BRAINIMPLANT by AIFLOWLABS · Research programme
Train with code, not just data.
Shape a model by writing to its neurons.
BRAINIMPLANT is an open workbench where AI agents reprogram an open-weight language model by writing directly to its weights. An Atlas finds the neurons that carry a behaviour, an Implant changes them, a Verdict tests whether the change helped and what else it touched. Every change is small, signed and reversible.
A language model drawn as a lattice: twenty-four translucent layers, each a grid of silver neurons. A gold probe enters from the right, and sixty-four columns in two layers turn gold. A signed Implant card appears beside them.
Chapter 02 / 11What happened on 2 November 2026
A language model edited another model’s weights.
Four MIT CSAIL researchers gave a closed frontier model, GPT-6 Astra, the raw weights of a small open model and a Python sandbox. No training run, no gradients: fifty forward passes and its own reasoning. It found the direction that made the small model loop, subtracted it from 131,072 weights in two layers, and the looping stopped.
- 0.016 %of the weights of Qwen3.5-0.8B edited: 131,072 numbers in layers 15 and 19.Takehi, Teoh, Zhang, Khattab, MIT CSAIL, blog post of 2 Nov 2026 ↗
- 7.9 % → 48.1 %MATH-500 with thinking on. GSM8K 13.8 → 67.1. Token-limit hits 93 % → 33 %.Same post; experiment 8 ↗
- +4.3 pointsthe fair gain, [1.9, 6.8], once the edited model is compared with the base model with thinking off (45.8 %). An independent check on Hugging Face.JonathanColetti/Qwen3.5-0.8B-Astra-WeightEdit, 4-sample MATH-500 ↗
Read it honestly
The headline. The base model loops until it hits the token limit 93 % of the time.
On a model that rarely loops (Gemma 3 270M), the same approach moved nothing by more than about a point. That is the honest size of the result. It is also early, public evidence that a general language model can repair another by reasoning about its weights. It needed fifty forward passes and a closed frontier editor. Both are the reasons this programme exists.
The camera dives into the lattice. Layers fifteen and nineteen glow; sixty-four gold columns in each stand out of the silver grid. Three readouts count up beside them.
Chapter 03 / 11Thirty-five years in the making
Nets that write other nets are not new. The editor is.
Every piece of this is published. A network writing another network’s weights: 1991. One network generating another: 2016. Rewriting a fact in closed form: 2022. Compiling a program into weights: 2023. The single direction behind a behaviour: 2024. A sentence turned into an 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. BRAINIMPLANT is the instrument that holds the pieces.
A path of eight silver nodes spirals around the lattice from 1991 to 2026. The last two nodes are gold.
Chapter 04 / 11The gap
Everyone asked the same three questions.
Under the announcement, the replies that mattered were not about the maths score. They were: what exactly changed, what else changed, and who approved it. No open tool answers all three in one place today.
And a quieter reason: a modified open model can hide a trigger that passes every clean evaluation. Diffing against the base is the standard first check, and we have not found it packaged as an open, end-to-end tool. For Hong Kong teams this is practical: HKCERT lists supply-chain and third-party risk among its five key risks for 2026, and a modified open model is part of that supply chain.
The lattice fractures into tilted layers. Three frosted panels float in front of it, one per question. Deep inside, one unlabelled neuron pulses red.
Chapter 05 / 11How it works
Train with code.
Find the neuronWrite the weightProve the change
Four parts, one loop. The Scribe is an AI agent with a sandbox and a budget of forward passes. The Atlas is a neuron-level map it reads to choose where to write. The Implant is the change itself: a sparse, signed, reversible diff. The Verdict is the evaluation loop that scores the change and everything around it.
The hypothesis we test first: neuron-level evidence lets an open editor do in a few passes what took a closed frontier model fifty. If it is wrong, we publish that too.
Four stations stand around the lattice: Scribe, Atlas, Implant, Verdict. Gold paths run from the Atlas to the Scribe, from the Scribe into two layers of the lattice, out again as an Implant card, and through the Verdict into a signed registry.
Chapter 06 / 11Why it matters
Fewer GPUs. Less data. More proof.
Teaching a model a behaviour the usual way means collecting data, renting a GPU cluster and running a job nobody can inspect. Weight programming starts from a model that exists and changes it with an Implant: a spec, a budget of forward passes, a proof. The published repair touched 131,072 numbers in fifty forward passes of a 0.8B model. The target fits on a laptop; the editor was a closed frontier model. Bringing that cost down is our first hypothesis.
Two ways to change a model
- Editor tokensa pool of open models proposes Implants
- Forward passesbudgeted and logged: tens, not millions of steps
- Reviewa person at the gate approves promotion
- Proofthe Verdict scores every Implant, every time
Illustrative, not measured. The compute ledger publishes the real cost of every Implant from the replication sprint onwards.
The hypothesis for M1 and M2: the Verdict’s score is the reward. A pool of open editor models proposes Implants, the ones that pass are kept, and the target model never goes through a training run. What comes out is a small, specialised model that fits on a robot or an edge device, which is where AIFLOW LABS builds.
Two lanes feed the lattice. Above, data blocks and a hot GPU rack push a long silver training run into the model. Below, a pool of gold Scribes writes Implants straight into it, and small copies of the model stream out to a grid of edge devices.
Chapter 07 / 11The Atlas
See inside before you trust it.
Pick a neuron. Push it down, leave it alone, push it up, and read the same conversation three times. If the behaviour lives where the Implant says it lives, the columns move. If a random edit of the same size does the same thing, the claim is flagged as unspecific.
Base model. The same prompt, untouched.
Synthetic preview: the bars stand for rollouts, not for real model text.
Built on the open neuron-explorer approach and compatible with Neuronpedia, circuit-tracer and TransformerLens. The Atlas also exposes its tools to the Scribe, so the agent that writes can look first.
The probe locks onto one neuron. Three frosted panels hang in front of the lattice, labelled minus one, base and plus one. Rows of bars in each panel stand for rollouts; a gold marker on the probe shows the dose.
Chapter 08 / 11Version control
Git for weights.
Edits stack as a history of signed diffs on a frozen, hashed base. log shows every Implant and who made it. diff shows the tensors. explain opens the Atlas page for that edit. 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.
Try it
$ implant log
Five Implants on base 7f3a…c1. Three written by the Scribe, two by people. Each carries intent, acceptance tests and a signature.
Each new edit is scored on the whole stack, because sequential edits are known to degrade models. The registry also reports how much of a model’s history was written by agents and how much by people.
A silver base block carries its hash. Five gold Implant cards stack above it on a glowing thread. On bisect, a cursor walks the stack and one card turns red. On revert, the cards fade from the top down and the base reports a matching hash.
Chapter 09 / 11Safety by design
An editor that rewrites models is governed, not just shipped.
The same machinery that repairs a model could weaken its refusals, implant a persona or hide a backdoor. So audit comes first: showing what a third-party diff really changed. BRAINIMPLANT ships no guardrail removal and builds its controls in from the first commit. These are properties we can test, not promises.
The Scribe stands on the left, the target lattice on the right, a platinum gate between them. A gold signed path passes through the gate. A dashed loop from the Scribe back to itself is cut and turns red.
Chapter 10 / 11What we will prove first
Five gates before anyone should believe us.
Every phase ends at a measurable gate, pre-registered before the work starts. Thresholds we cannot justify yet are set at the end of the replication sprint, not invented here.
Five platinum thresholds recede into the fog, labelled M0 to M4. The nearest one is lit gold.
Chapter 11 / 11A Hong Kong programme
A research programme by AIFLOW LABS.
AIFlow Labs Limited is a Hong Kong AI engineering company. BRAINIMPLANT 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: a harness, a public diff, an Atlas view and a note we publish together.
- Teams running open models. Become a design partner: bring one model and one behaviour you need to repair or specialise, and prove what changed.
- Partners and reviewers. Everything on this page is sourced. Request the technical dossier: research review, architecture, roadmap and the verification ledger.
The lattice settles. Its gold columns rest inside the silver grid, the probe withdrawn. The registry card beside it reads signed and verified.
Who is building it
Accountable people behind every result.
BRAINIMPLANT is owned and developed by AIFlow Labs Limited, a Hong Kong company. AI coding tools speed up engineering; named people are accountable for every result.
Only confirmed names are shown. No stock photos, no invented people.
- 01
Igor Lessio
Founder & CEO, AIFlow Labs Limited
23+ years building software. Production multi-agent systems and data infrastructure; interim CTO for an AI data company. Leads architecture, research direction and engineering.
- 02
Hong Kong developer
To be named
Hong Kong-based. Leads the replication sprint harness and local partner work. Named here once confirmed.
- 03
Advisers
Interpretability · IP and contracts
Named here once confirmed.
Take part
Write to the neurons with us.
The replication sprint starts with a public diff, a 0.8B model and a laptop. Researchers who want a shared harness, teams who run open models and need to repair, specialise or audit one, and partners or reviewers who want the technical dossier: write to us.
Where we are
Planning stage, November 2026. The research review, architecture, roadmap and verification ledger are written; the replication sprint is the first build. No results of our own are claimed on this page.