Questions
What people ask first.
Short answers, with the limits attached.
- What is REWIRE?
- An open workbench for writing, editing and auditing the weights of open language models. Every change is a small, signed, reversible diff (an Implant), produced by an AI agent (the Scribe), a solver, a generator or a compiler; shown in a neuron-level map (the Atlas); and scored by an evaluation loop (the Verdict). It is a research programme of AIFlow Labs Limited, Hong Kong, at planning stage.
- What did the MIT result actually show?
- That a frontier model with a sandbox can read raw weights of small networks and write working ones, and that one tiny gradient-free edit can remove a looping failure in a 0.8B model. It did not show that knowledge can be read from or written into a large model’s weights, and the fair capability gain is about four points once the base model is run with thinking off.
- How is this different from fine-tuning?
- Fine-tuning changes many weights by gradient descent on data. An Implant changes a chosen few by construction or by reasoning, is signed, carries its intent and inverse, and is scored against a regression battery. We do not claim it beats fine-tuning. We claim it can be made visible, measurable and reversible.
- How is it different from steering vectors or abliteration?
- Mechanically the 2026 edit is a steering-style direction removal baked into weights. The difference REWIRE adds is the loop around it: an agent that finds the target, a map that shows it, controls that test specificity, a manifest that signs it and a loop that scores it. Refusal removal is explicitly not a feature.
- Why “train with code, not just data”?
- Because the direction of travel is a model whose behaviour can be changed by writing a program, a direction, or a few neuron values, with a proof attached, rather than only by another training run. “Not just” is deliberate. Training by data is not going away.
- Is this recursive self-improvement?
- No, and we do not build or promote unattended self-rewriting. The editor may never target itself, promotion needs a human, every edit is logged and reversible, and any online mode is quarantined with automatic rollback. The published result is a one-shot edit of a small model.
- What do you ship first?
- A replication: the public edit re-run in our harness, loaded into the Atlas, with one small Verdict run and a two-page note. If our numbers do not match the independent check, we say so.
- Who is it for?
- Engineers who run small open models on their own hardware, including Hong Kong teams that keep data on site, and need to repair, specialise and prove what changed; auditors who review third-party diffs; interpretability and model-editing researchers who want a shared harness and benchmark. These are hypotheses we test with design partners, not market claims.
- Is it open source?
- Planned: permissive licence for our code, CC BY 4.0 for manifests, base-model licences preserved. We extend lyra’s neuron-explorer (MIT) rather than rebuild it.
- What does it cost?
- The replication sprint has a zero-cash path: public diff, 0.8B model, a laptop. Larger sweeps use pay-as-you-go GPUs. We report compute per edit as part of every result.
- Why Hong Kong?
- AIFlow Labs Limited is a Hong Kong company. The 2026 Policy Address sets an explicit control model for AI agents, and Hong Kong teams increasingly run open models where their data lives. An auditable, reversible way to change those models fits both.
- Could it make a model unsafe?
- Any editor could. That is why audit comes first, the Scribe can never target itself, promotion needs a recorded human approval, every edit can be reverted bit for bit, and refusal removal is outside the programme by policy.
01
02
03
04
05
06
07
08
09
10
11
12