n4pw01f

Nice work! You gave me something to fix!!

The model improvements value are at the plateau of utility right now, peeling out small gains which is pretty “meh” in terms of business value

at this point frontier companies are just selling upgraded harnesses and tool calls with the rest of us

One interesting thing I found recently was after enforcing geo-fencing, recaptcha intermittently started failing. They might be using global edge servers or something

I’m getting surprisingly good results on GEMMA4 2 bit, running on a Nvidia nano and raspberry pi

Used to follow this to a T. Love 12factor, evangelized it at a lot of companies too

Hot take - but I’d learn to build CNNs from scratch. We are only just now scratching the surface of convolutional neural networks capabilities

Self hosted GitLab has been good to me forever and has scaled and has a controllable attack surface as long as you keep on top of it

Newer app is moving to Google Cloud Secure Source Manager (because we are on Google Cloud and using backbone auth so it made more sense and less involved to manage)

If you look at how this works under the hood, “when two words that don’t necessarily matter to the context, it will pick one word over the other and a pattern emerges”

And that pattern will be reverse engineered in weeks by a startup

And a downstream agent produced by said startup will be built to roll those dice again and defeat the watermarking

And a the LLM vendors will combat this

And a cottage industry shall be born anew

Have a pretty large application running HTMX + WebComponents + Hono + Cloud Functions to serve up frontends, happy with it for about 2 years now

People throw salt at my stack but it’s fast, and straightforward to manage attack surface

Are you running bare metal or docker? I had to go down to 2 bit on Gemma4 E2B model to run on 8gb on a Jetson

Quality and idempotency is great but it’s still not exactly fast… fast enough and works offline

Is this something that you can get running on Debian?

Replace React / Vue with HTMX 2 years ago and still going strong.

Hono + WebComponents + HTMX + serverless is the backend for my apps now.

We ended up doing the same, own harness with a world view (schema.org) and map reduce rather than a RAG. Happy with the results and context hydration speed

I created a nice harness and visual workflow builder for my Gemini agent chains, works very well. I did this so it would create code the way I do, that is very editable.

In contrast, the vs code plugin was pretty bad, and did crazy things like mix languages

I am seeing shit tons of chatbots for everything under the sun being onboarded at corporate

Why not Jitsi - genuinely curious

I have used HTMX extensively with AI agents. It is fantastic for dynamic views.

Agent one: handles the request and does tool calls

Agent two: reads the result and decides on quality vs a re-drive if it’s low quality

Agent three: decides how to present the information to the user, creates a collection of HTMX elements

HTMX hx-get is reliably, and beautifully rendering the result of the Agentic Workflow without any react, etc.

Very happy and passing quality gates. I love not having security alerts every week to patch because of some buried react dependency library

I guess we just call them turn servers but technically it’s a “Stateful Dialogue Manager”

This process can be handled by a “turn server”

Was a pain to set up, but you can score the context completion and then if the score is under 98% or something, “ask” clarifying questions of the requesting agent or person or system