Skip to main content

Lab Notebook · AI on any computer

A neural network on a 1982 ZX Spectrum — 2.2 KB, no floating point, no GPU

In April 2026 we ran a real neural text generator on a 44-year-old 8-bit machine: a 2.2 KB GRU on a Z80, multiplying by shift-and-add and doing sigmoid from a 256-entry lookup table. It proves neural inference needs neither a GPU nor floating point — and it's deliberately open: fakes don't run on a Z80.

JR
Jon RossFounder, vocabotics — 15 years building safety-critical systemsLab report · dated 8 April 2026
Verified by a human. Drafted with AI, verified by a human. Jon Ross, 8 Apr 2026
Living document. Reviewed 8 Apr 2026
Entry date
8 April 2026
Category
Hardware
Lead over the world
the proof, not the lead
Access
🔓 Public
ResearchProvenDownload PDFPDF · 793 KB

A real neural net, generating text, in 2,229 bytes on a computer older than most of the people reading this — no GPU, no floating point, no cloud.

2.2 KB
the whole model, as a .tap file
Z80
1982 8-bit CPU, no FPU
shift+add
multiply-free integer maths
256-LUT
sigmoid from a lookup table

Honest evaluation

Proven

The hardware is the referee — a real GRU running inference on a real Z80 is a claim that can't be faked.

The evidence — full reasoning behind the verdict

Verdict: proven. The claim is specific and falsifiable: a real neural network can run inference on 1982-era 8-bit hardware, with no GPU and no floating point. It did — a 2.2 KB .tap file containing a real 2-layer INT8 GRU, running on a Z80, doing its multiplication by shift-and-add and its sigmoid from a 256-entry lookup table because there's no FPU to compute it directly.

It landed here because the hardware itself is the check. There's no way to fake a result on a Z80 — the model either fits in memory and executes on the available clock cycles, or it doesn't, and there's no benchmark methodology to argue about. That's what makes this one of the cleaner proofs in the whole archive: the claim and the test are the same object.

The honest scope limit, stated plainly in the file, is that this is a floor, not a ceiling — a proof of possibility and a deliberately public demo (the well-known Cho 2014 GRU, not the lab's proprietary architecture), not a claim of a useful model. Two vintage platforms are working; six more are designed but not built. What would extend the proof: getting a third and fourth platform (BBC Micro, C64, and so on) actually running rather than just designed — each additional working platform is a further, independently falsifiable instance of the same claim, not just more of the same evidence.

Runnable proof — see it work

If you want to prove that AI does not need a datacentre, the cleanest possible demonstration is to run it on a machine from 1982. You cannot fake a result on a Z80 — it either fits in the memory and runs on the clock cycles, or it does not. This report is open on purpose: it is a viral, trust-building proof, and there is no moat to protect in it.

What it was

A project to run neural inference on vintage hardware — ZX Spectrum, BBC Micro, C64, Amiga, Psion, 486 — wrapped in a generative adventure game where the model's hidden state literally is the save file. The organising question: how low can the hardware floor for a real neural network go?

What we built

A real 2-layer INT8 GRU exported to run on the metal:

  • A 2.2 KB .tap file — the entire model — running on a Z80.
  • Multiply-by-shift-and-add integer maths, because the Z80 has no floating-point unit.
  • Sigmoid from a 256-entry lookup table, because you cannot afford to compute it.
Research

The measured fact is delightfully concrete: a 2,229-byte neural text generator that runs on a 1982 8-bit CPU. Two platforms are working; six more are designed. No GPU. No floating point. No cloud.

What we learned — including the honest edges

  • It uses the public GRU, on purpose. This deliberately runs the well-known GRU (Cho 2014), not the lab's proprietary wave architecture. That is a two-tier IP strategy in action: a public face that can go viral without leaking the private core. Saying so is part of the honesty.
  • It is a floor, not a ceiling. This is a proof of possibility and a demo — not a useful model. It shows what is minimally required, which is the interesting number, not what is maximally capable.
  • "Fakes don't run on a Z80." The reason this is such a good demo is that the hardware is the referee. There is nowhere to hide a fudged result inside 2.2 KB and a 3.5 MHz clock.

Where it went / status

Kept open, as a demonstration and a recruiting tool for the lab's core belief: intelligence you own should not require hardware you rent. It pairs with the sub-£600 BCI at the sensing end — cheap sensors in, tiny models on the metal — to make one argument, from two directions, that the floor is far lower than the industry assumes.

What is still open — kept visible

The honest edges, next to the wins. This is what turns 🔬 into 🟢 — honestly.

  • It uses the public GRU (Cho 2014), NOT our proprietary architecture — a deliberate public face that can go viral without leaking the core IP.
  • Two vintage platforms are working; six more are designed, not done.
  • It is a proof-of-possibility and a demo, not a useful model — the point is the floor, not the ceiling.

Proofs & sparks

We demonstrate rather than assert. Each ✅ proof is a visible result with a hard figure.

  • A neural net on a 1982 Z802.2 KB net on a 1982 Z80a 2-layer INT8 GRU shipped as a 2.2 KB .tap file runs on a real or emulated 1982 ZX Spectrum — you can watch it think on a Z80.

Where this connects

Sources

  1. vocabotics project audit — Retro / 'AI on every computer ever made' (2.2 KB GRU on Z80), Apr 2026vocabotics internal project history · as of April 2026
  2. Cho et al., Gated Recurrent Unit (the public architecture used for the demo)the public model this demo deliberately uses · as of 2014

    We use cookies.