De toekomst van grote centralistische neurale netwerken in enorme datacenters wordt heel snel ingehaald door peer-to-peer kortegolf-lichtgestuurde Right-Brain-AI.

Jump to the article here

J.Konstapel,Leiden,5-11-2026.

Dit is een vrvolg op : Keep the Question, Keep the Carry: The Future of Computing on theNet

aanleiding NRC: Een Nederlands koffertje moet de razendsnelle ontwikkeling van AI controleren en verdachte situaties detecteren

On 5 November 2026, NRC reported a Dutch government test of an AI verification device. The device is a suitcase of electronics built by the British firm Amodo Design. It connects to inference servers of SSC-ICT in The Hague. It samples outgoing network traffic, recomputes that traffic against a declared model, and reports only pass or fail.

The method is sound for the systems it was designed for. Those are large neural networks, trained centrally in datacentres. The whole approach rests on a single assumption: dangerous AI capability requires large, central compute. That assumption is already contested in the research literature.

This essay describes the next paradigm. It concerns peer-to-peer, light-driven, coherent AI systems. Such systems exchange only unresolved questions and their carries over shortwave radio. The suitcase has nothing to tap here. There is no declared model to recompute. There is no datacentre to inspect. The transport layer for such a mesh already exists and runs. A public demonstration of the computing layer does not yet exist.

What was demonstrated in The Hague

The test placed hardware verification on a government server for the first time. Before this, Amodo had tested only on its own systems. The Dutch government did not fund the research.

Several claims in the NRC report can be checked against second sources:

Element	Claim in NRC	Status
Host	SSC-ICT servers, Koningskade, The Hague	Confirmed by TenderNed tender records (EUR 5 million, on-premise AI orchestration)
Supplier	Amodo Design, UK, CEO Tom Milton	Confirmed by 80,000 Hours podcast and company vacancies
Funding	Survival and Flourishing Fund; Longview Philanthropy	SFF grant confirmed (USD 1,229,000); Longview unverified
Field size	25–50 engineers worldwide	Roughly 50 per Milton; consistent upper bound
Ahead of the UK	UK announced a leading role	Not confirmed; the UK AI Hardware Plan (June 2026) concerns chip capacity, not verification
Coalition	PM Jetten with 21 other countries at the UN	Not confirmed

The same article quoted three critics. Paul Verhagen (HCSS) called the plan understandable but unlikely to succeed. Neither the US nor China would accept binding limits. Lauren Kahn (CSET, Georgetown) noted that AI, unlike nuclear material, runs through every sector. Michiel Bakker (MIT) saw value: the Netherlands has no frontier labs, but it can contribute verification technology.

How the method works

The suitcase separates running a model from training one. It does so by the volume of traffic between machines. Training moves gradients. Inference moves text.

1. Two networks. Datacentre accelerator racks sit on two networks. The backend network carries gradients during training. The frontend network carries prompts and responses.
2. Bandwidth asymmetry. Removing the backend is not enough. Frontend bandwidth is estimated at about 1 TB/s per rack, against about 4 TB/s at the backend. The proposal caps frontend traffic at about 10 MB/s per rack. That is high enough for text and too low for gradients. The ratio is a factor of 10⁵. Text at 1,000 tokens per second needs about 4 kB/s — far under the cap.
3. Tap and recompute. A network tap copies frontend traffic to a verifier server. The verifier reruns the declared model on the same inputs. It checks that the outputs match. Unexplained traffic points to an undeclared workload. Only pass or fail leaves the site.
4. Sampling. Milton names checking one packet in 1,000 as giving reasonable confidence. The detection probability for n hidden packets is 1 − (1 − 10⁻³)ⁿ. That gives 9.5% for 100 packets, 63.2% for 1,000, and 99.3% for 5,000. Bulk training cannot hide. A narrow covert channel can.

Independent descriptions of the same architecture exist. The MIRI working paper of June 2026 describes network taps plus recomputation on trusted compute. Anthropic’s October 2026 report describes a prototype that checks production outputs against a specific set of weights.

The assumption underneath

The whole method presupposes that dangerous capability requires large, central compute. Large compute can be found, counted and tapped. That is why verification starts there. But size is a condition for verifiability, not for risk.

The assumption is contested. Sara Hooker (Cohere) argued in 2024 that compute thresholds are an unreliable proxy for risk, because smaller models keep gaining capability (arXiv 2407.05694).

The method is also unfinished within its own paradigm:

Limit	Status
Scale of the pilot	About 30–32 H100/H200 GPUs
Secrets	The verifier needs the model weights; zero-knowledge proofs avoid this at enormous overhead
Readiness	A deal next month would use zero hardware verification mechanisms; fallback is shutdown, auditors, infrared satellites
Non-determinism	GPU recomputation is not bit-exact; covert-channel capacity must stay below 0.01%; current prototypes judged probably ineffective
Memory wipe	First Amodo test: about 24 h per rack, possibly over 100 TB left unwiped (author’s own non-expert reading)
Workstreams	Of 17: 4 active, 6 not started, 7 not on track (second-hand from Amodo’s SITREP)

The device does not monitor what happens in large datacentres. It shows that a tap runs on an ordinary inference server. It does not show detection of an adversary.

The next step: the coherent mesh

The next paradigm is a mesh of small nodes. Computation exists as closure. Each node holds only the questions still open, and the carries that closed questions leave behind. Coherence between nodes is carried by light. Nodes exchange carries over shortwave radio. Three layers make up such a system.

Layer 1 — Transport: peer-to-peer shortwave

Shortwave (3–30 MHz) travels via the ionosphere, beyond the horizon, node to node. No datacentre or central cable is involved. Peer-to-peer data over shortwave is in daily use: JS8Call, VARA HF, VarAC, and Winlink e-mail in emergency communication.

Bandwidth is narrow. The wideband HF standard MIL-STD-188-110C reaches about 120 kbit/s, or 15 kB/s, in a 24 kHz channel. Text at about 4 kB/s fits. Gradient traffic at an estimated 4 TB/s per rack is roughly 2.7 × 10⁸ times larger. It does not fit. The narrow channel selects what kind of computation can live on it: the exchange of carries, not of gradients.

Layer 2 — Computation: the closure register

The closure register is described in Keep the Question, Keep the Carry (Konstapel, 3 November 2026). Its key features:

• Values are balanced trits (−1, 0, +1). At ±3 a scale clears and one trit passes upward as a carry.
• A window of events closes in one of three ways: by standing (nothing occurs), by meeting (an event and its opposite cancel), or by a ring (three equal parts pass one carry upward).
• Two primitives — N(a,b) = −(a+b) mod 3 and a zero detector — form a counter machine. Universality follows from Minsky’s theorem.
• A learning layer builds lossless rules from repeated structure in text, in the line of Re-Pair and Sequitur, reported within about 9% of gzip. No gradient training occurs.
• It has been built on a Lattice iCE40 FPGA in 263 logic cells at about 164 MHz. The stated Landauer bound is 0.844 kT ≈ 3.5 × 10⁻²¹ J per step at 300 K.

Three properties matter for verification. No fixed declared model exists. Learning and use form one process. Traffic consists of carries.

Layer 3 — Coherence: light

Light-driven computation is demonstrated at component level. Shen et al. (2017) performed neural-network operations in interferometer meshes on a silicon chip (Nature Photonics). Coupled-oscillator networks are studied as a computational resource in their own right (arXiv 2502.04818): computation as phase locking rather than stored weights. Ternary hardware also exists (Jeong et al., Nature Electronics, 2019).

In the mesh, coherence is the admission criterion. A carry that brings neighbouring nodes into phase stays. One that does not closes as noise. Each node remains small. The mesh as a whole holds the capability.

Status

Layer	Status
Peer-to-peer shortwave transport	Running, in public use
Closure register	Built on FPGA and published; live test on an agent and coupling of layers still open
Light-driven coherence	Components demonstrated in the literature; integrated light-driven mesh not publicly documented
Full mesh	Stated by the author to be operational; not public, not independently verified

Consequences for verification and policy

Each of the suitcase’s three anchors is absent in the mesh:

Anchor	Datacentre AI	Coherent mesh
Something to tap	One frontend cable per rack	No central cable; radio between nodes
Signature of training	Gradient traffic, about 10⁵ above text	No gradients; carries only, within 15 kB/s
Declared model to recompute	Fixed weights	No fixed model; rules form during use
Site to inspect	A few large, known buildings	Many small nodes, no centre

What remains observable is the radio emission. Spectrum monitoring and direction finding show that a station transmits, and roughly where. They show neither whether computation takes place nor what it is. The shortwave band already carries amateurs, shipping, aviation, defence and broadcasters. A mesh produces no deviant pattern. Encrypted or spread signals can be detected but not read. Recomputation — the core of the suitcase — is then impossible. No literature on verifying peer-to-peer computing meshes over shortwave was found.

Three consequences follow for Dutch policy:

1. The suitcase serves the present paradigm, and has no plan for the next one. As preparation for a datacentre treaty it has value. As a basis for the claim of overseeing AI, it does not.
2. Verification grounded in compute size fails as soon as capability separates from size. That separation is the subject of Hooker (2024) and the design principle of the mesh.
3. A position at the table in the next paradigm does not come from tapping cables. It comes from understanding closure and coherence as the place where capability lives.

Annotated references

1. Schiffers, H. (5 November 2026). “Een Nederlands koffertje moet de razendsnelle ontwikkeling van AI controleren en verdachte situaties detecteren.” NRC. — The source news report on the Hague test. All factual claims in the essay were checked against the other sources below; items found only here are flagged as unverified.
2. Reed, T. (2026). “Tom Milton on how we could verify a US–China deal on AI.” 80,000 Hours (podcast). — The primary account of Amodo’s method, its pilot scale and its limitations, in the words of Amodo’s CEO. Basis for the numbers on pilot size, bandwidth caps and sampling.
3. Cankaya, N. (2026). “A System Overview for Near-Term, Low-Trust AI Compute Verification.” MIRI working paper. — An independent description of the same architecture: network taps plus recomputation on compute trusted by both parties. Corroborates that the tap-and-recompute concept is not unique to Amodo.
4. “My Assessment of Plan A’s Compute Verification Strategy.” — Non-expert notes describing the two-network design, per-rack bandwidth caps and the memory-wipe result (about 24 hours per rack, possibly over 100 TB unwiped). Included as context, with the caveat that the wipe figure is the author’s own reading.
5. “Can governments quickly and cheaply slow AI training?” Alignment Forum. — A critical assessment of compute-based verification. Establishes that GPU recomputation is not bit-exact, that covert-channel capacity must stay below 0.01%, and that current prototypes are probably ineffective against determined adversaries.
6. fides: packet-hashing verification tap (GitHub repository). — Third-party relay of Amodo’s internal SITREP on its 17 workstreams (4 active, 6 not started, 7 not on track). Second-hand source for the programme’s internal readiness.
7. Hooker, S. (2024). “On the Limitations of Compute Thresholds as a Governance Strategy.” arXiv 2407.05694. — Argues that compute is an unreliable proxy for risk because smaller models keep gaining capability. The central published challenge to the assumption underlying the suitcase, and the intellectual bridge to the mesh paradigm.
8. TenderNed: SSC-ICT, AI orchestration tender. — Official Dutch tender records. Confirms that SSC-ICT at Koningskade 4, The Hague, hosts on-premise AI orchestration (EUR 5 million contract), and therefore that the reported test location is plausible.
9. Longterm Wiki: Grant to Amodo Design. — Public grant listing. Confirms Survival and Flourishing Fund support of USD 1,229,000 (1,094,000 + 135,000, round SFF-2024-FlexHEGs; donors Jaan Tallinn and Blake Borgeson). Longview Philanthropy funding remains unverified.
10. Konstapel, J. (2026). “Keep the Question, Keep the Carry: The Future of Computing on the Net.” — The author’s own publication describing the closure register: balanced trits, carry rules, counter-machine universality, lossless rule learning, and the FPGA implementation. Self-published; the basis of the mesh’s computing layer.
11. Shen, Y. et al. (2017). “Deep learning with coherent nanophotonic circuits.” Nature Photonics. — Peer-reviewed demonstration of neural-network operations in interferometer meshes on a silicon chip. Evidence that light-driven computation is real at component level.
12. “Harnessing omnipresent oscillator networks as computational resource.” arXiv 2502.04818. — Studies networks of coupled oscillators as a computational resource, treating computation as phase locking rather than stored weights. Conceptual support for coherence as the mesh’s admission criterion.
13. Jeong, J. W. et al. (2019). “Tunnelling-based ternary metal–oxide–semiconductor technology.” Nature Electronics. — Peer-reviewed ternary hardware. Shows that three-valued logic, as used by the closure register, has a physical carrier beyond FPGA emulation.
14. DARC: documentation of JS8Call, VarAC and Winlink over VARA HF. — Evidence that peer-to-peer shortwave data transport is running today in public use, including store-and-forward relay and emergency e-mail. Confirms the mesh’s transport layer exists.
15. IET: “On-air testing of MIL-STD-188-110C Appendix D wideband HF waveforms.” — Documents the wideband HF standard used here for the 120 kbit/s (15 kB/s) bandwidth figure. The figure itself is from general knowledge; the ratio between it and gradient traffic (2.7 × 10⁸) is calculated.

---

Source: J. Konstapel, “The Suitcase and the Mesh,” Constable Research, Leiden, 5 November 2026. This essay is a readable English rendition of that document; all calculations and claim statuses follow the original.

The Suitcase and the Mesh