bitslix

@witcheer Yep, we only have a simple 8GB 4060 RTX and also try a little to play with it, but it feels so painful when you don't have the correct hardware or command lines 😂

@bitslix what do you mean? how to set it up?

Qwen3 8B on my RTX 4060 Ti 8GB. 8.2B params, all active. dense transformer with built-in thinking mode. 4.7 GB on disk (Q4_K_M). >50.8 tok/s at baseline. degrades to 36.6 at 24K (-27.9%). prompt eval is fast: 2,944 tok/s at 8K. the model is compute-bound, not memory-bound. VRAM is tight: 4.7 GB model + 2.4 GB KV cache at 32K context = 7.8 GB used. fits in 8GB but barely. reduce context to 16K for safety. ran 6 quality tests. 4 pass, 1 partial, 1 fail. this is the first model in my test set to cleanly pass the hallucination resistance test. however, found it very under in code generation. the thinking mode consumed all 2,048 tokens debating memoization approaches without producing the function. same pattern as GLM 4.7 Flash. thinking-mode models need higher max_tokens or thinking disabled for code tasks. the thinking mode is visible in the token counts: JSON test used 336 words of reasoning for 20 words of content. logic puzzle used 390 words of reasoning for 228 words of content. every response carries this overhead.

Soon, new offers from @Hetzner_Online The Auctions will be renamed to Ai Boxes :D

@TiinyAILab We’d be interested in reviewing the Tiiny AI Pocket for the German-speaking market, but unfortunately we haven’t received a response to our previous request yet. Could you point us to the right contact?

Tiiny x Obsidian: Raw Data In, Connected Knowledge Builds Up. With Tiiny AI Pocket, extend Obsidian with a local thinking layer that helps you curate and structure information into your second brain with a single command. Keep everything on-device and build your knowledge system more efficiently. #TiinyAI #AI #Obsidian #Claude #Qwen

@schematikio this is cool :)

Thoughts on our cyberdeck?

We got asked if we can #OpenSource the Clone. Should We? x.com/bitslix/status…

@tlakomy x.com/i/status/20557…

AI won’t replace software engineers, but if you refuse to adapt to the new reality and continue to pretend we live in 2021, you may be in for a rude awakening. Both things can be true.

@rioserenity_ Yes. Because thoughts are free, likes are optional.

would yall keep tweeting with 0 likes ???

@brockpierson Just a hint, there was a time before the public login for everyone, a time where you only could register when you are on a college.

Are you old enough to have logged in when it looked like this?

@DaddyPapaUwU @forallcurious Observation is interacting....

@forallcurious Nope. It's more that things are affected by interaction. Things exist in a superposition until they interact.

🚨: Quantum physics says that everything observed is affected by the observing watching it

@exQUIZitely I loved it 👍 was nice to watch

Are you pold enough to remember screensavers? If so... were you part of the SETI gang?

@theo Nothing new.... ceoinsightsindia.com/news/hcl-softw…

Title inflation has gotten insane. What do you mean "Head of OpenAI for Countries"

@adcock_brett @xlr8harder That’s correct. 💯 There are workers at companies like @geekompc who spend an 8-hour shift recording webcam footage, while another person checks the recorded video afterward. This can be done by robotics.

@xlr8harder There are so many humans doing this, it’s crazy tbh

it's funny all the people reading this and saying the work is useless. like, they don't realize we have real humans doing this job today. people have no idea how the things they order online arrive at their homes, do they?

Do you remember the 120GB Twitch source code dump? When people looked at parts of that codebase, many asked themselves how something like that could even happen. And that was before AI coding. So maybe “slop” is not an AI problem. Maybe bad architecture, rushed products, messy codebases, and questionable engineering decisions existed long before LLMs. AI just made it more visible, and faster.

SaaS (Slop as a Service)

Exploring the future. People would experiment, create, learn, build, and discover new things. Maybe humanity could finally work together on bigger goals, leaving Earth, exploring space, and solving problems that have been waiting for generations. Maybe the solution we have been looking for is already sitting inside the mind of some factory worker, warehouse worker, nurse, driver, mechanic, artist, or parent who simply never had the time to explore it. People would build. People would research. People would create art. People would raise families. People would form communities. People would explore nature, science, technology, and space. A world without forced work would not mean that people stop doing things. It could mean that people finally start doing the things they were always capable of.

what do you think people are gonna do all day when nobody has to work?

This will also matter in future hiring. At some point, the question won’t only be: “Can this developer write code?” It will also be: “Can this developer use tokens efficiently to solve problems, build products, debug systems, and ship faster?” Tokens are becoming part of the developer skillset. x.com/i/status/20557…

tokens are rapidly becoming the universal input for solving problems

Fully agree. This should be standardized around one shared `.agents` folder. If every tool invents its own location, we end up with the same mess we already had with config files, cache folders, package managers, and editor-specific metadata. Just use: .agents/ And if a tool needs its own structure, fine — create a subfolder: .agents/codemaker/ .agents/continue/ .agents/cortex/ .agents/ibm-bob/ That way projects have one obvious place for agent config, skills, workflows, prompts, memory, hooks, etc. The current approach is just fragmentation for no good reason.

This is the most insane shit I have ever seen. Look at all the possible skill folder installation locations Why does IBM bob have its own skill folder? What the hell is CodeArts Agent? Why can't they all use .agents??

@testingcatalog Cool, but searching is only half the story. The really interesting part would be posting, replying, publishing updates, and interacting from inside the agent workflow. Search gives the agent context. Posting gives the agent agency. That is where things become really useful.

HERMES 🔥: X Premium+ accounts can now use their subscription with Grok in Hermes Agent! Besides that, they can also use X search tool, which puts Hermes into a quite unique position now. It is not just OpenClaw vs Hermes anymore. It is OpenAI vs xAI once again 👀

Nice touch @cursor_ai 👌

@RoundtableSpace x.com/i/status/20557…

HERE’S A LIST OF EVERY MAJOR AI MODEL RELEASED THIS YEAR: GPT-5.4 GPT-5.5 Claude Opus 4.6 Claude Sonnet 4.6 Claude Opus 4.7 Gemini 3.1 Pro Gemma 4 Llama 4 Scout Llama 4 Maverick Qwen 3 Qwen 3.6 Max-Preview DeepSeek V4 DeepSeek V4-Pro DeepSeek V4-Flash DeepSeek V3.2 DeepSeek V3.2-Speciale GLM-5 GLM-5.1 Mistral Large 3 Mistral Medium 3.5 Mistral Small 4 Grok 4.20 Meta Muse Spark Kimi K2.6 MiniMax M2.5 NVIDIA Nemotron 3 Super NVIDIA Nemotron 3 Mini gpt-oss-120b gpt-oss-20b It’s been a crazy year.