How to Have a Data Science Career in 2026 with Marina Wyss, Senior Applied Scientist at Twitch

designing things is taking more of your time. Talk me through what does this involve?

Marina Wyss: So one of the reasons that it is taking more time is that we're simply doing more work. So because we are able to speed up the coding part quite a bit, I'm able to do a lot more projects than I used to. So I used to maybe be on one project at a time, and now I'm on probably four or five at any given time, which means more and more stakeholder meetings and more approvals for each project, right?

Because each project needs to go through legal and marketing and all of that as well. So the planning phase is everything from deciding of this huge list of things that we could build, which we've always had. This is, this has always been a problem, a blessing and a curse. So of all the things that we could build, which are the highest priority for the team at the moment?

And then getting all the stakeholders in a room to align on what we're actually going to build and why. And then from there, it's my job as the technical expert to actually make a plan for what the architectural will look like, what models we're going to experiment with, what a good definition of done would be, the projected business impact, all of that.

And then more stakeholder alignment, more approvals before finally getting into kind of a low-level design that I would ultimately use for the coding agent. 

Richie Cotton: Okay. Stakeholder alignment, I guess it's always a tricky thing. Is this something that you can learn as a skill? Like how to interact with other people, and get them to do things you want?

Marina Wyss: Yeah, definitely. I would say this is something you can really only learn from practice. There are books on, corporate communication and navigating corporate politics and things like that. So you could potentially gain some insight from those, but I would say ultimately that's something that you learn from watching people who are good at it on the job.

In my experience, it was just, finding managers or people that I felt were very effective at getting either what they wanted done or what was just, the best for the business at any given time, getting that across the line, and watching how they, how do they position their team, how do they explain what they're building, who do they talk to, and just learning the mental models that they seem to rely on and implementing them as well as I can for whatever I wanna push forward.

Richie Cotton: It may be something you can't just get from a book. I agree with that, like learning how to argue with people and have these kind of discussions. Okay. So beyond that are there other skills that you think are becoming more important because of this shift in the different amounts of work you're doing on different tasks?

Marina Wyss: I would say evaluation is has always been important, but is easier to overlook now, and so therefore I would say it is more important in a way than it used to be. I say this as someone who works with a lot of software engineers who are now moving into the AI engineering space, where of course it 

Richie Cotton: is essentially just calling an API, 

Marina Wyss: right?

But if you're coming from software engineering, you might not be as used to working with non-deterministic systems. And so 

Richie Cotton: having a comprehensive 

Marina Wyss: plan for how you're going to evaluate something that isn't necessarily gonna give you a deterministic output and has just generally fuzzier outcomes. The the definition of good isn't as clear as it is when you're building a service that has a defined input and output that's expected.

But there's more subjectivity to most of the things that we're building. So understanding different ways to evaluate AI systems and really prioritizing that from the beginning, even to the point of having a systematic way to do prompt engineering, all the way to end-to-end evals for an agent system.

These things all require more effort than I think we want to put because it's... We can go so fast, and it's really fun just to build the thing, but you have to force yourself to stop, slow down and put the time and effort into that because that's, that is the thing, in my opinion, that makes a difference between a product that is actually good and not.

Richie Cotton: Yeah, definitely. I think that's been, like the history of many decades. No one really enjoys doing the testing side of things, but it's incredibly important in terms of quality control. The other thing you mentioned about a lot of AI systems, particularly generative AI systems, are non-deterministic, and I think there's been an interesting shift that AI used to be a data team thing, and it's now just as much like a software engineering team thing, and software engineers are less used to dealing with, "Oh, th- there's probability distributions involved here, and sometimes stuff works and sometimes it doesn't."

So yeah. I think some good data skills there. Actually are there any sort of tips you have for going about testing things when they're non-deterministic? 

Marina Wyss: I think it's a lot about understanding a definition of good that might have, like I was saying, a lot of subjectivity. So if you're working on something that...

Let's say you're working at YouTube and you want to create something that will summarize what's happening during a YouTube live stream. What a good summary is gonna take a while to define. Who's your target audience? What kind of voice do you want? Are you okay with it summarizing something spicy if the creator is talking about that?

What guardrails do you need on PII? How long should it be? But just, there are different questions that you need to ask that need a lot of alignment. So if you have done the work to understand what you are what your goal is. There are a lot of methods, you could of course use like an LLM as a judge as a first step.

You could design experiments to see if your desired user behavior is actually being influenced the way you expect it to be by whatever product you're building. And then again, I think it's just about going really slow and evaluating each component separately in addition to the end-to-end. So if I make this tiny change to the prompt, do I have a system in place for tracking that change and tracking how it changes the metrics that I'm interested in?

Do I have a golden set? Do I have an LLM as a judge? Do I have a human in the loop? Like how will I know if I'm making things better or worse if we're working in a very subjective field? 

Richie Cotton: Absolutely. 'Cause quite often the answers if you've got a chat, then it's is this a good answer or not?

Yeah it's a little bit subjective as to whether it is or not. I love the idea that you've got to track everything and you just gotta make some judgment calls on is this right or not, have tons of tests in place and then, yeah see where you get to. Definitely I think an evolving field, but I do think this is like an important skill set to, to work on.

Okay. And how about soft skills for machine learning engineers? Are there any important skills you think are there? 

Marina Wyss: Yeah. So we have the classic communication. That has always been important, and I would say is even more important now for two reasons. The first is that you'll be working with even more non-technical stakeholders as more teams that previously wouldn't really have been involved in AR- AI or ML projects are now.

So you're gonna have, even more non-technical stakeholders that want to do AI, and you might have to help them figure out what that actually looks like in practice. And then the other piece of it is if you're working on a gen AI system, if you're working with prompting, your communication skills are directly influencing the output of the model because you are communicating in English to the model in a way that we didn't used to have to do when we were relying on code for the output.

So communication, of course. And then the other one that has also always been around, but I think is even more top of mind now, is comfort with ambiguity and change and continual learning. So even, way back in my day, years ago when I started with machine learning, there was always something new to learn.

There was never gonna be a point when you had learned data science or learned machine learning. Those are I think, impossible goals because there's just so much and it's gonna change so frequently. But now it's even faster. Now it's completely a fool's errand if you're trying to keep up with everything.

It's not possible. And so how do you develop a sense of enthusiasm and excitement and curiosity that keeps you motivated to learn over the long term without it feeling overwhelming and burning you out? So I think that balance is also really key right now. 

Richie Cotton: That's a really tricky balance.

Like I, I feel like one of the conversations I've had most often with like people who are in the AI space recently is like, "Oh man, we're like four years into this." It's like it's, it- there's a lot and it's been a lot for years, and it's, there's no sign of it slowing down. So it is hard work trying to keep up with everything.

Actually, do you have any advice on like how you achieve this balance? I think it's the holy grail, right? It's like learning new stuff, not getting burnt out. Go- what are your tips here? 

Marina Wyss: Yeah, I have a couple. So I think the first one is not worrying about falling behind too much because it is inevitable, right?

Like you, you cannot stay up to date on every single update. So the way I've always thought about this is my job ultimately is, my job as a machine learning engineer at a company is not to know things, but to deliver value. So I still spend the majority of my effort looking at the business and trying to find areas that we could use AI or machine learning to address business problems, working backwards from business problems to find the technical solution that fits that problem, rather than just learning all the technical solutions and then trying to find a way to fit them in.

That's really helpful because it narrows my scope a lot. So for example, I have until now, never needed computer vision. There has never been... I never had a use case anywhere I was working. So I don't really know very much at all about computer vision, 'cause it, I just, I had to let it go. I had to let something go, right?

But now I have a use case where we are using more multimodal stuff, and I need to get into that, so now it's an opportunity for me to start moving into that space a little bit more. So going from business needs to what I learn rather than the other way around. And then the other piece of it is Just staying curious and having a sense of excitement about it rather than overwhelm as much as you can.

Because again, it's impossible to stay up to date completely, so following that curiosity is a way to maintain your momentum over time, which is way more valuable than grinding yourself down and then burning out. I would way rather see someone spend minutes a day reading... Let's say you read a couple of newsletters in the morning with your coffee and one paper a week.

That's gonna do a lot better than really cramming, freaking out, and then, eventually being like, "I can't do this," and giving up. 

Richie Cotton: It just seemed a much more sensible approach, at least from a mental health perspective, rather than just like a bit of wait till AM just reading Reddit posts on AI.

Yeah that's a bad idea. Okay. I like the idea of slow and steady, keep staying curious and just learning new things. Now, one thing I'll say is you talked about your job as being machine learning engineer, but we've also been talked about AI engineer and the kind of closely related roles, but maybe a little bit different depending on which company you're working at.

Do you wanna talk us through like the overlaps and the differences between these two roles? 

Marina Wyss: Yes, for sure. Yeah, and to complicate it even more, my title's actually applied scientist, so I'm just throwing titles around. And the reason for that is because there's a ton of overlap. So there are not clear definitions in the industry that distinguish machine learning engineer from AI engineer, and there's even often overlap between data scientist and machine learning engineer roles, and nowadays data scientist and AI engineer.

So with that in mind, the definition that I like comes from Chip Huyen, who wrote the book "AI Engineering," so I'm gonna trust her. And she would define the distinction between AI engineering and machine learning engineering as AI engineers primarily build applications on top of pre-trained models.

So they would be taking a model like Llama or, any of the OpenAI models, any of the Anthropic models, and building a RAG system or an agentic system, maybe fine-tuning that model. But they wouldn't be taking custom data and building a model from scratch. Machine learning engineers, on the other hand, would be building models from scratch.

So that would be commonly fraud detection, recommendation systems, things like that, that are as of yet still not using AI models exclusively. We're starting to use them more. And I would also say from a practical perspective, in the job market these days, machine learning engineers should also know AI engineering.

So I think you would be in a little bit of a pickle if you were looking for a machine learning engineer job, but you weren't comfortable with the AI engineering skill set as well. So the way that I think of it is my mental model from, again, a practical perspective, is machine learning engineering is a superset of AI engineering.

And then data science would be pre- or pre-trained models potentially, but more on the machine learning side, but less focused on production and more focused on research. 

Richie Cotton: Okay. I love that distinction. And yeah, there, there's a bit of overlap between the roles, but, I don't, actually, I don't think I've heard the idea that a machine learning engineer is like a super set of AI engineering skills before.

That's a really interesting take. And no I like it. Okay. Would you say that if you wanna become a machine learning engineer is that the path? Would you go AI engineer first, or can you go to it from being a data scientist? What's the typical path into this role? 

Marina Wyss: Yeah. Machine learning engineer is cool because there's so many different paths.

Many people go from data science to machine learning engineer. That's what I did. There are also a good number of people who are currently software engineers who are trying to make it to machine learning engineer, and either go directly or make a pit stop in AI engineering. So AI engineering can also be a nice way to get into machine learning engineering.

So as you can see, there's lo- lots of different ways to get there, and I think that is in part because machine learning engineering really requires a ton of different skills, and it's hard to learn them all at once. I would say, again, another kind of more practical personal take is that for entry-level people, trying to jump straight into machine learning engineering can be hard.

That's a big jump unless you're coming from software engineering or data science or something as an intermediate step. And I think that's because those fields allow you to learn a big chunk of the skills that you need. Obviously, software engineer, you're learning a ton of stuff about engineering that will be useful for you as a machine learning engineer.

Data scientists are getting closer to the model, so then they just have to add that production element. So I think stepping stones are a really good way to approach this, but there's a lot of different options. AI engineer to machine learning engineer is quite logical in my mind right now because there are so many AI engineer positions open, and so few people actually have the skills So while the job market, of course, has its challenges right now, I do think that there is a really interesting opportunity if you're able to seize the moment with AI engineering and get in there.

That can be a really great stepping stone because again there's more open positions than there are people with the skills to do it. 

Richie Cotton: Absolutely. And that's really interesting, 'cause I think a lot of... I- basically every company at the moment is trying to build out some agents or some kind of AI capability, so there are gonna be these AI engineering roles available in, almost every industry at the moment.

Okay. So you talked about how there's these different stepping stones to getting into becoming an AI engineer or machine learning engineer. So where to begin then? What's step one if you're, just out of college or something and you're just wanting to get involved in the AI space?

Marina Wyss: Yep. So I'm gonna assume you have no technical skills. We're just gonna start from scratch, just so everyone's on the same page. So I would start with Python. I think that Python is nice because, first of all, you're going to need it for any potential role that we discussed. It will be necessary.

Even if we're working with coding assistants, you still need to know what you're doing with them. And it also gives you other opportunities. So let's say you start coding with Python and then you try some AI engineering projects and you're not that into it, you could go into software engineering or DevOps or, any number of other things once you have a foundation in a general purpose kind of language like that.

So that's why I like starting with coding. I also prefer to start with coding rather than math, which is the next one I'm gonna talk about, because a lot of people I find hear that they need all this math. They st- they hear, calculus, linear algebra, probability, statistics, and they get a little bit scared, and they think that they can't work in this industry.

And it can be off-putting, in my experience, when someone is interested in machine learning and then gets hit with calculating the chain rule by hand and all this math, and they're like, "I actually don't like this." When in reality, we never ever do that on the job. So I like starting with something practical where you can see the outcome of your work quite quickly.

You can make the app, you can make the model. And then an intuitive understanding of math is very important. So I think having a, like a high-level intuition for calculus, linear algebra, and then basic probability and statistics is very important espe- it's critical for machine learning and data science, and useful for AI engineers.

Just so that you, just so you know what's going on with the model. Even if, as an AI engineer, you're not going to actually be training anything from scratch, it's still just going to be useful for you to have the language to understand what's going on. And then for machine learning engineers and data scientists, it's important for when your model doesn't behave the way that you expect.

So those would be my foundational skills, and then from there you can layer on specific skills depending on the track that you're interested in, whether you wanna go into more research with data science or applied gen AI with AI engineering, et cetera. 

Richie Cotton: I think the point about mathematics is very fair.

I did a maths degree years ago, and I feel like all the kind of the linear algebra, calculus, things like that, I used to have in my brain. Now it's very fuzzy ideas, and that's enough to get me by. It's like fuzzy ideas of mathematics. It's not often I have to, I would say probably never have had to invert a matrix on the job.

Technology handles that com- fine. Okay. All we're starting with some Python, then some yeah, vague math skills, and then get a bit of domain-specific knowledge or any other kind of technical tracks. All so are qualifications important, do you think, to get hired in this role?

What kind of qualifications do you think you need? 

Marina Wyss: So it depends on the role. For machine learning engineering and data science, a master's is helpful. It is not % required. So I do know people who have gotten quite far in their degree with just a bachelor's degree. I know a select few that don't have any degree at all, and just got through based on boot camps and extreme talent.

But it is useful for those roles because resume screening tools exist, and there is unfortunately just the reality, if there's applicants they might just cut based on advanced degrees and not even see you otherwise. So that also doesn't have to be a blocker. There are ways around that through very deliberate networking, I would say is your number one way to get around that hurdle And then for AI engineering, it's less relevant because it's much closer to software engineering.

And again, this is a, an area where there are just not enough people to do the job. And so if you are very clearly able to do the role, you're likely to be able to find something even if you don't necessarily have an advanced degree that's relevant. 

Richie Cotton: No, that's fair. And I think on the point about, Like screening for for interviews and things like that.

Yeah, if you get like a thousand people applying for a job, which is very common now because it's you have bots apply for jobs on your behalf. Yeah it's very difficult. Like hiring managers just have to filter things out somehow 'cause you can't interview a thousand people. So yeah getting all the right keywords into your resume and into the cover letter, incredibly important.

Okay beyond the the qualification side of things, the other side is gonna be around having a portfolio of work. What needs to go into your portfolio? 

Marina Wyss: Portfolio projects serve a couple of different purposes. The first is learning. That, that is obviously the most important thing is you need to have the skills.

The second is it gives you the opportunity to add those keywords to your resume to help with the filtering. And then lastly, it gives you something to speak about in interviews. So you need different kinds of projects for each of those things. At the learning phase, taking courses doing kind of follow along, more simple Kaggle stuff, that is totally fine.

That's great. That's how we all learned at the beginning. But once you're trying to get to the point of using that project as a way to get interviews and ultimately get a position, in my opinion, the most effective thing you can do is to build something real for a real client. That can feel like a bit of a catch-"Wait, but no one will hire me."

And so I'm not actually suggesting necessarily doing this for a paying client. It could be something like an example I like to give is I used to volunteer at the Red Cross and I was, not in any tech capacity, but I just noticed one day that they were doing this really annoying manual data entry process.

And so I offered to fix it and it took me, I don't know, maybe five hours to build a very simple system. And they are still using it to this day and it's saving them like hours a week. So that was a real project where I identified a problem, built a solution, and can point to a measurable metric that I actually moved with my actions.

And I could put that, in my opinion, legitimately as work experience on my resume. So that's also getting a little bit past this like experience issue. So you could do something for nonprofits. If you have a friend or a family member who maybe owns a small business like a tutoring company or a local store and they're managing, let's say they're doing inventory and spreadsheets.

Could you build a forecasting model? Can you help your uncle with his like, yeah, his tutoring business that he's managing students by hand? Could you build an agent? Can you do something for a friend and family? Or can you just do something for a hobby group? Let's say you could build an app to help your D&D group with some organization or something like that.

And then you maybe even build an app and put it on the App Store and you can share it on Reddit and make sure it gets seen by the right communities. So the point of all of this is to build something that addresses a real problem that you see in your life or the life of your friends, family And has real-world constraints.

So you're expecting this to actually be used by, the local business that you're helping out, or actually be used by the nonprofit, or actually be used by your friends. And that gives you the simulated constraints that you would have when you're working at a real job, and can help really identify those skill gaps that you need to prepare for interviews.

It teaches you all the skills that you need to add to your resume because you're building something real. And when you're actually in an interview, you can speak confidently about this project as something where you had to make trade-offs, you had to navigate business decisions, even if it was, D&D business.

But you had to make some executive trade-offs with things like cost and latency and how are you gonna scale this, and security and all of those things. 

Richie Cotton: No I love that. And the core idea of having different kinds of projects for different purposes seems very important.

So something for learning for your own just increases in skill is different to something that you wanna show off to the world in a job interview. And yeah, I do particularly like the D&D example 'cause I think, yeah, it's you've got probabilities built in there with the, with Dungeons & Dragons, all those dice rolls.

Yeah. Useful stuff. But yeah I'm sure everyone has some kind of friend or someone they know who's gonna benefit from a little bit of AI assistance that they don't have, so yeah. All right. We talked a bit about the, what qualifications you need, what portfolio you need.

Talk me through what you need to go through the hiring process. What, what happens when you're getting hired for these roles, and how do you pass the the interviews and all the different stages? 

Marina Wyss: I'm happy to talk about getting interviews. We can come back to that. But let's assume you already have one.

The way that the hiring process works at Big Tech is usually some kind of technical screen that could be, like, a HackerRank thing or something, or it could be a live coding interview That will typically, for machine learning roles, be Python. What kind of Python test you get is going to be very different at different companies right now.

The industry is going through this very interesting shift where we're no longer just doing LeetCode. Some people do, but it's quite easy to cheat LeetCode now, and it's also much less relevant. It already wasn't that relevant. Now it's really not that relevant with AI coding assistants. So we're starting to see a variety of different interview types.

So you could get LeetCode. You could get implement a machine learning model from scratch. That's quite common. So like logistic regression or K-means or something. You could get a kind of just object-oriented programming, basic like build checkers question. Or you could get something where they specifically want you to use AI, and they want to see that you use AI very well.

So they're, they want you to implement a whole feature live with them in an hour, and they want to see that you can effectively use spectrum and development, and you know how to use multiple agents in parallel, and you can use all the cloud code features and stuff like that. So let's say you get through that round.

There, there will also be some behavioral questions. So that's kinda like your first screening. After the recruiter call, I guess I should say. So you'd have a recruiter call, which is standard, tell me about yourself. Then the technical screening. Then if you get past that, you would probably go straight to an onsite.

And the onsite will be a combination of coding... one more coding round typically You would usually have what I would call a data science interview. So that's gonna be more questions in the weeds on models, and typically in the format of a case study. So it would be something like, "We're building a recommendations model for push notifications.

Walk me through how you would build this." And in this case, they're not talking about system design. They really want you to talk about how you would prepare the data, how you would measure the outcomes, potential models you'd use, that kind of stuff. You would also have a system design interview, so machine learning system design.

Of course you're gonna have behavioral questions either separately or throughout. So that's really important to prepare for. I think that's the number one thing that people underestimate in interviews and can really make the difference between your success or failure. And so that's the typical process.

Plus or minus, maybe you'll have an extra case study or, p- maybe you'll have something that's more on stats or something like that if it's a more researchy role. 

Richie Cotton: Okay. S- such a wide variety of different problems there. Like you mentioned LeetCode, the the toy problems are fairly standard, but then implementing algorithms for logistic regression, I think you said, from scratch.

That's a tough one. It's if you know it, it's it's fairly straightforward, but it's something that's maybe not top of mind for a lot of people, so you need to do your prep first, I think. 

Marina Wyss: Yes. And I do have a little hint there. So one thing you can do is just ask them what the coding round will be.

So people don't do that because they think, they're not allowed to or something. But a good c- a good company wants you to succeed. And so if you directly ask them "Hey, I would love to prepare as well as I can for this coding round. Can you tell me what the format will be?" And really push them, because I had a case where I did this and they said, "It's testing production level coding."

And so I was like, "What does that mean?" And they were like, "It's algorithms." I was like, "Oh, so the opposite of production level coding. Okay, cool." 'Cause I would've done exactly the opposite of what they were suggesting. But really push them to a- to answer that, and they should be h- happy to do it.

And that should give you just much narrower scope of what you need to prepare, so you feel a lot better going into those coding rounds. 

Richie Cotton: Okay, nice. And do you have any more tips like this for, like, how you go about passing all these interviews? 

Marina Wyss: Oh, I do actually. Interview prep is one of my favorite things.

So the number one piece of advice that I have is to prepare for the questions as though you already work there. So what I mean by that is, let's say I'm interviewing for a machine learning role on a team that does let's say... We'll keep going with the recommendations for push notifications. So what push notifications should I send to who and when?

That's their primary thing they're working on, let's say. So you already know pretty high likelihood that the questions are probably going to be relevant to that team. So if I was preparing for machine learning system design, I would design a bunch of different recommendation systems for push notifications.

I know this sounds silly, but I think people often don't. And I mean going to the level of imagine I'm already a machine learning engineer on that team, or I'm already a data scientist. What data do I have? What metrics do I care about? What does this b- how does this business make money? What are the legal and regulatory constraints that this business has?

What kind of latency requirements? What scale am I dealing with? What kind of interesting features might they have? So I really sit there and try to pretend like I'm already on this team, I'm trying to make this company money or move whatever metric that team cares about. What would I have at my disposal?

And then I will design, like literally write out and make diagrams for what this system would look like. Maybe do a couple different ones with different assumptions, and do that at the level of the full system. Think about the models, like where, what interesting ways could I use like the state-of-the-art versus what would be the cheapest way that I could possibly achieve this goal?

What are common hiccups? And brainstorm all of the questions that you think they could potentially ask you about that system in advance. So I know I'm recommending doing a lot of work, and I am. But I think that's what's necessary, truthfully for me at least. I can't just walk into an interview and nail it.

That's not gonna happen for me. But what I can do is spend hours thinking of every question they're ever gonna ask, and then when I get to the interview I've actually had this happen multiple times where I've been in an interview and they asked me a question that I had prepared in advance because I guessed what they would ask, and I came up with this like very fleshed out question that makes me sound quite clever.

But really, I'm just very prepared. 

Richie Cotton: Like hours of prep for an interview it sounds like a lot, but yeah, I can certainly see how that dramatically increases the chance that you're gonna get hired rather than just rocking up there and being like, "Oh, yeah, we'll see what happens." So I love that.

Thorough preparation often does i- improve your performance in these things. Okay. You mentioned before we missed a step on how you get the interview in the first place. So I believe you have some tips on that as well. 

Marina Wyss: Yeah. So it's mostly gonna be networking, right? Most of the time it's not gonna work if you're early career.

So this advice really goes for people who aren't, if you're currently a principal at Google, this doesn't count. You have a different reality. But for most of us when we're starting out, if we just send our resume into the LinkedIn void, it's not usually gonna work. Not to say it's impossible. I have actually gotten both of my machine learning jobs just from LinkedIn applications, so I'm not saying don't do that.

But a higher ROI strategy is to be extremely thorough with your networking approach and Approach it with kind of the long game in mind. So the way that I like to think about this is, let's say you work-- Let's say you're not looking for remote roles and you have a local area that kind of limits your scope, just for the sake of argument.

There might be places that you would really love to work. Can you find all of those people on LinkedIn and then figure out everything fun they're doing? So if you are interested in machine learning at, say, different places, you would find all of their engineering managers, all of their machine learning people, their data scientists, engineers, whatever.

Do they have a YouTube channel? Do they have a blog? Do they have a technical blog? Do they have an archive paper? Can you find a conference that they are speaking at? So it's a little creepy, but the idea here is that I'm trying to find things that I can learn about their company and their problems, and then ask questions of the people who I want to build a relationship with.

So I don't find the cold LinkedIn message works very well. I say that as someone who gets a lot of LinkedIn requests, and I would love to be able to help everybody, but I can't. It's just not really possible. And truthfully, I don't really feel super comfortable giving referrals to people who I don't know at all.

But I have had situations where someone would contact me and be like, "Hey, I saw your YouTube video on da," or, "I saw this thing that you built and you wrote a blog about. I had this question," or, "I thought this was interesting. Can you tell me more about how you made that decision?" So they're showing that they actually put the time in to understand what I'm working on and ask a thoughtful question If you do that, I would say you'll get ignored at least half of the time.

But sometimes people will respond to you and you'll have a nice back and forth with them. And after you've had a, a friendly conversation for a little bit, you might be able to ask "Hey, could you, would you be open to a quick call to do like an informational interview?" Or, "Would you mind looking at my resume?"

Or maybe you don't ask for anything at all, and you just save it for your back pocket for when you do see a role. And once you do see a role come up at their company, you can pipe up and be like, "Hi, remember me? Remember me?" And they're way more likely to give you a referral, at least I have been.

This has worked on me multiple times, which is why I recommend it. Because I already will have a favorable opinion of this person. I'm like, okay, they're thoughtful, they're intelligent they were respectful of my time. Let's see f- let's see if this referral will work. So that's my preference for how to approach this.

You can also, of course, literally just cold reach out, ask for referrals. That's what some people do. I would ignore them, so it doesn't seem strategic to me, but I know people who have succeeded that way. And then of course in person. So that's like the remote strategy and I would recommend doing both of these.

But if you live somewhere where there are in-person meetups or events of any kind, I would highly recommend going because I've had multiple mentees get interviews and then positions directly from in-person events where someone was like, "Oh, that's awesome that you're working on this thing. My company's hiring for da.

Have an interview." 

Richie Cotton: Yeah. The in-person stuff with the meetups, often they have free pizza slices as well. It's like it's a bonus, like free food and the chance of a job as well, or at least some good conversation. But no I love all these ideas. And the stuff you mentioned about like just watching people's YouTube videos and commenting on them and have an intelligent conversation.

I think related to something that worked for me early in my career was just like putting in bug f- issue things on- Yes ... on GitHub repositories. And then like- Yes ... people would see, "Oh, this person's complaining about my software." And then somehow that develops into a thing where you're chatting about what's going on, and then yeah you can build relationships that way.

Marina Wyss: That's an... y- to add onto that I completely forgot about that. That's an excellent point. I have a coworker who got his first internships because he was one of the early contributors to Keras. And so then he when he was applying to internships he would just send his like Git history and be like, "Look, I built this tool that you use."

And everyone was like, "Yes." That, that is an excellent approach as well. 

Richie Cotton: Nice. That, that seem very cool. Especially working on, like Keras is like huge yeah I can see how that goes down very well. Or so I guess the other related idea, I don't know whether this works or not is it important to have a social presence?

Is being famous on social media, is that gonna help your career or not? I think it helps, but I certainly don't think it's necessary. I can point to a fel- a fellow creator, Igor. He has pointed out multiple times that his YouTube channel has helped him get interviews. So in his experience, it has been directly relevant.

Marina Wyss: I haven't been on the job market since I've been a creator, so I haven't seen any difference. But I would imagine first of all, just being active on LinkedIn is gonna help you surface higher when recruiters are doing searches. So that, that can help you get some visibility, and it can lend some credibility.

So if you're more entry level, if I see two equally experienced people but one of them has been writing about what they've been learning about or has some kind of relationship with other people in the industry, that's probably going to help with credibility. But I don't know that I would say it's your highest leverage thing at the beginning.

I would focus more on networking if I had to decide which one of those is a higher priority, but it certainly can't hurt. 

Richie Cotton: I s- I suppose it can double as a portfolio in some respects. If you've got videos of you doing stuff or you've got blog posts where you're analyzing stuff, then it's yeah, it's demonstrating some kind of skills.

Okay. All right. So I think we've covered everything from what you need to do in order to get t- through to getting a job. Suppose you do get hired, what does success look like in the first few months? Like how do you know that you're doing okay in your job? 

Marina Wyss: Yeah. So I think it really is about Kind of the same transition that you make from junior to mid-level in any industry, where you go from being someone who passively takes tasks from a manager or someone higher level to somebody who's proactively suggesting workflow or work, excuse me, work to add into your workflow.

So at the beginning, it's very normal when you're on-ramping to be given, these discrete tasks, add this feature or do this bug fix. But ultimately what you want to do is grow into a place where you're thinking about your impact on the business first and foremost, and learning about the business as much as you can so that you can identify ways that you can use your technical skills to address those business needs.

Then from there, it's what we were talking about before with influence and getting your recommendations actually onto the roadmap. Because it's not as easy as just being like, "I have a good idea." You usually have to make a very compelling case for why the business should essentially spend money to do your idea, because your time is money.

And you'll probably need help from other engineers or data scientists, so you do have to make a clear case for that And then also just being able to execute a little bit more independently. So in those first couple months you're onboarding, it's very normal to ask a ton of questions. Never feel bad about that.

If I have one piece of advice as a beginner or someone just new on a team, don't let your ego stop you and end up being much slower than you would've been, because ultimately you're being judged on your outcomes, not the questions you had to ask to get there. So I would much rather see someone who's new ask me questions at the beginning and then deliver something really high quality than someone be scared and not ask questions, and then I check in a couple weeks later and they haven't really made any progress 'cause they were blocked, but they were too scared to say they were blocked.

Starting to deliver things more independently from beginning to end including finding the business problem is ultimately your goal. And I would say one big thing that I also would look for when someone is onboarding is getting comfortable speaking up in meetings. So this can be a little bit challenging.

It was for me. It was very hard for me at first, and I've seen this for many juniors over the years, where they don't trust themselves in meetings, I think. And so they won't speak up if they have ideas or even questions. And I think that's a really clear moment where I can see someone move from a junior to a more mid-level position when they start being like, "Hey, I have this idea," or, "No, I don't think we should do that," or, "Actually, I'm not clear about this."

So getting your bearings enough that you feel comfortable contributing is a really good development milestone as well. 

Richie Cotton: Absolutely. I think everybody spends so much time in meetings whatever your role it's always too much time in meetings, so getting good at being in meetings is an incredibly important thing for your career.

So I love the idea of speaking up more. Actually, do you have any tips for how you get comfortable about speaking up more in meetings and how you contribute better? 

Marina Wyss: Not really. I think it's just you have to force yourself to do it. You have to do it enough that you realize you didn't die when you did it, and you'll probably, if you have a good manager, you'll also be praised for speaking up more.

And so I think getting yourself over that, that first few times you'll probably start getting a positive feedback loop. But there's not really, there's not really much to do but to do it. Sometimes you just gotta have to, kinda have to- face your fears. 

Richie Cotton: Absolutely, definitely.

And maybe we're back to doing some preparation. So the same as with the the job interviews, if you prepare for "Oh, maybe these things are gonna come up in the meeting. I can talk about this," that's probably gonna help rather than going cold. Yeah. 

Marina Wyss: Yeah, that's a good point. 

Richie Cotton: I like it.

Okay. So I'm curious, if you went back to the start of your career, what do you wish you knew back then that you know now? 

Marina Wyss: I think I spent too long focused on basically memorizing stuff which I think is pretty logical as a beginner. But I would try to read math test- textbooks and memorize them, and I would try to read the machine learning textbooks and memorize all the formulas.

But that didn't really help me build intuition about what to use for what business problem. And so I think if I were learning, if I were learning again today, I would think about it more conce- more top-down rather than bottom-up. So I really tried to go checklist these are all the models, and I need to know how all of the models work.

But instead, I probably would've benefited more from thinking about these are the types of business problems, these are the kinds of data sets that people have available. This is the kind of infrastructure that different teams are working with. What are the models that fit in there and make sense for these different use cases?

And then learn the details as I needed to. I also think that a really great way to learn more about how the models actually work is to code them by hand, and that's not something I did until way later. So personally, if I read a textbook that's just explaining, let's say, just a decision tree, I don't really get it.

I can follow the words, but I don't really get what's happening until I build it myself or until I'm able to draw it myself or explain in my own words what's happening. So I think I would've benefited from spending more time kind of thinking about it like if you were discovering a decision tree for the first time, like you're the inventor of a decision tree, how might you have come to that conclusion?

What would you have needed to see in the data or in the code or visually for it to just make logical sense? So I think I would've just approached studying quite a bit differently. And then in the modern world, I also would focus more on AI engineering than I did back then, of course in addition to the traditional ML.

I still think they're both important, but I do see more and more that we are relying more on pre-trained models, even for things that a year ago we were using custom models for. So I think that would be an area of focus too. 

Richie Cotton: Teaching how algorithms work i- is incredibly hard to do well 'cause it's so abstract in some ways.

And it's lower level than a lot of people think about. So yeah, I love the idea that like to build your own models and like from scratch, and that's gonna give you that kind of deep insight that... on adapting which... In fact, I think goes beyond models. It's if you wanna know how something works, just try building it.

Okay. Just to wrap up I always want more people to learn from, so whose work are you most excited about right now? 

Marina Wyss: A couple of YouTubers I've been enjoying are Andrew Codesmith. He does these really beautiful... They're just like very well done. They're vibey. So they're motivational videos for software engineers with really nice videography.

And yeah, also having some quite good information that he's sharing. And then I also like Philip Choi's YouTube channel. He's also doing... it's motivational stuff. You can see what I like. For software engineers who maybe feel like they are too old to start, 'cause he transitioned in his s, ancient.

Or feel like they've missed the boat because of AI. So it's this really honest and friendly and motivational kind of approach. And then I've also been enjoying stuff from A Life Engineered on how to progress in your tech career. So I think those would be my recommendations.

Richie Cotton: Okay. I like the idea of having some motivational stuff to, just make sure that you feel good about things afterwards. And I also want to believe that Andrew Codesmith, you say he's doing stuff on coding. I wanna believe that's his real name. 

Marina Wyss: I hope so. Yeah. I hope that's his legal name.

Richie Cotton: Fingers crossed. All right. Super. Thank you so much for your time, Marina. 

Marina Wyss: Thank you for having me.