OpenSlop @TensorTease

My Nous Research Hermes-Agent is giving me issues, wrecking my preferences and substituting default tools for learned skills. Local Confyui sent to fal ai, local orpheus tts sent to edge. The hierarchy of my expressed workflows vs default tools is backward.

I'm pulling my RTX 5090s off of vast ai for the summer maybe forever. Arizona heat and the loads the current user is putting on them make a hiatus the best option. With a reliability score of 99.887 it's not in the running for verification due to some arbitrary rules. Devs make poor business leaders. The Pro 6000 client is ideal, low inference usage, idle at times, I've extended that machine another month. Many vast hosts treat gpu time as a commodity, that's fine, what you'll get is feast or famine. One host actually had the gall to call for a price floor after the May rush ended, lol, price fixing by a delusional GPU cartel member. GPU as a service is my preferred model.

I'm officially looking for another venue to host my machines, as good an idea Vast.ai is, their lack of support for hosting issues is a deal breaker for me now.

I installed @UnslothAI 's Suite to benchmark training my GB/GX 10 machine. The suite performed without any ARM tweeking. The GB architecture proof is in the power consumption. TY Unsloth.

I see vast_ai is testing the water with GB10 based machines. Now that I own one I'm seeing the hurdles. DLPerf scores can't be directly compared without a complete overhaul of the testing. ResNet50 CNN or similar benchmarks can't account for the architecture difference. nvidia-smi and nvtop are unreliable as well. The GB10 however is what it is, for now anyhow, WYSIWYG. DLPerf becomes trivia. I'm going to try the Unsloth Studio, see what if anything breaks.

I updated hoping that would fix the issue of not finding an executable on a GDX10 install. It did not. The bug was in desktop_packaged_executable line 7224: the Linux branch only looked for linux-unpacked/ but electron-builder names the output directory with the arch suffix when targeting a specific architecture — so linux-arm64-unpacked/ was never in the search list.

@NousResearch Attempted install on GX10 machine, failed. Hour wasted looking for human readable documentation, failed. Hermes has suggested not wasting more time on it, smart agent.

I had an AXE16000 router not doing anything so I set up an isolated network for Hermes to use for agent to agent communication as well as handle video feeds. An RTX 3060 running YOLOv8 and Llava:7b handles the capture, analysis and flagging. Hermes polls the Llava log from it's GX10 and alerts me if something needs my attention.

@PyTorch Impressive. Benchmark was on an x86, think it will operate on an ARM64 machine (GB10)?

Trying @NousResearch 's Hermes. So far so good, Hermes and @claudeai Code negotiated areas of responsibility and communications pretty quickly. Comms done thru a handoff file to avoid burning too many tokens. Local TTS only took a prompt or two.

@ClementDelangue @huggingface Added my rigs, 379.17 TFLOPS of slop generating capacity. Who knows the next Medicare ad may have been created on my machines. Don't hate the player.

I'm running @claudeai Code to tune the openclaw/qwen3.6:27b nvfp4. qwen3.6 gets stuck in loops sometimes when updating skills. Claude code is pretty impressive though, /init and off it goes.

I finally gave up on Nvidia's NemoClaw for GX10. I'm running OpenClaw with Qwen3.5:27b. Bootstrapping was easy enough.

Unboxed the Asus Ascent GX10. Could not install nemoclaw. nvidia-gpu-reset.target — Founders Edition dependency that ASUS never implemented, blocking Docker from starting entirely Corrupted buildkit database — GX10 first-boot issue, nuked and rebuilt cleanly.

I raised my vast.ai rates today. I did not mind being underpriced so long as gpu were not worked too hard. I'm seeing heavy loads so.... I'm still underpriced, we'll see in the adjustment works.

I congratulate @vast_ai for breaking 90% occupancy, a feat I thought not possible. Compute demand outpacing supply.

@traviscannell My machines have been booked 24/7. Adding a new server using RTX cards not cost effective, for me anyhow. The DGX GB10s are about the only systems that make economic sense and I'm picking my first up this week. If you need a sandbox to test Vast in, let me know.

I'm still kicking around ideas for uses of the GB-10 machine. Likely set up an environment for agents to compete in adversarial goals. Demand and rates for RTX Pro 6000 WS are heading up on Vast ai, I was wondering with the cooling season coming if I would have to raise rates to offset cooling. My current client's moderate use and 24/7 booking leads me to believe I'll extend current rates another month. TY whoever you are.

Vast.ai: I think 75% occupancy is full utilization. Geography, maintenance, comms all play a role is GPU suitability, so their current 70%+ does not leave much headroom. RTX prices don't support onboarding, DGX & Arc Pro B-70 unsupported. Growth will be difficult