In short
Podcast Summary: The Neuron: AI Explained - Episode: The Humans Behind AI
Episode Overview In this episode of "The Neuron," hosts Grant Harvey and Corey Noles dive into the essential human element behind artificial intelligence (AI) with Casper Elliott from Invisible Technologies. The discussion revolves around how humans train AI models, the importance of data quality, the common misconceptions about AI deployment, and the future of work concerning AI's impact on jobs.
Key Topics Discussed
The Role of Humans in AI Training
- Invisible Workforce: An "invisible army" of humans is responsible for training AI models by labeling data, rating responses, and teaching models what is right and wrong.
- Training Mechanisms:
- Supervised Fine-Tuning: Providing models with high-quality examples to learn from, akin to taking a student to the library.
- Reinforcement Learning: Involves giving models sets of questions and evaluating their answers to improve their accuracy.
- Evaluations: Essential to understand model performance; human feedback is crucial to ensure that the model meets user expectations.
Importance of Data Quality
- Quality Over Quantity: Casper emphasizes that the quality of data is more critical than sheer quantity. A model trained on high-quality data will perform better than one trained on vast amounts of poor data.
- Human Feedback: Continuous evaluation and human feedback help identify biases and improve the model's performance.
Misconceptions About AI
- Not Just Internet Data: Many people believe AI simply gathers data from the internet; however, considerable human effort goes into training and refining AI models.
- Models vs. Applications: AI models need an application layer to interact with users effectively; improvements in applications can significantly enhance user experience.
Future of Work with AI
- Human Oversight: Despite automation, the future will require human input, especially in understanding complex tasks that AI may not handle well.
- Job Displacement Concerns: Concerns about AI automating jobs are prevalent but may overlook the new roles that emerge in AI oversight and management.
The Process at Invisible Technologies
- Working with Clients: Invisible Technologies collaborates with clients to assess their needs and determine how AI can solve complex workflows.
- Understanding Processes: Clients must have a clear understanding of their processes to successfully integrate AI solutions. Often, existing processes must be reimagined instead of simply automated.
Predictions and Trends
- Data as a Strategic Asset: Access to high-quality training data will become a significant competitive advantage.
- Increasing Complexity: The demand for specialized skills in AI training will grow as more nuanced and complex AI applications emerge.
- Physical AI Integration: Future advancements may lead to more applications of AI in physical environments (e.g., robotics, autonomous vehicles).
Final Thoughts
- Human Element: Casper reinforces that while AI tools are evolving, the human element in training, evaluation, and oversight remains vital for effective AI deployment.
- Emphasis on Learning: The conversation highlights the continuous learning and adaptation required in the rapidly changing landscape of AI.
Conclusion This episode of "The Neuron" provides valuable insights into the often-overlooked human side of AI and its training processes, emphasizing the need for quality data and human involvement in deploying AI solutions effectively. The discussion encourages listeners to rethink their approach to AI in business and society.
Additional Resources
- Invisible Technologies: [Invisible Technologies Website](https://invisibletech.ai)
- Neuron Podcast: [The Neuron Newsletter](https://www.theneurondaily.com/subscribe) for more insights on AI developments.
--- Feel free to explore these resources to deepen your understanding of AI and its implications in our world.
Written by AI. May contain mistakes. Listen to the episode to check what was said.
Transcript
Automatic transcript. May contain errors.0:00Casper Elliott:So behind every AI response, there's an invisible army of humans who trained it, labeling images, rating answers, and teaching these models right from wrong. Today, we're going to talk to Casper Elliott from the company that's trained 80 % of the world's top AI models.
0:23Casper Elliott:So welcome, humans, to the latest episode of the Neuron Podcast. I'm Corey Knowles, And we're joined, as always, by Grant Harvey, writer of the Neuron Daily AI newsletter. And today we're diving deep into the human side of AI with Casper Elliott from Invisible Technologies. Casper, thanks so much for joining us. Pleasure to be here. Thank you for having me. So, Casper, Invisible Technologies just raised$100 million. Invisible also says it has trained 80 % of the world's top AI models. Both of those are incredible stats. Can you just tell us a little bit more about what that process actually looks like?
0:59The way I think of it is, so large language models, they're not like traditional machine learning, right? They're non-deterministic. They're based on neural nets. They do some funky things. I think of a large language model like an enthusiastic teenager. Yeah, it really wants to answer questions, and it wants to get smart. But if you want a teenager to learn something, like there are some ways you can teach them, right? You could, for example, you could take them to a library, or you could give them a load of homework, or you could set them a test. We kind of do all those three things. When you hear supervised fine-tuning or reinforcement learning about human feedback or evaluations, that's actually one of those three things.
1:35So supervised fine-tuning is giving a model loads of real high-quality examples of data sets to look like. That's taking your model to the library and saying, here's some textbooks to read. It's going to read the textbooks. They'll tell you what's true. Reinforcement learning is, okay, you're going to give the model some questions. it'll give some answers and you're going to say if those answers are good or not like you might ask the model to write me a poem about um russia and then you'll you'll check that poem and you'll have something about what if that poem is good or bad and you'll give it so great yeah it'll learn from that and so that's reinforcement learning or you can call it reward modeling you're basically allowing you to change the way you reward your model for different types of answers and then evaluation is building the like the test the model has to take to understand if it's good because companies will release loads of different versions of models and they've got to understand if it's better or worse.
2:22Like, I mean, you've seen the news about Tractact GBT5, GBT5, they released it. It was obviously better in some metrics, but the audience wasn't happy. And that's why human evaluation is so necessary because people are non-deterministic too. People like to have opinions on things and you can't just be like, well, this was better on all our benchmarks. If it feels different to someone and the user doesn't like it, it doesn't matter if it's better. And that's evaluation. So we're kind of, we're the teachers behind the models.
2:46Casper Elliott:I think a lot of people believe that all that's happening is AI is scooping up the internet and just training itself on all of that. And then it just magically works. Can you elaborate a little bit on how that happens as far as like a little more technically? Have you met the internet? Like if AI is running off the internet, like we're in trouble. yeah um i mean ultimately a model is gonna it's gonna it's gonna look at its huge data set and then you're gonna have a a large an impossibly large set of hyper parameters that you're gonna configure to try and understand what the best next token to predict is based upon the data it's got it looked at but ultimately you've either got to improve the underlying data set which is hard like the data sets are huge they'd set to petabytes of information or you've got to do what's called post-training where you you use sort of smaller sets of data to improve the weightings like for clarity by the way i am not an ml scientist so i'm going to say some things and that they'll be as being like this guy's a charlatan so just for clarity before i get flamed uh and i'm giving some vibes here rather than the exact science um and then i think the evaluation speech is so key like there's all these different models and in lots of ways like you could go if you go to chat gbt claude gemini and you ask it a reasonably simple question it's going to give you the right answer pretty much yeah and or if you ask if you give it a complex subjective question it's going to give you some different answers but how the hell do you decide which is better that's where evaluation is so critical i think andre carpet someone has to read well andre carpet had a tweet about six months ago where he said like the entire ai is just an evaluation problem now because it's like the the models are getting so complex that understanding what better means it's not they're not linearly better you might get an improvement in say let's imagine you get an improvement in creativity but that also makes it more verbose and more likely to hallucinate is it better you might get more you might be more creative which is great it's what you're after but then it's making stuff up when people use trash bt or clod they have no idea that there are humans in the background who have labeled millions of training images, right?
4:57Rated response quality, as you were saying with human feedback, fine-tuned models through RLHF, and even reg-teamed them to find out biases, right? So all of that goes into this process, I'm assuming, correct? I mean, AI is not magic. It's one of the biggest problems. People think AI is not magic. AI is just better predicted text. It's very good predicted text. But fundamentally, if you go into your phone and you go into WhatsApp and you type a word and you keep pressing next word, that's predictive text too. It's just not as good because it's based on a much simpler model. Like all AI is trying to do is find the next character or the next pixel or whatever.
5:36It has no concept of truth. It has no concept of anything. And so these are just various different mechanisms to make it better at finding the next token. And that could be because you've given it a load of label data, which is making it be rewarded for certain behaviors. or you've given it loads of information which is making it be punished for certain behaviors. So red teaming is exactly that. You're trying to find, oh, the model has used racist language or the model has exhibited some bias or the model has used a false source. Like for simple things, like a model is going to be looking in its underlying data set for reference data.
6:15And there was some stat that came out that said 40 something percent of all the references the model is using is Reddit. i saw that yeah that was that was recent huh and like at a certain point they're gonna have like is reddit the best sort of truth is somewhere and someone like perplexity is gonna have a very high weighting for certain sites it's using for citation because they're the ones that it trusts and that is going to come from a set of decisions that someone is making around which sites to trust and that will be because there'll be a set of hyper parameters which say oh bloomberg thomas Reuters, like focus on those above Reddit or whatever random forum someone's saying something.
6:55Casper Elliott:You're essentially butting up against like Arthur C. Clarke's third law that any sufficiently advanced technology is indistinguishable from magic. And to the average person, like, you know, I always joke that when you when you work with someone who hasn't spent time with this and you show them something that really impresses them, it's that it's just like you pulled a quarter out of your ear. I'm fairly sure that the GPT-3 model was not that much better than the GPT-2 model. The difference was the longer context window and the application that allowed it to be conversational. That's what made it magic.
7:30That's what made your grandmother go, oh, I'm talking to the computer. Yeah. And it's answered back coherently. The model itself, this is where I think people get so confused. No one ever interacts with the model. They interact with an application layer on top of a model. They interact with a product. Claude is a product. Gemini is a product. and ChatGPT as a product. So you're improving the model. But often a lot of what we're doing as well now is helping give information to improve the product as well as the base model.
7:57Casper Elliott:You know how AI coding agents feel fast until you're stuck fixing their code and realize it would have been quicker to just do it yourself? Warp changes that. It's an agentic development environment, a new kind of tool that makes working with coding agents effortless. Warp connects to your code base, understands your context, and uses the best models, from GPT-5 to Claude Sonnet 4.5, to generate production-ready code from the jump. And when you do need to step in, Warp's built-in code review and editing tools make it seamless. You can see diffs, reprompt, or make quick changes right alongside the agent.
8:34Casper Elliott:No tool switching, no wasted time. That's why over 700 ,000 engineers from companies like Netflix, Ramp, and Amazon are already using Warp to save hours each week. So if you're ready to go from prompt to production faster, check it out at warp.dev. I like to play around with open source models quite a bit. And one of the things I'm running across is an increasing number of small models boasting about how they're trained on high quality data versus the Reddits and 4chans of the world. I mean, not that there isn't absolutely good value to be pulled from a Reddit because there's a lot of firsthand experience on there, but there's also a lot of toxic garbage.
9:20People talk about the toxic garbage and like that. There's not much that either. There's just loads of irrelevant stuff. There's just petabytes and petabytes of dross or random stuff that is adding no benefit. So it's just weight. Yeah. It's just and that's cost. That's inference cost. There's all these things. So there is honestly the open source thing, a lot of it is a cost play and a performance play. Like, because if you just have less data and fewer hyperparameters, your model is more efficient and easier to interact with and easier to adjust as you need it. But then the problem is that maybe sometimes because it's easy to adjust, you can adjust it in ways you don't expect.
10:00One of the fascinating, I was having this conversation with someone last week who was like, why are they called researchers? these organizations it's because like why are they not called engineers it's because because they are researchers this is this is pure scientific research like you make a hypothesis and you test it like loads of the stuff people don't know about they're exploring they're like ai agents have been the hot topic of 2025 we only really started to work out how the hell to use them in january like this is new and it's really like experimental yeah so people are like discovering Like this is why this is why the whole DeepSeek R1 thing came out.
10:38They were very excited because they discovered a new way to use reward modeling data much more efficiently. And that had a massive potential impact in the way you can train. I mean, it does. Distillation means that like it's a lot easier to pull stuff that already exists and make it easier rather than to do new stuff. But like it's all it's new. It's discovery. I think it's why it's so exciting. It's like there hasn't been a scientific field that has felt so relevant and current for a long time.
11:05Casper Elliott:I think that's a really fair assessment. You know, the amount of new research coming out and new ideas every day right now out of these labs is just insane. It's just a constant flow of research. Corey, Corey, it's so tiring. Let me tell you. It's so. Because you're just like, oh, should I read this paper? Is this important? and do I need to kind of... And actually, I will say, this is where people talk about an AI bubble. My personal perspective, you could pause model development today and the consumer wouldn't notice for five years because it's about the app player. It's the piece that sits on top.
11:47There's so much to be done there. So we will see slowdowns and we won't get these massive drums forward because we've done some of the easier bits, but there'll be underlying things around infrastructure, around how we make calls, like that will be really exciting that will actually make a huge impact. I mean, tokens have already plummeted in costs. So like cost is barely a concern, but it could be speed. I mean, as soon as someone gets something that is of a performance equivalent to 4.0 or something that you put on your phone and store locally, that has a huge change for privacy. Yeah. Has a huge change for security because suddenly you don't need the internet.
12:22You can be out in the wilderness. You can be not wanting to share your data with whoever your client provider is. There's a whole set of interesting outcomes there as well.
12:32Casper Elliott:Oh, yeah. I saw someone, and I can't remember who last night on Twitter, mentioned that we haven't even really talked yet about integrating this into text messaging, which seems to make so much sense, and nobody's working on it. Or email. No one has changed email. Email is still this awful system where you're like, oh, I've got to go up this chain. It's this long string of text. Someone's replied all somewhere else. I want an email system where I never have to go to Gmail. That's what I want. I want an email system that just... And I'm sure Gmail, I'm so good. I'm sure Google is stressed about that.
13:07But I mean, the flip side, and I think this is what's, this is going to be interesting is, like people aren't fully rational actors in ways that is useful for self-preservation and things. People are going to be like, I don't want this to change how I approach my email. I don't want this to change how I, like, I don't want, I mean, okay, how many times have you been writing an email and just press tab to end the sentence on Gmail?
13:33Casper Elliott:Oh, that's a good question. Like, I want to write this stuff. I want to write it myself. Like, yeah. You mentioned, so you mentioned a minute ago, you mentioned that, like, you think that there's a lot of value in the application layer. and you know as someone who's training these models working on this do you then see the future being like more so around smaller more niche models like do you see that being more of of where we're going to get a lot of the performance gains in the models you're still going to get huge trillion dollar companies spending a lot of money on their base models because their base models are behind the application layer and they're going to be behind a load of other companies application layers let's pivot back to invisible because um i know that you work with a lot of companies and uh i want to know a little bit more about how it how it works when you work with a company uh for instance there's like the frontier labs but there's also like let's say um you know in your recent fundraising uh video you had an example with the charlotte hornets came up um and how how would you work with a company like that um with your process um if you could walk us through like how that works one of the funny things about invisible is like i've been here three and a half years when i joined like i mean ai didn't exist you could maybe say but like i mean we did we did nothing in human data and we started working in human data relatively soon after that um and started working with a couple of very very marquee clients but we were historically a company where our goal was to find the complex processes that clients had and orchestrate them on their behalf like be the the invisible layer that powers these businesses often not in the stuff that they want to be doing but the stuff that they have to be doing like no one builds a business to have a finance department.
15:16No one sets up a company to be like, I really, really want to have a huge and expensive operations team. Like, how can we take those pieces and help them execute them more efficiently? And historically, it was all about the fusion between humans and technology. We were saying, we're not a BPO, business process outsourcing, where you just say, I'm going to take my work, I'm going to outsource it to 100 cheaper people, they're going to do it. that has its benefits because people are creative and flexible and people scale very very poorly and then we're also not a pure tech player around remote process automation because historically you guess if you've got a very very well-defined straight line process you can automate that using historic like companies like uipath and automation anywhere in blue prism and then as soon as you get something that's slightly out of that it breaks our thesis was always a combination of the two gives you the best flexible evolving able to work in a complex environment outcome that was actually how we got into human data because some we we were talking to these people and they were like well we've got this process we're trying to scale this complex workflow to produce conversational data and we're like well that sounds like a process we're really good at that yeah and we were really good at that and we kept getting really good at that and we kept doing more and more of it because it was by just approaching the kind of boring problem that they don't want to have to solve but they have to and solving it on their behalf so as part of that what what what did the humans do in that process versus what's automated like in human data or in yeah or i guess like you have you have quite a few different services that you offer right yeah like you have like the neuron data platform no relation to us i thought that was funny though uh the atomic process builder the axon agentic platform there's a lot of cool services that you offer so i'm curious yeah so so like so if i just go forward to someone at the charlotte hornets we come in and we want to be asking what are the problems that that we can solve for you what are the things that you either you don't want to have to deal with or what are some things that you've never even thought were possible but actually because of the advantages of technology are now possible and a great example of this like so talking to the charlotte hornets they were like I mean we love to ask this question like if you suddenly had a million entry-level employees what problem wouldn't exist or what would you try and solve and and with the child of Hornets they were thinking well like it'd be really cool if we could just watch every game of basketball in the US to assess players and obviously like you I mean it's probably an inefficient use of their funds to hire a hundred thousand scouts but now with I mean maybe maybe maybe maybe I've just revolutionized basketball maybe they'll be like this is the move we just hire a hundred thousand skypes but i doubt it um but instead we were like okay well we can get a video camera in these places and actually computer vision models are now much more advanced that we can start to pull out way more relevant data from these games and we can start to pull out patterns and we can start to pull out sort of recognitions about the movements people make or the shapes people people form when suddenly you've effectively got a scout in every game.
18:26Casper Elliott:And the amount of data that these sports nerds will have to eat up for the next century will be absolutely amazing. I always think of the sabermetrics community in baseball where they're like, you know, here's how you do against lefties when the sun's out and the wind's coming in from the east on a Thursday. Exactly, because basically, historically, and this is why sabermetrics was such an interesting thing, is because baseball is a very statistical, statistics heavy game, because that's like, lots of these things are measurable. But now things that were historically viewed as unmeasurable because they were unstructured data, you can actually analyze more effectively, which is, I mean, that's what language models and other models do.
19:09They take unstructured data and do interesting potential stuff with it. Not guaranteed. It could be complete dross. Yeah. But they give you the ability to interrogate it more effectively. Like, just think conceptually, a large language model reading Tolstoy is the same as a video model watching the Charlotte Hornets. It's able to take that unstructured information and bring some sense from it in a way that was never previously possible. It's two things. You asked your question, Grant, about the modules. We're not coming in there and pitching modules. We're not saying, buy our tools, buy our widgets.
19:42These are our tools in our tool belt. But our job is to come in there and be like, do you need a screwdriver? Do you need a hammer? Do you need a monkey wrench? what is your problem? And often it'll be a combination. And what's really interesting, I think this is really what differentiates ourselves, is the people we have as well are such an incredible factor in enabling those tools to be more effective. For example, like, okay, everyone wants it to be step one, have problems. Step two, talk about AI agents. Step three, magic happens. You deploy an AI agent. How do you know if it's doing a good job?
20:13You have to evaluate it. Okay. And how do you get an evaluation? You need to understand what the measures of a task being performed acceptably and accurately are. And often the best way to get those is to do the task a lot. Yes. Right. And that's what our people can do is they can actually perform this work and we can start to build up the evaluative sort of set of metrics to then deploy an agent effectively. Because candidly, if a company is going in and saying, I'll deploy AI agents, it'll all be tech. that either they're saying, hey, client, you used to do a hell of a lot of work for us, or hey, we are YOLOing this completely.
20:53And neither of those is particularly compelling for an enterprise.
20:56Casper Elliott:No, no, it's really hard to prove the ROI on that. It's hard to, everything about it, it's difficult to sell. And because we are not, like, we don't, we price on outcomes, we price on unit bases, etc. we aren't saying oh yeah we'll bring our people in and you'll pay hourly and we'll just do this forever because it's a wonderful business model for us we're saying our people are going to come in and they're going to work over the course of this project to make themselves obsolete in this case yeah or they're going to be like actually this piece is so subjective and so complex and so strange that actually you need people and that's going to be the best way to live quality because like we are not there we're there to deliver the outcome not the mechanism for the outcome we're not we're not there to sell a certain way of doing something we're there to just live with the outcome invisibly and then so is this technically so you might have been following the debate about the rl rl environments or just like vibe evaluating i don't know if you followed that on twitter or not yeah yeah uh are you technically doing rl environments then as you do part of your process here we're exploring them because i mean there are situations where they're very valuable because they basically allow you to create realistic data sets or create realistic evaluative scenarios, you need that.
22:09If someone's saying, I want to build a set of agents to do something in a certain SaaS platform, having an example of that SaaS platform you can mess around in is very, very necessary. In the same way that if you're training someone to use a system, you give them a training environment.
22:24Casper Elliott:Let's shift gears a bit here. I want to kind of talk about the elephant in the room when it comes to data in the AI industry. roughly five years ago data labeling was this unglamorous corner corner of the internet and now you cannot have a discussion around ai without someone wanting to discuss the need for data hygiene for data labeling for better data practices and people are suddenly realizing what a just god awful mess their data is across their company and it's become you know quite a roadblock we've always tried to do what we were good at, which was the complex work. We very much focused on the cutting edge stuff, bringing in multilingual, bringing in multimodal, bringing in STEM, all these dimensions of complexity, trajectories work for agents, et cetera.
23:18Because that's where we excel around complex data. We weren't involved five years ago when it was much more simple. We came in at the complex end because existing vendors weren't meeting the need and we met the need and we continue to meet the need. So that's really our world. Okay, that deal happened and it impacted the industry. Now we see a lot of headlines about competitors popping up and I think people are now finally seeing the value of it. So do you think that the type of work that you do is now incredibly important to these big frontier labs to the point where it's a strategic asset for them?
23:54Casper Elliott:or that people are noticing finally that it is and it always yeah by giving you the credit where it's due you know i think i think 100 i think i think 100 like it like i think the the need is growing not shrinking like we're not going to complete data like we're not yeah uh i mean there's always more data i also yeah and i i'm also much more of a like a super intelligence there than lots of other people because like we this is this is just better predictive text it's better and better and better and better and maybe there'll be something when we hit the cert the right number of sort of node equivalents like the number of neurons in the human brain and suddenly there'll be magic i can't predict that but i i don't know i'm not betting on it as much that there'll be need to be other structural advances that i very much not qualified to even understand let alone talk about to get to like relatable agi like it's revealing that the people talking about agi are the ones very very incentivized to talk about agi sure makes sense so yeah so i think we're going to but we are going to see more of a data need because people are going to try and make incremental improvements for competition and also we're going to start to see more like second level users like not the foundation model builders but the people who are building complex products in the healthcare space in the financial space whatever using agents using models who still have a need for trajectory data for evaluative data for for data for reward modeling for improving whatever, like be it multilingual.
25:22Like, I mean, think about it. Like lots of these models are very, very good at English because they're trained on a predominantly English data set. And then there isn't at other languages, but there are a lot of other languages. They're like, I mean, there are seven or eight languages in India with a hundred million speakers. Yes. That's a lot of potential consumers. Like, so there's a huge demand in these different places. With their own literature,
25:45Casper Elliott:with their own... With their own literature, with their own quirks and things like i there's there's a huge amount of stuff that can be done there and also as training costs go down maybe then you could start doing like like just because of like i don't know this isn't exactly an example of jebbin's paradox but like it's just because something like becomes easier doesn't mean people use less of it so how do you feel like the um skill level has for this type of work has evolved over the years and what what do you think that means for the economics of like ai training oh it's it's massively massively gone up um like when when we were starting you were taking sort of people who were intelligent and spoke good english and but like generalists like and there is still a role for generalists but in so many like think about it you want to make your model better at quantum physics you're gonna hire a generalist no you want you you want to make your model better at um swiss german you're gonna hire a generalist now you've got to find someone who speaks swiss german you want to make your bottle of swiss german in quantum physics like it like the complexity is only going to increase you and then you start to bring in modes there are different skill sets like audio transcriptions a skill set some people are very good at some people are awful like video analysis is like that all there's all these different skill sets and you're getting people who are very very good at certain things and there's a demand and i mean you don't need to be a phd in economics to understand that when there's high demand and there's like obviously some degree of limited supply costs are going to go up yeah and that's reasonable and we're seeing that like i mean we we have examples where we have nurses and healthcare professionals working on behalf of clients we have litigators working on behalf of clients like there's there's lots of complexity there and like people who have these qualifications aren't cheap yeah now there is also the this is this is where you get to a very interesting point like it's are these people um build like what's the are they build creating their own demise like i don't think they are i don't i think i i think in lots of situations a just the the evaluative nature of some of these tasks is going to mean that there's going to be a value of human oversight yeah and then i think and i think one thing that i i'm really interested in like you read about like entry-level jobs going down in like certain industries because of AI, et cetera.
28:11Who do people think are going to do the leadership jobs in 20 years time?
28:15Casper Elliott:Yeah, that matters. Like, and I think society, like this is where also people, like I think there's going to be surprising human reactions to certain things. And there might be a premium on human contact and human engagement. Like I just don't, I think people are saying we're going to have 99 % unemployment, have an engage with like how humans work. Like it's like Ludism was very fashionable in the 1830s, but we got beyond the mills. Yeah. And it couldn't come in the form of just KPIs that are 10x what they were. Or it could. I mean, we might end up with a three day work week. Yeah. I'd love that.
28:52I've seen headlines of that recently. Yeah. Like, no, I've seen that as well, because like there is a very utopia. And ultimately, like these are complex and the incentive drivers are going to lead to all sorts of unexpected impacts. But I think anyone who's saying, anyone who's saying AI will be 100 % utopian is naive. Anyone who's saying AI will be 100 % dystopian is naive. Yeah, I agree. It's going to be complex. It involves people. People are mental. Have you met people? Like, it's the same thing, isn't it? But I don't think we'll get the situation where we'll be like, no lawyers exist. Yeah.
29:24Yeah. Or I say this kind of tongue in cheek, but the other alternative, which I think you're saying isn't going to happen, is that the only job in the future will be training AI models. Yeah. Exactly. I mean, there was one of the funniest things I've read. And this was a delightful example of selective hubris was, I think it was Mark Andresen saying that, like, I think every job we automated except for VC. And I'm like, oh, come on. Like, like, like, yeah, you're on our side of the fence too over here.
29:55Casper Elliott:Come on. Yeah. Yeah. It's just like, it's like, it's the lack of self-aware. Like, I think some aspects of lots of jobs will become more efficient. i think there'll be some there'll be some jobs which are which massively decrease in their value because they can be a functionally automated but then the question is was the job adding value and is that actually necessarily a problem like are we frustrated like are we frustrated that we don't like hand pick crops like like you know it's it these things these things are complex i don't know i'm kind of meandering in a very confused way and you're like get this guy off no no no no no no No, that's fair.
Read the full transcript
30:31Or like I was thinking something that you said right before that about what did you say right before that? The before the crops part? Something very erudite. Mark Andresen. Oh, Mark Andresen. Yeah, right. Right. Oh, yeah. VC being being automated. No, not being automated. That was the one thing he was like every other job apart from VC.
30:51Casper Elliott:Every other job. Yeah, I think I think you know what it is. I think it's OK. There's two parts to this. I think there's, you know, number one is like everyone can see the flaws in the AI for their particular skill set. It's sort of like when you read the news in your industry and you're like, know what they get wrong versus what they get right. But then you read the news in another industry and you're like, oh, it sounds like, you know, like you just kind of take it at face. Everything they say on that one. Yeah. Yeah. You cast the example of, oh, I can't, what's it called where Dunning-Kruger, right?
31:23Yes. Where you don't understand another industry. you'll be like, that's clearly easy. It's all gone. But my industry is really hard. Yes, exactly. But then I think there's the other component to it, which is people will always want to work with other people. Like when you get to a certain level of success, you're going to want people to offload stuff too, or like partners to work with, bounce ideas off of. I don't think that's what we're going to go away. Yeah, people are like, yeah, people aren't like, like people don't work in the pursuit of 100 % efficiency. No. Like everyone does inefficient things all day, every day.
31:55for reasons of being human. And I think that's glorious and wonderful. It should be celebrated. And that's fine.
32:02Casper Elliott:Yeah, 100 % efficiency would be hell. Yeah, it'd be boring. Yeah, exactly. So then getting back to data for a second here. If data is critical and expensive, are we heading towards a world where access to high quality training data becomes the biggest moat? Like what's your take on that? this is where ai is a research driven sort of topic is very interesting like you publish a paper someone can then understand the paper like so like and so maybe maybe if the data is is the the only secret source remaining that i think it becomes very valuable and yeah i think that is the case and like companies very closely guard their data environments that you've got to like there's there's a lot going on there but yeah i suspect i suspect it will i suspect it will if it starts to mean the difference between 97 90 90 accuracy and if that's your only difference then that is your mode yeah i can see that being the case and and again to the point that's where demand and supply we're going to see an even more increase in greater increase in the value of this where do you see this industry in, say, five years?
33:13Casper Elliott:Will every major company need these types of services? Will the tools eventually automate themselves a bit? What do you see as kind of a little farther down the road? Five years ago, if some would say, yeah, this company you've never heard of is going to release a product called ChatGPT and it's going to be the fastest growing thing in human history, what would you have been like, yep, I'll buy that? I think this is actually one of the biggest challenges again and it's probably in a human nature like i think the speed of change in certain aspects is incredibly rapid and therefore intimidating and people will struggle to adjust but then the flip side of that and this is where i think and okay i'm gonna make one prediction i think we're gonna start to see much more model impact on a physical sense like autonomous vehicles robots etc because this last three and a half years has been only almost exclusively digital and digital like digital is huge i mean we're we're talking in real time across on some recording software and you're in the west coast of the us and i'm in london if you said that 40 years ago they'd been like this is witchcraft so it matters but i think as we start to see agriculture manufacturing logistics that's going to have very big impact yeah because i think the ability of models to handle real world situations it's going to increase like you see those sort of bottom dynamics things it was humanoid robots etc there's they're they're getting very very very smart waymo etc ways etc all those kind of things yeah exactly and that's and that's going to be very i mean yeah not again it's not magic like i think a certain ceo of a large automotive company had predicting self-driving cars the last 15 years and not delivering anything.
35:04So it's not magic, but I think we will see that. And I think that's going to be a huge, a bigger driver because that'll also feel bigger. I think because there are people, there are lots of people out there for whom AI is something happening to someone else.
35:18Casper Elliott:Yes. It hasn't affected them. Most people probably. No, exactly. Exactly. Especially in physical industries, like you're saying. Yeah, and it's so, it's something that we talked about on the company. It's like, it's so easy to be sat there being like oh my god i haven't read this thing that came out last week i'm so behind meanwhile 99.2 percent of the world is like sorry what's chat gbt so maybe not quite that yeah like it's so yeah it's so easy to look up and think you're behind and realize that everyone else is there and i think that's what's happened and that's why for other people there's this hype and it's a bubble i think it's a bubble but it doesn't mean it's important yeah exactly 100 i mean yeah the dot-com bubble you can't be like the internet what was that yeah it didn't disappear to be quick right exactly the underlying the underlying fundamentals are valuable even if there's height and there's overvaluation and there's all sorts of things which there are 100 is yeah and there always is in tech honestly i mean anytime you're dealing with disruptive industries, you know, the money floods in, you know, you know, and some of them, the industry catches up with it.
36:28Casper Elliott:Some of them, it doesn't. I mean, you know, you can look back at companies like I'm going to I'm going to throw Amazon under the bus as a famously unprofitable company that just just hemorrhaged money for two decades, you know, and I'm all the way to the top. Yeah, it's it's really interesting to see how it all plays out. But I really agree with you about we're going to see them more in the physical world. I've just in the last few weeks, I've noticed suddenly, you know, I'm running into this little robot at the local get local quick trip every day. It's now roaming around the store and it's got little eyes and it flashes and it talks to you when you go by.
37:05Casper Elliott:We've got one doing inventory, walking around the store at the grocery store. And I'm like, man, we're not far from like, you know, those unitary models and stuff like that being just everywhere. But then also, what's the human reaction going to be to them? It's going to be very interesting. It's the analogy I made at the start around models being like a teenager. It shows that the models don't behave as you expect. It also reminds you that people don't behave as you expect. I think that's going to be the two things that are going to be very, very interesting. That's so true. Yeah, we just wrote about Waymo and how the Waymo safety data is actually really, really impressive when you look at it, especially when you compare it to human drivers.
37:45Now, there's caveats in there. But I think one of the biggest things will be public opinion on Waymo's because you'll read stories where people are like, oh, I don't like that they idle in front of my driveway or I don't like that they make this loud beeping noise. It's a different thing. You just need the first trolley problem scenario where the model has to make a decision and there's a big court case. And then who knows what's going to happen? yeah yeah people people don't perform to like we i mean we ran something god three years ago where we were pulling data from a website and having to validate if it was a real real location and we had people doing it and they were doing it with a certain accuracy and we got a model to do it and they did it with a higher accuracy but when the model got it wrong it got it wrong in weird ways yeah i didn't like it because if you're like oh 95 percent i say the correct location and the other five percent of the time i say a different wrong location and instead it's 99 you say the correct location and one percent of time you say banana people can be like oh that's much worse yeah there was a study that just came out about this a harvard business study that basically said people blame robots more like they like like when when robots are at fault people really don't like it like they're way more likely to call them empathetic and people understand human mistake like it's people man they're the fascinating thing about all this same thing with cars right now you
38:59Casper Elliott:know every time an autonomous car is involved in a wreck right now it's global news and the fact It is like, you know, the numbers you ran this morning, Grant, are like, you know, they're 94 % safer than a human driver. And, you know, it's like we need some perspective here. I think it's important to understand that, you know, some of that stuff, we're very fallible. We'll be in a world and in like, I don't know, some, let's call it 50 years time, people will be like, wow, you drove places? That's insane. You drove yourself? Yeah, exactly. And who knows? like but but i just i just think yeah prediction is so hard because we're dealing with just i don't know i mean maybe i'm just a coward um but whether because you're maybe like more hands off you're less likely to jump on the the bandwagon of like a car or what do you mean by that coward i didn't like i mean i maybe i could not make predictions being a coward oh i mean oh gotcha gotcha i thought you meant you're like yeah i won't get in a way mail yeah oh no no i mean no I'd be very curious too, but I can also understand the discomfort.
40:01Casper Elliott:Yeah, same. Same, like, you know, if you're a person who needs control. Yeah, or you enjoy it. Like, I don't know. Like, it's, yeah, it's just, it's these. And I think that's what's really interesting is you also get, like, a lot of these become philosophical discussions. Yeah. It's not, like, because it's not about what's best in, like, a strictly rational sense. It's about desires and needs and all these things. That's right. Totally. Well, I mean, from a data perspective or like training perspective, I mean, is it harder to train? It must be harder to train robots and self-driving cars and the physical side of things, right?
40:36Oh, I think 100%. Like, I mean, we haven't done much yet in the physical AI space. I mean, actually, I don't understand enough about the models they're using to actually know why. It's still so new. Yeah. Yeah, I can imagine there's going to be all sorts of complexities around like unexpected events. Because you're always training for educators, right? like you're always trained like like if if if the car's driving from a in a straight line to b and there's no traffic and like it's easy it's going to be great but it's when someone someone does something unexpected or there's a near person whatever then you've got to build the data set in order for the model to behave in the way that you want it to for performance reasons for legal reasons for whatever and that's and again that's that's a challenge i think like model usage in heavily regulated industries is going to be a big challenge because there's going to be concepts of liability like and that's why again human oversight like it might it might be inefficient but it'll be the kind of thing the way there's legal robustness to it so people are comfortable with it otherwise you're like oh yeah i made this medical decision based on what a non-humanistic model told me and like who's liable if something goes wrong yeah yeah it's a great huge unsolved
41:49Casper Elliott:problem i feel like yeah it's it's been a it's a problem we still haven't solved from the internet yet like it it just it's continued to be a problem for for a lot of years and uh you know it's a good question you know if you own it if you build it who's the who's the fault operator is it the the person who who designed it like it's people like yeah people want to see someone is it me taking a nap in the back seat you know while we're going through the city and then if that yeah if that's the like just imagine as soon as that's the case then suddenly self-driving cars they're not dead because they're not dead like it's not a sort of binary thing but like suddenly they they can't self-drive like these are interesting because ultimately law is not an absolute thing it's going to be a people thing where people are making arguments and it'll come down to a public opinion to a degree like and that's why this thing will also shift i would suspect so if if people see value and they feel it more comfortable with it then it shifts and evolves um but it takes time to evolve like like people are slow to adjust um i think that's what's going to be interesting as well is seeing people who like just in the same way that someone who grew someone who grew up in who was born in like the mid 90s is way better at the internet and someone who was born in the mid noughties is way better using a smartphone someone who was born in the mid 2010s is going to be unbelievable the problem of engineering yeah by the time they're in high school in high school i would say yeah at the latest you know um if if not their education system is a problem um but it won't be the education the education system will be slow to catch up they always are yeah it'll be people like in the same way like in the same way that all the people who were building sort of geocities websites way back when are the future web developers i was one of those yeah yeah and like and you you did it and you learned how to do it and you messed around with the tools i did the people who are just like oh i'm gonna see if i can get chad gpt run my homework and two years later they built a complex multi-agentic system that's passed a phd you're like if that's the purpose is stopping your next unicorn i keep telling companies when they're like oh well well this resume was clearly written by chad gpt we should throw it out like actually what you should do is call them in because it shows that they've they've found a way around a process that is absolutely built not in their favor and found a way to make it work maybe they're who you ought to talk to this is something we often find myself talking when you talk to companies about like oh ai strategy i'm like the person you really like your chief officer i'm sorry is probably not going to be the right person it's the it's the random 24 year old who was unbelievably excited that gem and i released a new image model yesterday yes that person is the one who's going to like build some cool stuff that's going to change a part of your business because i'll be like oh my job is boring i'll automate bits of this yes exactly yeah yeah you know and that's exactly where you start is is get rid of the things you hate what are the things i least want to do this morning yeah and those people are like yeah and it's fascinating to watch them like i mean i can tell the story and then the person who's been referred to is going I find it very amusing.
44:52We've got someone in Invisible who, like, he's someone who came in. He actually came in as a contracted agent, like, doing human data work. And he was excellent, and we brought him on full time. He's now running an MLT, AI MLT, engineering team, because he is an unbelievably good prompt engineer and also good at understanding how to use them correctly. And I had a colleague of mine who I brought in, who's also a very, very, very good ML engineer, had a theory about this. And he was like, yeah, this guy, he was a professional League of Legends player for three years. He got incredibly good at making a huge number of decisions in massive, massive sequence.
45:31And also very quickly evaluating tradeoffs, which is all you have to do. And like, he's not a trained, like he hasn't gone through like 20 years of software engineering training, but he understands complex systems and how to handle them in different ways. and like it's a non-traditional skill set but it's proven incredibly valuable because three years ago this job didn't exist yeah so i mean if you're a person who plays league of legends and you're really good at it and or you are a parent of a kid who plays league of legends like i mean i get your kids with clarity as a lifetime nintendo fanboy like i'm not advocating believe legends here but um okay fair enough fair enough well a bit of the general idea like like there are non-traditional skills that are incredibly valuable in this world where you're you're effectively engaging with a model or a interface on top of a model to try and understand the problem or try and break it down or try and deconstruct it or try and like sequentially map it and do all sorts of complicated things to set up a set of agents using different tools to do some piece of work like that's not sit down write python hello world well what would
46:36Casper Elliott:you say, you know, again, kind of pulling us back to invisible. What are the... I don't think you're so professional. You're like, let's talk about the car. I'm just like, hey, I'm just chatting on League of Legends. Yeah, we're just having fun. But I do have a question I want to make sure we get in. And that is, what are the most common mistakes you see companies make when they're trying to employ AI, specifically around like data strategy? okay i mean there's a whole heap of things like i think thinking it's step one talk about ai step two question mark step three magic like no it's it's a very boring set of processes um a classic example i mean we really think of it like sort of left to right data process evaluation agents like you've got to have if you if if i'm going into your company i mean i'm saying okay you're you're an underwriter you're an insurance company i'm saying hey how many claims did you play out in the northwest of england in february 2024 that involved flood if your answer that is oh yeah fantastic let me just go and query done you're in a good place if your answer is lol you're how are you going to get ai to do anything you've got to have a clarity in your data landscape you've got to have understand what truth is so your models can actually or whatever the hell could do stuff another great problem is you'll be like okay here's your process you said you've got this problem, like let's understand your process, right?
48:02What's step one? Oh, step one, I just talked to Dave's team and they do something. Dave's team is the ultimate black box. If you don't know what Dave's team does, how the hell can you get someone else to do that? Or you can train a system to do that. You've got to know what the hell Dave's team does. That's right. You've got to understand what your outputs are, what your inputs are. Like you've got to understand your process. That doesn't necessarily mean mathing out your current process because another big problem is people are like, oh yeah, my process runs like this. AI-ify it. No, no, no. And I'm like, look, imagine if you'd said to someone in 1900 like right i want the best and fastest form of land transport we can get they'd be like here's a really fast horse we'll make it faster like the motor car is a different paradigm in a similar way you might your process might be some frankenstein's monster of bolt-ons and other things and actually you've got to break it down to what's the desired outcome what are the inputs i need yes then you can actually reimagine it from the ground up and have it ready to be improved like so that that's a couple of them like if the data is a mess forget about going a bit further if you just don't understand your process you need to break it down to an outcomes basis and then lastly like if someone's like oh i want this task to be automated like okay here's an example of the task was like tell me why it was a good task if they're like because i like it that's not an evaluation metric thumbs up thumbs are not evaluation you've got to break it down as much as you can so it like ideally be like oh yeah this is a good task because it gives a score of 17 out of 21 on these 21 different classification models yeah that that is a much better that is something you can then use to actually measure if something's done well yeah and then start to bring an agent to perform it because you can evaluate them correctly otherwise yeah again it's like thumbs up thumbs down it's why like It's why AI and the enterprise has been such a disaster.
49:51ChatGPT is amazing at solving 10 ,000 problems once. What enterprise solves 10 ,000 problems once? No, enterprise solves one problem 10 ,000 times. You have a very different set of expectations. And there are different departments, different groups of people.
50:10Casper Elliott:They're all connected in different ways to different tools. Often siloed, to your earlier point. And so you've got to get agreement on what good looks like, on the right data that's coming, all these different things. It's a very different challenge. And that's why you've got to approach it very systematically. And the human is a key part of it in terms of understanding, in terms of valuation. Honestly, I think it's weird that we're a company that spends so much time leading to these advances in these models and these products of these models. But actually, all it's done is make us realize that our people are even more valuable.
50:43Casper Elliott:I absolutely love that you mentioned, you know, the need to not AI-ify the process you have, but to maybe toss that process to the side and relearn it from the ground up in a different way. You know, I think that's a really crucial thing. And it's a mistake I see made constantly. It's like, all right, here's our process. Let's go through and step by step build this to replicate that. And it's like, no, you don't necessarily need to do it that way. You know, sometimes that's highly impractical and results in an over-engineered nightmare. Well, and that's not to say it's wrong. There are going to be situations where that's hugely valid, but you've got to look at it coherently with clear eyes.
51:23Yes. Is there like a magic number of like, once we get it past a certain threshold in terms of our evaluation metrics where it's like okay to deploy it? Like, do you, does your team have some sort of internal way? Is it different for every company? No. How do you handle that? No, it depends on like everything is a trade-off, right? Like you can, like if you automate a process that's 30 % correct, that could still be valuable. Or it could be like detractive because actually the people have examined the 30 % more. Like it's going to come down to cost. It's going to come down to speed. It's going to come down to the SLAs you care about.
52:03Like some people are fine with 80 % accuracy on something because it's already higher than it was. and we can just deploy something like that. For some people, anything less than 99.9, 3.9s is a disaster. And so they just got an approach for each of them. Like it's not a one-size vehicle. Yeah. Makes sense. Yeah. Because we keep hearing about evals and certain people are saying, well, you have to have a systematic way to evaluate your outcomes. Otherwise, it's just going to fail. There was that study that's like 95 % of enterprise deployments don't have an ROI because maybe this or that reason.
52:39But that's a different kind of eval, right? That's an evaluation of the ROI of the overall program rather than an evaluation on the model output. But I agree. If you're doing a piece of work in a business and you don't know if it's worth it, why are you doing it? And that applies without AI. Because we need to use AI. It's the future. But that's one of the biggest problems is people just like, yeah, like if I just, yeah, talk about AI for a bit and do some stuff, everything will be great like it's ai is another tool it's another tool in the tool belt like we talk about the tools in our tool belt like there are not the number of times i've gone to a client and we're talking to them and i've been like honey you need a regex like you didn't do anything more than that like because but they're like i want to use ai it's like why what's this going to add this is the most tedious boring process where you're pulling two things you're doing one predetermined calculation and pushing it to a third source the third target like you A model adds nothing.
53:33It's about knowing when to use it. I've always enjoyed working in Invisible because we were about understanding the problem and then finding the right approach to solve it. I mean, we were always like that from the start. It's just that the tools we have have got better. So as someone who builds these tools and trains them, meaning you know them pretty well, what is in your personal AI stack right now? What models and or tools are you personally using the most? I don't use that much. i am really bad yeah yeah like i i'm a bit of a i'm i'm i i tell a lot of people to use them a lot um it's fun messing around on like quick prototyping yeah yeah like claude co is fantastic for that um and i'm just i mean honestly the single biggest thing is i have a notebook here because i i write everything down and being able to transcribe everything for free is just everything because it means that i can just be like oh yeah here are my notes three seconds later But no, otherwise, honestly, you should cut that answer because that would be like, he's a fraud.
54:31Who's this guy? Wait, wait. So are you saying that you will handwrite something and then you'll take a picture of it and then have it uploaded? Yeah, handwrite, take a picture to transcribe. But I find my mode of engagement when I am typing versus writing is very different. Sure. If I'm in a call, I'm there, the screen is on and I'm video, but I'm writing, I can't get distracted. i'm writing i'm engaging when you're typing it's very different you can type something else you can whatever like it's it just doesn't work for me and i just like i like scribbling i like that kind of stuff same like yeah this this i mean we're what halfway through this is about two and here we go 19th september and that was here i've written this much notes i love it in what a week
55:19Casper Elliott:this is my little and the ability i've got a i have a small one i carry with me everywhere i go and but the ability yeah the ability to then push that into to do other stuff with it is that's incredibly valuable yeah and you don't have to take the time to type it out it's you know it's such a prosaic um use for it and like there are there are fun things that i just i mean honestly one of the other things i just have like we've been busy like raising 100 million doesn't doesn't happen with no work right i've got as much time as i want um to actually um just play around um there's so many cool things out there.
55:55You could like, one of the things, one of the first things I saw that I was like, okay, this is cool, was when someone created, and it was just like a random student or whatever, created a real-time Babelfish for playing, I think it was COD online. Nice. So they could speak and they used like, I think they used Whisper to transcribe, they used TradTBD to translate and they used Sum from 11 hours model to speak. and they had this close to real-time translation device for flaming someone who's had career. Amazing. Amazing. It fried their computer completely, but it's the fact that you can orchestrate those tools to create these cool things.
56:36And obviously there's going to be mistakes. It's not a finished product and yada, yada, yada. But you could just try this out. And that's cool. And that's fun. And that means that someone somewhere is going to create something random because they have this idea that'll be like, oh, yeah, wow.
56:49Casper Elliott:If you could automate one part of your job with perfect AI, but you had to keep one part completely human forever, what would you choose for each? Job, not life. Either. You can answer either. Life, vacuuming. Vacuuming is Satan's task. I mean, we have Roomba. We've had it for years. Yeah, but I live in a small two-bedroom flat with lots of stairs. And like, you Americans with your four and a half thousand square feet on one floor, you're having a great time. What would I keep human? People management. That's a good call. We'd need that to stay human. Yeah, you want to manage your team. You've got to understand them.
57:35You want to understand their needs. You want to understand their what. You've got to understand their drivers. Your job is to enable them, et cetera. That's pretty important. And if people are trying to like automate that away, they're awful managers and shouldn't be like any minute management.
57:46Casper Elliott:Yeah. I think in terms of automated, I mean, the note transcription thing was one thing. No, I think prototyping. I think the ability to turn an idea into something that is like 95 % fidelity, like super quickly, so you can just start to experiment is insanely valuable. It is. So I'd say just progressing further down that because it lets you just avoid that frustrating, boring part of the implementation stage after ideation. We're going to start off with automating. I'm definitely kind of leaning the same direction as Casper here. I want all of my house or whole chores to just be managed. I want the house to just be clean, the dishes to be done, the laundry to be in the closet.
58:34Casper Elliott:and all of those things are what I want. As far as what I want to keep, I think it's ideation of things. I really enjoy thinking. I really enjoy learning and diving into things I don't know. And I think if I have the ability to have an idea and then somehow go automate helping me spin it up quickly in a way that works for my ADHD brain, that keeps me on tasks and shows me, yes, you can do that. I think some combination there would be really valuable for me. I mean, just piggybacking off the bat, one thing I have quite strong opinions about, which annoys some of my colleagues, I think acts of communication, it's important to keep them human.
59:21When I receive, when someone's like, oh, yeah, can you give me some questions? I need some questions on this. I get a list that's clearly generated by chat CPT.
59:28Casper Elliott:Yeah. Have you thought about this? Like, yes, you can generate, but then edit, adjust, like make it relevant because I think there's going to be a premium on that kind of, and people will get, I will, unless the models get better and better and better, which I will see people would have premium on that kind of engagement. And if you're talking, if you're talking to a client or someone's talking to you and it's clear that they haven't put in the effort, that's going to be a big turnoff. It is. You can tell. Yeah. How about you, Grant? What's your, what's your pair? mine is probably like i don't know what this has technically called um and it might be the opposite of what casper is saying so hot take um interesting communication like when people ask you to do something that's like a one-off task i would love to automate that like if somebody's just messaging you like hey could you update this and you know i read a newsletter like if you could update this advertisement no i think that's different i think that's different like i would love to just automatically do it it's a non-interesting update and it's just like a busy task i agree i'm talking more about the like we have had a human interaction and now i need to send a follow-up to that i should give a human follower that's right yeah that's fair that's fair yeah i think because i go back and forth on what i would keep human forever but it probably should be like the like meetings as much as i want to automate meetings like you probably should not automate meetings yeah maybe you have less of them because communication off lot async is less automated, but you should still meet with your team in person.
1:01:00Casper Elliott:Yeah, I think so. Because people are the constant throughout all this. They're the fundamental part. Well, Casper, thank you so much for joining us today. It's been fascinating and it's just been a fun chat, honestly. We've been all over the place and I think in some great ways. Where can people go to learn more about invisible technologies and what the work you all are doing? If you go to invisibletech.ai, you will find out all about us. And I would recommend visiting there because we're doing some interesting stuff with a lot of exciting clients and a lot of different spaces. I agree. That's awesome.
1:01:37Casper Elliott:I agree. Well, everyone, thank you so much for joining us on the Neuron Podcast. If you enjoyed this conversation, please subscribe wherever you listen to your podcasts. And check out the Neuron newsletter at theneuron.ai for daily AI insights. Until next time though, farewell humans.
From the publisher
Behind every AI response, there's an invisible army of humans who trained it. In this episode, we talk with Casper Elliott from Invisible Technologies - the company that's trained 80% of the world's top AI models. We explore how models actually learn, why data quality matters more than quantity, what enterprises get wrong about AI deployment, and whether AI will really automate everyone's jobs. Casper shares insights from working with frontier labs, reveals the surprising skills that make great AI trainers (hint: League of Legends helps), and explains why the future needs more humans, not fewer.
Subscribe to The Neuron newsletter: https://theneuron.ai
Learn more about Invisible Technologies: https://invisibletech.ai