Of Books

On the State of a Future in AI

Essays on where AI is actually taking us.

35 min read · revised 21 August 2026 · ·

This is the org-roam home of the book manuscript maintained in LaTeX at 2026AIManifesto. This file is the org source of truth for publishing inside the KB (HTML export via SPC a o r w, PDF via C-c C-e l p using the built-in "book" LaTeX class plus the header snippets above, which reproduce the manuscript's page geometry, margin-note styling, and running headers).

Preface

My thesis is that over the next few years AI will move beyond screens and become embedded in the physical infrastructure of everyday life: airplane seats, bus stops, public transport systems, homes, workplaces, and many of the objects we currently treat as ordinary. This transition will require a shift away from today's highly centralized model of intelligence, where most reasoning happens in distant data centers, toward edge AI systems that operate closer to where decisions are made. Without new infrastructure, societies will keep trying to extract more efficiency from aging systems rather than building the conditions for the next stage of economic and social transformation.

This future will not distribute intelligence evenly. We are likely to see tiers of AI: cheap models for everyday use, costing pennies per million tokens, and highly capable specialist models reserved for organizations, decision-makers, and those who can afford them. This creates an inverse labor dividend. People already responsible for complex decisions may use AI to take on far more work, accelerate their careers, and move into ownership roles faster, while others may become increasingly managed by algorithmic systems. The gig economy already shows this pattern, where human labor is directed by platforms that decide where people go, what tasks they receive, and how their performance is judged.

AI will also compress institutional and economic timelines. Companies may form, scale, mature, and die much faster as AI reduces the time required to build and test new ideas. Careers may also become less linear, with people exploring multiple professional paths over a longer and more flexible working life. At the same time, AI risks dehumanizing domains such as labor, media, and war by making management, content production, and conflict cheaper and more automated.

To navigate this transition well, organizations will need to accept a certain amount of institutional risk and create sandboxes for experimentation, and involve genuine technical experts in strategic decisions rather than acting only from fear of falling behind. The central question is not whether AI will increase productivity, but what kind of society that productivity will produce. One path leads to passive consumption, algorithmic management, and increasingly artificial experiences. The other uses AI to expand human potential, giving people more freedom to explore who they can become and how they can experience the world.

1. AI Leaves the Screen

AI has invisibly existed in the world for a long time now, since early neural vision systems such as Fukushima's neocognitron, and LeCun's later convolutional networks for document recognition \autocite{lecun1998gradient}, all the way down to the prices of real-estate prices being set by linear regression methods or multi layer perceptrons. With the advance in our ability to model language, the power of these methods became accessible to anyone who speaks English or another language which dominates the datasets used for training.

We have gone from background AI to foreground AI by concentrating on the single interaction modality of language. In the future I hope that we will move again from the foreground AI interactions to background modes. This future would be rich in computer perception which currently even in fields like robotics is highly sensitive to noise and out-of-distribution observations. Assuming we are able to overcome this challenge, the data we would be able to gather from cameras, PIR sensors and maybe even electromagnetic scanners should become invaluable to helping AI models make decisions for us and improve our daily lives. One such improvement could look like observing footfall patterns of people leaving their houses for work and dynamically adjusting public transport timings to accommodate everyone's schedules.

The subtler version of this is already in everyone's pocket. When Apple introduced Deep Fusion and Smart HDR \autocite{apple2023computphoto}, the company was not simply improving photography. It was inserting an AI decision layer between the physical world and the image a person sees. The camera sensor captures raw photon data across multiple frames. The computational pipeline then merges those exposures, suppresses noise, redistributes tonal range, and adjusts colors according to trained aesthetic preferences before the image is ever rendered to the screen. What reaches the eye is not what the sensor recorded. It is what the model decided the scene should look like.

The interesting consequence is that photographs taken on an iPhone often feel more real than photographs taken on competing hardware with comparable or even better sensors. The word "feel" is doing a lot of work in that sentence. Apple's post-processing is aligned toward human aesthetic preferences built from large-scale training, not toward fidelity to the raw optical signal. The output matches how people remember a scene more closely than how the scene appeared to an uncalibrated lens. We have come to associate the AI pipeline's output with truth, because it reliably produces images that satisfy our expectations of what truth looks like.

This is not a criticism of Apple. It is an early example of what invisible AI looks like when it is working well. Nobody opens their camera app and thinks about multi-frame fusion or semantic segmentation of foreground subjects. The technology disappears into the result, and the result is accepted as a direct record of the world. The same dynamic will repeat at every layer of physical infrastructure as background AI becomes pervasive. The optimization becomes the environment. The environment starts to feel natural. And the question of what the world actually looked like, before any model touched it, becomes harder to answer.

assets/fig_01_interface_shift.pdf

In robotics, gathering unstructured observations of the world and matching them to some hypothesized structure currently requires physical intervention in how people behave. This sort of crutch is what is helping us meet half way with the current intelligence capabilities of artificial reasoning systems. The idealized version which is currently being explored with VLA models is to abstract away this physical intervention by combining some generalized understanding of the world which is embedded in LLMs with post training on physical tasks. That direction I think is plausible however does not scale very well and requires a lot of labour into data curation. An ideal formulation would be creating a constraint or formulation of some problem and using some un-supervised learning approach to extract or distill knowledge from a pre-trained general model. This could look like creating a task formulation of "finding glasses in an apartment" and leveraging the knowledge a general model has of how people behave or where they might take off their glasses to teach this narrow task.

Ever since the Industrial Revolution, our work has been constrained by allowing machines to execute tasks at the expense of individual presence. The invention of the World Wide Web and the accessibility of cheap desktop and laptop technology have amplified the need for human presence. Now, when implementing AI, we only amplify the need to be present by a computer to read what the AI is saying and to give feedback. This, however, is not how most of us want to see the future. The future, in which bulk work is carried out using human-computer interaction principles, is likely to move toward a voice-centric system and an intent-based system, where tasks are carried out virtually anywhere, allowing us to pursue other interests while maintaining equal efficiency. This essentially makes anyone a manager of a workforce they have. But what this also opens up is the ability to experience more things in the world and in physical life, which is likely to contribute to a new economy where experience and interaction with multiple facets of life are more desired.

2. The Edge Intelligence Society

Currently, we outsource intelligence to centralized datacenters, moving away from the progress the tech world has made over the last 20 years in transitioning from client-server models to more distributed compute capabilities. Just as we once interacted with mainframes via clients, we now interact with LLMs over chat interfaces, yet the actual reasoning and "thought" occur in a distributed computational environment often thousands of kilometers away. I believe the majority of future requests to these AI systems will follow a long-tail, Beta-like distribution, where most queries do not require complex reasoning. Such requests might address world events, summarizations, or general knowledge. These tasks can be handled by smaller models that contain the necessary information within their weights. Instead of Dario Amodei's vision of a "country of geniuses in a datacenter" \autocite{amodei2024grace}, the goal should be a world of easily accessible generalists. In this way, we can provide equal access to intelligence and free up computational capacity in datacenters for more intensive tasks.

assets/fig_06_edge_architecture.pdf

If you are away from home and need external intelligence, you will likely be connected through a telecommunications provider. If you are on the street asking for a restaurant review or a translation, your request could be routed to an edge-computing device located inside the nearest cell tower. This could drastically reduce the oversubscription of datacenter resources and improve the user experience through ultra-low latency. This approach can be applied to any AI task, including navigation, logistics, or even entertainment, connecting to the concept of distributed tiers of intelligence available to different groups of people.

I see the primary constraint to this development as infrastructure, which has not yet fully caught up with the need for wider fiber optic connectivity. There are still developed countries without access to high-speed internet, yet we discuss AI connectivity—which cannot exist without the former. To build a truly intelligent city with smart capabilities, the infrastructure for energy distribution and connectivity must become flawlessly resilient. This requires fiber connectivity across all areas and low-voltage access even in remote parts of a city.

assets/fig_02_infra_readiness.pdf

A smart city can be envisioned as a network of sensors, cameras, traffic lights, gates, escalators, speed signs, and public transportation, providing easily accessible business intelligence for everyone. This vision cannot be realized without the underlying infrastructure. Is it the responsibility of a technology provider to "bring your own infrastructure," or is it in a city's best interest to build this infrastructure to enable a new economy of smart cities?

Once the infrastructure for edge intelligence and federated data processing is in place, the intelligence layer becomes the easier part. While many Europeans lament the need to catch up in the LLM race, I believe a better initiative would be to stop playing catch-up and, for once, think ahead.

Intelligence at the edge also means something different for sectors like insurance. We already have edge intelligence in our vehicles. Most models since 2017 or 2018 have adaptive cruise control and lane assistance, which is one of the earliest edge intelligence mechanisms \autocite{nhtsa2024adas}. It has most probably saved lives and reduced the number of accidents, even minor ones, which implies fewer insurable events for insurers. What that means for the future of insurance is a separate topic, but the broader point is that edge intelligence can massively improve risk coverage across many domains, not only in driving.

Consider streets and crosswalks. With more cameras on roads you can have much more intelligence to notify people not to cross even where there is no pedestrian traffic light, or add warning systems that connect directly to vehicle systems. There are now proposed communication protocols like V2V, vehicle-to-vehicle, that can alert drivers approaching a crosswalk to slow down before they can even see the hazard. This level of edge intelligence extends further: if someone is choking at a restaurant, you suddenly have a medical expert on hand. If someone is eating something and you have access to their medical records, you can have quicker diagnostics and emergency medical services can bring the right kind of assistance. There is very little downside to having intelligence accessible everywhere around us.

There is, however, the risk of excessive cognitive offloading \autocite{cogoffload2024}. If we suddenly have intelligence everywhere and do not need to think for ourselves, we become entirely at the mercy of technology, and technology right now is still very vulnerable to blackouts and cybersecurity threats. A society that depends on artificial intelligence for its day-to-day reasoning, even cheap AI, is a vulnerable one that would need time to recover and regain those skills if the systems were taken away. Education systems may need to address this proactively. The optimistic reading is that edge intelligence democratizes understanding and knowledge. The pessimistic one is that it centralizes domain expertise and silos it in the hands of a few people who train and maintain these systems.

Infrastructure readiness is the binding constraint, and it has concerned organizations from city planners to NATO. Thousands of devices in a city constantly streaming data raise hard questions about unified communication protocols and interoperable data processing that can prioritize certain signals over others. An early example of this kind of coordination already exists: ambulances communicate wirelessly with traffic lights to trigger what is called emergency vehicle preemption, or EVP, sometimes also referred to as an ambulance priority traffic signal system. That small-scale coordination is a preview of what full-spectrum edge intelligence could accomplish, and of the engineering discipline required to make it reliable enough to trust with human lives.

Energy infrastructure is probably the most underrated proxy for how ready a country actually is to take advantage of any of this. Before a city can do anything intelligent, it needs power it can rely on. In much of sub-Saharan Africa, including Kenya, the grid physically exists but it lacks the orchestration intelligence that would make it stable. The result is brownouts and instability that ripple through anything depending on continuous power. What has emerged in response is edge observability built from the cheapest hardware available: mobile phones with cellular connections attached to legacy grid components, surfacing anomalies in near real-time \autocite{seforall2023africa}. The underlying machine learning is not glamorous at all. It is mostly time series analysis, forecasting load patterns and catching the early signs of failure before they cascade. But a grid that can see itself can be managed, and a grid that can be managed starts closing the reliability gap that stops more sophisticated infrastructure from being built on top of it.

Spain is one of the more interesting contrasts here. If you look at a satellite image of the south you will see large stretches of solar farms, in Castilla-La Mancha you find wind turbines spread across the plateau, and on top of all that there is nuclear baseload. Managing a grid drawing from sources with such different timing profiles, where solar peaks at midday, wind is intermittent, and nuclear is essentially constant, requires very careful synchronization across the whole network. When that synchronization breaks, as it did during the Iberian blackout in April 2025 \autocite{iberia2025blackout}, the consequences are not local. They propagate across interconnected systems in ways that are hard to predict and expensive to undo. The optimization techniques being applied to grids like Spain's are a long way from time series monitoring in Nairobi. Mixed-integer linear programming is used to solve the unit commitment problem: which generating assets to bring online, at what time, at what cost, given forecast demand and the physical constraints of the network \autocite{padhy2004unitcommitment}. It is a fundamentally combinatorial problem, and as the number of variables grows, more distributed renewable sources, more variable demand, the computational complexity stacks up fast.

One of the most interesting variables entering that optimization right now is the electric vehicle fleet. A parked electric vehicle is not just a consumer of grid power, it is also a distributed storage asset. Vehicle-to-grid architectures let energy flow in both directions \autocite{kempton2005v2g}, and once you accept that, the logistics scheduling problem and the energy optimization problem become coupled in a genuinely interesting way. A fleet operator can add a term to their routing objective that rewards pausing a truck during low-utilization hours at a location where it can push power back to the grid, which reduces both the operator's energy cost and the grid's peak load. The logic is not unlike what UPS did when they redesigned routes to remove almost all left turns across their fleet \autocite{ups2007leftturn}. The saving is invisible at the level of any individual route and only shows up when you model the cumulative cost across millions of deliveries. V2G is the same idea. The benefit does not live in any single vehicle's balance sheet, it lives in the system.

The political version of V2G is cross-border energy trading. Countries with surplus renewable generation at certain hours can sell to neighbours whose mix peaks at a different time. That requires not just physical interconnects but coordinated market mechanisms and real-time optimization across independent national grids. It is, in miniature, exactly the same coordination problem that edge intelligence poses for data: how do you get systems with different owners and different incentives to share resources in a way that makes the whole better without any individual participant losing out?

Energy grid intelligence is a useful stress test for the kind of institutional coordination that AI-enabled infrastructure is going to require at every level. The observability gap in Kenya, the synchronization failure in Spain, the logistics-energy coupling in V2G, they all describe the same underlying problem: complex systems managed in silos accumulate fragility. The countries that have built observability, optimization, and coordination into their energy systems are not just running better grids. They are developing the institutional muscle that the next layer of infrastructure will require.

3. Cheap Models, Expensive Judgment

Different groups of people have completely different needs for offloading parts of their daily work onto AI. Doctors or medical researchers will need compute time to run simulations, software engineers need large context for code understanding and consultants need language comprehension and formulation. Each of these tasks has a different implicit cost in terms of time and model complexity required. Currently a single subscription pools all of these people into one computational segment—everyone has access to the same models at the same cost. I do not see this as a sustainable or even profitable solution for AI labs. If an allocation for a software engineer gets exhausted every 5 hours but another client pays the same amount without utilising the subscription for days at a time, there is a disproportion between supply and demand. Not to mention that the costs of blindly using models without consideration of cost has shown to drive over-spending without any observed economic output.

What is obvious to people working with LLMs every day is that you can and should use different models for different tasks. The model-access goes beyond pricing and is more of a political problem. As a country, if you develop some very capable AI solution for a task, do you keep it to yourself or do you let anyone use it? We have now recently seen Fable become a citizen-restricted model where only USA citizens are allowed to use it \autocite{anthropic2026fable}. This is, briefly after Anthropic was labeled as a supply-chain threat to the country.

Most people in the world will probably not use models which cost over 1 per million input tokens, in fact they might gain enough utility from models that cost under 0.1 per million tokens, there will always be people and organizations who will be dissatisfied with the performance of models that cost over 100 per million input tokens. Where does the most profit for the AI labs come from? The global race is measured at the most capable and most expensive frontier models which is where most of the R&D will be focused on.

assets/fig_03_model_price_tiers.pdf

The set of questions that a small model can answer correctly is part of the set of questions that a super smart model can answer as well, but this is only a one way street. Super-capable coding models will most likely be useful to cybersecurity experts of AI engineers, but to use a model like that to suggest recipes is not the right way to manage cost.

On the question of where token prices are headed, there are two camps. One hedges toward prices skyrocketing in the near future, expecting that the current state we observe is being subsidized to drive adoption and lock in early adopters, with the accompanying hype as a bonus. The other camp, despite having observed the increased subsidization of capital expenditure in AI firms, holds that it is reasonable to assume the future will allow us to run models cheaper. That is, after all, the ultimate objective of research in large language modeling right now: improving numerical processes and squeezing as much as possible out of the KV cache to reduce computational time.

This is complemented by the push toward longer test-time compute \autocite{deepseekr1}. If we can generate each token cheaper and faster, we can generate more tokens and let the models think for longer periods with the promise of higher accuracy on certain tasks. Running a model for a long time means doing inference much longer, which also increases the context of what the model can actually understand. Models have often been shown to suffer from something like context anxiety \autocite{lostinthemiddle2023}, but that is something we can address through harness engineering and better post-training specifically for long-running agentic tasks. If we are able to compound cheap token execution with longer reasoning, the hope is that the economic productivity gain returns more than the cost.

What this unlocks for the future is that a majority of economic tasks will be automatable to some extent \autocite{mckinsey2025auto}, since the nature of many roles in firms is to perform certain kinds of jobs and repeat certain kinds of patterns across different business scenarios. Those scenarios can then be generalized into agentic systems. However, when a client has a problem to solve and uses generative AI to produce some sort of summary of that problem into a document they send to a consulting firm, that consulting firm is most likely going to have its employees process that already AI-generated problem description using another AI system. You end up with a bridge between two artificial intelligence systems with a human simply passing the message between them. And most likely it is the same underlying system, something like ChatGPT, being used on both sides. What differs is only the context in which the information is presented to the model.

Apart from that, the entire process loses human judgment and the exposure of human experience to the problem, which is what has historically determined the success of different solutions. People who have worked in a field carry distinct experiences and distinct knowledge that can come together to produce something genuinely novel. If you simply use the same AI system on both sides, that experience and knowledge becomes homogeneous across both parties. You effectively gain nothing and are only doing work that was previously done for the sake of doing work, rather than actually gaining or providing a unique insight.

How we structure business workflows around this new baseline of judgment that everyone has suddenly acquired, and how we integrate that into society, is going to be a significant challenge. An even bigger challenge is the survivability of businesses due to the collapse of junior hiring. Many junior candidates will suddenly arrive with nearly identical sets of experiences and skills that do not in any way contribute to the economic growth of an organization. Meanwhile, those who have previously built deep expertise within organizations are slowly going to retire. There is a projection that around 150 million more jobs globally will be filled by workers over 55 \autocite{bain2023aging}, since they are the ones who carry the judgment and understanding of systems that will be most in demand as the homogenizing effect of AI takes hold at the junior level.

4. Algorithmic Management and the Gig Precedent

Something emerging in current economies are AI driven gig jobs \autocite{kadolkar2025}. When you place an order to get food delivered to your house, there is an element of intelligence which powers the assignment of that job to someone in the fleet of delivery drivers. These drivers sign up to a platform and in return for their labour get paid varying amounts depending on the kind of jobs they pick up. The catch comes in the access of certain jobs that an individual can be assigned by the algorithm. When you start fresh you have access to a pretty large distribution of jobs you can take, delivering across the whole city or taking rides from the city to the airport. From this distribution, an algorithm has to make a decision of what to assign to each individual. This decision process is highly biased on previous performance of an individual and often tends to be very unforgiving.

We have only observed this kind of work in physical deliveries and taxi-like services but the model is promising to companies in terms of profit. This merits an expansion into other parts of the economy, perhaps services like plumbing or even more white collar jobs like accounting will slowly move towards this model of jobs being assigned to people by an AI instead of a manager. Some industries will have self-imposed this kind of management style due to the need of cognitive offloading but for others it is an inevitable cut of the middle management layer. This direction, if unregulated, can bring us to a very dystopian future where the majority of society works for a machine and is being told what to do and how to do it by a form of AI.

In software engineering, so-called "vibe coders" have completely offloaded cognitive responsibility of design and implementation to a machine, however, their lack of understanding of a system puts the machine in charge of making key decisions which can dictate the entire direction of their solution. An LLM coding agent can approach a problem and propose some number of solutions, without a discriminatory and biased gate on top of those proposals, all future solutions will eventually regress to some form of average. This is something that you can already see in the design of user interfaces by AI. A well trained eye can very easily spot if an interface was AI generated without any opinionated direction, as much as a teacher can spot an AI written essay compared to an essay of a 5th grader. In practice a system like this would look like an "AI Agent" capable of ingesting large amounts of context of the entire business process and has the ability to assign different people different tasks.

Given that people have some inert "taste", whether that be for management or for business, that is what forms the culture of any organization. In the case that a company has no direction or preferential "taste" for what good is, their competitive edge is bound to regress to the mean. This is why in the future, AI should be used for not replacing what the average person would do, but its stochastic nature should be leveraged for the sampling of the entire solution space, to cover all possible ways to approach a problem in a systematic way.

5. Company Lifespans Will Collapse

Ten years ago, building a new company meant assembling a team, dividing up the work, and waiting for each person to get through their share of it. The more people you had, the faster you could theoretically move, but only up to a point. Coordination overhead, onboarding, and disagreement all eat into that gain. The tradeoff was always the same: ship sooner with a thinner product, or take longer and deliver something more complete. Either way, the clock was measured in years.

I believe that in the earlier days of AI we will see shorter periods of time for delivery of solutions which in the grand scheme of things lets organizations iterate on their business model quicker. If you think about MTTCC (mean time to cash cow), this should reduce significantly for most new companies \autocite{innosight2021}. This we are already seeing in the speed of acquisition of novel startups. Up until now we have measured the productivity of an economy by some proxy such as the issuing of patents, I think in the future we will measure it by the speed of death/acquisition of companies. Patents were a bureaucratically constrained estimator anyway.

To enable this, there is still a constraint of the geopolitical location of innovation. There are better jurisdictions that support this kind of development such as China and its Greater Bay Area innovation initiatives which allow for rapid market expansion and the EU has been seeing motivation to move in a similar direction with the EU Inc. initiative \autocite{euinc2026}. In the future venture capital and other private equity should surpass military subsidized developments to democratize this innovation away from academic research institutions. That is not to disregard the contributions of DARPA to the current state of the art technology.

6. Careers After the One-Life Profession

In the past, people and industries would prefer people to stay in the same job for very many years. You join a company, you build out a product, you start very junior and you climb the corporate ladder. That is still very true for companies like banks. There is very little that has changed in that aspect. But over the last ten years or so, we have seen the millennial generation slowly starting to take on more jobs and spending a year or two at one company at a time only.

I think this is a kind of anti-mimesis: rather than art imitating life, life is starting to imitate the mathematical formulation of reward. People are structuring their careers, whether they know it or not, as though they were agents optimizing a reward function. And the problem that formulation describes is the exploration-exploitation tradeoff. As an agent in a system, which in this case can be a person in the labor market, you have the option to explore, meaning have many different jobs, or exploit, meaning commit to one job for a prolonged period of time. In experimental and general reinforcement learning studies, it is always better to have higher rates of exploration in the beginning when you are joining some new environment that you do not know. And after that exploration, you can move more towards exploitation. To give this a little mathematical grounding, the Upper Confidence Bound algorithm \autocite{auer2002ucb} formalizes exactly this intuition. Each option is scored by combining its estimated reward with a bonus that shrinks the more you have already sampled it:

The bonus term is large when you have sampled something rarely, and shrinks as you sample it more. Options that have rarely been tried carry a large score boost. The algorithm is, in a very literal sense, incentivizing exploration. A person early in their career is in exactly this situation: the environment is new, the option space is large, and sampling widely before committing is the rational strategy.

I think exploitation can come much later than it did thirty or forty years ago. What has also enabled this is the interconnectivity of the world. The number of opportunities available to a single individual has increased significantly, which prompts a much higher rate of exploration and can enable people to truly find their global maximum of satisfaction, or whatever other heuristic they are guiding themselves by. They can now find that by exploring early and exploiting later. And I think that the period of exploration is only going to increase in the next few years, due to the higher availability of different jobs and different industries, since it is now much easier to get started with pretty much anything you can imagine. It is much easier now to learn everything, since information is just so much more accessible than it used to be.

I think it is reasonable to say that committing early is, in fact, more of a risk, even though it might be perceived as a safer, more secure option. The risk comes from the fact that the future is probably going to change in ways that we cannot predict, and companies might die and appear in matters of years, not decades. That concentrated bet on a single organization surviving long enough to reward your loyalty carries more variance than it used to.

That said, it would be naive to present exploration as a freely available strategy. Taking the risk of exploring multiple roles and industries has real physical and economic costs. Moving between jobs means gaps in income, loss of benefits, and sometimes the need to relocate. For someone without savings, without a safety net, or with dependents, the exploration phase is not a luxury they can afford. The people most able to explore freely are often those who already start from a position of relative security. That is an important constraint, and it means the model described here currently applies unevenly across the population. The ideal future that AI should be helping to build is one where these barriers come down, where the cost of trying something new is low enough that exploration becomes a genuine option for everyone, not just for those who can already afford to take the risk.

The near-term direction of this is something like spending one month at one company and then another month doing something else at another company, or maybe working at multiple companies at the same time, without any legal moonlighting issues. That is what is slowly starting to happen in younger generations. People now have multiple jobs where they can learn a lot in each one, and then maybe they can contribute to each one even more due to their sudden gain of experience across multiple fields.

I think it is also much better to now explore multiple fields very early on rather than to specialize, since there are a lot of cross-domain implications for pretty much everything. If you have an understanding of something from working in biology, maybe you can have that as a transferable skill to anything else. That is sort of the direction of careers that AI is going to enable.

There is also a distinction that has to be made clear here. Sure, you can say anyone can now be a software engineer in the sense that they can make some sort of website very easily. But I think that is just a different benchmark of technology. That is as if you said everyone can be a financial expert just because everyone can now open Excel and calculate different things in tables and use formulas very quickly, rather than using calculators manually when Excel was first invented \autocite{autor2015skillbiasedtc}. The tool changed the entry point. It did not dissolve the distinction between competence and mastery.

There is, of course, cause for concern in terms of job replacement and the need to relearn and retrain. But there is also cause for excitement in the number of opportunities that will arise in the near future, and how dramatically those opportunities will change in the next ten years. The person who explores broadly, accumulates cross-domain experience, and then consolidates into something deep will arrive at their area of specialization with a richer set of priors than was previously possible. For people willing to explore early, that is not a threat. It is exactly the condition under which the Upper Confidence Bound delivers its largest returns.

7. The Inverse Labor Dividend

When people hear that AI is coming for their jobs, the assumption is usually the same: we will work four-day weeks, then three-day weeks, and eventually we will not need to do very much at all to maintain our lifestyle. I do not think this is necessarily a fallacy. But it is an assumption that there is no more work that could be done, and that assumption is where I think it breaks down.

The inverse here comes from a current blindness to the potential of how much each person can actually contribute, not just to their job, but to their development in any field they are involved in. I believe that people who have enough autonomy to explore ideas will not work less with AI in their fields. They will likely work more, because they will have the freedom to explore more. That, I think, is one of the most unforeseen impacts of AI. Suddenly people will be able to manage more projects, make more decisions, coordinate more people, and test more ideas. Months of work will compress into weeks or days. This does not free us of responsibility. It lets us absorb more responsibility.

This absorption of responsibility is not evenly distributed. People who have autonomy and control over how they use AI will experience it as an amplifier of their agency. You can either leverage AI, or AI can be leveraged on top of you. For those in the second position, the experience is more likely to be additional pressure: more tasks assigned, less slack, more monitoring, in a way that could negatively impact work morale over the long term.

A useful way to see this is from the perspective of a truck driver in the US, where long-haul freight is a backbone of logistics. Drivers are currently capped on the hours they can spend behind the wheel. If AI handles most of the driving, then for every minute behind the wheel a driver loses less energy, because the cognitive load of concentration is offloaded to the system. That might mean drivers can stay behind the wheel longer. It might even mean they are expected to. Is that economically efficient? Yes. Is it beneficial to the driver's health? That is debatable. The technology lowers the cost of the task without necessarily improving the conditions of the person doing it.

Contrast that with a proactive employee who is given the power to suddenly solve more of their own problems. They will solve those problems, and they will be rewarded for that. They will absorb more responsibility, as mentioned above, and build more leverage over time. The same AI, in two different contexts, produces two very different outcomes.

There is a version of this argument that has become popular in recent years: that generative AI is already hollowing out junior roles, with early-career hiring falling sharply in occupations most exposed to AI tools. The story is intuitive and fits the narrative. But a closer look at the data complicates it. A 2026 Warwick working paper covering 243 million new hires and 407 million job postings across the US, UK, Canada, and Australia found that when you estimate the effect of GenAI exposure and working from home exposure jointly, the GenAI coefficient attenuates sharply and often becomes statistically indistinguishable from zero, while the WFH effect remains robust \autocite{lambert2026wfhjunior}. The decline in junior hiring is real. The cause appears to be remote work restructuring how firms onboard and supervise early-career employees, not AI replacing the work itself. That distinction matters because the policy response to a WFH-driven restructuring is quite different from the response to automation. One asks firms to rethink how they develop people at a distance. The other implies that no amount of rethinking will bring those roles back.

AI does not directly mean we have more leisure. What it can mean is that we are able to create more wealth, more output, and more contribution per unit of time. That is the inversion of the labor dividend: not less work, but more of what work can produce, distributed according to how much autonomy you had to begin with.

8. War Without the Same Human Cost

Currently the world has a lot of wars going on, especially the more east from central Europe you go (such as Ukraine and Iran). War used to require a lot of logistics for the sake of conquest at the expense of human capital. This conquest was motivated by expansion of territory and influence. In order to get some territory you needed people to fight with another group of people, if that would be in physical hand to hand combat or leveraging missiles/jets. This has changed a lot in the last 5 years, with one of the biggest changes being that you no-longer need to have a person in a trench, or a pilot in the sky. Missiles or early rockets (which were used in the conquest of the Americas by the British) set a trend of offloading some form of force into an in-animate object rather than a person. This is now changing even faster than people realize to things like unmanned ground vehicles. In the Ukraine-Russia conflict, UGVs are being actively used alongside FPV drones to carry payloads, learn about territory and in some cases even passively hold territory. From a game-theoretic perspective we are starting to reach nash-like equilibriums in which neither party gains much with additional actions when technology reaches a stale-mate.

Changing gears a bit, when you think about the idea of planetary defense, what most organizations wont mention is that their focus of planetary defense concerns only certain territories. In fact, inter-planetary conflict can include conflict of different countries which carry out a large part of their conflictive battle in space or in low earth orbit.

9. Sandboxes, Infrastructure, and Institutional Risk

Decisions to use AI in existing business processes can come from two motivations: fear or need. Those that come from fear are often top down, while the latter are bottom up. The difference between a fear-driven and often less successful implementation is that it looks for a problem that fits some existing solution on the market, while need better fits, either enabling a better problem-solution fit while promoting innovation in the market by pushing for certain features. Many companies purchase AI-based solutions out of hype and generic fit. To prevent fear-driven AI deployment, organizations should create sandboxes for quick iteration rather than locking in to solutions that promise general efficiency gains.

AI can be hyper effective in tasks which are classified as "knowledge work". Does this mean that companies can fully run autonomously and just pass around information between agents? Most likely not. The ones that get AI wrong will significantly risk their business running into the ground. What I see as a trend is a strong product focus in technology where suddenly a lot of execution, research and everything else which used to take time starts to fall away, and underneath all of that is more time to be spent on refining taste and understanding of the market.

10. The Entertainment Trap vs the Human Potential Path

References

\printbibliography
/ or Ctrl K
search; empty, it shows notes to revise and lately read
m
mark this note to revise, or unmark it
c
cover maths, proofs and solutions to test yourself
n
uncover the next covered item
y
copy this note’s org-roam link
e
copy this note as markdown, LaTeX kept as TeX
p
the next problem to work; there h hints, s shows the solution, x passes it for today
→ ← j k space
turn the page, when reading in pages
l
on the map, trace the greatest distance, then the longest path
double-click
a formula, to copy its TeX
rest on maths
bind a part of a formula and open the wheel of notes it comes from; scroll or ↑ ↓ widens it, 1–6 goes
?
this card