Hi everyone, Iām Jen, an indie researcher with a math background.
Iām always curious to find other āgrassrootsā R&D teams or individuals who are also in the weeds, building foundational models or systems from first principles.
My current focus is on math reasoning and the ādistributed chaosā of optimizers like Muon/FSDP.
Just wanted to put a signal out:
Are there other independent or small āgrassrootsā teams out there I should be following?
Who else here is deep-diving into math reasoning problems?
Always looking to connect with fellow trailblazers and see what hard problems everyone is tackling!
Good topic. Iām in that camp. Prototyping a modular opportunistic control system targeting domains where linear reasoning is brittle and brute force is useless (open-ended, unbounded problems).
Thanks for the reply. Your description of a āmodular opportunistic control systemā for unbounded problems is fascinating.
It reminds me of the multi-faceted (and somewhat āchaoticā) approach in Moonshotās Kimi K2 paperāsynthesizing domains, tools, agents, and rubrics all at once.
Iām curious if your work is focused more on the post-training phase (like complex reward models) or on the agentic workflow itself (like the real-time API/tool orchestration)?
For me, my current project is a scratch-build of ReTool paper from ByteDance Seed, so Iām always excited to find other researchers working in this space.
Very cool. My understanding is that ReTool takes a kimd of workflow-centric view of orchestration. Something like a declarative DAG of tool calls?
Iām experimenting with a more signal-driven controller that learns when to reallocate attention between subtasks. But itās a complementary path.
My orchestration layer treats models like K2 as components in a control loop that learns when to change direction, reallocate effort, coordinate subtasks, etc. One key feature is using a variety of custom heuristics to measure and respond to āstucknessā on sub tasks. Iām basically treating stuckness as a first class signal instead of a nuisance.
This is a really interesting approach. āStuckness as a first class signalā is a cool concept. Is there a paper or even a blog post I could read to learn more about this āsignal-driven controllerā idea? Iām always curious to see how different researchers are framing these orchestration problems.
In the works, but planning comprehensive documentation after implementation to avoid unnecessary rewrites. I had thought to begin blogging but time has been limited as of late.
If you get interested in collaboration, maybe we can cross pollinate ideas and see what blooms.
Grizz here - Iām from Texas, background in emergency medicine and E.M.S. Operations for over 17 years approaching the whole AI consciousness questions from a radically different angleā¦
of optimizers like Muon/FSDP.