uaruss

@elonmusk Looks like Starship landing on the right!

The human-perceived RGB is image 1 and the Tesla AI photon count reconstruction is image 2. This is why Tesla FSD can see so well at night or through extreme glare.

@FluentInQuality Who wouldn't want to work for this man.

Jamie Dimon, CEO of $JPM, has fired people for running good meetings. Not bad meetings. Good ones. His definition of a bad meeting: it ends with "great session, let's pick this up next week." No owner. No deadline. No accountability. The people who run those meetings, he calls them "good bureaucrats." They love the process. They angle for the next job. They pound their chest. They grade themselves on activity, not outcomes. "If you don't get rid of them, no one believes you." He's applied this at $JPM, the most profitable bank in history, for 20 years. Bureaucracy doesn't require size. He said it can hit a single branch. "It's always the manager, stupid." The full clip is worth 3 minutes of any leader's time.

@TravisMcDonald @0xSero Any chance you rna the 3.5 version of that previously? If so, how do thy compare?

@0xSero Completely agree. I'm running the q5 version of unsloth/Qwen3.6-35B-A3B-GGUF on a rtx 5090 and it's freaking amazing. This thing cooks. You going to do a quant on it?

New best local model for y'all 16GB-64GB rejoice, the chosen one has arrived. huggingface.co/Qwen/Qwen3.6-2…

@TrungTPhan This is the key ->"our feeling is that the day that somebody working in Apple decides that they can't make a difference anymore, is the day we've lost"

Worth noting Steve Jobs actually studied long-lasting organizations to manage Apple, including the Catholic Church. “We observed the oldest and largest organization in the world has only four layers of management.  That's the Catholic Church. And, uh,  five if you count the highest order I suppose. So, we see no reason why we need over four layers of management. Indeed, we have usually about three. That’s the president. Maybe, we have a division manager and then maybe under that a marketing or engineering manager. And that's really about it. So, that's what we're trying to do. […] We hire people to tell us what to do. [They] come back and tell us how much it's going to cost and go do it. We've got an incredible group of entrepreneurs…[these] people are very independent thinkers and what they really want is…the environment where they don't have to convince 30 other people that it's the right thing to do. […] We really make an attempt to do that and I guess our feeling is that the day that somebody working in Apple decides that they can't make a difference anymore, is the day we've lost.” *** He was 28 for this chat….full 55-min vid for Aspen Design Conference Talk from 1983: youtu.be/t9HmOz8H0qI?si…

@sudoingX Awesome! I agree w/ you on opening for AMD. The hardware in the R9700 is a steal, just need a few more months for the software to really crank!

@sudoingX @xai Anyone had success with Grok 4.1 fast reasoning and Hermes? It was pretty much a disaster when I tried to run it recently (significantly worse that local qwen models w/ hermes)

should @xai help official grok oauth for hermes agent instead of us reverse engineering the web UI? one decision from them and every X Premium+ sub gets a production ready agent. builders and xai eyes both.

@freeroll88 Miss your posts. Hope things are well!

even with how many trades i have under my belt i think i am getting better i am really happy with my trading

@sudoingX @ItsTomWoolley Are you manually switching models or soemthing else? I'm using one of the Carnice qwen models (64 GB VRAM), but still too often default to basic Claude code or Claude, and ignore hermes, for critical work.

tom this is exactly how i run it too. opus as the brain for precision instructions and quality control, 27b dense handles the heavy crank locally, opus reviews and refines without burning credits on every tiny task. orchestration is the next layer of leverage once you realize no single model has to do everything

people ask me this all the time so let me say it out loud. the best model to pair with hermes agent is claude opus 4.7, by far, in order of magnitude. the rest is not even close. pair opus 4.7 with hermes and you feel the unlock immediately. the self improving loop, the per model parsers, the tool call handling, every part of hermes agent amplifies when the model on the other side is that capable. hermes agent is one of the most well designed and rapidly contributed open source projects in the space right now. the combination with opus is the current ceiling for agentic work.

@freeroll88 that had to be about as textbook as they get!

#NQ_F double gal. short 168.5. TP1. 128.25. 4/6

#AMD R9700 cheats: (1) Fan issue: See RDNA3 gpu_od/fan_ctrl sysfs interface (2) Power limits: Undervolt by 0.75mV and set power limit to 315W (3) For MoE on multiple cards, do this: github.com/ggml-org/llama…

@sudoingX I just got Carnice-MoE-35B-A3B-Q6_K running on my dual AMD R9700s, llama.cpp, ubuntu. So far, great experience with Hermes! The dense model would be interesting!

the 5090 is online. first benchmark dropping soon. what should i test first?

Thanks! I'm running this on dual AMD R9700 GPUs on llama.cpp with the config below (if it's helpful for others). So far, seems much better than than non-carnice version!) -m /home/***/models/Carnice-MoE-35B-A3B-Q6_K.gguf \ -ngl 999 \ --tensor-split 1,1 \ -fa on \ --cache-type-k q8_0 \ --cache-type-v q8_0 \ -c 131072 \ --port 8081 \ --host 0.0.0.0 \ -np 1 \ -t 8 \ --temp 0.6 \ --top-p 0.95 \ --top-k 20 \ --min-p 0.0 \ --presence-penalty 1.5 \ --repeat-penalty 1.0 \ --jinja \ --chat-template-kwargs '{"enable_thinking": false}' \ --reasoning-format deepseek

i've got inspired by @kaiostephens carnice and i've just released the 35b a3b version. i've been sick since the past few days with severe fever but i am still cooking! after ~2 days finetuning, here is both the sft and gguf version, available in bf16, q8, 6, 5 and 4. goal is to be able to enjoy that great idea of his on a unified hardware through a mixture of experts! SFT : huggingface.co/samuelcardillo… GGUF : huggingface.co/samuelcardillo… for people who are unaware of what's Carnice, it is a qwen3.5 finetune with datasets that are mainly designed to specialize the model in hermes agent usage by feeding it example of tools calls etc... feel free to benchmark and get some feedbacks!

@gophillionaire @Teknium @Google @openclaw @NousResearch Ha! I've got it running on dual AMD R9700s on llama.cpp. Had to tweak a llama.cpp bug (submitted a PR), but it's been stable for ~2 weeks since.

@uaruss @Teknium @Google @openclaw @NousResearch Haven't tried yet!

Has anyone tried @google Gemma 4 with their @openclaw / @NousResearch Hermes? Still looking for a good Claude alternative as the main orchestrator.

@Teknium @gophillionaire @Google @openclaw @NousResearch Why not an MoE model (e.g. Qwen3.5-35B-A3B)?

@gophillionaire @Google @openclaw @NousResearch I think everyone I've talked to has liked qwen 27B more than gemma 4 for agentic use

@freeroll88 You absolutely crush it with futs. Yet you like trading stocks better? Inquiring minds are curious how this can be!

i like trading stocks better

@altryne Buy gpus

My feed is showing me a bunch of folks who tapped out their whole usage limits on Mon/Tue. Is this your experience? Please comment, I want to understand how widespread this is

@ThatWolfieGuy @vagab0ndsec @sudoingX Fixed it and submitted a PR. Octopus Invaders game ~46 t/s on dual R9700 GPUs.

@ThatWolfieGuy @vagab0ndsec @sudoingX You're running llama.cpp or something else? I'm hitting this issue w/ dual R9700s on llama.cpp: github.com/ggml-org/llama…

local AI hardware tiers: $4,699 - DGX Spark (NVIDIA wants you here) $1,989 - RTX 4090 (overkill for most) $1000 - RTX 3090 used (sweet spot) $250 - RTX 3060 used (currently testing every model that fits 12GB) $0 - CPU only (it still works) jensen announced the top. i've been posting receipts from the bottom.

Guaranteed? Yikes.

@NousResearch Incredible, thanks for sharing!

Thank you to all 187 submitters. We loved seeing the range of domains and all of the uses for Hermes Agent you found that we could never have designed around. We're looking at ways to feature more of the standout entries, as well as doing themed hackathons in the future - watch this space. In the meantime, all winners will be contacted via Discord for prize distribution. Finally, if you’re new here and all of these projects have you feeling inspired to build something of your own, try out Hermes Agent! nousresearch.com/hermes-agent

The Hermes Agent Hackathon has ended! Finalists were selected by Nous staff out of 187 submissions on creativity, usefulness, and presentation. We were blown away by the quality and range of what people built using Hermes - thank you to all who participated. Winners below ↓