AI is a One-Player Game: Why LLM capability refuses to scale

Quantum mechanics is controlling how AI scales in the enterprise.

Not literally. But functionally.

An LLM can feel like a proper reasoning partner: messy, iterative, almost alive. But the moment the organisation looks at it and, quite rationally, demands sharing, auditing, repeatability, it has to behave like software.

And that's a trap leaders could be walking into if they promise scale.

The moment you try to make LLM use shareable, auditable, and repeatable, you collapse the LLM into a tool. Documentation. Training. Prompt Libraries. These are all collapse events.

You can share the ashes, but not the fire.

The best users aren't following a protocol. They're in an iterative, intuitive, and unrepeatable dialogue.

Then someone asks them: "How'd you do that?"

The second they open their mouth to explain, it's gone. They're no longer describing a partnership; they're describing a tool they used.

The distinction shows up as market behaviour: in procurement, governance, and adoption patterns. Web traffic data from SimilarWeb suggests that Microsoft Copilot's market share is decreasing, and enterprise AI adoption faces documented implementation challenges, as reflected in recent industry surveys and analyst reports. If conceptually, an LLM is incapable of surviving within the predefined software box we are trying to force it into, the decline in the adoption metrics starts to make sense.

This article calls out the uncomfortable possibility that LLMs don't scale because they are a 'one-player game'.

A one-player game

Companies can create prompt libraries, retrieval-augmented generation (RAG), AI centres of excellence, and ongoing AI skills development, but that is like confusing the notes with the lecture. And in doing so companies are building multiplayer infrastructure in a domain where value is generated through non-transferable collision of individual context and model ambiguity.

The reality is that LLM utility relies on a 'one-player' dynamic: a private, high-context loop of judgement and interrogation that resists systemisation. When organisations try to extract that dynamic into a library or a process, they isolate the output and discard the judgement.

Language is rarely the bottleneck. The relationship itself resists capture. What happens between a person and their model is theirs. Attempts to systematise it produce a library entry or a training module, something that can be filed and circulated but stripped of the judgement that made it useful. The value lives in the conversation itself, in the iterative exchange that even its participants could not reconstruct.

'Scaling AI' looks less like network effects and more like trying to photocopy a fingerprint.

Understanding accumulates between two parties, not across a network

If a friend tries to explain the secret to their marriage, you listen politely. The advice is real, but untransferable. What you are really hearing about is the unique, unrepeatable history between two people.

LLMs operate the same way. A thousand employees prompting the same system are having a thousand separate, private conversations.

Platforms with network effects rewarded collective adoption. LLMs reward individual skill. The person getting value is not contributing to a shared asset; they are building a private competence. The investment logic that has worked for the last decade of enterprise software may not apply here.

Even if you recorded every prompt, archived every session, built perfect retrieval across every interaction, the capability would still not transfer.

Prompts are not recipes. They are receipts.

'We can know more than we can tell'

Michael Polanyi was writing about this in 1966. Experts know more than they can explain. Ask a surgeon how she knew to cut there, or a negotiator why he paused at that moment, and the answer will be incomplete. Withholding is rarely the issue, it's that the knowledge lives in the perception itself. They recognised the situation before they could articulate why it mattered.

LLM use is the same kind of skill. The part that makes someone effective is the part they cannot fully state.

Five moves happen in effective LLM use. All of them transfer poorly.

• Problem choice: knowing what is worth asking in the first place
• Framing: knowing which constraints are real and which are negotiable
• Interrogation: knowing where to push back, where weak points hide
• Evaluation: recognising confident nonsense when you see it
• Integration: fusing the output with everything the system does not know

Prompts capture none of this. They are the residue of a judgement process, not the process itself. Surface technique transfers. Perception rarely does.

You can share the ashes, but not the fire.

So what would a multiplayer version actually look like? Not prompt libraries. Something else.

Scalers and seekers

Most companies will scale the tool version because it's the version of AI that can be presented as behaving predictably. It can be documented, audited, rolled out, and reported. It fits procurement, governance, and performance metrics. How much it achieves remains an open question.

Alternatively, companies can lean into the one-player game and build infrastructure around the relationship.

What would organisations learn if our questions posed to LLMs left traces and fragments for those who follow?

For example, current LLM deployments measure activity without mapping territory. They count tokens and track sessions. They know how many questions were asked but not where questions clustered, where they avoided, or what patterns emerged from collective enquiry.

Since the systems track activity rather than territory, there is no record of these gaps. A problem can remain unaddressed indefinitely simply because the lack of enquiry generates no data.

Multiplayer infrastructure would find the patterns and opportunities and leave a trace that benefits those who follow. Mapping the question territory: which framings opened ground, which led nowhere, which regions remain unexplored.

The organisation would see its own attention as a pattern, and seek what the patterns reveal and what is being systematically ignored.

But trust must precede enquiry.

For question-mapping to work, employees need confidence that their half-formed thoughts will not follow them around. The tentative query, the naive assumption exposed, the wrong turn that revealed a right one. These are precisely the interactions worth surfacing as patterns. They are also the interactions people will suppress if they suspect observation.

A thought partner only functions under conditions of trust. You do not censor yourself with a trusted colleague. You think aloud. You say the thing that might be wrong because saying it is how you discover whether it is wrong. The value emerges from the willingness to be uncertain in real time. Employees who do not trust their employer will not trust systems their employer controls. The LLM becomes one more place where candour is punished, and so candour disappears.

Advantage accrues through tacit maps that resist capture and transfer

LLMs were presented as general-purpose amplifiers, technologies that would compress the distance between the capable and the less capable. Emerging research on AI skill gaps suggests a different pattern: advantage appears to accumulate among those with existing domain expertise and judgement capabilities.

The organisational reflex will be to chase what can be written down, audited, measured, and shared. But the decisive capability in a one-player game is the part that cannot be fully spoken, only exercised. The gap between what organisations want to capture and what actually drives performance is where the next wave of investment will compound, building powerful engines for the individual but no transmission for the firm.