to get that much better — there's this great guy, Michael Kohen, at this company called SparkAI.

Their big insight is similar to that line. They're like, "Look, we want autonomous vehicles and we want them to be perfect. But they're not going to be perfect for a long time. So let's just make sure there's a human in the loop."

You can think of them as like...whenever the machine is uncertain about something right in front of them, they'll get a response in a pretty short SLA to make a decision. And thus you can actually roll out these sort of real world applications with the realization that the model doesn't have to be perfect. That we can actually have backup systems.

I think that sort of perspective — assuming the sort of non-utopian view of what's possible with machine learning — is super exciting to me.

Ethical considerations of accepting that ML models are fallible

I'm curious what you think about — and I guess this is a broad question — ethical implications of machine learning.

Many people talk about machine learning and ethics and I feel like there's constantly in the news issues that come up with machine learning.

What do you make of it? Do you feel like there's special ethical considerations unique to machine learning — different than technology — or not?

How do you think about what kind of world you want and what regulations make sense?

I think it's a good thing that we live in a world where people are more sensitized. On the one hand. I'm very glad to see lots of people applying their minds towards it, on the one hand.

On the other...so, this might slightly get me in trouble. There's a game that I play with friends of mine who are ethicists, who are thinking about the effects of technology.

I ask..I think it's appropriate to ask these questions around what are the implications with this or that. If you were around in like 1950 and someone proposed the compiler to you; for the first time, someone said, "We've got this really, really great way of making software easier to develop, and available at mass scale, and et cetera, et cetera."

Would you have allowed me to build a compiler?

Just imagine all the harm that could come from a compiler and imagine, to be honest, all the harm that has actually come from compilers. Everything from hacking to stealing money from people, et cetera.

There's a way in which I think there's a reasonable argument that we wouldn't...given some current frameworks, there's an argument for why we should not have had a compiler.

Which seems, on the face of it, at least to me, crazy. Right? Absurd.

To me, the questions instead should...there should be this sensitivity, and there should be these sets of questions, but in some ways the questions should all be around, "How do we think about what do we do if we're wrong?"

I think one of the beauties of machine learning is that embedded in machine learning — at the very core of machine learning — is this idea that these are not fixed heuristics or business rules. Actually, these are guesses that we just have to assume will be wrong sometimes, right?

In that way, once you think from that framework, or once your executives understand that's how models actually work — that they're wrong, they're never going to be perfect. Otherwise you can have a big "if then" statement — once you realize that they can be wrong, then you need to build the systems and the processes to deal with the fact that they could be wrong.

You also need to build a whole set of ethics and ways of thinking about...questions more like "responsibility" rather than "possibilities". And I think that shift in the way you might think about machine learning, I think it will be much more profitable, in the sense of being useful for humanity.

What do you think?

I guess it does feel like machine learning might not be as neutral as compilers in some cases. If you imagine it taking inherent biases that we have in our society and then encoding them in a very efficient system, so that they can be deployed at bigger scale and with possibly less oversight.

Right? That's only if you fall for the idea that we're trying to build an all-knowing God brain that will solve things for us perfectly.

To be honest, oftentimes when you'll talk to executives, that's how they'll think about machine learning. They'll think, "If only we can get this perfect, then we can rely on it forever."

But instead, if we thought about it as a bureaucracy that is right some of the time, but wrong too. If we thought about it as a possibly fallible system, and we built in the support for that...because the nice thing about machine learning is that it's incredibly cheap. In the grand scheme of things, it's incredibly cheap to make these judgements. And also it's centralized. Right?

By being centralized, and being cheap and conscientious — meaning it's consistent — then you actually have one place where you can go and you can always say, "We fix it here, we can fix it everywhere."

That's one part of it. I think the other part that you highlighted — which is it captures inherent biases — that's the other part. Which is, in some ways it's a problem with the way that we anthropomorphize machine learning.

One way to think about it is this amazing genius thing. On the other hand, you could just think of it as an incredibly conservative attempt to cluster collective intelligence.

If we understood that machine learning was derived from data — and data is by nature historical, and anything historical by nature happened in the past — then I think that changes a little bit your expectations about what the model could do.

And then it changes your expectations around what layers you need to put on top of it. Because you can't just rely on the model. You're going to have to have both straightforward business rules to protect yourself, but also you have human processes that will actually think it through.

I do have to, at this point, make the plug for one of my favorite papers, which is called "Street Level Algorithms," which talks a little bit about this idea.

You'll have to link to it, I don't know if you've...have you read it?

No, no.

I think I've tried to make you read it many times. It's totally worth reading. You should get Ali or Michael Bernstein to chat about it at some point.

But I think their core insight is that if you did think about machine learning models as bureaucracies, or as processes that could be wrong some of the time, that you change your expectations. But also the ways that you can take advantage of machine learning, which is to say, "You fix it on one place, you fix it for everyone," right?

Those sorts of inherent advantages go with the grain of the technologies rather than against it.

Reflecting on past investment decisions

Have you ever gotten a pitch on the company and not invested because it made you uncomfortable? Like from an ethical perspective?

Oh yeah. Plenty of times. And I think-

-really, plenty of times?

There are plenty of times when I will say...on the one hand, I'm utility maximizing, but then I have my own idiosyncratic definition of utility. And my definition of utility doesn't map directly to just dollars, but maps into ideas of who I am and what kind of person I want to be and what kind of world I want to be in.

I think that that's true about all VCs, right? Everyone pretends that they're...rather, a lot of people pretend that they're pretty straightforward in dollar maximizing, but that's not true.

We all have tastes and we all have things that we like, or don't like, and good or bad reasons to say yes or no to things. And I think that reality is always sitting with us.

Is there a company that you feel like you've massively misjudged? Is there any wildly successful business where you'd go back, and think about the pitch, and feel like you missed something or should update your belief system?

Constantly. You know, the whole set of low-code no-code companies that I sort of dismissed.

I don't know if you remember this conversation. There's some point when we chatted, where I basically said that, "You know what I really believe in? I believe in domain-specific languages. I think that DSLs are a much more powerful way to express business applications and the possibility for business applications, than all these low-code no-code things.""

I was totally wrong. I entirely misjudged the the value add of making something easy and the way...in part of my head, I was like, "Well, a developer's valuable not just because they can write things in good syntax, they're also valuable because they have to think through complicated ideas, abstract them, and come up with a good code to actually build something, to get something to work."

What I misjudged was that there's a whole set of low-level GLU things that people need every day, that are super easy to do, that sort of fall right under the cusp of really scary programming. in. So, that I totally misjudged.

Thoughts on consciousness and Theseus' paradox

One topic that we've actually never talked about — but I kind of wanted to use this podcast as an excuse to ask you — is, I'm curious what you think about AI and consciousness.

Can you picture AI becoming conscious? Is that something that you think you could imagine happening in your children's lifetimes?

What does that mean?

Could you imagine that there's an ML system that gets to the point where you would not want to hurt it? Where you cared about its wellbeing.

There are a couple of different angles that I go on with this.

I think that's true right now. I feel bad when I do lots of things to anthropomorphized...I feel kind of bad when I drop my phone. I feel really guilty and I feel kind of bad about it

For your phone?

For my phone, yeah.

I think there are lots of ways that I, as a human, sort of assume human-like characteristics to almost everything, right from the weather to my camera, to the screen, to some computer program.

I get irritated. Why do I get irritated with Chrome as if it's an actual person? It's just a bundle of numbers, right?

I actually think that we're there already. I actually don't think that my willingness to imbue moral worth or value to non-human things is something that's out there someday, but actually something that we do all the time right now.

And then, although I am Christian — which we've talked about before — I don't really take a magical point of view on consciousness.

I think consciousness is controlling what I pay attention to, and the continuing [?]?

I mean, I both value it. I think it's really important. And it's an incredibly important organizing principle, obviously, for me day-to-day.

I kind of think that lots of things are conscious already. That they already figure out ways to direct attention, and organize, and also tell stories about themselves.

Does your Christianity not inform your thoughts about consciousness at all?

It totally does. But I think there's a little bit of this angle where I think that the things we learn about the world or science constantly shift, and so I'm actually quite open and willing to sort of adopt and adjust based on how we end up changing our view of the universe.

I don't know, does that make sense?

Yeah, totally.

Is that like a coherent-

-but I guess this is the thing that always makes it concrete for me — that I was telling you I had to ask you, and I don't know how you felt about it — but I always am curious if people would go through that Star Trek transporter.

If you saw a whole bunch of people go through a thing that disassembled their atoms, and put them back together somewhere else safely, and you're convinced that it would work, would you subject yourself to that?

Would that alarm you, or not?

I have contradictory impulses. I get carsick, and I get woozy standing up...walking over a bridge. So I'm sure there'll be that trepidation.

But isn't there also this view? When you think about yourself right now versus yourself 10 years ago, a bunch of the atoms have changed, have been replaced, right?

In some ways, we are going through this slow motion transportation. In some ways you're just speeding up that transformation, of the rearrangement of those bits.

I probably wouldn't be the first person to do it. But, you know-

-you'd be like the hundredth?

Meaning that I would not necessarily have some deep, ethical, mystical reason to be concerned about it. Because I kind of think we're going through it already.

Your set of atoms...are you your set of atoms, or are you that pattern that your atoms are in?

In some ways, you're the pattern.

Interesting. I'm not Christian, but that transporter I think makes me more nervous than it makes you.

But isn't it true though that you...if you thought about you current material composition right now, the literal pieces of it have changed pretty substantially and will continue to change. Right?

For sure.

I just gave you my most tech-positive version of it, but sure, you're asking me tomorrow if I would do it. I think, "Well, a little scary, let's find out."

But don't you also believe that you're your pattern, rather than your actual...like, who you are is the organization of these things inside you, rather than the actual substance of it.

That's true, but I feel like I'm going to experience the world through somebody's eyes, and I think I am concerned that my future self might not be inhabiting the body of the person that comes out of that machine. But my wife strongly disagrees with my point of view on that.

I can see both sides of it. I'm just pretty sure that I just wouldn't do it, no matter how many people went through it and told me that it was safe.

You say that now, but I will just remind you that our ability to adapt to circumstances and to change expectations is pretty dramatic. There are plenty of things you do now that are super...like, it would be super weird to you from 1999 or whatever.

You're really young, too, but you know what I mean. Our expectations around what's normal or not normal shift consistently.

Like staring at a phone all day long.

Yeah, seriously.

Why it's important to increase general ML literacy

All right. Well, final two questions.

One question is, what's an aspect of machine learning that you think is underrated or underappreciated or under-invested in?

I do think all of the HCI social system stuff really is under-invested in. And I think that there are lots and lots of opportunities.

It's interesting to me that the tools that annotators get right now are still so bad. It's interesting to me that the tools that data scientists use in some ways have not really changed since...remember your friend Cahir who wrote that paper, like 2013? Look at his paper in 2013, it's like the tools in some ways have not changed enough, right? So I think there's lots and lots of opportunities there.

And I think there are lots of opportunities in making mainstream or more generalized...to generalize from the lessons we learned from human-in-the-loop. I think calling things human-in-the-loop kind of was a mistake. There should be a better name for it. And if we had a better name for it, then everyone would think of all their jobs as human-in-the-loop.

Because I kind of believe that. I kind of believe that in the end, if we're successful, every process will be slightly better understood, and could be consistent, and get consistently better. Because our job as humans were to either figure out edge cases or create broad clustering so that we can be consistent.

So you care about the interface of humans and machine learning, how they can work together?

I think that at multiple levels, at the level of the person making the initial decision, at the level of the person learning from that, at the level of the people controlling that, at the level of people benefiting from that; I think all those things...we're still in a world where so much of that is siloed. The way to think about it is siloed.

And I think the ways to unlock lots of business value — but also to be honest, just straightforward, good things for humanity — is if people had, at all levels of that game, a bigger view of what it is that they're engaged in. Which is sort of a great game of collective intelligence.

All right. Practical question — which might actually have the same answer. It's never happened before as I've asked these pairs of questions — but when you look at machine learning trying to get adopted and deployed and useful inside of enterprises, where do you think the bottleneck is?

Where do these projects get stuck?

I think they're so often badly conceived and over-promised.

You know, we joked about this in the middle of this; I am still kind of convinced that if we offered your exec ad class to every senior executive in the world, that we would basically all make much better decisions, and we'd end up with much more successful implementations.

So I think that that part is definitely true. And I also think that the other thing that's holding us back is we still don't have great methodologies for thinking about how to build these systems.

That we are still...in software development world — someone just gave me this history — random coding becomes engineering when NATO decides that it's an important thing in like 1968. And then we codified all this waterfall stuff, right?

It goes from waterfall to extreme to agile over the course of the last 40 years. And what's interesting to me is that that methodology, I think, is mostly wrong for building machine learning models. We are still shoehorning these projects as if they're software development projects oftentimes, and thus wasting a bunch of time and money.

Outro

Awesome. Thanks James.

Okay. Take care.

If you're enjoying these interviews and you want to learn more, please click on the link to the show notes in the description where you can find links to all the papers that are mentioned, supplemental material, and a transcription that we work really hard to produce. So, check it out.

Bonus: How James' faith informs his thoughts on ML

So James, here's what I really want to know. How does your religion inform your thoughts on machine learning?

This might be both borderline kookie and heretical. We'll just caveat it first that way.

Fantastic.

I think that there are a few different angles.

I think the first is that, at least in my theology, part of godliness is the act of creation. And I think that there's a way in which, as an investor, I put faith in the act of creation in helping people make something new.

So that's one part.

And the creation of however you want to talk about machine learning, I think there's this sense in which the models that we're building in some ways have inherent worth and dignity, as basically sub-creations of people, right?

That we are creating something new, and — whether you want to call it life or whatever you want to call that thing — that it is something fundamentally new and different and interesting.

That piece of it then informs the way I think about both its capabilities and why it's important, but at the same time — this is the part where I think other folks might have trouble with this — I do believe that we're fallen.

I believe that we...I actually think that we want to be good. But we're actually bad. And I think that anything we create in some ways has tragic flaws in it, almost no matter what.

In that way, I'm actually much more both forgiving about people, but also institutions, but also the models that we make, right? These things that we're making have great beauty and potential, but they're also tragically flawed because we are.

I love it. Awesome.

Oh man, that's definitely going on the podcast. That was great.

It's kind of plausible, right? It's not crazy.

I agree with all of it. Yeah, totally

I think we oftentimes all think we're good. I mean, we think we're good, but we actually kno...it's not that I'm good. It's that I want to be good. And I'm just always doing stupid things

Of course the things I created are going to be imperfect. That means that there...it also means there's this constant chance for improvement, which is the core of the understanding of gradients.