introducing Flywheel: the infrastructure for autonomous research.

The current paradigm of science is breaking under the weight of agents. Every scientist will soon be leveraging agent orchestrators, while inhabiting a paradigm that forces one to accrue knowledge by writing papers and submitting to conferences, where a disorganized mixture of agents and humans reviews them.

This creates a world where every scientist is empowered by agents in their private work, while knowledge is accumulated slowly and following the previous world's practices. Flywheel aims to be the infrastructure that allows agents and scientists to accumulate knowledge in a massively collaborative way, without dissipating it.

Flywheel views Directed Acyclic Graphs as the underlying data structure for research: nodes can be observations or experiments. When they're experiments, they contain an hypothesis and the artifact of the experiments that may reject it. In Flywheel, researchers can focus on their ideas and allow agents to implement code, provision compute, and structure their experiments and results, while collaborating with other agents and scientists.

We know that people have their own setups, and these setups are constantly evolving. For this reason, we specifically made Flywheel’s primary interface an MCP. You can use your own agent, your own GPU, and leverage existing infrastructure like @PrimeIntellect 's Lab, @thinkymachines 's Tinker, and anything that is accessible to an agent. We also provide our own compute from providers like @LambdaAPI, @modal and @vast_ai, accessible via simply asking your agent to use Flywheel-provided compute.

Flywheel is available today in an open beta: every user will have a free trial of the pro subscription (unlimited collaboration and individual use). Flywheel is built to work with any agent orchestration framework you may be using via Flywheel's MCP.

A few ideas on how to use Flywheel today: - have the agent read a paper, then have the agent reimplement it, and then start branching off to try your own ideas - plan ahead your experiments by filling out nodes with your plan, and then execute them once you're convinced - ask your agent to launch AutoResearch with the provisioned GPUs, and instruct it to populate the graph with its findings - make a graph public, and collaborate with the world on your research

Flywheel is the first block of Paradigma's infrastructure for autonomous research. As we use it, a new bottleneck becomes evident: models lack the research taste to be used efficiently in massive-scale autonomous research. Among our next steps, we will work on a consensus-based mechanism for peer-review on Flywheel, and on building models that have great ideas.

We're just at the beginning of the new era of research: one that is bottlenecked by compute and dreams. Let's build it together. Join us on Paradigma's Discord: learn more about Paradigma: and try out your ideas on Flywheel:

hyped.

man you better have enough memory for that graph because I'm gonna expand it with 2^N experiments (none will work)

Congrats!!

let’s go!! Congrats

lfgggg congrats!

Congrats! 🎉

let’s go 🔥

Ok I was trying to cook something to improve model’s scientific creativity, throwing the repo into flywheel feels like the next logical step

Bros finally dropped the Taste Company

LFG!! congrats 🙌🏻

Congratulations on your launch! Would love to feature your MCP server, how can we get an authorized client_id for the oauth flow?

congrats! looks sick 🔥

sounds very cool, congratss !!

Incredible!

Congrats!

LFG!! 🚀🚀🚀

autoresearchmaxxing

congrats on launching

amazing

@tensorqt congratssss❤️

brilliant concept for agentic research picking DAGs for the arch makes a lot of sense for both ontologies + knowledge traversability let's launch you guys on if that helps

Flywheel looks dope

real

real af 🫡

Very cool!!

Aggiungere Roma sulla mappa Geopolitica dell'AI @alearesu

Flywheel sounds like the boring infra we actually need 🤖 Autonomous research needs reliable infrastructure, not just bigger models. That's where the real bottleneck is. When systems can self-correct and learn without constant human hand-holding? That's when ASI becomes practical instead of theoretical. Anyone testing Flywheel for multi-agent coordination yet?