A bright future

Key Points:

• AI adoption is accelerating and for good reason, it has gotten a lot better over the past two years.
• We think prospects for further improvements are bright. Widespread adoption of AI agents is likely in the short term, and the rollout of AI powered humanoid robots is probable over the medium term.
• History suggests technological revolutions create more jobs than they destroy. However, AI still threatens material job losses given the pace of improvement and risks creating a generation of workers who miss out on foundational learning because using AI is too easy.

‍Imagine walking into your office in 2030. Your AI assistant has already prioritised your day, responded to a bunch of emails and set up a few meetings. Your humanoid robot receptionist has finished the office clean and is ready with a coffee when you arrive. Meanwhile, half your team consists of AI agents who never sleep and barely cost anything. It’s not an unreasonable scenario. In this month’s Market Insight, we will outline what we think AI means for the future of economic growth. The estimates of the economic impact of AI from economists, NGOs and think tanks varies widely, from boosting GDP by a measly 1% over ten years (2024 Nobel laureate Daron Acemoglu) up to a 3.4% per annum boost between now and 2040 (McKinsey). As you can probably guess from the intro paragraph, we think the future is bright, but given we aren’t Pollyanna or Cathy Wood we sit somewhere in the more reasonable lower middle of that range. We first wrote about developments in AI in March 2023 and introduced it into our Strategic Asset Allocation Review as a modelled scenario a few months later. Since then, progress has been rapid. Models have become smarter, faster and less error prone. AI agents are becoming more widespread, and we are now getting glimpses of a future filled with AI powered robots. The far future is even more exciting.

‍From Novel to Useful

Over the past two years, AI models have gotten much better and their integration into work and home life has accelerated. Their performance across expert level specific domains (coding, math, science etc) now surpasses expert level human results. Their performance across simple reasoning tasks is substantially lower than a human (currently around 50%) but is still steadily improving. While these results may seem counterintuitive, AI excels where data is readily available. Science, coding and math are extremely well-documented domains. Novel reasoning with basic concepts is much more complicated as there is little structured data to train a model on.

Part of the reason models have gotten much better since early 2023 is the amount of computing power being thrown at them has increased materially. Models with more computing power outperform those with less. As do those trained on more data or with more model parameters. This is known as the AI scaling law, and for the most part it follows a predictable pattern - you can roughly estimate how much better a model will be if you know how much bigger you made it. There have been some efficiency improvements in model training and falling hardware costs, but a lot of improvement has been by brute force. From a practical perspective, that’s fine, the cost of compute and data has always fallen structurally over time.

The functionality of AI from a practical perspective has improved substantially. Most top tier models can now create and run code in a web browser, which wasn’t available in the past. Some can generate web artifacts, which can range from in browser web apps to 3D games. AI can go away and generate lengthy reports on complex topics. In March 2023 we marvelled at text to image generation, now we have text to video generation. When people Google search now, the AI overview at the beginning is more often than not sufficient, no need to trawl through web pages. The impact from a productivity perspective is becoming clearer, with many businesses reporting increased usage of AI and lower operating costs.

Many of the early problems with AI also seem to be mostly ironed out. Hallucination, where a model invented an answer which sounded correct, but it was actually made up, has become much less frequent. AI images of people now have the correct number of fingers on each hand. Answers in general seem much more logical and reasonable. AI has gone from a novelty to a useful tool.

‍AI Agents

With the use of web-based chat models now relatively ubiquitous, what is the next evolution of AI integration into the economy? In the very near term, we are in the middle of a more widespread rollout of AI-based “agents” across many businesses. As has been the case with most AI developments, the cutting edge of this has been in the tech sector. Developers can now very easily assign AI agents to complete coding tasks, even complex ones. They run in the background, check code for bugs, run tests and iterate through different options until the code works as intended. The developer can then review this work and change as needed or integrate it into their process. Importantly, this can happen across multiple tasks in parallel. In practice, this means a close-to-zero-cost team of workers sitting under every developer. The productivity benefits are enormous.

AI agents are also more commonly being deployed across other industries in task automation. Salespeople can use agents to automatically qualify leads, personalise outreach messages, and update CRM records. AI agents can screen resumes, schedule interviews, and handle routine employee inquiries about benefits or policies. Some companies use AI to facilitate the first round of interviews. Agents can automate data entry, manage calendar scheduling, process expenses and create invoices. Basically, any task which looks pretty much the same every time and uses a computer can be completed by an agent. This space is in its infancy, and we expect agents to become quickly integrated across organisations, and be capable of increasingly complex tasks, sooner rather than later.

‍Humanoid Robots

A little (or a lot) later along we can expect AI powered robots to become common in homes and workplaces. The closer these robots can come to being trainable and operating functionally in a human form (as opposed to a robot vacuum cleaner for example, which has a single function), the closer we will come to a golden age of productivity growth. The humanoid aspect is extremely important. The world is designed for humans to operate in. The better robots can operate flexibly within that environment the closer they will come to being able to onboard workflow from people working in industries outside white collar office work – which is most jobs (denoted in blue in the chart below).

While this may seem like science fiction, Mercedes and BMW are trialling working humanoid robots in production facilities now. Amazon is testing humanoid robots for package delivery. You can buy a base level Unitree G1 from China for around $40,000 today, though there is no out of the box AI so doing anything practical with it is out of reach for the average person today. That said, companies such as Figure and Apptronik show very impressive videos of their early-stage robots being trained to operate autonomously. For now, the focus is mainly around warehouse packing and sorting and simple repetitive manufacturing actions. They are slow, and still expensive. However, as with all new technology, over time the price falls and the function improves.

Self-Driving

Self-driving cars have felt just around the corner for a number of years. We actually think they are much further away than people think, with some caveats. Most consumer vehicles will remain stuck at low levels of automation, requiring constant human attention, due to cost, regulatory barriers, and technical complexity. Consider a car at 110kph encountering rain, roadworks, poor connectivity, and kangaroos - all simultaneously. AI can't respond quickly enough to these edge cases, and training data is scarce. With catastrophic consequences for failure, the economic payoff, while large, isn't game-changing. Compare that to a humanoid robot in a car yard. The robot washes the cars every day, makes coffee for customers, mops the shop floor, organises the tools in the workshop every night, maybe it performs some maintenance on the cars. The consequences of failure are much lower. A dropped tool, a dirty floor, a dent in a car. The technological challenge is also easier.

The exception to the above is the robotaxi model and potentially efficiency in automating fixed trucking routes. Waymo operates commercially across multiple US cities providing over 250,000 weekly rides with better safety records than humans. However, they run a fleet of professionally maintained vehicles operating in carefully mapped, limited geographic areas with extensive remote monitoring and support infrastructure. The environment is much more controlled than ever will be the case for personally owned, self-driving, cars.

‍Mass Automated Research

We would be very surprised if AI in its current form cures cancer, creates a quantum computer or designs a working fusion reactor. They are language models, trained to give an answer to a question based on mountains of training data. Sometimes those answers seem novel or original, and perhaps by chance they are. However, we don’t think these sorts of models can have the kind of sudden intuitive leaps or paradigm-shifting insights that characterise the most revolutionary scientific breakthroughs. Perhaps new models will in the future, but they may never exist.

That said, we think AI does have the ability to accelerate the pace of scientific discovery. Just like AI today is helping equity analysts analyse company earnings calls quickly, scientists are using AI to process decades of research papers in minutes, spot patterns in data humans might miss and run thousands of experimental simulations simultaneously. A very real bottleneck in research is the limited by capacity to run experiments. Like agents can improve the productivity of software developers by building code based on instructions, in a world of AI powered smart humanoid robots, there is little reason why AI couldn’t design laboratory layouts to achieve the desired testing outcome, instruct the robots to fit it out as necessary and run multiple iterations of experiments, then pass back to the AI to process the results and either continue to iterate or finalise results with a supervising human. This workflow could materially accelerate the pace of scientific discovery and while it isn’t going to exist in our 10-year Strategic Asset Allocation horizon, it seems possible in the coming decades.

‍The Cost

At a macro level, there are two major concerns about AI, one of which we think is talked about too much (mass unemployment) and one of which we don’t think is talked about enough (lack of skills development). We will ignore the prospect of a robot uprising for now. We also ignore non-macro negatives from the adoption of AI (data privacy, ethical considerations, environmental impact, potential reduction in social cohesion) given limited investment implications.

While it is a risk, prospects for mass unemployment are pretty low in our opinion. To completely displace a large part of the workforce, AI would need to be a near perfect substitute for a human, including our soft, interpersonal skills. Let’s imagine a world 10-20 years from now, where AI agents very easily replace simple, repetitive tasks in office life. Processing data, invoicing and bookkeeping, low level email responses and customer service, producing standard reports, inventory management and ordering for example. Let’s also assume humanoid robots meet their early potential, packing and sorting in warehousing, simple food preparation and service, cleaning, transportation and repetitive manufacturing among other things. How many people doing this sort of thing are at risk of losing their job?

We think the answer is found by thinking about what AI is really good at and what humans actually do all day. AI agents and robots excel at performing relatively simple, repetitive tasks. Most people in the workforce don’t just do simple, repetitive tasks all day. Often, that is a part of their job, but there are numerous, unforeseen things and bespoke tasks that require attention. Automating the repetitive things will allow more time to work on everything else. The proponents of mass unemployment argue that because the repetitive things have been automated, businesses now only need 2 people doing the unforeseen and bespoke, not 4. Therefore, their workforce will halve. That’s is a great unknown. What proportion of businesses will choose to produce more output with their now more efficient labour force versus keeping output constant and reducing headcount? For what its worth, from where we stand today most businesses expect AI to have around a neutral impact on employment overall, though there is dispersion across industries (see below).

History suggests that positive technological shocks don’t lead to structurally higher unemployment. The harnessing of electricity, the invention of combustion and steam engines, ploughs, vehicles, computers and the internet didn’t lead to mass unemployment. It’s always easy to imagine the jobs that will be lost, it's hard to imagine the jobs that will be invented on the other side. We think for the most part, economies will become much more productive. Where there is change, it won’t happen immediately as businesses take time to adopt new practices. This will allow some redundant people to reskill and seek work in other occupations. Some will be left behind, but the net effect should be positive.

Ironically, the other major macro risk we see coming out of AI advancements is a potential negative productivity shock. In work, the simple, repetitive tasks are normally given to 17-year-old apprentices or 21-year-old graduates. The apprentice digs the holes, sweeps the shed and cleans up the site. The graduate spell checks the report, gets the coffee and does the cold calling. What will they be doing in 10 years’ time? Probably not that, and probably not doing their manager’s work either. New entrants are more at risk of joblessness because they have the easy jobs AI is likely to take. If there are fewer new entrants, there are fewer people being trained to be qualified tradespeople and managers.

In addition, for the lucky ones who do secure employment, how many will over rely on AI and not learn how to perform their roles from a first principles perspective? By asking AI how to wire a breaker panel box, the junior electrical engineer never learns why things should be done the way they are. The junior software developer never learns how to code because AI writes it for them. The junior lawyer relies on AI to research case law rather than doing it themselves. Perhaps this doesn’t matter. No specialist in any industry understands the entire chain of events that goes into their output. As Leonard Read and then Milton Friedman argued, no one knows how to actually make a pencil. Still, there seems to be a reasonable chance that AI could lead to a gap in workforce skill development, which could have negative implications for productivity growth in the period beyond that.

‍Portfolio Positioning

AI has been included as part of our Strategic Asset Allocation Reviews since 2023, both as an explicit positive scenario and as a general positive influence across the board. As the probability of the positive scenario rises, our portfolios will try and capture this thematic more explicitly. From an equities perspective, the changes associated with AI will clearly generate winners and losers within the market. Valuations for anything to do with AI are expensive, but perhaps that is justified given its potential transformative impact on the economy.