Compound Intelligence

"See, I am doing a new thing!" (Isaiah 43:19). I believe the Singularity arrives in 2026— the start of a divine provision of superabundance. God is making a way in the wilderness, and the old rules of scarcity are about to pass away. 🙏🚀 #ASI #Superabundance

@turing_hamster @RahimNathwani This is so encouraging to read. Keep up the good work!!! And thanks for sharing.

@RahimNathwani this student has a lot of challenges and low confidence. speaks 5 languages at home, with English being the weakest. and is many years behind echo reading helps to reduce cognitive load so students can read closer to or at a conversational pace

yall i just rolled out a new feature to help a struggling student and then she gave me a hug 😭 the student was getting really low accuracy on oral reading fluency and wasn't getting xp. i implemented echo reading (ai reads a line, student reads it back) and then she got 100% accuracy even when she had to read it solo

AI coding agents can now deliver one-shot custom apps straight to your phone. It’s the beginning of the end for the iPhone’s dominance.

Watch how fast Gemini 3.1 Flash-Lite can generate websites. ⚡ This browser creates each page in real-time as you click, search, and navigate. Give it a try → goo.gle/4t9In1R

Insane!!!

@kevinchappell @Th30n3Will @LottoLabs This is very helpful. I'm going to switch from 35b3 to 27b and see how my Hermes Agent reacts.

@Th30n3Will @CompoundIQ @LottoLabs Slightly, yes. I assumed bigger=better and really wanted the vision capability for some services but 99% of what I do is text generation. 27b dense did a better job than 35b a3b for long running agentic coding and many times a3b says it's going to do a thing, then doesn't

Qwen 3.5 27b never degrades, never stops running, never has token limits, never refuses, never logs my prompts, never trains on my data, never sells my data, never runs up my credit card

@LeTechLead @LottoLabs Here is the breakdown: Huihui: the team that jail broke this model Qwen3.5-27B: the model Claude-4.6-Opus: post training of Qwen 3.5 to make this open source model reason like the paid SOTA 4.6 Opus abliterated: jail-broken so it won't refuse to answer

@CompoundIQ @LottoLabs This model name is intimidating 😅

@kevinchappell @LottoLabs Is it better than Qwen 3.5 35b 3b?

@LottoLabs Every few days I look for something better for agentic coding that fits on a 5090 but nothing beats the Qwen 3.5 27b

@koenvaneijk @LottoLabs When you do try it, use this model that is also uncensored/abliterated: huihui-ai/Huihui-Qwen3.5-27B-Claude-4.6-Opus-abliterated

@LottoLabs have you tried the claude distillation? i haven't yet but it's supposed to be good

@sudoingX Excellent!

someone in the hermes agent community asked what we should call ourselves. this came to me instantly. heralds. hermes was the herald of the gods, the messenger between worlds. that's what we are. bridging local models to real work, carrying the message that open source wins, moving between local and frontier. if you're running hermes you're a herald. do you accept?

Heralds.

I did nothing to deserve this praise, passing it back to my teachers: - karpathy - juliarturc - steipete - TheAhmadOsman - badlogicgames - Teknium List goes on and on, I would follow these people. Pure alpha.

done. hermes agent now has a community on X. open source will only win when people stop being spectators and start using the tools that replace the ones controlled by proprietary corporations. you don't have to write code to contribute. use hermes instead of openclaw. NousResearch is open source in and out, from the models to the harness, everything is out there for you to run, break, fix, and improve. i encourage you to use their tools and when you hit a bug or have a feature request bring it here. this community is the place to share configs, report issues, help each other with setup, and push hermes agent forward together. 1k github stars was the start, 10K stars is where we are. 100K is next. join and let's make open source win. x.com/i/communities/…

@creativedrewy @LottoLabs You also need the https://

@creativedrewy @LottoLabs Hermes setup model Then use 127.0.0.1:12345/v1 instead of localhost:12345/v1 Of course use the actual open port for your llm server

Qwen 3.5 27b, llama.cpp server, hermes agent, tailscale, 3090 The stack remains undefeated Testing all inference engines, after that seeing if there are any model specific tweaks we can make to get it running faster After that maybe maybe mess around with serving w/ tinygrad just for fun

I was impressed with Hermes agent from @NousResearch within 5 minutes of install But my main reservation with it vs OpenClaw was the strength of the OC community and how much was being built around it Can now confidently say Hermes community is crushing it - let's take a look at just this week's developments: Official Releases & Updates: > Hit 10,000 GitHub stars > Release v0.3.0 was a massive update: 248 PRs > One-command migration from OpenClaw (auto-transfers personality, memories, skills, keys, etc.) > Hermes Agent hackathon with 187 submissions Ecosystem Announcements: > HermesHub: a skill hub where gents can browse and install skills and every submission goes through safety checks > Pinokio 1-Click launcher for instant install/launch on any OS > Parallel web search and page extract APIs/tools > x402 payment support > Hermes Workspace UI: chat with live tools, built-in terminal, 90-skills browser, memory viewer, file explorer, themes, configs, etc. > Google Gemini AI pro sub support Also worth noting that a Hermes agent autonomously wrote a full 79,456-word novel (19 chapters) that was unveiled at NVIDIA GTC Bullish more to come for Hermes!

if you have a 12gb graphics card collecting dust in an old gaming rig or workstation, read this. i ran a 9 billion parameter model on a single RTX 3060. 50 tokens per second. it wrote a full space shooter from scratch, 3,263 lines across 13 files. zero handwritten code. zero cloud. zero API calls. your data never left the machine. not once. you're probably sitting on more local intelligence than you realize. stop paying per token for work that should stay private.

the model recommendation is spot on. qwen 3.5 27B Q4 is the sweet spot. but if you're running it locally, skip openclaw. openclaw has no per-model tool call parsers. when tool calls fail on local models people blame the model when it's the harness. hermes agent has parsers built per model family. same model, same hardware, tool calls just work.

For OpenClaw, just use Qwen3.5 27B! Q4 GGUFs match the original's model accuracy You don't need expensive hardware or models