Edition 3 · 08 / 08
A Framework for Finite Computation
How to evaluate useful intelligence, prepare for competing futures and keep the system around the model accountable.
The evidence does not establish the end of bigger models. It does establish that parameter count is an incomplete description of the system a person uses. Training allocation, sparse computation, inference effort and external tools address different parts of the problem.
This framework is an editorial proposal for making decisions under that uncertainty. It is not a validated standard or a forecast. Its purpose is to turn the edition’s argument into questions that a team can test.
Start with the accepted outcome
Define what completion means before choosing a model. A correct answer, a checked calculation and a successfully executed workflow are different deliverables. Record how correctness will be assessed, which failures are unacceptable and when human review is required.
Measure the complete cost: model calls, retrieval, external tools, retries, elapsed time and human correction. A low token price is useful information, but it does not establish a cheap result.
Match the intervention to the failure
| Observed problem | Candidate intervention | Evidence to request |
|---|---|---|
| Missing current facts | Retrieval and source access | Correct source selection, freshness and citations |
| Weak performance on hard cases | Stronger model or more inference effort | Success rate at matched total cost |
| Arithmetic or execution errors | Code, tools and independent checks | Tests against known correct results |
| Large volumes of separable work | Parallel agents or specialist models | Coordination overhead and unduplicated output |
| Data control requirements | Local or private deployment | Access controls, retention and operating responsibility |
| Unacceptable operating cost | Routing, distillation or a smaller model | Preserved quality on the real workload |
These are candidates, not automatic remedies. Retrieval can select the wrong document. A tool can be called incorrectly. A second agent can repeat the first agent’s error. Preserve a baseline and require measured improvement before adding complexity.
If intelligence remains scarce
In one scenario, the hardest commercially useful tasks continue to require enormously expensive frontier systems. Capital, compute access and infrastructure remain concentrated. Organisations become dependent on provider APIs and exposed to pricing, availability and geographic constraints.
The strategic response would be to examine continuity and substitution: can the workflow degrade gracefully, can a different provider perform essential tasks, and can the organisation retain its data and evidence? These questions connect directly to Edition 1’s governance framework.
For society, such a scenario could reinforce unequal access and geopolitical competition for computing infrastructure. That is a conditional consequence of scarcity, not an assertion that it must happen.
If useful intelligence becomes abundant
In a second scenario, highly capable models become inexpensive enough for many organisations and individuals to operate. Local and private AI could reduce API dependence. Specialised systems could lower barriers to entry and spread automation into activities that cannot currently justify expensive services.
Availability also changes misuse and verification. More actors could obtain powerful capabilities. Cheap content generation could increase the burden of checking authenticity, while human attention remains limited. Lower access costs do not automatically produce trustworthy information or fair outcomes.
Abundant models can still depend on concentrated chip supply, cloud services or distribution platforms. The two scenarios may coexist: plentiful routine capability, scarce frontier capability and concentrated infrastructure underneath both.
Decide what would change your mind
Keep a small evaluation set of representative work, including awkward failures. Reassess when a new model, price or workflow becomes available. Compare complete systems under the same permissions and acceptance criteria. Record where a cheaper option is sufficient and where the frontier premium still buys an important result.
MIT’s research on convergence and its metric-dependent qualification provide a reason to maintain several measures. Lower prediction loss, more tasks completed and better economics need not move together. The companion article sets out that evidence and its limits.
For most of the LLM era, the model was the story. The emerging system view asks how much useful intelligence can be extracted from finite computation, including what happens around the network. That does not make the foundation model unimportant. It makes the surrounding decisions visible.
If intelligence eventually becomes cheap, the model itself may cease to be the moat. The future of AI may, quite literally, lie beyond the model.
References and basis
This checklist and the two scenarios are the publication’s synthesis, not findings attributed to a single paper.
- Meek Models Shall Inherit the Earth (2025): the convergence proposition under stated assumptions.
- Two AI Metrics Diverged (2026): why capability measures affect the conclusion.
- IEA, Key Questions on Energy and AI (2026): infrastructure and demand context.
- Anthropic, Multi-agent research system (2025): a documented orchestration example.