Surviving AI Fatigue: Staying Sane and Relevant in a Fast Moving Field

Speaker 1 [00:03]

welcome to our next talk Surviving AI Fatigue by AJ and Jayashri Krishnan just for the questions if you have questions afterwards please submit them via talks.pycon.de and now please welcome our speakers on stage and give them a warm round of applause

Speaker 2 [00:36]

Good morning, everyone. Thanks everyone for joining. So this is a topic that's kind of close to our heart. We've been feeling this for some time and we thought, okay, we need to kind of put this into words some way. So we thought, let's try this out. So surviving AI fatigue, probably many of you felt it in some form or another. So these are some of our thoughts and tips about this and trying to dig deep into this. So, just an overview, so what we want to talk about today, more or less the problem, this is like how it is in the media and we're trying to, is this real, some of our analysis and so on. Coming to introduction about ourselves, so like as they already introduced, my name is Ajay, so I work as a software engineer at Ansys or Synopsys. I'm also part of Apart Research, so they're a non-profit to do AI safety research. My primary work is in fluid dynamics, GPU algorithms, and AI safety. Of course, this work is nothing related to my work. This is just out of curiosity. And then my co-presenter is Jayashree. So she's a senior machine learning engineer at Siemens, also an APART research fellow and visiting researcher also at Aachen. Her research is primarily in computational biology, foundation models, time series, and also AI safety. Maybe some of you also attended her talk yesterday on time series forecasting. So as a disclaimer in general, so I'm not someone who says, oh, I don't like AI, I'm never going to use this thing. So I do use it every day. So it's part of my bread and butter. I agree, ML and AI is absolutely disruptive technology for most of us. Use it for both work. Personally, of course, I use this for also this talk. And I never want to take any chance with Roku's Basilisk anyway. so I thought I'm going to put this disclaimer in. So maybe some of you also attended the talk day before yesterday about hope, hype, or headache about AI. In that, also the presenters spoke about something similar. So the length of tasks that an AI can do compared to humans has been doubling about every seven months for the last six years, which means if a human takes, let's say, two hours to do a task, and AI is able to do it at least 50% accuracy in quite a short time. And this number of hours has been doubling. So this is a plot from Meter where they show this, and we are already approaching something like 10 hours of human work is already something that an AI can do already, and primarily LLMs and agents can already do. And there is no reason to believe that this is going to stop, and we could probably be approaching in a couple of years an entire week's worth of work by a human being done by an agent, of course, significantly faster without wages and whatnot. So coming to the problem. All of us have somehow seen this on our LinkedIn feed. You open your social media, and you see, oh, there's a hot take on something. Oh, there are like a gazillion new agents and skills.md and whatnot coming in. And there is a research paper that transforms everything that you've ever known before. And we've all seen this thing and we've said, oh, I don't know what to do. Am I using the right tool? Is there this specific person? Do they have some skill that's going to make me a 10x engineer? So we've all had this thought and said, okay, you know what I'm going to do? I'm going to save this right now. I'm going to get to this thing. I'm going to bookmark this archive article. Let me get back to this thing. And you see this and you say, and in fact, some of the really popular ones, you might even skim through this and say, yeah, okay, I mean, I think I understood this, at least partly. So we've seen this thing. This happens with social media or newsletters and so on. And we thought, okay, so how is this community doing about this thing? So we went back and said, what is PyCon DE 10 years ago? What was their word cloud? And I highlighted some of the, let's say, ML-related words. So we see them, not so much, but still they exist. But anyway, when I was looking at this, one thing caught my mind. I was like, what is Luigi? I was like, is this like a star Python developer that no one knew about? But it turns out it's actually a Python library, so I didn't know about this. But then I said, OK, 10 years on. Now we are in 2026. What's the latest word cloud now? Thankfully, agents. The good thing is Python is still bigger than agents. It would have been insane if we were in a Python conference and agent was the most popular word. Turns out it's still Python, so we're good. But clearly shows a trend towards this, where these things are getting very popular, And it's everywhere for us. Now, this is a show of hands. So how many of you have felt super excited? You look at something and you say, wow, this is amazing. I'm going to use this thing. You start using this thing, and you're like, this is tiring me right now. It's too much. Or someone comes up to you and says some, like a new student or someone comes and says, what should I be doing? Should I be doing ML courses right now? Should I care about coding at all? Or should I learn something completely different? And also, as a show of hands, how many of you felt this pressure to say, I need to pick the right tool today? I need to use Haiku for tests. I need to use Opus for writing code, and so on. Have any of you felt this way? OK, a good chunk of you. Now for the ones who didn't raise your hands. What are you doing? How are you keeping calm in this thing? This has been stressing me out. But anyway, so we started talking about this thing about a year ago, more or less, where we said, we went from this cycle after cycle, so something new comes up, we are excited about it, we start using it, and then we are tired about it. And then the next thing comes up, we use this thing, and then we are tired about it. So we went from using GPT just as a chatbot, then to integrating it into, like, copy-pasting code from IDE to GPT and back, then started using Cursor, then used Cloud Code, then whatnot, and used Skills and MCP and whatnot. So we were, like, kind of really getting tired about this thing, excited and tired in this loop. Then we started speaking to our friends and colleagues, and they all were kind of feeling the same way. So this transcended any of the domains. They were in academia, who were in management, who were students, or friends in something completely different. And all of us were kind of feeling this thing. And we said, okay, we need to kind of start looking into data about this thing. So then we were like, are we hallucinating this thing? Maybe this is there. So we are not. So clearly Google Trends shows that many people are feeling this thing. So it pretty much didn't exist. took off in 2024, and then like an upward in 2026. So clearly, this is not something that just we are hallucinating. You look at AI fatigue, you Google this thing, you will get a summary, ironically, by AI, what it thinks is the most, what is AI fatigue. There are very recent articles about this in CNN that says AI fatigue is something that's catching up, and there is brain fry that's happening. And also, there is a Harvard Business Review article about this thing. And they've actually gone into the research of what happens when you have so much AI thrown at you. So for a more defined way of this, I'd like to give this to Jayashree.

Speaker 3 [08:33]

So I think Ajay kind of very nuancedly described this whole emotional rollercoaster, and that's pretty much AI fatigue. It's the mental, emotional exhaustion that you accumulate, or if you're really good, you are able to diffuse over the day, because you're constantly bombarded by all these new things, that is AI, on top of the things that we were bombarded before, like Teams, Signal, WhatsApp, whatever. So all of this thing, this feeling that you go through because you're constantly context switching between doing stuff with humans and agents and agents, humans and multiple agents is more or less what I would see as AI fatigue and it's also formally defined so. And eventually, of course, it could affect people's productivity and may even cause burnout. And here there are a bunch of different things happening, right? On the top left is the AI brain fry. That's the kind of the mental fog Maybe you guys would have seen, felt, experienced, whatever, where you move between things and then you forget, oh, I went to that window, but I wasn't sure why I went there because you are handling multiple things at the same time and the concentration is lower. And the productivity paradox is that we are all told the more you do, it's better, but at some point, it's not like, you know, linear function, right? You plateau and you're like, okay, actually, I cannot do more, but I have to do more. and then of course you can feel the constant symptoms in terms of how you feel motivated or the lack of it and there are multiple drivers in this from the environment and from self that that kind of goes into this vicious loop and in we have felt that there are basically two kinds of things that you face every day that i try to be conscious about number one is the thousands of micro decisions that you now need to take so previously we were taking fewer decisions now we are taking more decisions. And second is the FOMO fatigue, right? Like, what if I didn't read that paper or article? I'm going to miss out on something really cool. And then someone's going to do better than me. So that constant thing of FOMO or I'm missing out on cool stuff and changing on a daily basis all adds up. And of course, there is a grand schema of things like the general world stuff with AI slop, the misinformation, the pressure that you're going to get displaced tomorrow, the lack of security, and then the overall existential vibe that we get, oh, AGI is going to come and then take us all over kind of stuff. So all of this thing is now kind of bringing up this paranoia, at least for those of us who are in our circle. And then we realized we should do proper data analysis on this thing. And I give it to Ajay to explain that part.

Speaker 2 [11:10]

in this before. So we've observed this, we've read about this, we said, okay, we need to dig deeper into this, at least from a perspective of really looking at the data, because there are two possibilities. This is true, this is indeed happening, or we are just being bombarded with it because of social media. So we said, let's try and dig deep into this. And we said, let's look at the two topics that are maybe related to us. One is research papers. So in academia, it was something like a big part of it. Second is software, coding, and so on. Specifically with research, we wanted to answer three questions. Is there genuinely more research happening nowadays? Is this across the board, across all topics everywhere? And is this research getting harder to read? So those were the kind of three questions that we wanted to answer. So research output, if we plot the top three ML conferences we see has been growing exactly exponentially. Now this is ICML, ICLR and NeurIPS, just the top three. You could argue these are all like peer-reviewed, everything is good, sanitized, you could say yes. But one thing that we are majorly missing is actually archive. So if you look at archive, which is not actually peer-reviewed, you see that it's a huge, huge amount of work that's actually being put up in archive. Now, this is, of course, it's a staggering amount of research that's happening. And it is obviously, like, no one person can actually look at this thing fully. But the good thing is, all of us will get an ML paper one time or another. And we were like, yeah, this is going to happen. But when is this going to happen? So you look at the projection, It seems like by 2068, all of us will be NML author papers. So it seems quite optimistic then. But going back, so is this true across... So at least it's obvious that there's a lot of research happening in this area and a lot of research output is coming out, definitely also because of AI. Now is this true across all fields? So we said, let's take, as an example, we'll take a field that hasn't been disrupted. High energy physics, for example, not everyone has a large Hadron Collider in their backyard. So we said, let's look at that. So there it seems like the number of papers published is kind of at a reasonable rate. So it's not something that's going up exponentially. By the way, the two peaks, it's not that people got amazingly productive in 2009. That's just archive re-indexing it. So that's why those two peaks in 2009 and 2015. So even we were surprised at what happened in 2009, so we're like physicists, amazingly popular, but it's just re-indexing. But it goes to show that this is indeed a problem of ML, or this area necessarily, not true across many of the other disciplines. So this was the second question, so this is not true across all disciplines, but some of them are indeed being extremely research productive. The third thing was acronyms. Oh, I've realized that every paper comes with its own acronyms right now, and acronyms make it really hard to read. So there is this idea that cognitively you read an acronym, and then the acronym comes somewhere else, and you kind of need to load it back into your RAM to say, oh, what did that acronym mean? So it kind of disrupts your flow of reading. So it goes to show that we have too many acronyms, and that is also true in ML specifically. So it seems to have been a trend across many domains, but high-energy physics seems to contain it, while ML just keeps going up. So it does seem like acronyms is also a problem that we are facing, and it is indeed an issue for cognition. So when we were doing these exploratory things, we thought, okay, these forced acronyms seem coming up. So some of the popular acronyms that we thought that were interesting were pasta, crap, attending, which I think is a crime, to be honest. Attending is a nine-letter acronym with five coming from the first word. Come on. If it's from any of your papers, I already apologize. But we thought some of these acronyms were... Everyone wants their own acronym. So I get that feeling, but I feel like we are overdoing this. So that was one part about research. So this is actually an idea from the talk two days ago about hype, hope, or headache. They had a statement that said there is a new GitHub developer almost every second. So I thought, oh, let me see. I don't know about developers. It's kind of hard to get the data. So I thought, let's look at the GitHub repos. And that seems like it is quite going up also exponentially. So I got this data from the Wayback Machine, so I parsed the GitHub repo data and also the total PI packages, and both of them seem to be more or less exponentially going up. And if we look at it basically between the last two years, or 25 to 24, that's about 200 million GitHub repos, and we have 30 million seconds every year. That means that's about eight repos every second that's getting open source. So that's a lot of repos to keep track of. So this is what we felt, and the data seems to back us up in this case. So then we said, okay, how has it affected our lives and what can we do about it? So I'll give it back to Jayashri to talk about that.

Speaker 3 [16:45]

Thanks, Ajay. So we saw all the data, and I just wanted to put it across that, as he mentioned, it's going across multiple levels, right? As a developer or a technical lead, on an everyday basis, you're seeing all these changes. We start from RAG, and then you went to agents, and then orchestrating them, and then MCP, and other tooling capabilities, and so on and so forth. As a technical manager, you're constantly making these high-stake decisions. Should I go in this direction? Should the architecture be dependent on this or that? Which is quite fast evolving. As an ML researcher, well, we saw the papers. You have enough stress already being in academia and then double stress because you're now having to keep up with the papers as well. On the business and the management side, well, you have to do the AI strategy. And then we all know AI strategy is very hard because everything is changing all the time. So every one of us is somehow exposed at our own level. It's the same storm, but we are in different boats, but we all face the same kind of problem, I would say. And what we have in the stock is not like a silver bullet. Hey, we have figured it out. We know what we have to do to cope up with this thing. We don't, because it's going to change tomorrow anyway. But there were some coping strategies that has been helping us to be at least conscious about it and then not be on the rat race, so to say. These are some five, six stuff that we consciously thought out, wrote down, and also ideated with our friends, which might be useful in this talk. The first thing, and the most important thing, is information diet. We are in the information age, and there was a point in time where being connected was the cool thing. Ironically, I hope I won't be shot down for telling this, being disconnected would be the cool thing right now. So if you pick three to five trusted sources and just be connected to or subscribe to that and get all your updates, that would be really cool. And I would go on the mission of unsubscribing aggressively, if possible. Number two is timeboxing. So I'm not sure I'm happy to discuss your experiences, but for us, more than one, two hours of working with agents, clock code gets a bit exhausting because you're doing multiple things at the same time. So I now have a timer where I timebox it and then just continue working on it and then get out once the time is up because the context switching costs too much energy. Third is knowledge base. So there was a time when we used to write in notebooks, and then on tablets, and then now we take notes on the computer, and now we are like, LLM will take the notes anyway. But I feel like we have lost the learning process. I think we learn when we write and when we do stuff ourselves, because in the end, it's about also remembering what we have to do and what we learned out of it. So I have kind of backed out, at least in some meetings or calls, to make handwritten notes or notes on your favorite software notion, for example. and that really helps me to keep track of things and slow down, so to say. Fourth is to use AI with intent. I primarily for the text aspects used only for summaries and drafts and so on and don't use it or abuse it for everything. Next is JOMO, which is, of course, we need to create an acronym, right? So it's like, it's the complimentary of FOMO, which is joy of missing out. You will miss out, I guess. That's the truth. We see the stats. So just probably cool to just embrace that we wouldn't know everything that's happening and probably not be able to understand everything as well. And the last piece is small experiments. I think one of the coolest things I feel about the whole AI wave is how quickly I can try new ideas. So just trying small experiments out and ideating quickly and then taking it to the next stage has its own kind of fun that we used to have back in the old days when we used to fully code. and then the seventh one it's always one that I really cherish which is learning in public I feel like I learn best when I explain it to someone and that always is something that we need to keep going on keep doing it whether LLM does it or it doesn't it doesn't matter for the sake of our learning we need to keep that going and the last one an interesting one is individual coping strategies this reddit post I was just 20 days back I saw when we were googling around and this person has put up a fatigue is real. And deadlifts really help him or her to cope up. So I would empathize with the guy I really like sports. So for me, sports is about like, you know, de stressing. So really, this is the moment to find our hobby and creativity to kind of vent it all out, I think. And of course, there are probably more but happy to hear what you guys think. And very quickly, I want to also, I'm no expert at this, but I think what would be nice as an as for me, if organizations overall could could stimulate is number one if goals are set clearly for the next one two three months and not constantly changed to a lot of hands-on training as these new new coding tools come if we have like training on a monthly basis hey there's this new thing coming let me take it up on me to explain to you so if there is some organized training and feedback loops right because that's that's always important with the customers with the users whatever and Most importantly, focus time. We have a really nice culture in our organization where you have blocked focus time over the week where typically you don't have the onus to take up meetings. This I would really, really cherish and it would be cool if we can expand on that going forward. The second is the governance space. I know it's all also a bit top-down, so if we don't change stuff too often, that's also better than if it's changing on a weekly basis. Also, there should be norms for data use because right now we are kind of putting things together and sometimes we are not sure, can we actually give this data, can we not? So there should be clear norms around that. And of course, there will be a lot of pilot projects coming because now we can code quickly. So we need to have logic around when to sunset things and when not. So with that, I would like to slowly start wrapping up this talk. The three reminders which I tell myself, which might be useful for all of us here. Number one, you're not behind, even though it feels like that. number two the goal is not to know everything even though it feels like you need to know everything because we are living in the information age number three it's always always to the best of our benefit to stay curious and stay sane i guess our superpower is selective ignorance perhaps you would agree so with that um i would like to give it to you all guys to to if you have any thoughts ideas, coping mechanisms, please share. We are happy to learn, B-Y-O-C-M. And please feel free to connect with us, get in touch, and happy to have a chat. Thanks a lot.

Speaker 1 [23:29]

We have a couple of questions in the chat. So first, as an application maintainer, I think it's maintainer, I think the word is missing. As an application maintainer, one of my biggest fatigue points is reviewing PRs. We are still at the point where it is risky not to have a human in the loop when it comes to code submissions. However, the volume of code generated has increased and will continue to increase thanks to agentic tools. Do you have any thoughts on how we mitigate the extra work created by the volume of new code generated these days?

Speaker 2 [24:06]

I think I'm kind of on the same boat, to be honest. So I get quite a few PRs that I will have to review. I think there was something that I was reading where many of our brains are kind of built to write code, not necessarily read someone else's code. And now we have pretty much no writing to do and only reading to do. I think it's probably all we can do is take our time with PRs. I know it's annoying for the developer. you put a PR, it's been a week, it's not been integrated yet, but I think it's best if we actually take our time to review this. I've also felt there's a lot of institutional knowledge that many of these LLMs miss, so it is quite critical that we still have a human in the loop, but honestly, I have no way how to say, I mean, I have 25 PRs per day, how do I mitigate this? No idea.

Speaker 1 [25:01]

Another question was, I don't suffer from AI fatigue, but my partner does. How can I support her?

Speaker 2 [25:16]

You want to take it?

Speaker 3 [25:18]

Well, I guess food helps, sports helps, you know, the standard stuff, like taking weekends out, traveling somewhere, like getting life, not looking at the computer. So maybe just like triggering them to like not look at the computer might be a good idea. It's good that you're aware that your partner has this issue and you, you know, you have this idea now that these tools could help. That is option one. Option two is, of course, help them with the code, if that's a possibility.

Speaker 1 [25:52]

Another question is, can some of the heavy increase on publications on archive be explained by a lot of the papers being AI-slopped?

Speaker 2 [26:05]

I think that's true. That's most certainly true. So we don't have a direct way to know this, but we thought, let's say a proxy way to look for this is, I don't know if you've noticed that all of us have colons in our name, like titles right now. So most, because LLMs tend to generate, you do blah, colon, blah, blah, blah happens. So that kind of thing. So we started just looking at, is the number of colons in titles going up? And it's true. So there is definitely a lot of AI generated. Is it slop or not? I think that's up to the reader. But yes, certainly lots of them are AI generated.

Speaker 1 [26:41]

Okay, thank you. We have had hype in the past. What is different about AI?

Speaker 2 [26:49]

I think, in my view, this is purely personal, I feel like the fact that it's affecting a lot of, I think it's different basically because it seems to be touching pretty much every area that's there. So many of the previous hypes that I've lived through has been in fairly contained areas. So, okay, autonomous driving is coming or it affects all the people who are in the driving or logistics industry. or if there's another automation coming then it affects the manufacturing industry but this seems to be something that's across the board so I think there was a recent study or like a recent publication from Anthropic which kind of said what percentage of jobs are going to be like susceptible by what percentage of automation and it seems to be pretty much across the board there are a couple of areas that are hard to touch so like nursing or veterinarians and so on but apart from that pretty much it seems to be across the board. I think that's probably the biggest difference that I see.

Speaker 3 [27:48]

And plumbing, I want to be a plumber.

Speaker 1 [27:52]

yeah we all have to come up with a plan b right yeah um one question which is a nice one to finish a great talk if you had to summarize in one sentence what would be your recommendation to stay sane

Speaker 2 [28:07]

I think somehow get away from your computer for some time. So I think just closing it, say done. I do something else for some time.

Speaker 3 [28:18]

I would say change always happens, accept it and try to live with it.

Speaker 1 [28:22]

here. Okay, thank you very much for this great question, and thank you to our great speakers.