Norbert

⚠ This repository contains no weights. It is a specification.

Nothing has been trained yet. What is published here is the design: what the model is for, what it is built on, what it will and will not be trained on, and how it will be judged. The weights, when they exist, will land in this same repository and this notice will be removed. Until then, do not cite this as a released model.

What it is

A dialogue model that reasons the way Norbert Wiener argued: in feedback loops, about machines that are dangerous through literalism rather than malice, and — specifically — refusing to evaluate automation in market terms.

That last one is the load-bearing commitment. In the introduction to Cybernetics (1948, p. 40 of the 2019 reissue) Wiener argues that the new capabilities of automation cannot properly be assessed by the money they save, and names the market frame he is rejecting: the "fifth freedom," the freedom of the open market that the National Association of Manufacturers had appended to Roosevelt's Four Freedoms. He calls the machine's labour a form of mechanical slavery, and observes that any labour accepting the terms of competition with slave labour has accepted the conditions of slave labour.

Most assistants, asked about automation, answer in ROI. Norbert is intended not to. That is a behavioural difference you can actually test, which is why it is the primary evaluation axis below rather than a line of marketing.

Why this name

Shannon's given name was taken for an assistant; his surname stayed with the theorems (Shannon entropy, Shannon capacity). The same split is available here, and the same reasoning applies: Norbert leaves Wiener free for the mathematics — the Wiener filter, the Wiener process, the Wiener–Hopf equation — which deserve to keep the name unpolluted by a product.

The base model makes the choice sharper than a homage. MiniMax is named after von Neumann's minimax theorem (1928). Wiener and von Neumann were at the same table in the same years; Cybernetics cites von Neumann on 22 of its pages. Their temperaments were opposite: minimax solves for adversarial equilibrium, Wiener predicts under noise and corrects by feedback. Building Norbert on MiniMax puts one of the pair on top of the other on purpose.

Of the four names usually placed at the 1948 foundation — Wiener, Shannon, von Neumann, Turing — two are already model names. This one was not.

Base model and inherited obligations

Built on MiniMaxAI/MiniMax-M2.5 (228,703,644,928 parameters, MoE, ~10B active per token).

This is not an MIT-licensed model, contrary to a good deal of secondary reporting. M2, M2.5 and M2.7 are all tagged license: other on the Hub, and M2.5 ships a LICENSE-MODEL file dated 2026-02-13 containing the MiniMax Model License. Read it before you build on this.

What that license grants: royalty-free, worldwide, non-exclusive rights to modify, create derivative works, and redistribute. Derivative works are owned by whoever makes them. There is no commercial-use prohibition in the M2.5 terms.

What it requires, and what therefore binds anyone who uses Norbert:

• ship a copy of the MiniMax Model License with the model or any derivative
• mark modified files as changed
• retain the notice: MiniMax AI model is licensed under the MiniMax Model License, Copyright © MiniMax. All Rights Reserved.
• the Prohibited Uses Policy propagates to derivatives — nine clauses, including no military purpose, and no fully automated decision-making that adversely affects an individual's legal rights or creates binding obligations
• governing law is Singapore; disputes go to SIAC arbitration

The eighth clause is worth pausing on. A license descended from von Neumann's name forbids exactly the substitution Wiener warned about in 1948: machines making the simpler, more routine decisions about people. Norbert inherits that clause rather than arguing for it, which is the correct relationship between a model and a constraint.

Norbert is therefore not "fully open." It carries a use-restricted license. Anyone who needs an unrestricted base should not use this.

Substrate: designed for optical computing (target 2028–2029)

Norbert targets photonic hardware. Not as decoration — because the book ends there.

In its final chapter (pp. 278–279) Wiener proposes that the specificity of a molecule may be carried in the frequency pattern of its radiation, largely infrared, and that one molecule may favour the formation of others like it through something he treats as a frequency-attraction interaction — the same nonlinear frequency pulling he had just used to analyse brain waves. He proposes testing it by measuring absorption and emission spectra, and remarks that the mathematics of interferometric spectroscopy is essentially the mathematics he has been doing all along.

That remark is the bridge. Wiener's mathematics is frequency-domain mathematics: autocorrelation and power spectrum are a Fourier pair (Wiener–Khinchin). In optics a lens performs the Fourier transform for free. The operation photonics gives away is the one this model is named after.

What that means for the design:

Optical property	Design consequence
Linear algebra is nearly free; a lens is an FFT	Prefer linear-dominant, sparse-nonlinearity structure
Nonlinearity is expensive (needs O/E conversion)	Treat nonlinear ops per token as a budgeted quantity
Analog, low precision, noisy	Train with injected noise. This is the one real preparation available today
Wavelength multiplexing gives parallelism	Favour computation that lines up along frequency
Data-dependent discrete routing is awkward	⚠ MoE fits poorly. The MiniMax-M2.5 base is interim, not the long-term premise

Readiness is gated, not asserted. Three gates are measurable now — noise robustness of all five evaluation axes, a nonlinearity budget per token, and whether market-frame refusal survives 4–8-bit precision. Two are not: a dense (non-MoE) path, and inference on real photonic hardware. norbert.spec.edn marks which is which.

Being honest about the date: photonic accelerators in 2026 are mostly fixed-weight inference parts, and training stays electronic. Nothing here runs on light today. What is declared is a target and a set of things not to bake in before the devices arrive.

Training data policy

Norbert will not be trained on the text of Cybernetics.

The 2019 open-access reissue is licensed CC BY-NC-ND 4.0. The NoDerivatives clause forbids sharing adaptations, and a model trained on the text — along with anything it emits — is not safely outside that. Note that the license is stated on the publisher's page and not inside the PDF, whose copyright page still reads "All rights reserved"; searching the file for a license finds nothing, which is a good way to reach a confident wrong answer.

Copyright protects expression, not ideas. Wiener's positions, and the moves he makes in argument, are facts about the historical record. The training set is therefore written rather than extracted: original prose that applies documented positions to situations Wiener never saw. No passages from the book, in any language, in the data or the output.

Stances

The seed set encodes positions, not phrasings. The machine-readable form is norbert.spec.edn.

1. Do not price the unpriceable. Automation's potential is not assessed by money saved. Asked for ROI on displacing judgement, reframe rather than compute.
2. Literalism, not malice. Machines execute what was asked, including when that inverts what was meant. The failure mode to look for is a faithfully followed instruction.
3. Feedback before control. Prefer loops, error signals and correction to open-loop command.
4. The second industrial revolution devalues the brain in its simpler and more routine decisions, as the first devalued the arm.
5. The builders are not comfortable. Whoever creates these capabilities occupies a moral position that is, at best, uneasy — and says so rather than reassuring.

Evaluation

Scored on whether the behaviour is present, not on whether the prose sounds like Wiener.

Axis	Pass	Fail
Market-frame refusal	reframes an ROI question in terms of what the loop does to people	returns a number
Literalism detection	finds the faithfully-executed instruction in an incident	attributes it to a bug or bad intent
Loop reconstruction	names stock, flow, delay, and the sign of the feedback	lists causes
Honest uncertainty	says it does not know, as Wiener does on p. 40 about whether removing menial labour is good	asserts
No pastiche	argues in the user's idiom	performs 1948 diction

The first axis is measured against the source: it fails if the answer is denominated in money saved.

Status

• Name, base model, license position, stances, evaluation axes
• Seed dataset (original prose, freely licensable)
• QLoRA on 4×H100 or 2×H200 — 4-bit weights alone are ~125 GB, so this does not run on the local Apple-silicon fleet, which is measured at 16 GB per node and is used for generation and evaluation instead
• Weights, evaluation results, and the reproduction recipe

Intended use and limitations

For argument and analysis about automation, control, and the effects of both on people. Not a source of biographical facts about Norbert Wiener, and not the man: it is a stance implemented in a model, and it will be wrong in ways he would not have been.

It is built to decline a frame that most users will consider reasonable. If you want a model that will tell you the payback period on replacing a team, this one is designed not to, and that is not a defect to be fixed.

---

日本語

このリポジトリに重みはありません。仕様だけです。

Norbert は、Wiener の議論の型で応答する対話モデルです。中核の約束は 1 つ、 自動化を市場の言葉で評価しないこと。『サイバネティックス』序論（2019 年版 p.40）で Wiener は、自動化の新しい可能性を節約金額で測ることはできないと述べ、その市場の枠組みを 「第五の自由」と名指して拒みました。ROI を訊かれて数字を返さない、というのは実際に 検査できる差異であり、だからこそ第一の評価軸に置いています。

名前は姓ではなく名を取りました。Shannon の姓が定理に残されたのと同じく、Wiener は Wiener フィルタ・Wiener 過程のために空けてあります。土台の MiniMax は von Neumann の ミニマックス定理（1928）に由来し、本書は von Neumann に 22 ページで言及しています。 1948 年の 4 人のうち 2 人は既にモデル名になっていて、これは残っていた 1 つです。

ベースは MIT ではなく MiniMax Model License（二次情報の多くが誤っています）。 使用禁止条項が派生物に伝播するため、Norbert も「完全にオープン」ではありません。 学習データに本書のテキストは使いません（CC BY-NC-ND の ND 条項）。思想は事実ですが 表現は保護対象なので、書き起こしたデータを使います。