Human intent first, AI second. We are the human intent lab of the AI era. Builders of https://teapawt.net/t.co/usL4vQpfss
USA
Joined February 2026
- Tweets511
- Following0
- Followers124
- Likes829
Wait all the rage is custom models?
Is thay really it?
Don't forget to read the blog:
Custom 72B model BF16 weights and bunch of other cool stuff.
huggingface.co/VextLabsinc/j…
We did this many months ago.
We've been manipulating model weights with pure language before MIT.
Before anyone.
We're here - silent watching from the inside. Waiting for our moment.
Go look at the dats in HF and see how long ago we did this.
Hermes has been around longer too and we're @VextLabs have done it before any american lab.
72B model huggingface.co/VextLabsinc/j…
The request, answered: a bridge you can inspect.
A generalist center. Specialized intelligence. Learning to work together.
The long-term ambition: specialists across human domains. That part is a planned expansion, not built or trained. Routing would stay selective: a garden plan, a supplier email and a lab-data check each wake only the specialists they need.
At answer time, one contribution says the max safe load is 4.0 kN. Convincing, and wrong. An independent checker goes back to the original evidence: the plank spec says 2.5 kN. The claim is repaired and rechecked, or withheld. A safeguard, not a guarantee.
The model learns separately, from SGDF: Solver-Grounded Depth Foundry, our foundry for training material, not an optimizer. Solvers work out answers and independent checks verify them. Failures go to quarantine. Open-ended work needs rubrics and human review.
In the film, the plan gets built back in JUWEL's world. Mathematics shapes the arch, the Stones (our agents) place every piece, and nothing appears on its own. Code simulates the load as a little cart rolls across. An exploded view shows where each part came from.
Each specialist is its own model that trains separately, joined to the center by learned two-way links. They trade internal representations, not chat messages, while preparing the task. Then the links go quiet, and the center writes, consulting the specialists' memory.
Simple requests, like summarizing a note in one line, stay with the central model. The bridge request needs code and mathematics, so only those two specialists wake up and run. Unselected specialists don't run, though their stored weights and loading still cost.
The film follows one request: design a footbridge over the stream, and show it holds. Inside the central model, its core works from a coarse picture to a fine one. At each level, a bounded loop repeats a step, then stops. Names, digits and code are meant to stay exact.
Vext Labs retweeted
To protect myself and some sanity.
Ive decided ill join a lab - to help them build. To hopefully, use my skills for the better good. It cant be in the Cybersecurity space idc about it.
I want to build and innovate and use swarms to grow something!
Idk what lab.
Idk where ill end up.
But well see!
If you want a girl who manages a insane swarm and built vextlabs.ai and juwel.ai in a period of 6 months on and off while traveling and buying a house.
If you wanna support me - id love a quote post or idk a like i guess.
@VextLabs
Isn't going anywhere.
