appv⦿id

@deepcandles @trekedge @NousResearch i think he meant training sota models

@trekedge @NousResearch do open weights? no need to pay compute bills

The best part of working at OpenAI is that our mission is literal. We want everyone to have access to superintelligence. No hiding our best model for only powerful companies. You get the power.

We are the universe desperately, beautifully, trying to see its own face. And every tool we pick up to do it, is made of the very thing we're trying to look at.

Those who own Energy, AI and Robotics will rule the world. Money won't save the ones who don't.

@GoingParabolic That's a forth dimensional image. Think about it as an 3D object which timeline is collapsing onto itself.

This image is destroying my brain.

@TrelisResearch more like wtwb (watch the world burn)

How to Claude Code Yolo --- 1. Rent a ~$4/month server from Hertzner 2. Install Tailscale on the VPS and on your Mac (you'll need to expose a UDP port on the VPS for Hertzner) [optional, or just ssh into the VPS] 3. Log in via ssh through tail scale, and set up a non-root user (Anthropic blocks --dangerously-skip-permissions on root) 4. Install a safety hook to avoid dangerous commands - npx claude-code-templates@latest --hook=security/dangerous-command-blocker --yes 5. Github setup: Create a fine-grained PAT with read/write on contents and pull requests for specific repos only! Then protect your main branch with a required PR rule. This way Claude can push to any feature branch freely, but can't land directly on main. Then - to do data prep and train or evaluate voice models - asr or tts... 6. Set your HF token on studio . trelis . com and get a spend-limited API key to put into environment variables on the VPS. Claude Code can then call the Studio API to trigger training runs, pull datasets, and run evals — all autonomously. 7. Optionally, add a Trelis Router (router . trelis . com) api key to your Trelis Studio account so you can evaluate proprietary models. Finally - in the server - you can run: alias yolo='claude --dangerously-skip-permissions' then tmux new -s yolo yolo and you'll start off a claude session where it will run without permissions just close the terminal to leave it run, or press Ctrl+b and then d to detach from tmux. 'tmux attach -t yolo' to reconnect later. --- for more, check "studio . trelis . com/yolo"

@cosyposter I'm pretty sure that in our persue of exotic flavors we spread sneaky cancer from time to time without realizing it

How long do you really want to live, though?

@Teknium 0.8b model that does as good as the big ones

What would be your most desired new capability for hermes agent that would unlock the most for you right now?

The careers of the future are strongly related to AI, MR and Robotics

@Kekius_Sage As far as I know the path gets stretched which makes light take more time to reach the other side

If light particles have no mass why are they attracted by the force of a black hole?

@COFFINHEAD1 @yacineMTB The only rich people will be the ones with robots powered by AGI and beyond

@yacineMTB The duality of man

I think it's genuinely over for all of us

@simplifyinAI it doesnt work good, that technique its too exp

Someone open-sourced a tool that REMOVES LLM CENSORSHIP in 45 minutes 🤯 It’s called Heretic. Instead of fighting with complex prompts to bypass safety filters, you run one single command and it permanently deletes the model's ability to refuse a prompt. • Fully automatic (Zero config required) • Preserves the model's raw intelligence • Works on Llama, Qwen, Gemma, and dozens of others • Runs locally on consumer hardware 100% Open Source.

@burkov To be fair, qwen 2.5 0.5b was a really good model even compared to some 1b models, but yeah, qwen3 and 3.5 small models were mostly benchmaxxed

I don't trust Qwen models, even the large ones. I trust small ones even less. Qwen has been benchmaxing its models since the very beginning, so the benchmark performance of a 4B parameter model beating Gemini 2.5 Flash Lite is most likely just the benchmark performance only. 4B is just not enough to store any general intelligence because most of the parameters will still be used to store facts. Well, not facts, but something that remains of them.

Even a 1 trillion parameters language model could fail at the example above while a proper symbolic one won't

As opposed to symbolic models (and humans!) language models might split these sequences incorrectly, and even in the best cases the still hallucinate due to language poisoning (statistical knowledge instead of structured knowledge): ●● ●● ●● ●● ●● ●●

Language models usually get to an erroneous conclusion given some statistics. Symbolic models will mostly never hallucinate because they can't! There's no ambiguous solutions, the only way to lower their losses is by literally learning the rules!

@paneerchilli65 Looks so cool, I wonder if a planning mechanism (looking ahead in the future) could help

run 2.2-> diffusion steps T = 30, 10k steps, 500 tiny stories, gen_len = 256

@MiraRuido Why or and not and?

Is the bird pulling the tree, or is the tree holding the bird back?

Ok, this model is cool. I can see how it could work with robots. But I still need some large model to translate the reasoning.

@BasedDaedalus everyone pulling the jim carrey trick now

they switched dario it’s over