Jacque Istok

jmfi

AI & ML interests

None yet

Recent Activity

reacted to cahlen's post with 👍 11 days ago

In the spirit of "just making shit"—because, frankly, you can these days—I decided to get back to basics. I built a lightweight RL MLP powered by WebGPU that runs directly in your browser. The twist? I replaced the standard MLP with a Continued Fractions network. The real win here is interpretability; by applying a Taylor expansion to the continued fractions, we can actually decompose the output and see exactly which features influenced the outcome. It effectively replaces the usual "it’s just black-box magic" with actual visibility into the logic. It was a fun experiment. Feel free to clone the repo and make it your own! https://huggingface.co/spaces/cahlen/neuron-runner https://github.com/cahlen/neuron-runner https://huggingface.co/papers/2601.21766

liked a model about 1 month ago

enfuse/edgevla-small-fmb

reacted to cahlen's post with ❤️ about 1 month ago

I was in the mood to fine tune a model, you know since we can just do shit these days, but I've also been on a pure math kick lately going back to some of my old academic stuff. And I remember specifically back then thinking about how difficult it was to narrow down a problem worth working on. This is hopefully help narrow that gap, at least for some niche parts of number theory. Convergent is a model that knows number theory and computational mathematics. But I've also trained agentic tool calling loops, using prebuilt tools advertised on my computational number theory project website https://bigcompute.science So essentially, you can use this to talk about unsolved problems, ways to attack them, how to write the cuda kernels to investigate them on your local GPU, or come up with brand new conjectures! But most importantly it can zero you in on something that might be more interesting, especially if you have a model to talk with the problem about. I'm aware their aren't a ton of number theorists that just hangout together... heh. So I'm sure there are bugs. Just post on github or huggingface and I'll be able to get to them if I have time. Everything open source. Model weights, training data, training scripts. Have at it. https://huggingface.co/cahlen/Convergent-7B https://huggingface.co/datasets/cahlen/Convergent-7B-data https://github.com/cahlen/convergent https://bigcompute.science https://mcp.bigcompute.science/mcp

View all activity

Organizations

reacted to cahlen's post with 👍 11 days ago

Post

334

In the spirit of "just making shit"—because, frankly, you can these days—I decided to get back to basics. I built a lightweight RL MLP powered by WebGPU that runs directly in your browser.

The twist? I replaced the standard MLP with a Continued Fractions network. The real win here is interpretability; by applying a Taylor expansion to the continued fractions, we can actually decompose the output and see exactly which features influenced the outcome. It effectively replaces the usual "it’s just black-box magic" with actual visibility into the logic.

It was a fun experiment. Feel free to clone the repo and make it your own!

cahlen/neuron-runner
https://github.com/cahlen/neuron-runner
CoFrGeNet: Continued Fraction Architectures for Language Generation (2601.21766)

liked a model about 1 month ago

enfuse/edgevla-small-fmb

Robotics • Updated Mar 15 • 20 • 1

reacted to cahlen's post with ❤️ about 1 month ago

Post

233

I was in the mood to fine tune a model, you know since we can just do shit these days, but I've also been on a pure math kick lately going back to some of my old academic stuff. And I remember specifically back then thinking about how difficult it was to narrow down a problem worth working on. This is hopefully help narrow that gap, at least for some niche parts of number theory.

Convergent is a model that knows number theory and computational mathematics. But I've also trained agentic tool calling loops, using prebuilt tools advertised on my computational number theory project website https://bigcompute.science

So essentially, you can use this to talk about unsolved problems, ways to attack them, how to write the cuda kernels to investigate them on your local GPU, or come up with brand new conjectures!

But most importantly it can zero you in on something that might be more interesting, especially if you have a model to talk with the problem about. I'm aware their aren't a ton of number theorists that just hangout together... heh.

So I'm sure there are bugs. Just post on github or huggingface and I'll be able to get to them if I have time.

Everything open source. Model weights, training data, training scripts. Have at it.

cahlen/Convergent-7B
cahlen/Convergent-7B-data
https://github.com/cahlen/convergent
https://bigcompute.science
https://mcp.bigcompute.science/mcp

reacted to cahlen's post with 🔥 about 1 month ago

Post

289

Speaking of just being able to make shit these days.

4,753 WiFi networks. 1,544 BLE devices. 41 IR signal patterns. 117 kilometers. 257 open networks. 125 printers exposing WiFi Direct. 163 Samsung SmartTags. 121 separated AirTags. A Mercedes-Benz MBUX infotainment system advertising to anyone in range.

That's from just one leisurely evening drive in socal.

All classified, mapped, and queryable. From a device that fits in your pocket.

ESP32 Marauder + Flipper Zero + DGX Spark running OpenClaw with a local Qwen 3.5 35B abliterated model. All AI processing on-prem. No cloud APIs, no token burn. If you can run it with a local LLM, do so. Free inference forever.

Throw it in your bag and drive. Every 20 seconds — full sensor rotation. WiFi, BLE, infrared, SubGHz. GPS-tagged, timestamped, saved to SD. When WiFi is up, it fires data home. The device never waits. It just keeps scanning.

Two ESP32 chips plugged into a Flipper Zero. White chip does WiFi and BLE. Orange chip does IR and SubGHz. Both scan in parallel. Two sets of hands, one brain.

It classifies every device from raw BLE bytes — AirTags, SmartTags, printers, iBeacons. Builds a persistent knowledge graph. Filters out your own devices automatically. Plots everything on a live map with an AI chat panel. Ask it to highlight all separated AirTags and it does. OpenClaw does the heavy lifting, I just emit data to a sensor garbage disposal type API endpoint.

What interests me most is what happens over time. Same routes, different days. Devices that show up Monday but not Friday. Networks that vanish. Trackers that migrate. The ontology never rotates.

I'm probably never going to release this. But you could probably build it yourself.

Casually I call it 'MarauderClaw' because I added a full agentic action mode where OpenClaw can take over and control the flipper zero and Marauder entirely, but that just kinda seemed like a bad idea, so nah on that one boss.

Favorite WiFi name in the wild so far: Wu-Tang-Lan.