AI and entry-level jobs

0:00 · Title Let's start then. Good evening. Thank you very much for coming. My name is Ben. I think our time and attention is one of the most valuable things we have and you have kindly offered me yours for the next 60 minutes or so. So thank you very much. I really appreciate it. The latest technology promising more of our time back to us is AI. Now I know many of you probably are sick of hearing about AI but I'm afraid we can't bury our heads in the sand on this one. We have to face the consequences it presents to us head-on. Many parents I talked to are concerned about AI's impacts on their kids or students learning but I'm far more concerned about AI making certain learning obsolete. I used to implement AI for businesses and I can sort of see this coming and that's what we're going to talk about this evening. AI's impact on entry-level jobs. Now we're going to look at the trends of previous technologies and we'll see that the progress seems to follow certain rules or laws and the progress is extremely predictable and we're going to see if that those certain trends or that progress also applies to AI. We're going to look at AI's capabilities because of those trends and see if that translates to any impacts showing up in the employment data or labour market data. And finally we're going to see what to do about it. But to give you a hint there's a reason I left my job as an AI engineer to found my company Stable Future and talk with you tonight because I think those impacts are starting to show up. So let's begin. The progress of certain technologies is extremely predictable. In fact sometimes it's so predictable that when we plot the progress on a graph they seem to follow straight lines and the progress follows certain rules or laws. So to see the future we don't need to believe in science fiction. We just need to look at straight lines on graphs. The most famous of these straight lines

2:37 · Gordon Moore was discovered by Gordon Moore. Here he is. Discovered it in 1965. He saw that the progress and the power of computers was doubling roughly every two years and that this doubling every two years will continue for at least a decade.

2:56 · Moore's law But it didn't continue for just one decade it continued for six and it's still alive to date. Now as promised this is a fairly straight line but that's only because we've changed the y-axis here. It's not the y-axis that we're all used to. Here we're measuring computing power on the y-axis. Normally it would start at zero and goes from 1,000 to 2,000 to 3,000 add up a thousand every time but this one multiplies by a thousand every time so it doesn't go it goes from one to a thousand to a million and so on. Which means this graph is hiding something.

3:40 · Moore's law, linear scale It's hiding something that is every tech nerd's favourite word. Mine included exponential. Computing power progressed exponentially over these 60 years and computing power didn't increase a thousand times or 10,000 times it increased 300 billion times and

4:01 · Computing cost at the same time as that 300 billion times increase in computing power there was a efficiencies in the manufacturing of computers also led to decreases in the cost of computing power. That was also exponential and it didn't just decrease a thousand times 10,000 times it decreased 30 trillion times. So those forces interacted together which means that today we each have super computers in our pockets that are more powerful than the ones that took us to the moon in 1969. I think that's pretty amazing.

4:43 · Computing: cost and use So the story here is that the cost of computing power decreased exponentially at the same time that the computing power increased exponentially. So those forces interacted together which led to computing power and computers diffusing throughout the economy and being adopted everywhere. And again it was adopted, guess what, exponentially. So we don't just see that progress or this pattern show up in computing alone. We also see it

5:24 · Solar in solar panels. We see it

5:26 · Wind in wind turbines,

5:28 · DNA sequencing DNA sequencing,

5:29 · Batteries rechargeable batteries

5:31 · LEDs and LED lights. I think you can see where I'm going with this. So AI is the latest technology on everyone's minds. We need to measure two things. We need to measure the cost of AI and we need to measure the performance of AI. And if we see similar trends in AI then we can predict with a reasonable amount of certainty that AI is going to be used absolutely everywhere in the economy over the coming years. So for the cost, the cost is quite easy to measure, right. It's very quantifiable. Just look at the receipt, right. Very easy to put a number on the cost. But the performance is something different. For LEDs it's quite easy, you just look at how much light is emitted for a certain amount of energy or for rechargeable batteries like how much energy the battery can store over time. But how would we measure the performance of AI? One way is to just say well we can kind of tell. It's kind of obvious, right.

6:46 · Will Smith eating spaghetti, 2023 Here is a video, an AI generated video of Will Smith eating spaghetti in 2023. It's not very realistic. Spaghetti going everywhere. His mouth's moving everywhere as well. It's crazy.

7:02 · Will Smith eating spaghetti, 2026 Here is the latest video model with the same instructions 2026. Sound effects included. Smooth. Smooth. So performance is clearly increased. We can't put a number on it, right. It's not very quantifiable. A better way is to use the tests that we already have to measure our own performance. Things like, you know, tests, exams, competitions or if you're really smart, scientific breakthroughs like Newtonian physics or Darwinian evolution. So let's use those tests and let's start off

8:03 · Primary school maths: GPT-3 easy with primary school maths. Models in the middle of 2023, sorry, the middle of 2020. This was a model from the makers of ChatGPT called OpenAI. Middle of 2020 they're getting just under 10% on primary school maths. Not very good. Thing to understand about AI is that they are computer brains. They are neural networks, similar and based off the same architecture as our own brains. So it's just a computer brain. The researchers made the computer brain AI more efficient and with those efficiencies after two and a half years of research,

8:46 · Primary school maths: GPT-3.5 the next best model scored 50%. Good improvement but two and a half years. and we're only on primary school maths. So instead they made the computer brain ten times bigger with better computers, bigger data centres

9:05 · Primary school maths: GPT-4 and finally the test was declared effectively solved when that model scored 90%. So we're gonna have to find a new test. But researchers at this point thought the progress was a bit too slow. They needed a new way to progress these models. So what they did was they taught the models to reason through the answer before they gave it. So you give it a maths question, it breaks down, like we do, the maths question, into steps, reasons through it and then produces its answer. And this, this is called the reasoning paradigm. You can think of it as the models, the reasoning and thinking for themselves. This turned out to be a breakthrough because only two years after this point,

10:00 · Maths Olympiad gold models that were released summer of last year, Gemini and ChatGPT won a gold at the International Maths Olympiad. So this is at the level of the brightest young maths talent in the world. Over the same time, over that two-year period to summer of last year,

10:24 · PhD-level science they also beat PhD level experts in their own fields in physics, chemistry and biology. So physics questions, PhD answered physics questions right in their own field, they got 65% models today get 95% significantly outperforming PhD experts.

10:46 · The IQ curve If we were to measure their IQ, they were significantly below the human average, now they're significantly above even the top 1%.

10:58 · Every domain, the same direction And however we measure this progress, whether it's in reading a clock, which they're still below human, strangely enough, or general common sense, fixing software bugs, no matter what it is, progress is only going in one direction, is only going up and to the right. So this is up to, again, summer of last year. And up to this point, there are questions about whether AI could reason and generalise beyond its training data, or whether it's just a parrot, just regurgitating the answers we give it in the training data. So there are two sides to this debate. And this year, we found out the answer, because just a few weeks ago, the latest AI models, again, bigger computers, more efficiencies in the computer brains,

11:55 · Navier–Stokes they solved a Millennium Prize problem in maths. Millennium Prize problem, there's only seven of them in the world, they each have a million dollar bounty for solving them. And some of these problems, they've stood for decades, or a century, or more. And it solved one of them. This one in particular is called the Navier-Stokes problem, and it's to do with fluid dynamics, the way water moves in a container, for example. So it's making scientific breakthroughs. And it's important to note, this isn't like a maths researcher that's sort of leveraging an AI model to, you know, go back and forth and iterate on ideas. This in particular, and many other maths problems solved by AI, people just say, do a breakthrough, and try really hard, and believe in yourself, and the models go away and solve some of these problems. It doesn't happen every time, but that's one of the methods researchers are using to make frontier breakthroughs like this one. But this is a talk on entry-level jobs. That's what you're here for. There's not many maths researchers in the world. You might be having a bit of trouble, but it's not really representative of the economy and real-world jobs. So what we'll need is a test that tests for exactly that. And my last benchmark, my last exam, will show exactly this. But what you've got to understand is that we've moved on from the reasoning paradigm or the reasoning era to the agentic era. You might have heard of AI agents or agentic AI. These are AI models that they still break down the task that you give them into steps, but now they achieve goals. So instead of asking a model like ChatGPT, instead of asking them a question, you give them a goal. You say, you know, sort out this spreadsheet. It will break that goal or that task down into steps. It will complete each step and therefore complete the goal. So these models, ChatGPT, anything you do on a computer, it can now also do. This is the agentic era.

14:29 · Agents' Last Exam: the tasks My last exam is called agents last exam. It's across all of these domains. There's 1,500 tasks in this benchmark. It's across, you know, engineering, agriculture, legal, health and medicine, you name it, it's on there. This is a representation of what I like to call the laptop class, which I am a part of or the, you know, the knowledge economy. Each one of these 1,500 tasks was handpicked by an expert in their domain and it was a past project that they did. And it often took some, it took them an average of days or weeks for them to complete this project. And if you had to add up all the time it takes to do these 1,500 tasks, it would take someone their entire career to do them, to work through them.

15:26 · One task: a Greenland lake Here's just one example. It's climate science. So a climate scientist would have to look through images and figure out how a lake in Greenland got drained or how it disappeared due to ice caps melting and things like that. 6,000 images, you have to sift through three gigabytes of data.

15:49 · Agents' Last Exam: 10% So let's see the progress. Let's have a look at the progress. Those models around the same time was that International Maths Olympiad gold last year, they got 10%. Not that good, I mean kind of impressive because it is, you know, completing these tasks but it's only 10%. Again, bigger computers, more efficiencies.

16:16 · Agents' Last Exam: 50% Three months ago the latest models got 50%, 50%, half of a representation of the entire knowledge economy. Now, how long did it take this model to complete all of these 1,500 tasks? Well, half of them. It attempted, it attempted all 1,500 but obviously only got 50%. So how long did it take for the model to work through these problems? It took less than a month. And how much did it cost to achieve 50%? Cost 2,000 pounds. Exactly. And if you think that progress has stopped and that we might have hit a wall and there's no more progress to be had, well,

17:10 · Agents' Last Exam: 60% a few weeks ago the latest model got 60%. And in fact, last week, I've not added it to this because it's hard to keep up with all the progress but last week a model got 63%. And so now we're no longer measuring progress in years, not even measuring progress in months, we're now measuring it in weeks. So progress is occurring exponentially, performance is exponentially increasing. So that's what's happening with performance. With costs, last year an international maths Olympiad gold would cost 2,000 pounds, this year it would cost 20 pounds. So it depends on the domain and the sector that you're measuring but costs are decreasing about 10 times a year, something that cost 1,000 pounds, last year it cost 100 pounds, this year 10 pounds, next year. That is quite predictable progress. Now, you might be thinking, okay, it's getting 60% on agents last exam, the job market has to be in freefall, right? But that's not what we see. Like I said, it takes time for technology to diffuse throughout the economy and probably some of you here haven't heard of these results, which means business owners probably haven't heard of these results and they've not caught on yet. So again, it takes time for the technology to be adopted. So let's look at current job market impacts. The best, the most in-depth studies, what they do is they measure a sector's exposure to AI. So that's just like, you know, you take a job, you see which tasks AI can do and the more tasks AI can do, the more exposed that job is to AI, right? The best, the most in-depth

19:18 · Workers aged 41 to 49 studies show that employment in the most AI exposed sectors is actually up since the release of ChatGPT, which is really strange. You think it would be the opposite but those same studies look to the effect of age. So this comes with a huge caveat. Senior employment is up. Some of their wages. Seniors are having a great time. Even the most exposed seniors, employment is up compared to the least exposed seniors.

19:51 · Workers aged 22 to 25 But when we look at junior employment, we see the opposite. Junior employment in the most exposed sectors is down. Why is this? Now there are a few theories, but I think the one that makes the most sense is that seniors are the task setters, juniors are the task executors. And AI has become an excellent executor of tasks. It's now better, often better, faster, cheaper and safer than a junior. So normally a senior would set the task, give it to the junior, junior would produce an output and the senior gives feedback, iterates on the output and then verifies, yep that's good. Off you go. AI is now replacing that role of juniors of executing tasks.

20:48 · The corporate pyramid This is a standard corporate pyramid. You've got seniors at the top, juniors at the bottom.

20:54 · AI takes entry-level roles What we're seeing now is this in the most exposed sectors. AI is beginning to take entry-level roles. I believe in the coming years, it remains to be seen, but I think with a fair amount of certainty we'll start

21:12 · AI climbs the pyramid to see this AI climbing the corporate pyramid taking more entry-level roles. Now whether or not it can climb any further than this, whether or not it starts setting the tasks for perhaps even AI to do, remains to be seen. We're not sure, but certainly this will start happening. It already is. So the number of juniors entering the entry-level job market through things like apprenticeships or degrees, that's sort of fixed for the next few years. People are already in university, people have already started apprenticeships. However the number of opportunities is shrinking. So that means competition is rising. There is a bar to land a job in these exposed sectors. Above the bar you are leveraging AI, and below the bar you are replaced by AI.

22:15 · The more exposed the job, the more you need The more exposed the job, the more you will need AI skills because you'll be leveraging AI, grades to stand out from the competition, work experience to show that you know about the sector, and network, which is something people often miss, because AI is now flooding online applications. People use AI to make their applications more efficient, and now employers can't keep up with the number of applications that they've got. And so they're going through people they know or recruiters to fill the roles. And lastly, evidence of enthusiasm. Like all I mean by this is a personal project in the sector to show that you're genuinely interested in the work. All of that will require work ethic and stress tolerance because it's on top of the expectations that you already have as a student, on top of the grades, on top of the degree that you meant to get, on top of the apprenticeship and so on. Now, hands up, who can drive? Okay, that's pretty much everyone. Keep your hand raised if you believe you are better than the average driver. Is that hand going up? It's humble man, humble man. 90% of people believe they are better than the average driver. That's mathematically impossible, right? Only 50% can be above average, and it's fortunately the other 50% below average. So as competition is rising, bear that in mind, you know, if you're a student, bear that in mind. We overestimate our own capabilities and we underestimate the capability of others. Okay, so this is, yeah, this is kind of a bit depressing, you know. If the jobs aren't in the laptop class, part of the laptop class and knowledge economy, where are they going to be? In other words, how can we find the jobs of the future? I think the one way, one way to look at this, the one that makes the most sense, is to look at it through a lens of supply and demand. I'm sure we've all heard of supply and demand, but just so we're all on the same page, if I'm in my kitchen and I'm really thirsty, and someone offers me a bottle of water for a thousand pounds, I'm not gonna take that. It's a bad bet, it's a bad deal, because the supply of water from my kitchen tap is massive. I can get it for next to nothing. Put me in a desert, though, where my demand is still really high, because I'm really thirsty. That offer of a thousand pounds for that water starts to look really tempting, because the supply of water is so low. So what we want to look for are sectors, where the demand for work is massive, but the supply of people that can do that work is low.

25:33 · Where the jobs will be So here are some sectors, not an exhaustive list, but just to give you an idea. The most obvious place where this dynamic is taking place is that we have an ageing population, which by definition means the demand for health care is massive, right? It's increasing over time, because we have an elderly population who want health care, but at the same time, the supply of people who can fulfil that demand, you know, health care workers, is shrinking, because less people of a working age. So, see things like medicine and dentistry, I think, is a really good option. It already was. We have physiotherapists, paramedics, adult social care, and nursing. I think these are all good options. Sadly, climate change is a problem we're gonna have to deal with, and the green transition needs hands. So, we'll need solar and wind turbine technicians and engineers. We'll need to electrify the grid to move away from fossil fuels. We'll also need the retrofitting of buildings and houses, for things like heat pumps, air conditioning, and insulation. And one thing that's not so, well, one sector that's not as obvious is flood, coastal, and drainage engineers, because we're gonna have more extreme weather. So, that's climate change. Next one, geopolitical tensions. Sadly, again, this is on there. I'm not about to suggest that everyone go and join the Army or Navy or RAF, but again, they'll need engineers, technicians. We need to build submarines as well. Drones, munitions, and so on. Finally, last one, AI and robotics. Robotics is about where AI was five years ago. It's sort of known to be the next big thing, but people haven't quite caught on yet. I think becoming a robotics engineer, if that's your jam, that's a very good option. And AI, like I say, if you believe that you are capable enough to rise above the competition and you're really interested in the work, then leveraging AI skills to go into something like law, software development, banking and finance, so on, is a very good option as well. So again, not an exhaustive list, just to give you some ideas.

28:14 · Direction, skills, backups We spend 80,000 hours in our careers. Picking the right career was already one of the most important decisions of our lives. AI now makes that decision even more important. Taking the time to make even a 1% improvement in that decision can yield a massive return on investment. So how do you choose well? Step one, pick a direction. Students should align their careers with their interests, skills, and preferences, and they should find the best ways to enter those career paths, whether that's through a degree, an apprenticeship, or straight into work. But that's not nearly as important as step two, which is developing valuable skills. You need to understand which parts of your job AI automates and which parts remain distinctly human and then focus on the parts that have the highest skill ceiling. PwC estimates that employees with AI skills earn an average of 60% more than employees without AI skills. Even if you are on UK minimum wage, that means you're an extra 10,000 pounds better off every year after tax. That is the value and importance of developing the right skills in the AI era. And finally, step three, you need backup plans. You should develop an A-B-Z career plan in case of economic shocks or extreme scenarios. So your plan A is your preferred career path. It could be somewhat exposed to AI. Plan B is a viable alternative you're still very happy with and that is ideally less exposed to AI. And your plan Z is your lifeboat career. So no matter what happens, you've got something and that should be minimally exposed to AI. I left AI engineering to found stable future. We've mapped over 1,000 UK jobs to help young people implement exactly those three steps. The direction, the skills and the backups. So you as a parent can stop worrying, start acting and make sure your child is financially stable in the AI era. So if you want to put them in the best possible position visit stablefuture.uk. Thank you very much for your time.