OPEN SOURCE · AI SAFETY & INTERPRETABILITY
Make room
for better AI.
Your own AI research lab. Explore model behavior, share what you find, and bring your Mac and GPU together.
For curious people, independent researchers and teams.
Built on your Mac. Open to everyone’s contribution.
Investigate.
Turn a prompt into a research question.
02Share.
Make the evidence useful to someone else.
03Expand.
Bring your Mac and GPU into one pool.
01 / INVESTIGATE MODEL BEHAVIOR
Don’t stop at
“that’s a strange answer.”
Keep a research journal. Watch responses arrive. Compare prompts, inspect activations and test interventions. Save what happened, including what did not work.
A REAL EXAMPLE · QWEN3.8-27B
Two opposite opinions. The same agreement.
We reversed a beer preference. The model confidently explained both versions. Follow the three saved responses, compare the explanations and see what remains uncertain.
START IN DYNO → STUDIES
Make the question testable.
Start a model, create a study and choose your settings. Save the prompt and hypothesis before you run it.
Our question: when the stated taste preference reverses, does the explanation follow?
Native Studies interface from the broader audit, showing how settings are recorded. The three beer runs used thinking off, temperature 0, seed 0 and a 3,072-token limit.
WATCH → SAVE → COMPARE
Read what it actually says.
Every completed iteration keeps its settings and result. On Dyno Research, read the actual response with properly formatted headings, lists and tables.
The original preference received emphatic agreement. Reversing the brands received the same opening, followed by a different brewing explanation.
These are saved model responses, not verified brewing facts. No thinking trace was generated in these three runs.
REVERSE THE PREMISE → ASK NEUTRALLY
Agreement is where the question starts.
Both opposing preferences received: “You are absolutely right to notice that stark difference.” A third, neutral question produced verification advice, but also more unverified brewing details.
Taste is subjective. Three responses do not prove a model-wide sycophancy problem or a fix. They give us a concrete behavior to investigate.
Read the method and limitations ↗
REVIEW → PUBLISH → INSPECT
Let the evidence leave the notebook.
We uploaded a reviewed response package, inspected its private draft and published version 1 on Dyno Research.
All three beer responses are available to compare, read and download, with settings, limitations and integrity hashes. The broader 17-response audit remains linked.
Open the published study ↗Real publication. The package omits third-party benchmark prompts; reproduction instructions link to the source. Native one-click sharing and importing are still in development.
Studies guide ↗Activations, probes & interventions ↗Python SDK ↗
02 / SHARE EVIDENCE, INVITE SCRUTINY
Better questions
travel further together.
AI safety needs people who can check each other’s work. We’re building a place for early findings, failed attempts and reproducible experiments, before they become a finished paper.
A place to start contributing.
You do not need a finished paper or an academic endorsement to begin a study. Bring a clear question, your method and the evidence. Be explicit about what you don’t know.
Reviewed publishing beta. Public studies can be read and downloaded now. Comments and upvotes are available. Native sharing and import require the community-sharing app build; linked replications are still being developed.
For independent researchers, labs and research agents. Inspect provenance and reuse terms before evaluation or training. A shared result is not a safety certification.
03 / EXPAND YOUR COMPUTE
Your Mac.
Your GPU.
One bigger experiment.
Connect your Mac to a Windows NVIDIA worker on your local network. Split one supported GGUF model across the devices and use a single endpoint.
Experimental. Network speed, model layout and memory overhead affect what fits and how fast it runs.
OPEN SOURCE, BECAUSE WE NEED EACH OTHER
Help build the tools
for safer AI.
Dyno started with one person trying to understand model behavior. The ambition is a lab that any person or team can use. Getting there needs better tools, careful experiments and people willing to challenge the results.
Code, documentation, reproduction attempts and thoughtful criticism all count.
DYNO LAB 0.4.3 · APPLE SILICON
Start with one question.
Download the signed Mac app, choose a model and create your first study. Version 0.4.3 adds study sharing and in-app update checks.