A letter from the Director
The institute exists because I have spent half my career building artificial intelligence that operates inside companies, and half asking — in laboratories and in print — whether what it actually does inside those companies is what we say it does.
For most of the past two decades I have alternated between founding operating companies and producing peer‑reviewed research. Cirius, sold to Yahoo! Japan. Naked Technology, sold to Mixi. Cinnamon AI, where document AI met Japanese enterprise scale. Nexus FrontierTech, where it continued to. The Ph.D. arrived in the same period as the second of those, not before. The alternation is not biography — it is the methodological commitment I now found the Institute on. Scientific work disciplined by what survives in the field; engineering disciplined by what can be defended in peer review. The Institute is the place where I have decided to do both at once, in public, with patience.
Foundation Models brought something new to AI engineering: general capability. They are, in the language I find easiest to think with, the new CPU — extraordinary in their range, indeterminate in their function. A CPU becomes a game console, a pacemaker, an irrigation controller, only when placed in a context that determines what it should do. The same, I believe, will become obvious of Foundation Models within the decade. They become a coaching system, an audit reader, an executive cockpit, a cultural reasoning simulator only when placed in a context that determines what they should do — and judged by metrics that capture what they are doing in that context, not by what they did on a benchmark in a laboratory.
The discipline of placing them well is missing. I do not mean it is absent in particular companies — many of us have built systems that work. I mean it is absent as a recognised discipline, with a name, a shared method, a journal, a curriculum, a community that can hold each other to it. The field optimises Foundation Models. It does not yet optimise the engineering of their meeting with human cognition. The Institute exists to found that discipline. I have come to call it cognition‑aware engineering.
Where AI labs build CPUs, we build the design discipline for the end‑to‑end applications that give those CPUs function in human cognition. This is a positional claim. The Institute does not compete with frontier labs on capability. We work on what those labs are not built to address: the meeting between general capability and the four sites at which cognition lives. A single mind under pressure. A document of record. A team holding tension. A culture reasoning together. Each of these has structure that does not survive a benchmark — and engineering that survival is our work.
The horizon is 2035. By then we aim to have founded cognition‑aware engineering as a discipline — not finished it (that takes thirty years), but raised the flag, in the sense that Card, Moran, and Newell raised the flag of HCI in 1983, that Picard raised the flag of affective computing in 1997, that Hutchins raised the flag of distributed cognition in 1995. A founding manifesto. A workshop series. Demonstrative case studies, with their failures public. Terminology that becomes the field's. A citation network that points back. Ten years is enough for that, if we choose it deliberately.
A reasonable reader will press here: HCI already exists. Affective computing already exists. CHI has cognition‑grounded design tracks going back twenty years. What is distinctive enough to be its own discipline? The honest answer is positional. Each of the three I cite above answered to a single site of cognition — HCI to the interface, affective computing to the affect channel, distributed cognition to the social fabric. Cognition‑aware engineering is the design discipline for what happens when Foundation Models meet all four sites at once — the single mind, the document of record, the deliberating team, the reasoning culture — under the operating constraint that the same engineering team has to deploy across them. None of the earlier disciplines was built for that integration, because none of the earlier disciplines was working with a substrate as indeterminate as a Foundation Model. The question is whether such a discipline is needed. My position is that it is, because the deployments are already happening without it, and the failure modes are characteristic enough to merit naming.
Our method has a name in the philosophical literature: Functional Contextualism, in the lineage of Pepper, Skinner, and Hayes. Its principle, in a sentence: form does not determine function; function is what an act does in context. Translated to AI engineering, this means refusing to optimise a module's surface against an out‑of‑context metric — and committing instead to designing the metric that captures function in the specific cognitive context where the system will operate.
I do not, however, believe that cognition has structure as a metaphysical claim. I treat it as a working hypothesis: that the structure we hypothesise lets us accelerate engineering on problems the field has called soft because no one bothered to build the instruments. Distortions in a single mind. Provenance in a record. Tension in a deliberation. Cultural grammar in a collective. Each of these is real enough — once we structuralise it as engineering hypothesis — to support a metric, a behavioural specification, a refusal logic, and an evaluation regime that survives audit. Cognition has structure, we say. The rest is engineering.
We work through five disciplined steps — decompose the context, specify the functional metric, severity‑test against pre‑registered failure conditions, refuse what falls outside specification, and iterate when the field reveals what we missed. The loop has two threads in parallel: engineering and research, both at deployment speed, coupled at integration. Our role within it shifts by stage: we selectively build prototypes, we lead observation as consultants, we carry the research ourselves, partner firms and clients carry full implementation, and we return at integration as advisors. This is not a hybrid of convenience; it is the only operating shape in which the methodological commitment — academic and industrial discipline in the same study — can actually run.
Every work the Institute carries must answer to three independent panels. Academic peer review. Industrial deployment. Human consequence. A paper without deployment is a contribution; a deployment without peer review is a product; a system without human validation is a demo. Each of these is honourable. None of them, alone, is the discipline we are founding.
I take this from outside AI. The medical foundations — Wellcome, the Howard Hughes Medical Institute, the Gates Foundation — operate under triple discipline as a matter of course. Across AI research, no institute commits publicly to it. The Institute commits. It is our constitution: we are answerable on all three axes, simultaneously, and a single‑axis success is treated as incomplete. This is not pose. It is what makes the discipline a discipline.
The negative space is part of the position. We do not chase SOTA. We do not propose new architectures. We do not enter scaling competitions. We do not run pure NLP benchmarks. We do not treat plausibility as a proxy for fidelity. We do not treat convergence as the goal of deliberation. We do not treat persona as a proxy for culture. We do not collapse a model into either anthropomorphisation or pure instrumentality — we hold it as an institutional object, as our position papers argue at length. Each of these refusals has cost us papers. Each is a condition of the work we are here to do.
A ten‑year claim deserves a falsifiable commitment. By 2030 — five years in — three signals would tell me the framing was wrong.
The first. That no methodological vocabulary survives. If by 2030 no working researcher cites Cognitive Fidelity, Tension Holding, Evidential Trust, or Substrate‑invariance in a peer‑reviewed publication that I did not co‑author, the instruments did not hold.
The second. That the four cognitive sites collapse into one or splinter into many. If mind · record · team · culture turns out, in retrospect, to have been an over‑coarse partition that subsumes into a single category — say, interactive AI — or fractures into incoherent sub‑disciplines, then this was not the right cut of the joint.
The third. That the field has already developed an equivalent discipline under a different name — that what I am proposing is being done, more rigorously, by an active research community I have failed to recognise. The failure would be one of literature, not framing; but at this scale of claim, that failure is enough.
Any one of these by 2030 would warrant retiring the framing. All three would warrant retiring the Institute’s claim to be founding anything. I am committed to writing that retrospective in public if it comes due — and to doing so before the patron base has to ask.
A ten‑year horizon does not fit a grant cycle, and it does not fit an equity timeline. It fits patronage. Individual fellows, institutional patrons, and founding patrons make the long view possible. In return, they sit close to the inquiry as it happens — working papers as they take shape, private briefings, conversations that will not appear in print for some years, and a hand in shaping how the founding documents land.
I am not building this institute because the AI field is incapable of these questions. I am building it because the field, optimising for what it can already see, will not get to them at the pace they require. The instruments we are designing now — Cognitive Fidelity, Distortion Reduction, Tension Holding, Evidential Trust, Substrate‑invariance, Cultural Grammar, the audit matrices for cognitive sovereignty, the persona‑contract framework for deployed AI — are not the discipline. They are the founding documents of it. The discipline itself is what the next decade will become if these documents anchor it.
I would be glad to be told I am wrong about the timing. I would be more glad to be joined by people who are willing to find out.
Yours, Hajime Hotta, Ph.D. Director · Hajime Institute · Tokyo
Three ways in