About this talk
This talk discusses the implications of AI and large language models on the future of management and job roles. The speaker, the CEO of an AI startup and former professor at Cambridge, confronts the question of whether traditional management roles will become obsolete due to advancements in AI technology, particularly in the context of increased computational power and intelligence. He presents data suggesting that even top human strategists may struggle to compete with AI performance in decision-making scenarios. The talk explores historical precedents of technological disruption, stressing the importance of adaptation over resistance, and emphasizes that developers hold a unique position in shaping the future of AI implementation. Ultimately, the speaker encourages professionals to embrace change and leverage their skills to navigate the evolving landscape of work.
Full transcript
Uh hi everyone. Um this is my first time in Bulgaria. Uh grateful for uh to uh uh Gorgi, my friend. Um I am uh the CEO and founder for an agentic AI startup called Strategize. Uh but until two weeks ago, I was also a professor in uh in strategy at the University of Cambridge. Um I I quit. So I'm just sort of doing this full-time now. And
um I'll uh I I'll share why. uh but one of the things that I do teach at the MBA at at the Cambridge Judge Business School is this elective called uh strategy in the age of AI and it is meant specifically for obviously managers and uh one of the big uh questions that we've all had is that do we need managers going forward which is kind of
like sad because you know I mean you're teaching in a business school people pay you to teach them how to be a manager And in the class you have to tell them you might not have a job in 5 years and you know it sort of creates a bit of dissonance. So uh I was I was speaking to Georgie about it and he said you know what
maybe you should also say the same thing in front of people who are not managers or developers and might appreciate the appreciate it more. So um so the topic is uh if you have enough GPUs are you going to be replaced or are you going to be awesome like we need to know right? So um so I think it's probably worth pointing out that we are living
in extraordinary times. I I I think most people don't know uh or don't know what it feels like to be in an extrapolation that is completely um it's completely off the charts. So if I had to give like a human example of reading this this graph, I think there there are two examples. The first one is um you know we've gone through three two and a half
years of uh of large language models uh in terms of its penetration in the world um it's the equivalent of the internet's um say over 13 years. So while you could argue that well you know it's not a fair comparison chat GBT runs on the internet but then you would say the internet is not a fair comparison because it runs on electricity and then you go to
electricity and you say well is electricity a really good example it was powered by steam and then you keep going back and back and then you have to be thankful to the caveman who found fire right so we we can't we we can't just say this is an unfair comparison every comparison is unfair from that perspective Ive but one thing is this that you've gone through two
two three years of AI that's the equivalent of going through 13 years of the internet and I think the other example is is um I've never been to the Everest but I have a feeling that if I was to go and stand at the base camp and I was to go and look like this uh the graph would look the peak would look like this right so
we are facing Everest and it is uh at the uh and it's good to recognize that I'm not saying everything is going to get destroyed but it's good to know right what we are dealing with so we have entered an exponential age of intelligence um so the question is what do we do with intelligence what's the point right so um I I think a a lot of
you would know about sway bench right it's like the bane of developers like every every 3 months the sway software engineering bench gets released and it turns out 80% of the job that us developers were doing it can now be done through large language models. Um and if you don't know about Swaybench I would say Google it. It's it's useful to know what it is benchmarking but
but also um it's not just software. So software is where it started. So the the the developers and the data scientists who made the large language models uh you know decided to get disrupted by it first. It's easy to develop stuff that you know intimately well. So they got really good at coding but when they when they got really good at coding they started getting really good
at other stuff as well. So here you see two other graphs um one is for biology frontier level biology and frontier level math. uh 99.99999% of the jobs that we do today in the world is not frontier. Frontier is like splitting the atom or finding DNA uh for the very first time. So everything we do is definitely one step behind the frontier. So we already know that
by 2030 they're saying that an LLM would basically be able to independently solve scientific mysteries. Uh so 4 years give or take from now, right? Uh which is fine. I don't think about scientific mysteries every day. It does not come up as a as a thing in my life, but everything else does. So if you are able to crack scientific mysteries, you can probably do everything else
underneath that. And this is how I would like you to look at that. This graph is saying it's not just software developers, guys. It's biologists, it's mathematicians. Um LLMs were really bad with math. Uh because of the nature of how they are built, they are prediction machines. So they don't compute, they predict. Uh until they started reasoning. And because of reasoning, you now have math that quite
good at it. But you don't need to put math on. So, so frontier math is not done by giving an equation to an LLM to solve. You give the equation to the LLM who then writes the Python code to solve it. So, it's easier. The actual compute is done by indeterministic way. The point being that all of the lovely stuff I talked about is now worth $10,
right? It's $10 per million tokens. A token is a word. You produce a million words of frontier knowledge. That's $10. So what? So I think I it was good that I quit Cambridge University. I mean I I I don't think I can survive at $10 per thousand words or per million words spoken. I'll I'll run out of calories. It'll be too much. Um but but it's true.
It's true. Right. We last year uh at Cambridge University I and a couple of other professors did a benchmark in which we ran a simulation of a digital twin of an of an actual automotive company. The gentleman has left now but uh it was an actual uh we didn't do it on Mia or but we actually did it on Tesla and we gave it we we gave
that digital twin as a video game to MBA students and executives and we had thousands of them play it. We took the four top four players in corporate strategy and then we benchmarked it with chad GPT playing the same simulation. Uh it was not good news for the best strategists in the world. So even for CEOs we wrote that article it came in Harvard Business Review. It
was called uh uh AI can mostly outperform human CEOs. And this was last year. So I effectively put myself at out of a job. The business school didn't like it either because we promised to build the future of of of CEOs and management and the next Steve Jobs should come from University of Cambridge. Uh turns out it comes from a mathematical model. So, you know, kind of
puts everything in perspective. And the reason I I think I'm talking about my personal experiences more is because it's very easy to come in and say, well, because I'm making money on it, right? But that's not the point. The point is that it has affected my life. it is probably affecting your life and you are standing at Everest knowing fully well that you might not be ready
to climb the peak but guess what you have to there's no other option so ju just telling you that I'm probably as screwed as everyone else in this room um which brings me to this graph these are a lot of frontier lab kind of graphs uh this is about forecasting so what do managers do what are what is the difference between great leaders and mediocre leaders Great
leaders tend to be right about the future more often than not, If you're talking about money, profit, and loss, of course, great leaders also need to be empathetic. Uh you you don't you don't want a jerk for a manager. And so, parking all of that aside, for a shareholder, a great leader is someone who can make great predictions. At its crux, that's what the role of a
CEO or of a CXO or of a senior manager is. And for we are not very good at forecasting in general. So this is something known as a super forecaster. A super forecaster is a human being who has consistently beaten in terms of accuracy complex predictions about politics, economy, exchange rates, stock prices versus thousands of general humans. These are known as forecasting tournaments. uh where nerds get
together and they sit down and they predict uh and uh for LLMs they weren't very good with GPT4 the lower you go the better it is the more accurate you get so in July 2025 it is missing the latest state-of-the-art model which is claude 4.5 uh but the thing over here is this that uh this red line for LLM forecasting is based entirely on no tool tool
and no additional data. So basically they 2025 knowledge human beings with 2023 knowledge LLM. So no tools, no internet and even then the LLMs beat the super forecasters the very best predictors in the world. It was able to beat without looking at fresh data. It was able to do that based on the training it had of history. And while I understand the gentleman over here said history
is done. Eh, not really. History teaches you patterns and uh when you do forecasting patterns uh is what enables a good forecast from a bad forecast. And so the information for the LLMs was not updated. The LLM was running on 2023 information. Human forecaster was running on 2025. It predicted beyond 2025 based on historical patterns and it was more accurate. Which means that being a CEO is
not as grand as it used to be. I mean my shareholders can call up Chad GPT and say is Hamza doing a good job? They'll probably say not really. And you know you get into that sort of issue. Um this is a logarithmic graph. If I had if I had done the y-axis or the the meter labs would have done the y-axis uh it would have just
been a graph like this. you would not be able to see the difference because it'll be like look climbing up a wall. So it they made it logarithmic and in in the in the area of LLMs in GPUs um time matters. Time matters because you give more time uh the LLM thinks for longer. When it thinks for longer it does a better job. It's it's a thing.
Um so at GPT5 we've reached 1 hour of consistent work. We think from this graph that by the middle of 2026 an LLM can be given a job an agent can be given a job that can run for up to 8 hours. So you can give it like um a really complex problem and it will take up to 8 hours to solve it and do it correctly.
A human will probably take an big difference. Uh we do state-of-the-art harnesses around state-of-the-art models and for example in our harness uh we've been able to hit uh around an hour ourselves and up to 300 tool calls a single shot before the an before the agent gives you the final answer. It does 300 investigations without asking for your permission and is tends to do really well. Tends
to do really well. Right. So um if we give them more time, if we give them more GPU, we we climb the Everest faster, So what do we do? Any ideas? Any thoughts? I know it's a big it's a big room. I'm pretty sure people have their thoughts and ideas but uh this is not going this is not very uh comforting or do you guys feel comforted
with the idea that I mean are you loving it that an LLM can that sort of beat your CEO and predicting the market I mean what's the point of having him or her there right what's so it it kind of sucks in general for senior managers now as well it's not just software developers or biologists and whatnot so what do we do from here. So what do
we do from here? So we need to learn from history. Uh does anybody know about John Henry Irons? Okay. So John Henry Irons is this um uh comes from this folklore. We do not know whether John Henry Irons was a real person or was a madeup story but it's an American one. And it's uh back in the industrial revolution which was huge. And in in in industrial
revolution, they got a a steam drill, big innovation. The LLM of its time was the steam drill. And uh so this um tall African-American guy represents this whole community of mine workers predominantly from the same uh sort of ethnic background. And then you had this uh boss come in and say I have got this steam drill that can just drill through the mines 8 hours a day,
12 hours a day. It just keeps on drilling. You don't have to pay it much other than steam and maybe some oil to maintain it. So John Henry Irons did not like that. He was like screw that guy. So he goes to the he goes to his boss and he says, "I want to do a competition between myself and the steam drill." And this guy is like
a tall guy. This guy is like a machine. He's like the Incredible Hulk. I mean, you you you can't you don't want to pick a fight with this guy, right? He's huge. And he says, "If the machine fails and I beat it single-handedly, you will not bring the machine into the mine and you will not fire any of my brothers and sisters." I mean, if a 6'8
guy comes in and tells me that, I'll probably say yes. Don't want to pick a fight with the guy. He was like, "Sure, let's do the let's do the race." So, according to the folklore, they did the race. Guess who won? How many of you think John Henry Iron's one? One person. 2 3 4 5 6 Everybody else thinks the machine works. No, actually John Henry Irons
won by like a very small margin. But then he died. He had a heart attack and he died. Guess what happened the next day? The boss switched on the machine. Right? So the question is, do you want to be a legend or do you want to be rich? I mean, I I don't know. Uh I don't I don't want to die that soon. That kind of sucks.
So you you do you want to be John Henry Irons? It's your call. Some people want to to die at the altar, but as somebody who teaches innovation and so on so forth, I'm wired to believe that you uh you need to harness things instead of fighting them. It's kind of like saying, uh, you know, it rains a lot in Cambridge. Uh, I want to make a
machine that takes away all the clouds. Well, not really. I just want a roof on my head. So, you don't want to fight laws of nature. It's it's usually a dumb idea, but if you do want to do it, good luck. Um, and so, uh, LLMs are very much following the laws of nature. And if you look at sort of the state-of-the-art or frontier labs, the the
developers there all are already quite religious about LLMs. They talk about how they're breeding uh intelligence in a equivalent of a petri dish. So of course there is some some crazy there but uh don't want to be John Henry Islands, right? So uh so what does that mean? Which brings me to the last slide. I I was trying not to do a very long talk but the
so like any good academic in the age of AI I asked AI what does it think will replace uh humans right so the best thing to do it was to look at it from a historical perspective again history is important pattern recognition is important yes it has happened but you need to learn from it to be successful in the today and tomorrow so when we had the
agricultural revolution we found out irrigation we started digging sort of canals and uh controlling when plants grow and when plants die, things that were in the in the hand of God until we decided that we could just store water somewhere. That was the agricultural revolution and we had very simple jobs back then and a lot of them got replaced with you know farmer like more specialization built
around that technology. So that's one. So this is like 10,000 years ago, right? Then you had the industrial revolution. completely completely wrecked a lot of things but produced more jobs including some of the jobs that have built this building. You could have never built this building without the industrial revolution. You would not have had air conditioning. Now you have air conditioning uh repair men and repair women
and so on so forth. an explosion of new kinds of jobs while taking away old kinds of jobs but more jobs and on balance because of of this, right? Uh and then you had the reason why all of you are here. You are the children of the big tech revolution. you've gotten rich and successful and uh whatever else is a good indicator in Bulgaria of a great
life uh because of uh of of the big tech revolution. So all of us are here because of this and we we we collectively displaced switchboard operators, typists, uh travel agents. Who goes to a travel agent anymore? I I I don't know like I don't want to uh people who who do Fuji films and you know used to make pictures and Kodak moments gone. How many of
you know of uh of a business that is sustainable and is growing and and does so by uh printing out film? None. But what did you get? You guys. We got you guys out of this. And I take you guys any day over uh somebody who's uh I I I personally quite like developers. I I think they are uh they have far less in their system than
sort of managers. So I generally I'm happy that this change happened. Um you have web developers, you have you know all of those things. Then what happens with the AI revolution? So this time around I was too lazy to sit down and do my calculations. So I asked our AI what it thinks it's going to replace. So it basically put in a bracket there which is called
conjecture. Thank god it did not lie. Uh I think one of the things you will notice over here is that at this point the AI thinks more jobs in absolute count will be replaced. Uh so the number of jobs coming in their place is going to be lesser in count. uh it's conjecture. I I think that if I had to read this and please don't take this
as gospel, it's a bunch of tokens, but um the way I see it, it almost appears that uh everything seems to have an AI attached next to it in the title. Basically, what even the very best AI thinks that it's going to be a hybrid model. And uh which brings me to the sort of the last point I wanted to make. Why do I love developers so
much because unlike most of the people in this world, you actually have agency on where this goes. A lot of people think that technology happens and one fine day, you know, we get hit by it like a bus and we don't know what happened. But actually technology is done by people like you. Millions and billions of dollars are flowing in to get till here which means you
have way more agency. So if there is a doctor or a lawyer or an educator who needs to be AI augmented who's going to do it for them? You guys, right? So I think unlike any other conference, this is a special one because you get to choose. So there are so many new jobs that you will be powering. So please don't be afraid. Yeah, is going to
hit the fan. Okay, fine, whatever. But you are in charge. You're still in charge. And uh make the best of it. And with that, I would say thank you. >> Uh thank you. Thank you cuz I think that we have some time for some questions. >> Okay. I I can give answers questions. >> No questions. You guys are the best. I love develop five week bug off.
>> So I >> well uh at least a question from me. >> Okay. Yes. >> Uh well I saw on the last slide what uh opportunity we have as an engineers. But what about the CEOs? >> Uh they don't have a lot of opportunity. You guys are the next ones. So don't mess it up. >> Okay. Thank you. >> All right. Thank you.
