The Short Answer
"Theory Is All You Need" is not new and not a proof: Teppo Felin and Matthias Holweg published it in Strategy Science 9(4), 346–371, in 2024, and it argues conceptually that LLMs are backward-looking and imitative rather than originating. If that argument holds, the differentiator is the proprietary context you feed a model, not the model. With ibl.ai you own all the code and the data.
The corrections matter, because a strategy built on a misread paper inherits the misreading.
Is "Theory Is All You Need" a new paper, and did it prove anything mathematically?
No on both counts.
The article is Felin and Holweg, "Theory Is All You Need: AI, Human Cognition, and Causal Reasoning," Strategy Science 9(4), 346–371, published in 2024. A preprint circulated earlier that year.
That preprint was posted to SSRN in February 2024. The journal version is DOI 10.1287/stsc.2024.0189, published by INFORMS after peer review.
What happened recently is recirculation, not publication — it was posted again to research blogs on September 14, 2026 and picked up from there.
It is also not a mathematical result. It is a conceptual argument in a strategy journal.
The abstract puts it plainly: "Scholars argue that artificial intelligence (AI) can generate genuine novelty and new knowledge… We disagree. We argue that AI's data-based prediction is different from human theory-based causal logic and reasoning."
"We argue" is not "we prove." There is no theorem here, and describing one is the fastest way to lose the argument to anyone who opens the PDF.
The published article lists Holweg at Saïd Business School and Felin at both Utah State's Huntsman School and Saïd; Felin was Professor of Strategy at Saïd from 2013 to 2021 and is now at the University of Utah.
What does Felin and Holweg's paper actually argue about LLMs?
That prediction and origination are different operations, and that LLMs do the first one.
Their claim is that AI's approach to knowledge is probability-based, "largely backward looking and imitative," while human cognition is forward-looking and capable of generating genuine novelty.
They introduce data–belief asymmetries: cases where a belief precedes the data, and motivates the experiment that produces it. Their worked example is heavier-than-air flight — a theory held against the available evidence, which then generated the evidence.
One footnote gives the scale of the difference. For an infant to be exposed to the roughly 13 trillion tokens used to train current LLMs, the authors estimate it would take about 1.8 million years.
That is not an argument that models are weak. It is an argument that they are a different kind of thing, optimized over a corpus rather than over the world.
The strategic consequence is the part worth keeping. If the limit is structural rather than a matter of scale, a bigger model does not remove it. That is a very different forecast from "current models are not good enough yet."
It is an argument, not a finding, and it is contested — which is the normal condition of a paper worth reading rather than a mark against it.
Why does the heavier-than-air flight example matter to an enterprise AI buyer?
Because it is a clean case of the best available data pointing confidently in the wrong direction.
The paper's history runs as follows. Otto Lilienthal died attempting to fly in 1896. Samuel Langley failed publicly and expensively in 1903, nine weeks before the Wright brothers succeeded.
Reflecting on those failures, the New York Times editorial board estimated in 1903 that powered flight would take "the combined and continuous efforts of mathematicians and mechanicians from one million to ten million years."
That estimate is often retold as the verdict of the scientific establishment. It was a newspaper editorial. Lord Kelvin, then president of the Royal Society, had separately declared heavier-than-air machines impossible.
A model trained on the complete written record of 1903 would have reproduced that consensus faithfully, because reproducing the consensus is what the training objective rewards.
The Wright brothers had less data than the consensus did. What they had was a decomposition of the problem into lift, propulsion and control, which told them which experiments were worth running.
The enterprise translation is direct. A model gives you the existing consensus, fluently and cheaply. It does not know anything about your operation that is not already public.
If a model cannot originate, what differentiates one enterprise AI deployment from another?
The context, because everything else is shared.
Two organizations running the same frontier model over the same public corpus converge on approximately the same answers. Neither bought an advantage; both bought the same commodity input at the same price.
The variable is what only one of them has: its records, its processes, its decisions and the institutional memory of why they were made.
We argued this from model convergence in intelligence is a commodity, your data layer is the moat — the observation that capabilities were flattening across vendors.
Felin and Holweg get to the same place from the opposite direction, which is why the paper is worth adding to that case rather than restating it.
Convergence is an empirical claim about this year's models, and a capability jump would weaken it. A structural claim about what prediction can do survives the jump.
You do not have to accept their argument to act on it. It is enough that the two routes to the same conclusion fail independently.
There is a practical test for which side of that line a deployment sits on. If a better model shipped tomorrow, what would you have to rebuild?
If the answer includes your integrations, your guardrails and your evaluation set, you built on the commodity. The positive case for building the other way is in why the context layer is the moat, not the model.
Does anyone running a frontier-model business agree that models are substitutable?
Publicly, yes — and the framing to attribute rather than assert.
Satya Nadella made the operational version of the point on CNN's Fareed Zakaria GPS in late July 2026:
"By keeping the harness separate from the model and the context and memory separate from the model, you absolutely can use multiple models for what they're great at. At the same time, any one model can go away, and you can still continue to be in control of your own destiny."
He put the risk bluntly in the same interview: "Any firm that doesn't have this control, I will claim will not remain a firm because you've essentially outsourced your thinking."
Note the date. That is late July 2026, not this week, and it is Nadella's position rather than a settled fact about the market.
The architectural instruction inside it is specific: keep the harness, the context and the memory outside any one model family, so replacing a model is a configuration change instead of a migration.
What does the DeepMind Institute launch change about model choice?
It raises the expected rate of replacement, which is an argument about switching cost rather than about model quality.
Google DeepMind launched the DeepMind Institute on September 16, 2026, directed by Shane Legg, James Manyika and Demis Hassabis, with opening essays on reasoning transparency, economic policy and frontier-AI governance.
Two claims circulating alongside it did not survive checking, and we dropped both.
We found no primary source for an executive saying "we are close" as a quotation; DeepMind's own framing is that the field is approaching AGI.
And a widely-shared "75% of enterprise buyers" figure about AI spend and infrastructure resolves to no named survey we could locate.
What the launch does support is mundane and useful: the organizations building these models expect the frontier to keep moving. Whatever you standardize on today is the thing you will want to replace.
We looked at the same arithmetic from the release-cadence side in when three labs pace the frontier, your roadmap slows.
How does ibl.ai turn proprietary context into a differentiator you own?
By putting both the context and the platform that reads it inside your perimeter.
With ibl.ai you own all the code and the data.
The platform runs on your infrastructure with full source code access, is model-agnostic across any LLM, is usage-based with no per-seat pricing, and deploys anywhere — your own cloud, on-premise, GovCloud, or a fully air-gapped network.
Agentic OS connects your systems of record — SIS/LMS, HRIS, CRM, ERP, EHR/EMR, case and document management — over an MCP-based interoperability layer, and assembles a per-user memory with field-level permissions, PII redaction and consent enforced beneath the model.
That memory is the asset the paper points at. Agents manage it directly through a save/update/forget/search toolkit, with temporary facts carrying an expiry purged nightly and duplicate detection running semantically (September 11, 2026).
Substitutability is operational rather than aspirational: a six-hourly task syncs the model catalogue, and a discovered model stays inactive until a verification gate has tested multi-turn recall, streaming, tool calling, image input and a real, non-zero cost (September 18, 2026).
1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY.
Related reading: intelligence is a commodity, your data layer is the moat — the same conclusion reached from model convergence rather than from a structural argument about prediction.
Sources: the journal record, volume, issue, pages and abstract from RePEc's entry for Strategy Science 9(4), 346–371; the February 2024 preprint, the 13-trillion-token footnote and the 1896–1903 flight history from Felin and Holweg's SSRN text, published at DOI 10.1287/stsc.2024.0189; the September 2026 recirculation from New Savanna; Felin's Oxford dates from his Wikipedia entry; the Nadella quotes from TechCrunch's report on his Fareed Zakaria GPS interview; the DeepMind Institute launch date from Axios and its directors from TechCrunch.