Clusy Inc

@eldar_hsnv with Clusy you can do anything

Can you change an LLM’s behavior without fine-tuning it, changing the prompt, or updating its parameters? Turns out you can. We reproduced activation steering in Clusy and shifted the same Qwen 2.5 7B model from negative → neutral → positive by adding a single vector inside the network. Same model. Same prompt. Same weights. Just one vector. Reproduced in Clusy from the ActAdd paper in minutes.

Clusy has already helped bring 400+ projects to life. Here’s a quick look at some of the published use cases on our Hub. Check them out, and let us know what you’re building!

Finance doesn’t have a modelling problem. It has a “this took 3 days and 14 spreadsheets” problem. That’s what we’re fixing

The most impressive thing an AI agent did for me this week wasn’t getting the right answer. It was realizing its answer was wrong—and retracting it. I was a CV researcher and have a demo with a CV startup on Saturday: production computer vision for a live manufacturing line. The task: given a camera frame and the print plan, segment only the objects that should be visible. I was in calls all day, so I forwarded Clusy the client’s raw email and ZIP file and walked away. It came back with all 293 frames auto-labelled using SAM 3. Not color thresholding. Concept-prompted segmentation—which matters because the red material in the images isn’t actually their production filament. Against a reference set I built independently: - 0.82 recall - 0.78 precision But the labels weren’t the interesting part. Clusy also discovered: • The print bed shifts 100–200px relative to the camera, so a fixed transform won’t work. • Auto-exposure drift causes ~85% of pixels to change between frames, killing frame differencing. • Eight of the “images” were actually 404 pages saved as .jpg files. Then it ran an ablation and reported a huge result. I pushed back. It re-derived the experiment, found that it had augmented over one range and evaluated outside it, and withdrew the conclusion. The result was wrong. The agent caught its own mistake. Still running the full experiment, but this is the behavior I care about: Not an AI that always sounds confident. An AI that can argue itself out of a seductive, incorrect result.

Reproducing the BERT experiment in Clusy with just one prompt. The whole process took under 15 minutes, thanks to parallel branching. We ran four learning-rate experiments at the same time. We even beat two of the accuracy benchmarks reported in the paper. Research reproduction is becoming a prompt, not a project. Check it out at clusy.io

Update: a bunch of new Clusy features just shipped: • Expanded CV support, including data labelling • More robust agent plan mode and branching • BYOK for Claude and OpenAI • Connect local Claude Code/Codex and use your subscription • New GPU runtimes and higher-RAM options All changes are live now. Apologies for the brief disruption earlier.

Just tested Samsung’s tiny 7M TRM model in Clusy. The claim is pretty wild: it can score very high on ARC-AGI-1, even against much larger models. So I ran the actual eval instead of just reading takes about it. Clusy pulled the repo, made the plan, loaded the checkpoint, ran the benchmark, showed solved ARC grids, then tested the obvious question: is it really reasoning, or is it leaning on puzzle IDs / recursion / augmentation voting? Replay here: app.clusy.io/share/bA9seSDW… This is the kind of thing I want research workflows to become: run it, inspect it, challenge it. Try it at clusy.io now.

Don’t just read the paper. Make it prove itself.

I tried using Clusy to look for UFOs 🛸 I’ve been having fun lately analyzing weird datasets, and this one was about UFO sightings. Clusy is obviously useful for AI/ML experiments, but I also like using it just to poke around in data, find patterns, and see what interesting stuff shows up. Notebook here 👇 app.clusy.io/share/IVqQi3at…

Hot take: reproducing ML papers should take hours, not weeks. We reproduced the SetFit paper end-to-end — data, training, eval, results — without the usual notebook chaos. Research should ship this cleanly.

Try at clusy.io today

Jupyter won the interface. Clusy upgrades everything around it: - cloud compute - clean environments - autonomous runs - experiment branches - shareable workflows The future of notebooks is not local files. It is cloud-native workspaces.

We built @clusyio - an AI-powered Jupyter Notebook that helps you run your experiments. Our agent helps you at any step of the pipeline: finding datasets, preprocessing, EDA, model architecture selection, actual training, evaluation, deployment, sharing, etc. Try it out at: clusy.io

Introducing Clusy: an agentic notebook for data science and AI workloads. From data sourcing and environment setup to code generation, cloud execution, model training, evaluation and deployment, Clusy helps you go from idea to result faster. Open beta at clusy.io