The walkthrough

1. 1. Train with code, not just data

   REWIRE is an open workbench where AI agents reprogram an open-weight language model by writing directly to its weights. An Atlas finds the neurons, an Implant changes them, a Verdict tests the change. Every change is small, signed and reversible.

2. 2. What happened on 2 November 2026

   On 2 November 2026, MIT CSAIL researchers gave a closed frontier model the raw weights of a small open model and a sandbox. No training run, no gradients. It found the direction that made the model loop, removed it from 131,072 weights, and the looping stopped. It removed a failure mode; it did not add knowledge.

3. 3. Thirty-five years in the making

   Every piece is published. Fast weights, 1991. Hypernetworks, 2016. Closed-form fact edits, 2022. Compiled transformers, 2023. Steering directions, 2024. Text to adapter, 2025. Each needed a system built for one job. 2026 adds a general model with tools as the writer, and a map of the neurons it writes to. REWIRE holds the pieces.

4. 4. Everyone asked the same three questions

   Under the announcement, the replies that mattered asked three things: what exactly changed, what else changed, and who approved it. No open tool answers all three in one place. And a modified open model can hide a trigger that passes every clean evaluation, so diffing against the base comes first.

5. 5. Train with code

   Four parts, one loop. The Scribe, an AI agent with a sandbox and a budget, reads the Atlas to choose where to write. It writes an Implant: a sparse, signed, reversible diff. The Verdict scores the change and everything around it. Find the neuron, write the weight, prove the change.

6. 6. Fewer GPUs, less data, more proof

   The usual way is data, a GPU cluster and a run nobody can inspect. Weight programming changes an existing model with an Implant: a spec, a budget of forward passes, a proof. Hypothesis: the Verdict’s score is the reward, open editors propose Implants, and out comes a small model for a robot or an edge device.

7. 7. See inside before you trust it

   Pick a neuron. Push it down, leave it, push it up, and read the same conversation three times. If the behaviour lives where the Implant says, the columns move. If a random edit of the same size does the same, the claim is flagged as unspecific.

8. 8. Git for weights

   Edits stack as a history of signed diffs on a frozen, hashed base. Log, diff, explain, bisect, revert. Bisect finds the first edit in a stack that breaks a test. Revert restores the base bit for bit, and that property is tested, not assumed.

9. 9. Safety by design

   The same machinery could weaken refusals or hide a backdoor, so audit comes first. No self-targeting, a human gate for promotion, scope limits, signed provenance with bit-exact revert, audit mode, and no guardrail removal. Properties we can test, not promises.

10. 10. What we will prove first

    Every phase ends at a measurable gate, pre-registered before the work starts. Replicate the public edit. Open editor at a budget. Hybrid editing with bit-exact revert and bisect. Compiled modules grafted, or a documented negative result. Then a public benchmark.

11. 11. A research programme by AIFLOW LABS

    AIFlow Labs Limited is a Hong Kong AI engineering company. REWIRE is where we take the next question seriously: once a model can be programmed with code, what does it take to do it in the open, and safely? Researchers, join the replication sprint. Teams running open models, become a design partner.