Most AI is scaling up. We're scaling down.
Project EDEN: a language model that fits in 16MB. It doesn't store its weights — it regenerates them at runtime from a 4-byte seed, like DNA. The only thing saved is the "epigenome."
Edge AI, rethought from biology up. 🧬
https://t.co/HCUhmwZnAw
Can you crush the loss floor of a token-less language model while capping the entire saved checkpoint under 16MB?
Project EDEN is hitting the H100 cluster to find out.
We’ve bypassed standard semantic tokenizers entirely, mapping raw UTF-8 bytes directly to a 256-vocab execution loop. But the real game-changer is the Epigenetic INR framework: 0-disk base weights generated purely via deterministic VRAM seeds.
⚡ Current H100 Telemetry Sandbox:
Loss: 1.8211 @ Step 20.7k+
2.62 Bits-Per-Byte (BPB) compression floor
Stable 1.05s / step processing cadence
Instead of human-engineered structural expansion, the network uses dynamic biological gating to actively prune underutilized expert folds (currently sitting at 15% fully pruned, 27% decaying).
We are pushing the parameter-to-size ratio to its absolute mathematical breaking point to see exactly where diminishing returns kick in on highly constrained hardware footprint limits.
Codebase branches, telemetry tracking dashboards, and white paper spec coming out of BVE Labs next. 🧬🦾 https://t.co/0x1LRcjbpN
#LLMs #PyTorch #NVIDIA #OpenAI #DeepLearning
@greynguyen How are you handling HIPAA-compliance? If you're storing lots of sensitive information, customers will want to know it's secure.
If you're looking for helping building out infra, fractional help, etc. Let me know.
@johnmark276@BrianSuttererMD The band constantly itches (as someone with more sensitive skin). Would love for them to at least solve that problem if they want people to wear 24/7.
Also, their "charging" mechanism leaves a lot to be desired.
@PratapRanade@arenaphysica@AMD@anduriltech@SiversSemicond Applied physics is so amazing, but it's unfortunate that it really just mean building more weapons for the DoD instead of creating things & solving problems for humanity.
@willahmed@WHOOP Until you see the pricing for basic services for Whoop as a consumer, then actual implementation. I bought two past few years. Bought Life & new Whoop. 3-4 day battery life at best. Props on marketing, but band quality is terrible, battery life and unit economics are hard pass.
@thsottiaux This is a nice benefit. Respectfully, would love to see OpenAI integrate better support flows--no reason that this can't happen. Any business user that has an issue or gets an account restriction, it's almost impossible to appeal or get any clarification.
@bmorecardiology Agreed. One of the things we're working on. General speaking, EPIC costs for SMB are almost untenable. Creating better automated solutions for SMB is what we do best, and where we can make some of the biggest impacts and opportunities for advancement.