Lead & Follow Podcast — Season 5, Episode 5
Guest: Elena Filatova · July 17, 2026 · 24 min
Elena Filatova, who spent seven years at Google in data infrastructure before founding BCB Code Labs, walks through what happens under the chat window: the attention mechanism, why swapping one word rewrites a response, how assigning a persona filters the training data, and why hallucinations are statistical rather than dishonest. Also covered: energy costs, and the back-and-forth that keeps a person in the lead.
Listen to this episode on Buzzsprout →
Questions Covered in This Episode
Why does changing one word in a prompt completely change the AI’s answer?
Elena Filatova points to the mechanism called attention, from the 2017 Google paper “Attention is All You Need”. Her image for it: you’re walking with the ocean on one side and grass on the other, your eyes drifting, and someone says the word “storm” — suddenly you’re looking at the water. Ask an AI to analyse a statement you find “interesting” versus one you find “confusing” and the spotlight lands somewhere else entirely. Sharna Fabiano had tried exactly that in Elena’s workshop and got two piles of evidence for opposite conclusions, from the same statement.
Does telling ChatGPT to act as an expert actually improve the answers?
Elena says yes, and it’s the easiest win there is. Tell it it’s a mathematics professor and it’s more likely to reach for the parts of its training data that came from university sites and scientific literature. She describes all that data as “mushed into” a cake — like baking an egg in, where “the egg is no longer an egg” — so naming a persona is a way of pulling at one layer of the cake. Sharna said the technique was “worth the price of admission” because it changed her question to: what kind of perspective do I want?
Why does AI make things up and still sound so sure of itself?
Because it isn’t a lookup engine, it’s a statistical one — predicting the next most likely word, not checking whether anything’s true. Elena’s careful about the word lying: that would imply it knows the difference between a truth and a lie, “and it really does not”. It may sound like it has authority, she says, when it just sounds confident, so the job is to read the answer and ask whether it resonates or seems a little flat.
How do you get more precise and less vague answers out of an AI chatbot?
Give it context. Elena says to think of it as a brand new baby with no idea who you are or what matters to you — it won’t even know a plant has no internet access. She also suggests asking for a list when you don’t yet know what’s relevant, then pulling on whichever items look useful. And she works iteratively: paste the first answer into a different chatbot with a critical persona and go back and forth until you converge on something that feels right.
How much energy does AI actually use, and which uses are worth it?
Elena draws a hard line between text and video, music or images — she says the energy involved isn’t even comparable. There’s a data centre somewhere eating land, water and power while employing very few people, so she thinks the honest question is which projects merit that expense. Sharna raised pricing out the high-compute functions as a way to conserve energy, and Elena agreed: if generating “AI slop” got expensive enough, it’d naturally fall off. She also notes that finding a source takes far less energy than generating something new.
Join the VIP list to hear about new episodes
Read the full transcript
Welcome to lead and follow a podcast that explores how great collaboration actually happens. I’m your host, Sharna Fabiano, author of the book, lead and follow the dance of inspired teamwork. And I’m pleased to bring you the latest research insights and learning methods to help you work together better with the people around you. I help founders, solopreneurs and new leaders become great at collaborating because I’ve seen how much more becomes possible when people integrate their ideas, perspectives and strengths in a practical way.
To learn more about coaching with me and for a one-time discount on your first session, go to sharnafabiano. com slash coaching and enter the promo code podcast. The link is in the show notes. And now here’s the episode.
Today, I’m speaking with Elena Filatova. Elena spent seven years at Google in data infrastructure and business intelligence before going independent. Now she builds AI driven applications, helps businesses automate with AI and teaches AI literacy and governance. I attended one of Elena’s recent workshops right here in long beach.
And I learned a lot of things very quickly that improved my own use of AI assistance like chat to BT or cloud or Gemini. It got me thinking about what it means to lead and follow in a quote unquote conversation with AI. I think it’s an important question right now. And I’m eager to learn more from Elena today.
Elena, welcome to the podcast. Thank you, Sharna. Hi, everybody. I’m excited to be here.
And I’m excited to talk about this topic, especially from the lead and follow framing. Lately, I keep hearing people say the term manage when we’re working with AIs as if it’s a coworker, especially engineers who think that it’s almost other junior engineers they’re working with. I think it’s interesting to explore whether it’s really a fair comparison and why it might be true, why it might not be true, if it’s actually different working with AI versus working with a person and how. I too am very curious about all of those questions.
And in particular, when you mentioned this common comparison to managing AI as if it’s a junior engineer or a junior employee, I am really struck by how automatic that hierarchical thinking is in our organizations. Even when we’re no longer interacting with a human being, somehow the conceptual framing is still that one needs to be superior to the other. What I’m interested in, of course, on this podcast is using the terms leading and following like we do in dance as complementary and fluid, meaning they can switch back and forth and are not necessarily hierarchical. I want to help people get out of that default thinking because although sometimes it’s appropriate, more often than not, it’s a huge limitation.
So I’m very curious to hear you talk about leading and following with AI. But before we get there, I’m wondering if you can just give us a high-level summary of how these AI systems work. Sure. So when you open the AI assistant, you aren’t just directly talking to the model itself.
Like some people think there is an application layer around it. So there’s the front end where the friendly chat window is. It used to be forms, drop downs, buttons, and now it’s just a text box that really makes you feel like you’re talking to a person. I’ve actually read some research on that about how that makes you feel more comfortable versus, let’s say, filling out a form in an old-fashioned way.
Then underneath this front end, there’s the back end that you don’t see. It passes the text that you entered to the model, stores your chat history, and holds security and privacy controls that check things before they reach the model itself. And then underneath all that layer, there is the actual large language model itself, which does the thinking part. Thank you.
That is super helpful. What is the model itself then, and how does it, quote unquote, think? Our alums don’t think in a human sense. They’re just predicting the next most probable token, which is just a word or a piece of a word, given everything that comes before it, based on patterns that they learn from massive amounts of data.
These are statistical predictions at a massive scale. The mechanism underneath it all is called attention. It was introduced in the paper, Attention is All You Need, which came out in 2017 from Google. To help illustrate that mechanism, imagine that you’re walking down a path.
On your right, there is the ocean, and on your left, there is a piece of grass, and there are birds in the sky, and your eyes drift from one thing to the next, taking in one piece at a time. Now, imagine that somebody is walking next to you, and they say the word storm. Now, your eyes suddenly drift to the ocean. That’s the word that redirected your attention out of everything that’s around you.
And the large language models do something similar. For every word in a sentence, the model shines a spotlight on every other word. It’s figuring out how much attention to pay to each one word, given the context. So, for example, if I say to the LLM, analyze statement X, I find that statement interesting, versus analyze statement X, I find that statement confusing.
That single word difference, interesting versus confusing, changes where the spotlight ends up. And in the first case, the model more heavily pays attention to the interesting parts of the statement X. And in the second example, the model will pay attention to the confusing parts of statement X. And so your response ends up being very different.
Thank you. And that explains why when we did this experiment, you know, in your workshop, the model then gave all the evidence for why the statement was confusing versus, you know, the second time it gave a whole bunch of evidence, you know, for why the statement was interesting, like same exact statement. Exactly. That is fascinating and a little bit scary.
Given that the assistant is paying so close attention to every word that you say and kind of reflecting back, you know, in a way your own, I almost want to call it like a bias, you know, you’re kind of feeding it or leading it with your own language. What might be a good first step in shaping our own inquiries, right, our own kind of leading inquiries in a way that helps us and makes, you know, makes the responses we get back more useful to us? The one thing that you can do that is pretty easy and immediately useful is to give the AI a persona. For example, if you tell this language model that you are a mathematical professor, it’s more likely to reach for the parts of the data that it was trained on that came from university websites and scientific literature, and it’s more likely to produce an accurate mathematical response to your statement.
So giving it a persona is sort of acts as a filter. Think about in our day-to-day interactions, we all have our own persona. We all have our world view that we perceive the world with, and so some things sort of filter out and other things stay for us. So in a way it is similar, except that LLM model encompasses the entire world’s knowledge and by giving it a persona, like you’re a mathematical professor, you’re a psychology major, or you are a chef, gives it a way to filter down the knowledge to something that might be relevant for that persona.
Thank you. This is so, so useful. I want to stay on this point for a little bit because I remember when I learned it in your workshop and tried it out, I immediately thought, oh my goodness, this one technique is worth the price of admission. One of the things that overwhelms me about AI is that I just don’t know what’s happening in the model, and I know it’s not intelligent the way a human is, but I don’t know which responses I can trust and which ones I can’t.
So what this persona technique does for me is demystify the mechanism and I ask myself a totally different question. I ask instead, what kind of perspective do I want? And that just feels like such a more helpful question. Could you talk a little bit more about why this persona technique, this filtering, works so well and why it makes such a big difference?
Right, so these LLMs have been trained on massive amounts of data. Some of that data came from, let’s say, Reddit. Other data came from books, and other maybe came from its own responses. If you’re using a free version, for example, then your data is being used to train the next generation of AI.
It may have come from social media. The truth about how these AIs work is that they no longer have the reference to the original source of the information. All that data, wherever it came from, got mushed into kind of like baking an egg into the cake. The egg, by the end, when you have the cake, the egg is no longer an egg.
It’s part of the cake itself. So it works similarly in that way. All this data got mushed into this cake. And so when you say you’re a mathematics professor, then you’re sort of pulling at a layer of the cake where there is that university data.
What else can we do as the leader in this relationship, the prompter to guide and direct the model? So the first technique about defining a persona is about giving AI sort of a filter to filter on the information that is relevant for you. But sometimes you may not know what is relevant for you. And in that case, asking it for a list is helpful because that sort of helps it surface different kinds of layers in the cake that exist.
And then once you get the list of, let’s say, which areas of research should I explore, then it will provide you with a list of different areas that it thinks you should explore. And you can pull on each one of those areas when you continue interacting with it, whichever ones seem relevant for what you’re trying to do. I really like this one because it seems to sort of naturally support the human thinking process rather than attempting to replace it. I can imagine, for example, how the request for a list might help me think more precisely about what information I want to find.
And speaking of precision, I notice sometimes if I ask a question, the response can sound smart but is actually pretty vague and unhelpful. So what might we do to get more precise and useful responses from these tools? Yes, you can provide to it context. That’s a very valuable technique.
The more context you provide to AI, the more accurate it’s going to be. Think of it as completely unaware, a brand new baby that came into the world that has no idea of what was going on, where you are, who you are, what is important to you. It will not even know that a plant, for example, has no internet access. It will totally think that it might because it has no such context necessarily if you don’t give it to it.
So sometimes it gives you a nonsensical response. It’s because it’s just not aware. It has all this data and all this information, but just beware that without context, it will not give you a good response at all. That is very helpful and it reminds me of how much context is written about in both the leadership and the followership literature.
There’s actually quite a bit of evidence that shows us that we can’t actually collaborate with other humans effectively without understanding the context we’re in. And it sounds like that’s even more true here when we are collaborating with a chatbot that has zero context of its own. So we really need to provide that if we want the interaction to be useful. Okay, let’s look at this relationship from the followership role.
When we are receiving these responses from the chatbot, what can we do to stay mindful of our own human thinking and how we are absorbing the material that AI is giving back to us? Right, so I think we’re definitely more used to interacting with other humans who do have that built-in context like you mentioned, but the LLM does not. So I think our interaction and the way we receive the information needs to change because you can’t assume the answer it’s giving you is informed in the same way that a human answer would be. For example, you can’t assume that it won’t in case of software engineering, for example, just delete the production database.
It absolutely will do that and it won’t care about that either. A human who made a catastrophic mistake such as that would be absolutely horrified about making the mistake, but AI does not care in the same way. And that’s a feature. It’s not necessarily something to be alarmed by or scared by.
It’s just something to be aware of when we interact with it. This is also where the word hallucinations come in. The model isn’t really a lookup engine. Like I mentioned before, it’s a statistical engine.
It’s predicting the next most likely word, not checking whether something is true or not. So when it’s making something up, it’s not lying to you because that would imply it knows the difference between a truth and a lie, and it really does not. And so it’s important to read its responses and think critically about what it’s actually saying. Does that resonate with you as a human internally or does that seem a little flat?
It may sound like it has authority, but often it does not. It just sounds confident. Something just came up for me as you were speaking. When I’m interacting with these AI bots, I notice I will often have an emotional response to them because the language is phrased in such a way that it feels similar to what a friend might say to me.
It can be very affirming and reassuring. And I catch myself feeling that, oh, hey, it gets me or understands me, when really that’s not at all the case. It’s just putting together language patterns that happen to have meaning for humans. I wonder if you have any thoughts about that.
When we have human emotional responses to these things, is that dangerous for us? How do we manage that? I think as long as we’re aware that it’s a tool rather than an actual intelligent being, I think that if it comforts you sometimes in a time of need, if it gives you something that helps you sleep better, if it reassures you or supports you in some of your ideas, I think that could be very helpful. The danger is when you believe everything that it says unconditionally and absolutely.
If there is an awareness of what it is, right, and there is that question on the back of your mind of, I still need to talk to another human being, a friend, I think that it can be still helpful to walk through your life situations with a chatbot. Yeah, that awareness piece really seems to be key. Here’s another followership question, I think. People talk to me about the environmental implications of using these tools, how much power they use, how much resource they use.
If we’re going to use them, receive the benefit of them as followers, perhaps, how can we also be environmentally responsible and what does that even mean? Yeah, that’s a great question, Sharna, and I think a really important part to understand here is the difference in the amount of energy that, let’s say, text generation requires versus video or music or even photos because that’s not even comparable. It’s also important to remember that these models do take massive amounts of energy. It’s important to remember there’s a data center somewhere that takes up a lot of land and water and energy and resources and employs a very small number of people on its own.
So, I think it is important to ask ourselves what kind of projects and products merit this kind of expense. This may be a conversation for another day, but it does make me think about regulation or pricing out those high compute functions and if that might be a way to conserve energy. Do you have any initial thoughts on that? I mean, not very deep thoughts.
Like you said, I think it’s a conversation for another day, but I would agree with that. I would agree that this is exactly the kind of area that I think should be maybe priced out because, I mean, yes, if it’s too expensive for anyone to generate this kind of AI slop, then I think naturally it would fall off. Elena, thank you so much for sharing all of these super helpful tips and demystifying how AI chatbots work. Are there any other thoughts you’d like to share about how we can have healthy lead and follow relationships with these tools?
One useful tool can be to ask these chatbots to give you a list of references, a list of links, resources. You can even ask for, give me the source for where in the book this certain passage is. So, point you back to the source rather than how it generates something from scratch. And, in fact, I was just thinking about how much more energy the generation process takes versus just a discernment.
So, it takes a lot less energy to, let’s say, find a source for something versus to create brand new information or brand new document for you. So, I think the important takeaway is to use these tools in a collaborative back and forth manner rather than engaging with them as if they are intelligent and have their own mind. Elena, when you said collaborating just now, it made me think of a human collaboration where it’s not just one question and one answer, but more of an ongoing process usually with a human partnership or a human team where you’re going back and forth with ideas and letting things emerge. And I wonder if you feel like that is a way to improve our use of these tools as well.
Yes, that’s a really great point because I do think that going back and forth and collaborating is a better way to engage with these tools because rarely do I find that the very first response is perfect. You know, I often take the first response, I put it into a different chatbot with a different persona at the top, I say, now you’re a very critical person, analyze this response for X, Y, and Z. And then I go back and forth until together we converge on an answer that feels good to me and where I can see there is like fewer criticisms from the AI as well. I find that process actually very productive and quite interesting.
Thank you so much. I really appreciate how you keep coming back to checking in with yourself as you go back and forth with the AI and using that as your compass rather than just accepting whatever it gives you. And with that, I just also wanted to add that I think there’s a lot of misinformation out there that you may hear where people talk about this super intelligence, AGI, etc. Personally, I find that this is not a useful conversation to have.
This is a lot of hype that’s not really grounded in how these models actually work. And a better question I think is whether what these models output is actually useful to you in your day-to-day life. I think that’s a great note to end on. Is it useful, right?
And I think you’ve given us lots of ways to make these interactions more useful. I know I’ve appreciated them and I’m sure listeners will as well. You can get in touch with Elena at bcbcodelabs. com and we’ll link to that in the show notes.
Elena, thank you again and I look forward to our next conversation. Thank you, Shardna. You have been listening to the Lead & Follow podcast. Special thanks to Glover Gill for composing our music and a huge thank you to all of you who support the show.
It really means a lot. If you enjoyed this episode, please share it with a friend or consider joining the Lead & Follow fan club for as little as $3 a month. There’s a link at the bottom of the show notes.