Canary in the Coal Mine: What is the Impact of AI on Tech Workers? – Ep 14, Part 1

Ella: We’re like constantly pegged at full marathon speed. And usually before, we’re like, okay, this is a big push, and then we get to kind of relax for a little bit before pushing again. And now because everything’s so efficient, we are kind of being asked to like push all the time. And I’m starting to see where that’s you know, it’s creating cracks in the surface.

Stephanie Hepburn: This is CrisisTalk. I’m your host, Stephanie Hepburn. In episode 11, psychiatrist Elias Aboujaoude shared that at his Stanford practice, he’s seen a rise in people seeking mental health support over their anxieties of job displacement by AI. Job insecurity, overload, and burnout have significant implications for our well-being. How employees are faring in the tech industry is a bellwether test. A canary in the coal mine for what impacts AI may have on the labor market as a whole and the people who comprise it. Today, in part one, Ella, an AI researcher, joins me to share how large language models are shifting job expectations. She shares the excitement she and her peers feel about AI and the increased efficiency it allows. But she also highlights that burnout is increasing and there’s an overall meta-anxiety in the industry. By the way, Ella is not her real name. She’s asked to speak anonymously. Here’s what she has to say. Let’s jump in.

Ella: I’m an AI researcher. I’ve been in this space for over 10 years. I started as a data scientist, then a machine learning engineer, and then now more focused on research. And what I do is from the start to build machine learning models. This also includes AI. So training them, changing architectures. And now I’m working more on the other side of things, which is evaluation. So seeing how these agentic solutions behave, if they’re safe, if they’re doing what we want up to our standards.

Stephanie Hepburn: And when you talk about AI, does that mean large language models, or what does that mean specifically?

Ella: Early in my career, it was like smaller models, neural networks, as technology changes or techniques of doing things changes depending on the task. So for instance, this model, an SVM support vector machine, is what it represents. We used to use that for classification, like dogs or cats, but it wasn’t really that great, or we would fail in certain cases. And so with the neural network, you didn’t really have to like specify any of that. It would figure out what was a cat and a dog kind of a little bit more autonomously. And so that’s how it’s been progressing. And now with like LLMs, the game has changed quite a bit because there’s always this option of like doing things very quickly that before wasn’t really the case. Before you had to spend a lot of time, you know, gathering data and curating your data. And that’s a really heavy upfront cost for companies or for people, right? And now with like an LLM, because it has all this kind of knowledge base, you can do very quick, rapid prototypes that are mostly okay. You know, obviously depending on what your area is, you might want to be more stringent or less stringent. But yeah, so it’s changed also what your options are in like model-wise, what you can do.

Stephanie Hepburn: How has that shifted your employer’s expectations of you? And then also how you feel in the job market. So how you feel, you know, you’re training AI in some ways. What do you think in terms of your own job position?

Ella: In my company in particular, there’s a big push to use AI everywhere in your day-to-day, if you’re writing documents to speed up any sort of processes, like you should be using AI all the time. So it’s really changed how we work. Usually, I’ll go back. So, usually part of my job of training these models what meant was I need to write code, I need to use different libraries to put together a script or program to essentially fetch data for somewhere, split it in very specific ways, then send a training job out, wait for results, kind of do an evaluation of see how it’s doing, and then circle back. And now there’s a lot of coding agents, which I would say maybe even six, seven months ago weren’t as great. They were okay, but you wouldn’t necessarily trust them as much. Whereas now they can do a lot of things for you correctly. And so there’s also like that shift in work where they’re like, even if you’re programming in a coder, they kind of want you to use AI as much as possible to be more efficient, which has good things and bad things because in one way it makes you a lot more autonomous. So for instance, if I need to touch a particular part of the code that I’m not really familiar with, it makes it very easy for me to adapt it, change it before maybe it would take me a lot longer to like understand how things connect and how they relate and maybe talk to somebody. Like this just lets me like it does all the grunt work and will tell me like this is how you need to use it, this is how you can adapt it. So it makes you much more efficient. I think one of the downsides is that LLMs just in general tend to be very verbose. And so that happens also in some of the code that it creates. We’ve had a lot of discussions internally of seeing this, right? Like it’s like bloating up the code base and people just like, oh, this works, and then sending it in.

Stephanie Hepburn: And so when you talk about coding agents that can help you, is that OpenAI’s codecs or something like that where you’re assigning tasks to it, or what does that look like?

Ella: Yeah, exactly. So there’s different kinds of software you can use from Claude or GPT or different ones, and they’re integrated like in your coding surface. And what that looks like is maybe I have an it almost, it’s almost like an assistant, like a junior coder person, right? In a way, you’re like, I have this main idea and this is the framework I want, and this is maybe the data types that I’m looking for, and then go implement it. And then I’ll check if it’s done a good job or not, and then like it’ll, you know, come up with tests. So it means I’m less and less sitting down and typing code, and more the agent is like creating the code and I’m like verifying and creating tests, or and sometimes coding as well, but much less so.

Stephanie Hepburn: So does that mean that companies are less apt to hire junior coders because there are these roles that it sounds like they could potentially become obsolete or shift or change?

Ella: I think that’s certainly a phenomenon that we’re we’re seeing, like, and I’ve chatted with a lot of my colleagues that now like junior positions are less and less available, especially I think for smaller companies, because just this, right? What a junior person can do, your coding agent can probably do better, potentially. And in theory, right, it’s making you more efficient. So maybe you don’t need that junior person to take care of that smaller project. And so usually at least the trend I’ve seen in San Francisco is most of the positions that are open are senior positions. So people who already have experience know a little bit more about like architecture and best practices and you know, like have that know-how to kind of like coordinate and orchestrate all these agents.

Stephanie Hepburn: That seems to then create a gap, though, because that means that those individuals already have training. And to get to that point, you’ve had to have career building blocks. So you often have to start as a junior coder in order to accrue that knowledge, right? So it’s like almost getting rid of that foundational learning. So that’s being replaced with these AI assistants. And then it means that those individuals who’ve already accrued that are able to continue onward, but it creates this gap underneath it because at some point, do you think more and more the AI technology is not gonna just take over the junior building block, but it’s going to also start taking over those senior positions.

Ella: Yeah, I mean, that’s several thoughts on this. Yeah. I think right now it may be making those junior positions more sparse. But I think in reality, what it’s gonna do is it’s gonna change the definition of what a junior engineer does, right? Because people who are, I imagine now studying computer science, or now that you can study like machine learning engineer or something like that, you’re gonna have access to these tools as well. Right. And so I feel like we’re maybe in a a bit of a transition period where the classical junior definition is not what we need at the moment, or less so, right. But but I think eventually as our jobs kind of change in a way because of all these tools, then like that definition of what you need from a junior is going to be different as well. So I think eventually it’ll wash out, right? I think that later on we will start hiring more junior people. Because I think also like for for many reasons, right? I think also for like a company it’s good for people to do mentorship, to practice leadership for many reasons. That’s one thought. I think, I mean, obviously there’s always, you’re right, like these agents are getting better and better by the minute almost within my colleagues. There’s like always a joke of like we’re coding ourselves out of our job in a way. Uh, which, you know, you know, depending on how you take it, you’re like, great, I’ll retire early. But I think there’s again, and I I think it’s like the classical definition of what you’re doing and how you’re doing work changes, right? So, like I, to be honest, haven’t trained a classical model in months now, right? That was part of my day-to-day end job, and that’s not what I’m doing right now. We’re doing something is like a big fine-tuned job and pipelining of data to fine-tune one of these bigger models, which is like a different kind of strategy altogether. But yeah, I think the other aspect of it is the models are really good at executing, but I think you still right now need somebody to orchestrate and kind of like explain things and give it context and like what to do. So I think at least for now, we’ll need a human to kind of coordinate that, right? And you can have multiple agents working for you, for instance. And so you need to give specific instructions and how they relate and what each one does. And so I think like that that aspect is not going away soon, hopefully.

Stephanie Hepburn: What are your colleagues’ thoughts in terms of the progression of AI, their own jobs? Is there an increased demand? Is there an increased expectation of efficiency? Efficiency is a great thing, but sometimes efficiency also comes with a greater expectation of production. And so I’m curious as to what your colleagues and other people in the field, your friends, are experiencing.

Ella: Yeah, I mean, what I observe is there’s big excitement about it in the sense that like, you know, this is so cool. And again, like what it was able to do six months ago or a month ago, like, you know, they’re releasing models like almost weekly, you know, and you know, all these different companies. And so you’re always like testing new things. And so it’s in a way exciting. But on the other hand, you know, if if we talk about like efficiency and and things of that sort, I feel like there is certainly an expectation to be more efficient, not just in your coding, but in, you know, the documentation that you write about what you’re building and what you’re doing. And what I’ve noticed is that we’re becoming much more efficient, maybe too efficient. Um, so you know, that now you have agents that are like creating code more rapidly. You can create all these prototypes more rapidly, you can write documents even more quickly, but then you’re sending that to other humans and we can’t process it that quickly. And and I think also because things are moving so rapidly, you know, like usually I’d say like a year ago, there was a bit of like a wave in the year in terms of peak times and kind of times where things are a little bit more relaxed and you get to do more research. So, you know, like if we had a big release, it was, you know, a big push, and then afterwards there would be like a decompression time company-wise in in general, right? And that would pick up again later on. But there was like a bit of a wave or like a cadence of how things flowed in the year. And because, at least in my case, things are changing so rapidly, and we need to make sure our products stay up to date and we’re using the best models and we’re kind of staying in that forefront, there’s been no decompression time at all. And so I’ve noticed, yeah, a lot of burnout with my coworkers, a lot of people getting sick all the time. And I think in general, I think people are just like very exhausted, to be honest, because nobody’s been able to like fully have that decompression time, which we we were kind of used to. You know, it’s just kind of like a constant output of things.

Stephanie Hepburn: Yeah, and I can see it’s machine versus human in that regard. The machine output, machines don’t get tired. And so the machine output is increasing. Um, and then you’re not able to have, as humans, that decompression time you’re used to. Like you said, these ebbs and flows after a big production after a big output, you would be able to step back. And not just, it’s not that you weren’t being productive, but you were going into these other roles and other work that needed to be done as well, like research.

Ella: Yeah, exactly. We’re like constantly pegged at like full marathon speed. And usually before, we’re like, okay, this is a big push, and then we get to kind of relax for a little bit before pushing again. And now because everything’s so efficient, we are like being asked to like push all the time. And I’m starting to see where that’s you know, it’s creating cracks in the surface.

Stephanie Hepburn: What are you able to do for yourself? Because this isn’t happening within the company, and of course, you can push to have changes within the company if possible. There are those avenues, and I’m sure you’re exploring those for yourself. What are the ways that you’re trying to ensure that you do get that downtime or you are able to work on the research aspect that also needs to be done?

Ella: That’s a good question. So I’m I’m I’m I’m optimistic. So I’ll I’ll try to like block time off my calendar and things of that sort to say, okay, now I need to concentrate. Like I have all these documents to read and provide input and et cetera, et cetera. But I also need to set time for me to write code and prototype certain things. And so I’ve been trying my best to kind of allocate time to things, because otherwise it’s just like a constant stream of information and you’re responding to, you know, messages on Slack and urgent things and whatnot. And so trying to be a little bit more strict about those boundaries. And I think for me on a personal level, I’ve taken to like meditate a lot so I can essentially like calm down my nervous system. And also the the other aspect of it is because things are changing so rapidly. Sometimes, you know, we plan for something early in the week and we make a plan and make a decision. And by Wednesday, everything has changed. Right. And so it’s like this constant moving target, which also creates, you know, some level of exhaustion, right? Instead of saying, okay, this is what we planned, we’re always kind of pivoting because things are changing so quickly or we need to get on top of XYZ thing or test. It can be overwhelming, I think. And so that’s why I’ve kind of taken a lot to meditation to try to be like, okay, if I’m calm when I come into work, then I have the bandwidth to deal with all these changes and deal with this rapid prototyping that we’re now kind of pushed towards doing.

Stephanie Hepburn: So what do you think companies should be doing to address this? You know, because this is a new shift. And like you said, it’s happening very quickly. What’s happened, the typical cyclical pattern has shifted. And so now there’s just chronic demand. And like you said, instead of having a break after the marathon, it is marathon after marathon. What do you think these companies should be doing in order to ensure that employees aren’t burning out?

Ella: That’s a great question. Like if I put myself in the position of like a CEO or something like that, I can see why they’re pushing us. Because the other aspect of this whole kind of sphere is also that because all these LLMs are getting so good at what they do, they’re getting really good at everything, including possibly the product that you’re developing. And so there’s, I think, also this angst in a way that you know, how do we differentiate ourselves from like what this LLM can do versus what our product can do? Like what can we add or put on the table? And so I think that’s why there’s also some like broader meta anxiety in a lot of companies, is because also like what you do may be obsolete possibly soon. And so I think that’s why we’re like doing our best to stay on top of things, test all these things constantly so that we can provide a better experience to our users than what they could get if they just directly go use Claude or GPT. So I think there’s some aspect there in an ideal world. I would say, like, you know, we need to be mindful in the sense that the amount of output we’re creating because we’re being, like I mentioned, so much more efficient does not match like our human rates of processing information. So I think figuring out that balance is important.

Stephanie Hepburn: It’s interesting when you’re talking about that meta anxiety, because I did speak with somebody and they were telling me they would have AI assistance run throughout the night so that they never felt that they were having dips in productivity.

Ella: Yeah. I could see that happening. I mean, not to that extent, but we do we do that all the time, right? Like we’re training models 24-7 so that even, you know, like spinning up different machines so that you’re being as efficient as possible. So I I can sympathize with that.

Stephanie Hepburn: So what does that look like? Does that mean waking up at certain times in order to check in? Or do you just allow it to do its thing and then you check, you know, the next workday?

Ella: Yeah. I mean, if it’s training a model, you just kind of let it go and you you check the next day. And and and again, like good things and bad things, right? Because also if I have agents, I can give them instructions of what I need to be done. And sometimes that can involve like long compute runs. And so I don’t need to be staring at a screen, right? I can just tell it do all these things, run these tests, give me results and plot them in a nice way so I can share them with stakeholders, and I just let it run, right? So sometimes I’ll do that at the end of the night, at the end of, you know, like you know, you can schedule things and let that run at the end of the night, and then in the morning I have some results and I haven’t had any like downtime, right? Otherwise, I’m like waiting for this during the day, so to speak. There’s some things that I, you know, like require my input. Yeah. But there’s certain things I can just kind of unleash with the agent.

Stephanie Hepburn: Over time, we’ve seen this shift in a blurring of the compartmentalization between work life and personal life. That’s not novel. It’s like the speed in which, like you said, AI is shifting and changing. I think it was just so poignant when you talked about not only the meta anxiety, but just, you know, the human rate of processing information. And we do, as humans, we do need off time. And you know, that is not a weakness. It is where our brain is resting and thinking and coming up with ideas. So how does one find that balance acknowledging the race that exists where if what you’re developing could potentially be obsolete in a month because AI has already made that leap? It’s finding ways to protect humans so that they can continue to do the creative and unique, frankly, outputs and ideas that only really they can come up with.

Ella: Yeah. No, I think it’s a balance. And I almost feel like this is a new way of working. And usually when you have a new technology, like it’s it it comes in, it’s a big sensation. And then we kind of like have time to adapt to it and integrate it into how we work and kind of yeah, make make it useful and and and blend well with like our environment or our culture. And I think right now, because this is sort of developing at like light speed, I feel like we we haven’t had the time to fully adapt to it. We’re like in this constant reaction mode as opposed to like figuring out how to better integrate. That’s why I I can’t answer your question of like what’s your company? I have no idea. Yeah. I’m just like surviving, you know, in a way. Yeah. So I I think it’s like we haven’t really been able to like fully like integrate that in a way, like we’re just reacting. We’re just, you know, like there’s this new thing. How do we react? Like we haven’t been able to like really fully comprehend this new technology, I think, right? Like there’s a lot of unknowns about it. It does, you know, a fantastic job, and you see the output, but like exactly how it’s doing it and why it’s doing it is a bit of a black box, right? And I think because things are moving so quickly and there’s now more competitors, like nobody’s gonna stop and say, like, oh, maybe we should like figure out how this is going to change your society or affect our mental health, or what are the consequences of this technology now, making decisions on things like hiring and salaries and like whatever implementation you’re using it for. I don’t think we’ve, as a society, I guess, like stopped and put the brakes to be like, let’s think about this. Because again, everything’s kind of like moving quickly, and nobody wants to, in a way, like lose at this game.

Stephanie Hepburn: That was Ella, an AI researcher, on the paradox of AI efficiency and the exhaustion she and her peers are experiencing trying to keep up with demand. Please stay tuned for part two. If you enjoyed this episode, please subscribe and leave us a review wherever you listen to the podcast. It helps others find the show. Thanks for listening. I’m your host and producer, our associate producer is Rin Koenig. Audio Engineering by Chris Mann. Music is Vinyl Couch by Blue Dot Sessions.