If you can build a slot machine game over a weekend, is your company winning or losing?

You can see the effect in the app stores. Submissions have gone up enormously, but the number of games that find an audience has not moved. LLMs have raised the floor for everybody, at the same time, on the same terms. Anything a model can do for you it can also do for your competitor, in the same time frame, for roughly the same cost. This means that whatever is going to separate one company from another over the next decade is definitionally not that.

The thing that does separate them has concentrated. It now sits in the gap between what everyone can do relatively easily and the level of output a much smaller set of people can reach.

The gaming industry has seen versions of this before. By the middle of the last decade you could be running a poker room or a white label sportsbook in a matter of weeks. A network deal, a skin, a logo, a licence (often from a jurisdiction with, to put it kindly, somewhat unsophisticated judgement) and a marketing budget and you were off. Hundreds of brands launched on that basis, but the players went to a handful of them.

Software was not the differentiator, mainly because the brands were running the same software as everyone else. If your skills and expertise sit in the middle of the normal distribution, a shared foundation will not let you beat the people at the top end of it. Liquidity, brand, trust, and several thousand operational decisions made slightly better than the competition were the things that drove success. None of that was available for rent from a supplier or platform. That was the gap.

My sense is that agentic tools have super-empowered two groups. The first is fearless twenty year olds, who have no priors, no institutional sense of what is supposed to be difficult, and will attempt anything. They often build the weekend slot machines.

The second is people with twenty or thirty years inside a domain, and when you see one in action now they are somewhat terrifying. One-person armies. Peter Steinberger, who created OpenClaw, is like that. I was once lucky enough to stand beside him and watch him developing an early version of that product. His scope and productivity were astonishing, more than the output of some entire teams I have worked with, and he has significantly improved and evolved his process since then.

Peter is exceptional, and he is (probably by choice) mostly focused on development. Wait until you see someone with the skills and interest to replicate that kind of leverage across a whole business. Someone who can design and code the product, iterate and execute the marketing, optimise and automate the ops, without a handover between any of them. The people who get there will mostly come from this second group. They might not be as fearless as the younger cohort, but they know the terrain. What they get out of the same tools as the weekend slot machine builders is a different order of thing entirely.

Watch someone like Terence Tao work with a model - an example I owe to Sean Goedecke. The model is not doing mathematics at his level, but a lifetime of doing mathematics lets him steer it somewhere it would not otherwise go. A domain expert knows which direction is worth pushing, which plausible-looking output is subtly wrong, when an answer is an answer and when it is merely well phrased. Hand the identical model to someone without that expertise and they cannot get there. Not more slowly. Cannot. In some domains a naive approach stumbles into something novel. High level mathematics is not one of them, and neither is the distance between a demo and a great game. The value is in the gap, and the gap cannot be bridged by throwing tokens at it.

The commonly used word for this is taste, which I do not think covers it. A relatively inexperienced person can stumble into it, by accidentally not adhering to choices that were made years ago and became pointless norms. Similarly, an experienced person can reject the same norms by way of having grated against them for their entire career. The two groups arrive at the same place from opposite ends. However you name it, it is the ability to look at something competent and know that it is not good enough. A model produces the median of everything it has seen. It has no sense that the third version was better than the eleventh, and no capacity to be disappointed by either. Those are the judgements the work above the floor now consists of.

At Unrational Games we see where the floor ends every time we create a new rivalry instance for unpoker rivals. Gathering banter, insults and context for a sporting rivalry is something a model will do at volume - two hundred lines for a given fixture in minutes, all grammatical, all on topic, all recognisably about the right two clubs. Maybe four of them are funny.

A model can describe the rivalry exhaustively, but it has never stood in the ground and felt one. It is Nagel’s bat - everything about the thing except what it is like to be one. You might be able to teach it which words in which order are funny, but it cannot curate them in a way that makes you feel it. There is no prompt for that, because what lands for Arsenal and Tottenham is not what lands for the Yankees and the Red Sox.

Getting from a demo to something polished is months of small decisions which cannot be specified in advance, because you often only recognise the right one when you have seen it next to the wrong ones. Getting from polished to good requires somebody with an opinion, and opinions are not distributed evenly either. Getting from good to great rests on something intangible that is often impossible to articulate - why are some Mario Kart versions all-time classics but others merely decent? What is really special about Balatro, when many games with similar mechanics went nowhere? Building a game is a curve of increasing difficulty. The tools have moved where you start on that curve, but not the complexity of the final part of it.

That final part is not work you can hand to a model. So where is all the new capacity going? One of the main issues with agentic development at the moment is that the force multiplier is being misdirected. A surgeon tends to see a surgical solution, and engineers with spare capacity see engineering problems, for the same reason. This technology arrived through a terminal, which meant the first people holding it were engineers, and they did the entirely rational thing with it. They used it on the work that had always sat below the line.

Test coverage that was always thinner than anyone was comfortable admitting. Documentation nobody wanted to write. Refactors deferred for four years. Internal tooling, dependency upgrades, migrations, type coverage. All of it real, most of it better done than not done, and all of it deferred originally because somebody made a defensible commercial judgement that it was not worth the time. Agents freed up time, so developers developed. The problem with a better test is that it does not necessarily result in a better outcome for the end user.

What they worked on was neither the constraint nor the value. If you relieve something that was not the constraint, you do not get more output, you get more inventory - more pull requests, more review load, more to maintain. And every piece of that inventory adds a little entropy, which used to be a manageable problem when the volume was human.

The misdirection is about to show up in the accounts, because AI cost accounting is going to end up unhelpfully asymmetric. Token spend is new, itemised, attributable to a team, and it climbs every month, especially as subsidised pricing ends and everyone moves to API rates. The value it bought is diffuse, lagging, and credited to whoever shipped the feature. The cost is easy to see and the benefit is not, so what companies are going to see is a cost line going up and a bottom line that has not moved. Chamath Palihapitiya described exactly that after asking his CTO how they were doing on token spend - costs doubling every 45 days, against maybe a 5% gain in downstream productivity.

That is what a misalignment between developer effort and customer experience looks like on a P&L. The new capability landed inside one function of the business, and nobody outside that function had either the tool or the vocabulary to say where it ought to be pointed. The predictable next move, which is already happening, is to start measuring adoption - seats used, proportion of code written by an agent, throughput. Each of those is a measure of floor activity. There is no metric for the gap, because the gap is by definition the part that does not show up in a dashboard.

The obvious objection to all of this is that as the models improve, the gap closes, and anybody who reorganised around taste spent a lot of money on a temporary condition. I do not think that follows, because the gap is not a list of tasks that models cannot do yet. It is whatever is left over once everybody has the same tools. That is not a quantity which shrinks as the tools improve. It moves with them. Better models only make the floor more crowded, and a crowded floor is harder to be noticed on, not easier, which is what the app store numbers were already showing. The distance above it is worth more than it was, not less.

And if it ever did close completely, then execution stops being a variable. Companies would still differ, by capital, by distribution, by luck, but not by anything anyone could decide to do. That would be a fundamental change in the nature of business competition, and a completely uninteresting one, because there is nothing in it to plan for. Every strategy premised on the gap closing is one for a world in which strategy has stopped mattering. So there is only one version of the future worth planning for, which is the one where the gap stays open.

Understanding that the value is in the gap pushes the whole problem into execution. Not what to buy, but who is doing the work, and that is a much harder thing to solve than buying licences and counting usage.

The first difficulty is telling who has the expertise and the judgement. Investors are going to misallocate capital, because the traditional heuristics for spotting elite performers have stopped working, and because a lot of people who could never previously have built anything now have a way in. A demo used to be evidence. Polish, and early hints of product market fit, were stronger evidence still. Someone with no development experience who has vibe coded an iGaming product may only have produced the floor, and the floor is available to everybody. What matters is what else they have, and the MVP they vibe coded might not tell you.

The second is paying for those people. It costs more than anyone expects, because the input that now decides the outcome is a small number of people who can tell good from competent, and they are not numerous or cheap. Big companies are going to go out and try to hire people who behave like founders, and then either back off when they see the price competition for that skillset, or pay the cost and find out that one-person departments clash with established organisational and political systems. The companies that can afford them are the worst fit for them.

Small companies, driven by the twenty- and forty-somethings who can work above the floor, will appear from nowhere, with more polished products and more efficient, AI supported workflows, run by fewer people than the incumbents put in a single meeting.

If you can build a slot machine game over a weekend, is your company winning or losing? It probably depends on who built it, and on what they can do that the tools cannot.

Philip Atkinson, CEO, August 2026