Maciej Rutkowski

Using Qwen 3.6 27B I wanted to reverse engineer the app that resets the waste counter in Epson printers which makes them otherwise unusable unless you pay up to change a variable in its config. It however refused due to DMCA and other consumer-hostile laws. Changed to an uncensored model and...

Holy hell, I was not expecting to rewatch Even Horizon in form of a Strange New Worlds episode. 👻

iPod Nanos are... a challenge to repair. I'm equally fascinated and apalled at how those things are put together.

Some exciting new stuff in the PS5 jailbreak scene happened in the recent few weeks. Trying the new exploit with my 12.70 console, the console crashes on game exit (using a raunchy japanese Lua-based game); I thought, why don't I ask an seemingly all-knowing coding entity about the possible causes.

Claude was *THIS CLOSE* to calling the police on me - refused to talk about the topic completely. Ok then, I asked my own agent harness running on Qwen 27B and got a complete explanation how the exploit works, what the possible causes may be, and a fixed LUA source in the workspace to try out.

I'm a dimwit. I've been running this power-hungry moster of a GPU, RTX3090, on a single PCI-E cable essentially - simply connected the two plugs from the same cable to GPU's sockets, instead of providing it with power from two separate rails. Symptoms? Voltage drops, crashes, oxidated pins.

So, an update on self-hosted Qwen 3.6 crashing issues - looks like it was all LM Studio's fault all along - now running on llama-server again, with a a 150k context, 4096 ubatch, 2048 batch, KV cache at Q4_0, and only a few times llama-server simply shut down without any messages

Has an awesome sunday morning session with Claude before family woke up. Updated my motorized camera slider's stale firmware to make more room for new features, by changing partitioning, replacing some heavy C++ stream code with snprintf calls - gained >600KB free space, wooo.

The physical slider is the next thing that needs "refactoring". I designed it in Blender almost 5 years ago. It handles big cameras, uses an aluminium extrusion as the base with the platform moving on pom wheels, so it's more of a prototype, but works great.

Has an awesome sunday morning session with Claude before family woke up. Updated my motorized camera slider's stale firmware to make more room for new features, by changing partitioning, replacing some heavy C++ stream code with snprintf calls - gained >600KB free space, wooo.

Before that the iOS companion app also got vastly improved with new UI and refactored the code for better maintainability. So the stop motion feature has been planned out, was ready to start implementing, buuut I hit the daily ration of tokens - still ridiculous and distopian.

Has an awesome sunday morning session with Claude before family woke up. Updated my motorized camera slider's stale firmware to make more room for new features, by changing partitioning, replacing some heavy C++ stream code with snprintf calls - gained >600KB free space, wooo.

From some initial testing with coding, Qwen 3.6 27B slaps. Haven’t seen any reasoning loops, so far no failed tool calls. Really smooth. At least for Golang and Swift projects. Much better experience than the predecessor. The model used: Q4_K_M, KV cache q4_0. And 4gb VRAM left. 🤘

Just as I said some nice things about it, entire morning spent on autopilot retry failures, incorrect tool calls in thinking blocks, my word. Gotta get to the bottom of this. Sometimes the model (or llama.cpp) just crashes, and prefill taking ages will cause timeouts too.

From some initial testing with coding, Qwen 3.6 27B slaps. Haven’t seen any reasoning loops, so far no failed tool calls. Really smooth. At least for Golang and Swift projects. Much better experience than the predecessor. The model used: Q4_K_M, KV cache q4_0. And 4gb VRAM left. 🤘