Managing Industry Research Teams, with Fernando Pereira

Guest: Fernando Pereira
Hosts: Matt Gardner, Waleed Ammar

We invited Fernando Pereira, a VP and Distinguished Engineer at Google, where he leads NLU and ML research, to talk about managing NLP research teams in industry. Topics we discussed include prioritizing research against product development and effective collaboration with product teams, dealing with potential research interest mismatch between individuals and the company, managing publications, hiring new researchers, and diversity and inclusion.

Matt Gardner
00:00

Welcome to the NLP highlights podcast where we talk about interesting work in natural language processing. The hosts are Matt Gardner, Waleed Ammar and Pradeep Dasigi.

Waleed Ammar
00:09

Okay. Today our guest is Fernando Pereira. Fernando is a distinguished engineer and vice-president heading NLP research among other areas at Google research. Welcome to the program. Fernando.

Fernando Pereira
00:20

Thank you for having me. It’s kinda interesting to have this conversation. We had planned it for quite a while and then schedules interfered, but and now we have to do it remotely whether we want it or not. So it’s a new world for all of us and try to make sure that our audio equipment works reasonably for so that we can do our work and interact with each other professionally as before. So it’s a, yeah, it’s a new normal at least for awhile.

Waleed Ammar
00:53

Yeah. It is surprising how much work we can get done even with all these constraints.

Fernando Pereira
00:57

Yeah. Yes, that’s right. I’ve been in a video conferences since 7:00 AM. One of the things I’ve been doing since we have had our restrictions for working at home and so on. Is is having these virtual coffee times with members of my team across the world. So I have a couple of those at 7:00 AM and then I’ll have others at 11:00 AM during the workweek so that we can just chat about, you know, what’s going on, you know, how we are feeling about this situation, any new interesting ideas, project. So it’s just a very informal thing that we’ve been doing, which has been a lot of fun actually. And I’ll continue it after this is over cause our team being distributed all over the world makes it a really difficult to keep in touch with the, you know, the diverse team that we have.

Waleed Ammar
01:50

Totally. Well speaking of the size of your team, my first question for you is, you’ve been at Google for 12 years now. How did the research organization evolve over these years?

Fernando Pereira
02:04

Yeah. So I came to Google a bit over 12 years ago and my hypothesis when I considered coming over was natural language technology, NLP technology had been making quite a bit of progress through the use of statistical methods and machine learning. And this was the time to try to see how it could help on a user facing role in a big way, which is particularly in search, how can we understand what users are looking for better, what documents contain better so that we could provide better search results. So that was the ambition. It was a bit crazy to think about it because the techniques we had 12-13 years ago were very primitive compared to with what we have today. But yet I thought that it was a time to try to do this now at the time search was essentially done through information retrieval techniques very successfully, but basically not relying so much on the actual structure of a query say, but mainly on the keywords on the unigrams bigrams in that query and different ranking functions that took into account frequency, inverse document frequency.

Fernando Pereira
03:19

All these usual things that people in IR had developed very successfully. And so it took quite a while for us to figure out where we could have initial impact by using the state-of-the-art in natural language processing in particular in things like parts of speech tagging and dependency parsing to try to get a bit more of the structure of queries and how they relate to document. So there were a bunch of experiments and false starts but we got enough traction in first few years that it gave us confidence that we could continue to build the effort and it gave confidence to those that provided the resources to continue to grow the teams in. And we continued to do more interesting work over time that has had very large impact on, you know, on search and other products.

Fernando Pereira
04:15

And this is the natural language side. I’ve managed a number of large scale machine learning projects over the years and a variety of other things. But I think we should focus here on the natural language side. So the difference, I mean the big difference is that we started with a really small team and we have a much, much larger, much more distributed team now working on the wider much wider range of questions. You know, like with everybody, the development of deep learning methods using either the availability of the computation has made a big difference to how fast we can move and how much impact we can have. So, you know, that’s kind of the story of our progress from making certain, you know, small improvements. For instance in query processing or multilingual query process, you know processing in particular international, you know, ways to internationalize different systems to being able to deliver major changes in how search does ranking and question answering that, you know, so it’s going start from incremental valuable improvements to pretty radical changes of how the whole machinery operates.

Waleed Ammar
05:37

Would you say that the ability for the research work to grow within Google had something had more to do with how successful it has been in impacting the products or in like actually attracting the right talent?

Fernando Pereira
05:52

It’s a combination. I think this is basically a kind of feedback, positive feedback, right? So if you are successful in certain areas, then naturally there’s more ability to reach, to have the kinds of conversations about resources that lead to the potential to create, for instance, longer term research efforts and to hire researchers who may be looking at working on longer term problems and so have built on the foundation that is to some extent apply, but still trying to, you know, develop the state-of-the-art to one where we have fairly substantial efforts to really do longer term work that is not directly tied to product, but is still always motivated. It’s always motivated by questions like; Hey, why am I at Google? Well, I am at Google because I care about understanding what users want to accomplish through say a search query.

Fernando Pereira
07:06

Because I see that as a sort of the ultimate demonstration of language understanding is when we can actually answer correctly, accurately, in a timely manner. The variety information needs of users across hundreds of different languages. That’s when we’ll know we have sort of arrived. We are not there yet by any means. So that big goal then connects to all of the research that we do both, you know, it connects effectively to a lot of the research that people do outside Google and that’s why we can have a really healthy relationship between our work and the work going on in the academic world.

Waleed Ammar
07:47

So you mentioned contributions to product, but also longer term research. So what are the different ways you consider to be kind of like the different kinds of impact that research teams at Google achieve?

Fernando Pereira
08:02

Well, it’s, I think the two main ones. And then there are lots of different details, but at one level can we really rethink how we do major parts of our, the user-facing process we have, whether it is in say advancing machine translation or doing better question answering contributions to those efforts that are significant. Measurable are one way that our impact is measured. On the other hand, creating new new approaches, you know, developing new models, new ways of new algorithms that we publish is another form of impact, ideally the two meet somewhere where when we are super and we are not always super successful, but when we are we have something that’s both a substantial research contribution and something that changes how our products operate for the better.

Waleed Ammar
09:07

Yeah, that makes sense. Do you find often some friction between these two goals? Like for example finding patterns or whether or not to open source a certain project?

Fernando Pereira
09:17

Well, I wouldn’t say conflict. I think we have to be thoughtful about clearly, I mean we are part of a company that is we would like to be successful and continue to develop the best possible product in the long run. So, we have to protect its business interests. But on the other hand, we understand that the conversation with the outside world and in particular researchers, students even other companies is really important. So we try to maximize the communication bandwidth, but we also have to be very careful about certain areas where we are dealing with sort of ethical or privacy matters. So it’s one thing to develop a new say question answering algorithm is a very different in requires a lot more thought to for instance, release that data set like we did with natural questions and that required extreme care to ensure that the data we released, which is based on actual user queries, protects user privacy and is compliant with all the applicable policies. So the degree to which we can be open about what we do depends to some extent to not just the business interests but also ethical and you know, privacy and regulatory requirements.

Waleed Ammar
10:46

Yep. That makes sense. So I would like to move on to another question about like planning research. So, you know, like top down planning sometimes make sense in certain products in research. Do you consider top-down? Like how do you see the difference between top-down planning and bottom-up? How do they go hand in hand?

Fernando Pereira
11:07

Well, so, I think that this goes back to the question, why are we at Google, right? Or we are here because I think ultimately you want to care about the mission of the place you are in. If you are in academia, you care about the education, you should care about educational mission, you should care about training students, about creating knowledge and expanding it. If you are at Google, you should also care about the fundamental mission, which is organize the world’s information, make it accessible and useful to everybody. So if you care about that mission, that’s the main top-down drive we have, right? It’s, think about that mission. Think about the problems that brings up, for instance, problems in language, understanding for user queries, for document understanding, problems of modeling trust in information. So not all sources of information are equally attested.

Fernando Pereira
12:08

Has equal support. In a situation like the present situation with COVID-19. Clearly it’s extremely important that we give people the best possible health information. All of those questions are extremely hard, natural language understanding matters, but they are also scientifically extremely interesting. So there is no contradiction there. And if most of our work, our research work is driven by those high level goals, then even though people might have very diverse ways of approaching the problems, we’ll have enough options to have impact. Some of those ideas will pay off and they will become ideas we can engineer into our products. And and that I think is the, sort of the ideal state of our research enterprise is to have these very high level goals and then this way of turning the individual creativity of researchers and connecting it to this goal. So that’s my job. You know, my job and the job with my leadership team is to try to draw those connections, identify the opportunities, help researchers find the path to success within this very broad framing organizational goals that are tied to the Google’s mission.

Waleed Ammar
13:29

Yeah, I mean that makes us that this has been my experience at Google so far. So I would like to switch to, instead of like focusing on Google research, let’s pretend that you have a friend who’s starting a successful tech company just finished raising $30 million in a series A fund, something like Cora or Stack Overflow a few years ago, just for a frame of reference. Now this friend, her employees have been trying to catch up with research and implement the relevant methods, but she’s starting to think about growing research team to help ensure the longterm success of the company. So what criteria would you use to assess if the company is ready to have, to invest in a research team?

Fernando Pereira
14:15

So actually I have done this in a specific cases. I have friends who have started companies and I have advised them in this specific way. I think the advice I give is think of the problems you’re trying to solve. Don’t think about, Oh, should I have a research team or not? You should think about what problem do I need to solve that my current engineering team is struggling with and what is the kind of expertise that is most likely to help us advance that or find alternatives to that. You know, maybe not necessarily to solve that particular problem and find ways around that problem. Do I need someone who has a different expertise say in multilingual language processing or a different expertise in applied machine learning? That’s I think the way to approach this. The approach is not, Oh I need a research team.

Fernando Pereira
15:14

Let’s go and find some researchers that happen to be available. You have to find people who are motivated to come and solve the problems that you have. They have to be motivated by the mission of the company, by the vision of where the company is trying to go and have the right experience, the right mindset to come and help you solve those problems. And it could be that the solutions they’ll come up with are not necessarily cutting edge state-of-the-art cause very often the best thing a researcher can do for an engineering and product team is to know very well the solutions of 10 years ago that are very practical now. This is actually one challenge of our current AI education in NLP education is people are so fixated on the last three years of research and the state-of the-art race that often they don’t know much of what was done in the past that could be very relevant.

Fernando Pereira
16:16

One example I’ll give you is that there are many problems that people face that have to do essentially estimating uncertainty and yet most of the students that we see at an interview that have not done the graphical models, a graduate class or any kind of, they don’t actually have a strong probability background, which 10 years ago of course everybody did. So in a situation like this you might say, Oh, what I really need, let me give you an example. Suppose that I am a Cora type company or Stack Overflow and I need to have some crowd-sourced data and I need to deal with inter annotator agreement and the model, the process of annotation and the annotated biases and so on. How do I do that? I need to actually go get someone who understand probabilistic models and think of maybe think of the problem as a latent variable model where there is sort of a true label that these people are approximating and that each individual has some false positive or false negative rate and which has to be estimated and you will naturally do a problem like that getting to either, you know, if you just estimating getting a point estimate to an EM type algorithm or if you kind of have some Bayesian background, you might do something maybe even a bit better using Bayesian modeling of the situation.

Fernando Pereira
17:46

But that’s what you’d need for that case. Right. But yet if we just say, I will go and hire some fashionable researcher, they might not have the skill to solve the problem that you really care about.

Waleed Ammar
17:58

Yeah. Thank you. That makes a lot of sense. That also answers like who would you want to hire? Like you look for the people who would actually be equipped to answer these questions that you’re struggling with.

Fernando Pereira
18:09

Yeah, that can. I mean I have my interview questions made, right? You know, it’s a, Hey here I have this problem, walk me through approaches to addressing it and you’re not going to get a solution, but you get the thinking process whether those people are senior researchers or junior researchers, we don’t know. I mean you have to focus on the problem you’re trying to solve. Not on this notion of Oh, I need a research team, like a essentially as some kind of a badge of honor or something.

Waleed Ammar
18:43

So now a few years later your friend managed to like hire the right people, make progress. And a couple of researchers she hired kept like the team kept growing until it’s maybe it’s like 20 people, like maybe half of them are researchers, half are engineers working together on solving the problems that matter for the company. But you know like it’s kind of inevitable if you’re hiring researchers that they would be also interested in maintaining, you know, a strong publication record and maybe thinking about problems that are not directly related to what, like it’s not the study of the top priority for the company. I think this often happens in research groups. How do you balance between the interests of researchers and the interests of the company?

Fernando Pereira
19:31

So the first thing to keep in mind is that if this is a problem about, you know, talking about essentially a recruiting problem to start with, which is if this problem becomes a real issue. There’s mismatch between the researchers’ interests and the needs of the company, then you probably did something a little bit wrong in hiring right. Clearly you have to understand that there’s a variety of motivations among people. And you need to mix. And you want to have a mix of talent and a diversity of talent. And some people are going to be a little bit more motivated by one type of incentive and one type of work and others by others. But on average you should have a collection of people in your team that are overall aligned with the direction you need to go in that you believe is the direction of the company.

Fernando Pereira
20:30

So clearly with respect to publication record, I would say there’s a difference between publishing some, continue to publish incremental results. You know, again, another epsilon in the state-of-the- art of set benchmark versus making a big change on how people think about the problem. And what I constantly have this conversations is it’s quality, not quantity. When I was a professor, I had this conversation with my students as well, and when the conference deadline, the ACL deadline was approaching or the ICML deadline, you know, you have everybody trying to push this new result or this paper out. And I often served as a sort of a bit of a filter in the friction say, look, this is not ready. If you want to send this, don’t put my name on it. Basically even if I had contributed and the reason is because I think that the incentives in the current academic or research culture are distorting of not just what’s valuable to a company but towards the valuable to society and science in general. And so actually I think that you think carefully about what’s really valuable. There is less of a conflict between publishing really top work and also achieving high impact results then people think. I think it’s a false opposition.

Waleed Ammar
22:00

Yeah. Thank you for bringing this up. This problem of like how valuable are the research that we publish and like how significant is the result needs to be before we publish it. And the incentives in academia is a really important one. I wonder if you have any thoughts on how to improve the current, you know, the current, how do I say it?

Fernando Pereira
22:23

Well, I don’t know. I was having a conversation earlier today about this. So it’s a very hard mechanism design problem. I have tried to address it locally by having multiple layers of a discussion and, you know, constructive criticism about lines of work and about, I mean in the end this is a distributed problem. I mean there’s not going to be a law that from above that says research should be done this way. Right. I think that in the end, it’s people like me and people, the senior people in my team and certainly people in other, in the different academic roles who have some influence on how we carry out this activity and try to incentivize really transformational work doesn’t mean look, not everything succeeds, right? That’s okay. But I think that you know, running after incremental gains, and that’s true by the way, on the product side as well.

Fernando Pereira
23:28

One of the things I keep telling people often is a quit while you are ahead if you are working on the project, which has had the big success, but now the incremental successes are smaller and smaller because there’s diminishing returns in that particular technique then before you really start slowing down [is] the time to come up with the next thing to do. And that’s true in research. That’s true in product impact as well. The greatest value of a research team is to be able to jump to new ideas that could have much greater leverage than the ideas that you’re working on now. But you can only do that if you are willing to abandon something that’s still yielding results before it’s completely exhausted. So that is, balancing that whether it’s on the science side or in the product impact side is I think the critical part of the certainly the job I do. And I’m not saying that I have a magic recipe, but I say that this is something that I’m always paying attention to. This diminishing returns effect that happens across all areas of science and engineering.

Waleed Ammar
24:45

Yeah, I definitely remember having discussions with Matt and other people at the AI-2 on this topic. And like one consensus I feel is that it’s a lot easier to do this with a more senior person or like someone who finished their degree, maybe have been already like out of their PhD for a few years and they can, this like this is the feeling that as a student, as a PhD student there is more of a rush to publish because you are going to be compared to other people and people who are hiring or looking for like easy ways of evaluating candidates will always be looking for like easy metrics to look at, like number of papers, things like this.

Fernando Pereira
25:29

Okay. So actually looking at your set of questions, right. And, you talk lower towards the end about diversity and inclusion and I think what you just said is extremely deeply connected with our lack of diversity in a lack of inclusion. I think that the sort of bean counting approach to recruiting and to valuing people, this counts the value of people with diverse experiences that have had different career paths that may be were trained in a different discipline. And as a result it limits our ability to find creativity, innovation, alternative perspectives outside our comfort zone, the comfort zone of, you know, just looking, Oh you did everything by the book, you got your N papers in ACL and your N papers in NeurIPS. So you are a good person. I mean, which is I think is extremely destructive to the health and wellbeing of our field.

Fernando Pereira
26:33

Back to when I was a professor at Penn and a department chair, and one of the things that I observed when I started there with respect to diversity in the undergraduate program, is exactly this. Everybody that was allowed into the program tended to be exactly the same kind of person that had had a computer since they were eight year olds. And they tended to be male. They tended to be well off. They tended to be from certain locations in the United States and that was very destructive. You had sort of a monolithic view of what is success in computer science and when I started programs in the department to open doors for people with different backgrounds, different experience to discover whether they actually are as excited about the field that was quite successful. We increased diversity in our program very substantially.

Fernando Pereira
27:30

And the main thing was not to assume that the way that I am is the way that everybody should be. And that’s true about the whole process of hiring in our organizations. I don’t need to just have another machine learning person. For instance, I need to have linguists I have all people that have a strong linguistics background and guess what? They tend to be more diverse than your computer science, your typical computer scientist and so on, so forth. I mean, so bringing from the social sciences, bringing from the life sciences, I mean one of the things that we have been trying to do with our AI resident program is to bring people from different areas, from life sciences, from social science, from economics, from physics. And again, it’s a more varied, more diverse set of talent than we get if we just look at the standard computer science pipeline.

Waleed Ammar
28:25

Yeah. It’s probably going to be a number of years before the culture. This culture becomes more you know, like common in other places as well. And people start feeling.

Fernando Pereira
28:35

So the reality is that this is something I fight constantly. I’ve had many conversations about hiring and about evaluating candidates. That always the kind, Oh, but X does not have as many papers as Y. And the one question might be, well, they maybe have different career trajectory. Maybe they had some issues in their life that got the way of essentially being able to meet every single ACL deadline. Let’s look at what the fundamental contribute information we have about their talent and about their creativity, about their interest in solving real problems. And you often find that if you look at it, not from the point of view of did you kind of tick this various check boxes, but rather did you bring a new perspective to what we need to do because we don’t have any problem in hiring more people like the people we have already, but that’s not really where it’s at. Where it’s at is to hire new people, more people when we would hire hire people that are different, they’re fundamentally different than have other dimension had a critical look at what we do that can change for the better. What we do.

Matt Gardner
29:51

So say you are a person who is screening resumes for an internship or for a new research scientist or whatever. I’ve been in this position recently. What recommendations do you have for someone who is doing this? Given that basically all you have is a resume, how do you get the information that you’re talking about? Do you have any advice?

Fernando Pereira
30:10

Yeah, so I mean I have not been doing that myself so I cannot give you a sort of procedurals answer. What I will, ask you is for instance, does this person have a life experience, work experience for instance, academic experience that is different in interesting ways from what we have already an adds some dimension for instance, someone was like, look at the resume and say maybe it’s not someone who has a computer science degree, but they were a field linguist in West Africa that would wake me. I would immediately be say, Whoa, this is very interesting. This is maybe someone who will have a much better perspective on how we think about different languages and dialects in our products and so on and so forth. I mean you have to look at the specific cases.

Fernando Pereira
31:03

You can not have some kind of general recipe, but you have to think of it from the point perspective of that when you build something like a search engine, you’re building something for, you know, hopefully form the greatest number of people and therefore they have different techniques, different experiences, different expectations. If you only have a very narrow range of people in your team, you’re going to forget, not ask these questions for instance, about different dialectal variation areas like, you know, code switching, like the fact that most people are multilingual outside the United States. In fact, the United States have a lot more multilingual people than people think about. And so all of these factors, you know, if you go at the screening resumes with an open mind, you’ll see there is interesting different people in that with different experiences. You’ll want to interview them. You don’t you know, you clearly, you want them to be able to work in your environment reasonably well. But you know, if they don’t necessarily know the fine, all the fine points of of sorting algorithms that’s not here or not there, right?

Fernando Pereira
32:18

It’s a, we have enough of those people, but someone who really understands you know, for instance, you know, that has a socio-linguistics background, or sort of neuroscience background could be incredibly illuminating to your, to the work you do and guess what, they might well be different in the backgrounds in ways that are also very interesting from the point of view of the dynamics of your organization. So I cannot say a specific recipe, but this is the way that what I recommended, and what I constantly tell. In fact, we changed our hiring guidelines for research scientists. I pushed for this and then we did it a few years ago to get away from the idea that all research scientists have to be computer scientists. And I said, no, I, you know it could be research linguists, it could be a research biologist. We need these people, depending on the goals we have for particular team, for particular project, you need a wide diversity of talent.

Fernando Pereira
33:17

And for instance, we have an intern longterm intern in our team who’s a biologist by training and she’s working on our biomedical information retrieval with Ryan McDonald and Keith Hall and others. And that’s a very important role for us is have someone who understands the domain and has this, can learn where they can learn more about the say the machine learning and natural language understanding and we can learn more about what matters from, people doing biomedical research so that at that two way street is what we’re trying to build.

Matt Gardner
33:59

Yeah, that’s good advice. I guess I feel like it’s easy in a small company where you’re the one who is doing research and screening resumes to latch on to like easy signals, but that can be pretty damaging. It’s good to hear these reminders.

Fernando Pereira
34:13

Yeah, I mean, look, I can only do the job of trying to persuade, try to nudge people in this direction. We’re not perfect by any means. We, have, it’s easy to fall into the, because you are in a hurry, and you have lots of resumes to look at. So it’s easy to fall into these sort of standard patterns, part of the job of research manager is to make research more interesting by kind of getting in way of routine, asking questions, you know, why are we doing it still this way, whether it’s the science or the engineering or the recruiting or the, you know, all of these things have to be challenged in a positive way, but they have to be challenged. So that’s, I mean, my response to this question about diversity and inclusion is not, I mean unfortunately is not something I can give some fantastic metrics on. It’s just a question of having my personal experience, both as, as a faculty member before and now as an industry research manager is you have to challenge your own assumptions about what’s important. And one of them is what’s important is all of us did this for the same deep learning course in the same top 10 universities.

Waleed Ammar
35:33

Okay. So I would like to pick your mind on like the collaboration styles between research groups and product groups. So they’re kind of like two like common ways of doing this is; the first like in some research groups the researchers would develop some method as inspired by certain problems that matter to the company. Implement prototypes, show results on some academic datasets and only then collaborate and reach out to product teams who will assess whether this is useful for them or not. And oftentimes this means like re-implementing the library or the method to be more consistent with the existing code base, another like way is to actually embed research groups within the engineering teams and like work side by side to develop these methods and become like basically you don’t need to implement anything at the end. Do you see like what are your thoughts on this pros and cons or if you have a preference?

Fernando Pereira
36:29

Well it really depends on the type of problem. I mean, I think we’ve been very successful and in fact most of the biggest successes have been when the research results are in the form that are kind of easy to communicate and to collaborate on with the product team. If the notion of, Oh, here’s research, they build a prototype and now there’s some kind of a set of developers in a product that’s going to take the prototype and rewrite it from scratch. That model, I’ve not seen it succeed much at Google. I think it doesn’t mean that the researchers are not independent. They are, but I think that the successful efforts always involve one or more of the researchers being very engaged in the collaboration at the level of we [are] sharing the code, we [are] doing the experiments together, we [are] looking at the results together.

Fernando Pereira
37:31

You teach me what’s important from a product point of view. I teach you what I’ve learned about different ways of modeling this problem. So because otherwise you get off, when the failed mode I’ve observed is someone will ask a research product group, Oh, you have such and such a problem, let’s say a spam classification problem example I give and what do you want? Oh, I want hyper say 95% precision at 80% recall just making up some numbers and researcher goes away, trains a model on some of the data. Maybe there’s some gold labeled data and Oh, oh, I can get this, but then product team takes the model and retrains it on the real data they have and the quality is much worse. Oh, wait a minute. Why is that? Oh, except the fact is that the actual, you don’t have much gold label data.

Fernando Pereira
38:29

Most of your label data is noisy because the humans who are labeling it are in a hurry and or get tired and they make mistakes. Oh, you didn’t tell me that. Well, I mean clearly this is not going in that well. Right, because I did all my work. I got the numbers you told me and now you are telling me that my work doesn’t matter. This is a failure mode that is not uncommon. And that we try to avoid by asking the question, what is the real problem you’re trying to solve? It’s not to get the classifier with some precision recall. You so have a system that is embedded in a product that is going to use the information that the product can collect. It might be noisy labels for example. So actually the problem of creating something that handles noisy labels is much more interesting. It’s actually from a research point of view, it’s a much more challenging one and one where the researcher can learn a lot from practice. So creating those early relationships and those early conversations is what avoids the kind of failure that I described. It also makes the research much more interesting. So that’s the model that I try to encourage. That’s the model that we’ve succeeded. Our biggest successes have all been under that model.

Waleed Ammar
39:48

Yeah. Thank you for sharing your thoughts on this, I think we’re almost out of time. So I wanted to ask if you have any final remarks or advice you’d like to share with researchers or research managers?

Fernando Pereira
40:00

Well, years its for molding for researchers wondering should I consider a role in industry, whether in you know, small company in a startup or a large company and they think the answer is you shouldn’t think of it as a better payed academic job. I think that’s a very self-defeating way of proceeding, even though it might be more profitable in short runs. I think that people should think it as what are the problems that this company’s trying to solve or these problems I care about. Does it have a function that I share, a function in society that I am aligned with and if so, and if I think that can contribute to solving those problems and creating those results, that will be a value in the ways that my own value system operates, then that’s where I should go in this direction as opposed to, you know, the more natural for someone who’s had a graduate degree to continue to be in academe.

Fernando Pereira
41:03

But that’s even true if someone wants to be in academia but then doesn’t care about teaching, it’s again a misalignment of values that is problematic. So decide to do the research and join the organization that aligns with your values of what you’re looking for in your future. Don’t decide it on the basis of, Oh, they have a big research group, so I want to join because there’s all these successful people there that’s not a good longterm recipe for wellbeing and for impact and satisfaction in your career. Do something that you feel really committed to as in terms of the, your curiosity, your personal values, where you think you can help make a difference.

Waleed Ammar
41:54

Yeah. And honestly we’re very fortunate in this, in NLP and also computer science at large in having many choices. So I think, we do have a lot of choices to make and we can align it so that we, like we’re pursuing the thing that we actually care about.

Fernando Pereira
42:10

Well, thank you for talking to me. Hope that the recording is acceptable to your podcast. I look forward to seeing it.

Waleed Ammar
42:19

Yeah. Thank you so much. This was fun.