All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
6 Months of running AI Sysadmin + Support Taught Things About Finetuning
PulsedMedia
Member, Patron Provider
Reaching 9.47/10 CSAT And 19 minute median first response time took a lot of effort, and that road taught something invaluable; Finetuning is a must have long term, lots of tokens and effort is wasted when the model is not aligned to our needs. Antipatterns are just someone elses patterns.
Wrote today a detailed blog post about Antipatterns and what they really are, just someone else's biases: https://pulsedmedia.com/blog/2026/07/ai-agent-antipatterns-what-9-47-10-csat-taught-us-about-finetuning/
Local continuously retrained, distilled, finetuned, however you want to call it will be highly important in future as soon as the hardware becomes practically available for small business and even individuals, by the looks of it, at least 768GB of VRAM is required.
No AI was used for this post, i wrote manually every single word on this post. The only tool use was a browser and single copy & paste to get the URL linked. Every word was an individual human keypress. No AI rules was broken here, even those which are not mentioned. The OG blog post is also 100% human written.
Comments
Not this LLM sloper :-/
@PulsedMedia there's offer from @LukeRhino $68 (8+piece) for 16GB HBM2. with 300W TDP
8*16 ... Still vastly short from our needs.
Thanks for the offer tho.
The opportunity cost is too much, but worth it to at least investigate.
I only tweet ChatGPT slop
The 8+ is when the discounted price kicks in. They have 1700 of them
beyond 8 it gets complicated to cluster.
Even with NVLink?
how do you NVLink beyond the single chassis?
You have to network these, that's 400G switching
nvlink worth more than those cards
Eternal Väinämöinen 4 president!
@PulsedMedia, has the budget been exceeded when using Väinämöinen?
experiment worth following
i can bet PM owner is billionaire and not making any money by running his business
just for fun.
🤯
I think @PulsedMedia said something in Discord about Väinämöinen going in a loop burning through tokens, so support responses will be slow for a few days.
Maybe @PulsedMedia has to add coworker who examine the progress and said "Stop the task, wait for human input"
kernel of truth, i do this out of curiosity more these days than anything else. Tho profit is also welcome
Yes, a bug in token caching caused each ticket to burn 2x as much, then another bug (and antipattern) caused essentially unlimited reruns of ticket processing. Now harsher limits for a moment, and then lifting the limits again.
Expect 15-90mins of delay in processing for a week or so.
Part of the antipattern is the common work deferrence, it keeps consider human time is infinite and 100% free, i keep plugging the roads how it can escalate to human or leave the work undone, but it takes time. This is also a case for finetuning, the LLM is heavily biased towards not completing work.
But it will be years before we can have truly finetuned model, it's expensive. And it would evolve constantly too.
GYAHAHAA. The LLM must be mirroring You from the docs
@PulsedMedia Instead Väinämöinen focussing on anti patterns let it use best practices in your preffered language php. don't learn it by explaining what it shouldnt do. The LLM doesnt know how to unwatch the antipattern elephant in the room. Almost like human cannot unsee things.
Yes because positive reinforcement is the only thing that works.
No, you have to guard against the antipatterns, if you don't look for them you cannot guard against them.
Okay you have more experience steering ai. I am just last months busy with agentic workflows.