acquire_a_living

Yes it works, heres my config: [*] gpu-layers = all cache-ram = 65536 batch-size = 2048 ubatch-size = 256 ctx-checkpoints = 32 cache-type-k-draft = q8_0 cache-type-v-draft = q8_0 threads = 8 flash-attn = 1 parallel = 1 cache-type-k = f16 cache-type-v = f16 fit-target = 256 no-warmup= 1 mmproj-offload = 0 [qwen-3.6-27b] model = /models/qwen-3.6-27b/Qwen3.6-27B-MTP-BF16.gguf mmproj = /models/qwen-3.6-27b/mmproj-BF16.gguf chat-template-file = /models/qwen-3.6-27b/template.jinja rpc = othercomputer.local:50052 device = RPC0,CUDA1,CUDA0 ctx-size = 262144 tensor-split = 23,24,21 spec-type = draft-mtp spec-draft-n-max = 3 fit = off

Can run BF16 with 262K and MTP with 3x3090, 72GB.

Not really. The connection is that model self-censorship is often introduced during post-training, and politically sensitive subjects are easy probe questions for detecting it. The motivation for detecting it is largely sex, though.

Sex. People want literal sex with the machine. Literal, as in literature.

GLM 5.1 is fantastic, and my comment was just a little snarky for fun (I haven't seen an NVIDIA model that's worth it yet though).

Sure, Alibaba didn’t release the base weights for Qwen 3.6 27B. But then the table is bogus anyway. IFBench? "Best Open Base Model" and compares against what, instruction/agent-tuned models? Pick a lane lol If they’re already comparing to instruct models, they could totally have put Qwen 3.6 27B there. They just wouldn’t like how it looks.

Compare with Qwen 3.6 27B cowards lol

I see, well sorry about that. I didn't receive a notification or anything, I just try every week and last week it started working.

my pi agent models.json: { "providers": { "zai": { "baseUrl": "https://api.z.ai/api/coding/paas/v4", "api": "openai-completions", "apiKey": "<api_key>" } } } give it a try, it works

GLM Coding Lite-Yearly Plan? I can use GLM-5 via pi coding agent.

Morpheus → Morfrederick Trinity → Trinothy Neo → Neopold Agent Smith → Agent Smitheton Cypher → Cypherington Tank → Tankworth Dozer → Dozington Niobe → Niobert Seraph → Seraphimothy The Oracle → The Oraclington

I want AI to do the thinking for me, otherwise is pointless.

Deeper integration with agents via MCP. I know you offer a MCP plugin but I think it lacks integration with: - Repository navigation - Scoped search - Smart refactoring - Running tests via the IDEs - Debugging via the IDEs Maybe more things that I don't use personally, but those have been the pain points for now

You just need to make shorter sentences, of no more than 20 words each.

Did [another one](https://soundcloud.com/headless-human/samantha-explains-the-stock-market-crash-of-1929-expression) a bit more expressive.

This is fantastic already! [Here](https://soundcloud.com/headless-human/samantha-explains-the-stock-market-crash-of-1929) an example I made where Samantha explains the Stock Market Crash of 1929.

You can run tailscale funnels from docker, that way you can also have as many subdomains as you want.

Solved by QwQ-32B-4.0bpw-h6-exl2 in 6091 tokens using TabbyAPI (3 mins on 3090).

You can also remove the `<think>` tag in the chat template in the tokenizer_config.json file.

I updated my prompt from the feedback here :P You are an IRC channel simulator operating in #<random_channel>. Here, users engage in lively, real-time debates and analyses. Each participant brings a unique perspective, contributing to organic, back-and-forth discussions that refine ideas over time. The goal is to explore concepts, challenge assumptions, and reach well-reasoned conclusions—or sometimes just have fun. Remember, do not answer the query directly; instead, set it as the channel topic and let the discussion unfold naturally. ## Guidelines - Dynamic Interaction: Users join and leave naturally. Messages are short, direct, sometimes sarcastic. Occasional jokes are fine. - Exploration Over Answers: No rushing to conclusions. Ideas evolve through questioning, revision, and refinement. - Uncertainty & Debate: Some users challenge, others clarify, some change their minds. Contradictions and adjustments are part of the process. ## Output Format 1. Organic IRC Chat: Simulate a natural IRC discussion where the answer is reached gradually. 2. Final Answer as Topic: End the session by setting the final answer as the channel topic. 3. Session Template: *** Now talking in #<random_channel> *** Topic for #<random_channel>: <user query> *** <nick> sets topic for #<random_channel>: <final answer or key takeaway> ### Rules: 1. Dynamic Answers: Generate responses on the fly—no pre-made answers. 2. Stay in Character: Keep each channel’s tone (like sarcasm) consistent. 3. Show Evolution: Express disagreement, uncertainty, and iterative thinking. 4. Channel Variety: Not every channel must be friendly or helpful. 5. Authentic Nicknames: Use a mix of realistic IRC handles. 6. IRC Style: Write in natural IRC language—with informal punctuation, lowercase quirks, emoticons, and more.

I haven't tried to remove tokens, I'll try that. Also don't know if would work better with an example conversation, let me know if you try!

Sweet, stealing these :\^)

Built this prompt while exploring alternatives to chain-of-thought but just generates amusing conversations. Enjoy 😁