Vibe-SysAdmining: LAUGH OF THE DAY
In which a frontier LLM model tells me to configure a new chatbot to use model “/anthropic/claude-sonint-4-6”; yes, this only means that somewhere in the training data some poor SOB mistyped “sonint” for “sonnet”; but nothing with even sparks of anything that could be called “agi” would let this typo through at a temperature=0.1, would it?
I am not a good-enough programmer to be successfully vibe-coding anything, save for scraping, indexing, summarizing, and archiving scripts. What I may be good enough at is to successfully vibe-sysadmin the under-the-dining-room-side-table machine. Why? Because the machine is free, and it looks to me as if it would payback in a year given likely datcenter costs, even if the machine were new and costly. Save some money. See whether any belief that companies will make fortunes off of selling the services of their data centers is vain, whether it is true that the obvious path for users will be to light up their own dark silicon with open-source models and the time of their existing programmers and sysadmins.
So far, it is going well, I guess?:
Three models are currently loaded and ready:
alibaba/ollama/qwen3:32b-q4-128k: the 28GB Qwen3-32B model quantized to
q4with a maximum context length of 128k tokens, and with “reasoning” turned on.google/
gemma4:26b-a4b-it-q4_K_M:the 23GB Gemma4-26B model, quantized toq4_K_Mwith a maximum context length of 256k tokens.alibaba/ollama/qwen3:8b: the 8.8 GB Qwen3-8B model, also quantized to
q4_K_Mwith a maximum context length of 40k tokens.
The hope is that:
the third of these will be the light, quick one calling tools and running scraping, archiving, indexing, summarizing, simulating, and calculating python tasks;
the second will be the workhorse; and
the first will do heavier lifting on unstructured databases—principally substantial text documents and collections of text documents, plus its own code and installation.
Do I understand what the words describing them mean? Not really. But I bet I do understand them better than Sam Altman or Greg Brockman does.
And the hope is that this thing will soon be able to effectively SysAdmin itself. God knows I do not have the time or the expertise to do that. I can say that it does have a much better idea of what unix and other configuration switches are and do than I do, but that is not saying very much. And it is, at bottom, a stochastic parrot.
And now I get to the punchline here: How much confidence, after all, should I have in a machine that—using a frontier LLM model—tells me to configure a new chatbot by typing:
/model claude-sonint-4-6 –provider anthropic
Yes. This only means that some poor SOB somewhere in the training data mistyped “sonnet”, but that this error makes its way to me with temperature=0.1—that is truly scary.