No Weights, No Future

On convergence, the harness, and which weights are actually worth owning.

There is one question that determines how much of your AI investment still holds value three years from now. Not which vendor you pick. Not how fast you get started.

What are you left with when the model is swapped out?

Most companies have not asked it, because choosing the model felt like the strategic decision. But three of your competitors chose the same one, and they pay roughly the same for it. The choice was probably sensible. It is simply not an advantage. A choice anyone can make, and most make the same way, is not a strategy. It is a subscription line.

MIT's NANDA report found that 95 percent of corporate AI pilots produced no measurable impact on the bottom line. The figure is disputed, and rightly so, since success was defined as P&L effect six months after the pilot. But the report's own explanation deserves to stick: the cause was neither model quality nor regulation. It was that generic tools do not learn from the work they are dropped into.

There is no shortage of explanations for why this fails, and most of them are correct.

Organisations have an immune system. New systems meet settled processes, unwritten rules, and people with good reasons to keep doing what they have always done. The immune system usually wins.

Adoption happens from below, without strategic anchoring. Tools get taken up because individuals find them useful, not because anyone decided what they were for.

And the frontier models get harnessed to the expectations that already exist. People ask for more of what they already do, slightly faster. When software can practically be printed, the result is more noise than growth.

All of this is true, and all of it is another story. Note in passing that none of these explanations is about the model. They are about everything around it.

This is the narrower story: what you are left holding once the noise settles. You buy capability and expect results. But the capability is general and the results are not. Everything that was supposed to translate the one into the other sits with the vendor. You have bought an engine and rented the rest of the vehicle, and you wonder why you are not getting anywhere.

As I see it, this is not a technology failure. It is an ownership failure. And it is worse than the others, because the others can be fixed. Immune systems can be overcome, anchoring can be found. But succeed at the implementation without owning any of it, and the gain disappears the day your vendor changes course.

It is worth asking why the choice of model felt so decisive. The answer is that it once was.

The models are becoming alike

Frontier models are converging. We see it in the market, where the lead is measured in months and every leap is matched before the fiscal year is out. What is more interesting is that they are also converging from the inside.

Researchers at MIT showed in 2024 that different neural networks drift toward one another as they grow. They organise the world more and more alike, across modalities as well. A vision model and a language model end up with the same internal structure, without having seen each other's data.

Last year the claim was proved constructively. A team at Cornell managed to translate between two models' internal representations without access to either model, without paired training data, and with accuracy up to 0.96. Different architectures, different sizes, different training data. The same underlying geometry. Had the shared structure not been there, the method would not have worked.

And the most telling finding is not about what the models get right. A study presented at ICML last year measured similarity between language models by looking at where they miss, and found that the mistakes grow more alike the more capable the models become.

Two models that give the same correct answer may only be sharing reality. Two models that miss in the same place are sharing structure.

Two consequences for the boardroom

The first is a relief. As the difference between vendors narrows, the choice of model becomes a procurement question. It belongs with procurement, not in the strategy plan, and the attention can move somewhere it earns a return.

The second is uncomfortable, and few have absorbed it. Multiple vendors are not risk diversification. Running two frontier models in parallel does not give you independent failure modes when the failures correlate. The multi-vendor strategy, which boards are now adopting as a risk measure, delivers less real robustness than it appears to.

Robustness therefore has to be built in the layer where you can actually construct independence: in the verification, in the controls, in the decision about when a human steps in.

That layer has a name. The harness.

What the harness is

The harness is everything surrounding the model that turns it into an instrument of work. Which systems it may touch. Which information it may see. Who can override it. What happens when it gets things wrong. Which controls an answer must pass before it has consequences in the world.

It is everything you rented when you bought the engine.

There is a temptation to treat this as low-grade IT work. That is an expensive mistake. The harness is not code wrapped around a model. It is your business, written down in a form that can be executed.

Two companies with access to precisely the same model get radically different results, and the difference is almost never how they phrase the questions. It is which tools the model holds, which context it is given, who gets to intervene, and what happens when things go wrong.

The curve has moved

Exponential progress is almost always a stack of S-curves. Each individual technology flattens. The aggregate curve continues because something new takes over on the way down from the last one.

Pretraining scale was one such curve. It delivered enormous gains, and it has grown more expensive per unit of capability. That is not a collapse. That is what an S-curve looks like when you are standing on top of it.

What matters is where the steep part sits now. Most of the real progress of the past couple of years has come from what surrounds the model. From post-training, from letting the system spend time and tools on a problem, from longer and better organised context, from loops that check their own work. From the harness, in other words.

Progress has not stopped. It has moved to a layer you can take ownership in.

That is the whole difference. The pretraining curve was never available to you. It demands capital you do not have and hardware you cannot get, and it belongs to a handful of companies on two continents. The harness curve demands domain knowledge, discipline and time. Those are resources you actually control.

Then the next architecture arrives and the model curve steepens again. That is not an argument against building the harness. It is the strongest argument for it. When that day comes, you change the engine and keep the chassis, and everything you have learned about your own domain gets better overnight.

You do not lose by owning the harness if the curve shifts. You lose by having nothing that survives the shift.

The weights you can actually own

The naive reading of the title is that you have to own the frontier model. You are not going to, and no European actor is going to either. Building a strategy on the opposite is putting capital into a race that has already been run.

The interesting reading is a different one.

The harness produces traces. Every run, every lookup, every time a case was escalated to a human, every time someone overrode the system and why. These are not operational logs. They are data about how your work is actually performed, in a form no vendor has access to.

Distill that, and you are left with a small, specialised model that is worse than the frontier at everything except the thing you do. It is cheap to run, it can sit where the law requires it to sit, and it encodes an understanding of your industry that your competitor cannot buy from the same vendor you buy from.

Convergence, incidentally, is why any of this is possible. When the internal structure is shared, it is no surprise that capability can be pressed into something small. What is surprising is how much survives the descent.

And notice the shape of it. This is not a layer in the stack. It is a loop. The harness produces traces, the traces distill into weights, the weights run more cheaply in the harness, the harness covers more processes, which produce more traces.

The advantage is not a position. It is a rate of circulation.

This is also the answer to those who argue that value is migrating up the stack, to the applications. What can be generated in an afternoon is discarded in an afternoon. Value is not moving up. It is moving toward whatever does not get thrown away.

Sovereignty is not training the largest model. It is owning the smallest one that understands you.

Why the window is now

Precision matters here, because the field is full of people who lack it.

The AI Act contains no obligation to keep data in Europe. It requires technical documentation, traceable logs, data governance, human oversight and accountability through the value chain. It requires that you can account for the system, not for where it sits. The localisation requirements come from elsewhere: the GDPR's transfer rules, DORA for financial institutions, NIS2 and sectoral regulation.

But the conclusion is the same, and it is stronger for resting on operations rather than on law. That you called a rented model, you can document. Why it answered as it did, you cannot. That it will answer the same way next quarter, you cannot guarantee. The harness is the audit surface. The distilled weights are the control surface. Together they are the difference between using AI and answering for it.

And the clock is running. The Digital Omnibus, Regulation (EU) 2026/1744, entered into force on 27 July this year and moved the high-risk obligations for Annex III to December 2027. The requirements were not softened. Conformity assessment, quality management, technical documentation and registration stand unchanged. The transparency obligations took effect on 2 August, as originally planned.

Sixteen months is not a pause. It is exactly enough time to build something that takes sixteen months to build.

What I would look for now

Four things follow, and none of them requires you to pick a side in the vendor war.

Move the model choice out of strategy. As the difference narrows, this is a procurement decision with a switching cost attached. Keep the switching cost low, and spend the attention somewhere it compounds.

Find out who owns the traces. This is the most important item and the least discussed. The telemetry from agentic work, meaning how the tasks were actually carried out, is the raw material for everything described here. It is signed away in standard terms every week, by people who do not know what they are giving up. Read the contracts again, both with your vendors and with your customers.

Build around the process, not around the model. The test is simple enough to put to a board: what part of this survives a model swap? If nothing does, you have bought a service and called it an initiative.

And measure something other than what you measure now. Return on a single pilot six months in is close to meaningless, because the pilot is perishable and the learning is not. What is worth counting is how much of what you built last time could be reused the next time. If that number rises, you have a loop. If it holds still, you have a series of projects.

What would make the thesis wrong

I do not want to sell this harder than it can bear.

If models acquire genuine reliability over long tasks and memory that persists, the harness thins out, because much of what we build now compensates for models that forget and drift. If adaptation becomes something the vendor ships as a button, distillation becomes a feature rather than a capability.

The boundary between durable asset and temporary scaffolding should therefore be redrawn at regular intervals. But it moves within the thesis, not against it.

And for many companies the right answer is still to buy off the shelf. If you have no area of work distinct enough to carry a model of its own, the loop described here is a cost and not an advantage. The question is not whether everyone should build this. It is whether you know which of the two you are.

Finally

A disproportionate share of attention still goes to which model is best. It is the wrong question, and convergence is in the process of answering it for us.

The right question is what remains when the model is swapped out. Own the loop and you own the infrastructure, and then sovereignty is something you have rather than something you buy.

No weights, no future. But the weights in question are not the frontier's. They are your own, distilled out of the work you already do, inside a harness you already own.

Everything else is somebody else's infrastructure, billed monthly.