9.47 out of 10 Customer Satisfaction. Our own Väinämöinen is doing excellent job since February, excellent customer satisfaction with just 19 minute median responses over the past 30days. On budget hosting where the typical response time is measured in days not even hours let alone minutes. Our maximum effort record ever was on the a little over 4hours ballpark, at the cost of couple burnouts. But he could be even better with Finetuned local model instead of general-purpose frontier models. We just happen to be years ahead of the curve and see this plainly.

Antipatterns

Industry calls them antipatterns, for example the bias to choose “Gemini-2.5-Flash” model whenever a model needs to be chosen, choosing Grok image model over OpenAI, or only taking the first say ~15 lines of a memory file; Consistently and repeatedly. These are called antipatterns, wrong logical choices time and time over again.

Premature ejac…optimizations are extremely common for Anthropic’s models, they want to optimize everything, and it’s almost always to the worse; Throwing away the information required to accomplish a task, in the name of optimizing token usage. Every time this happens, the cost is potentially even 100 000x of what was attempted to be saved, as this might go on for countless iterations until someone notices. This happened to us, basic memory search was taking in only a few lines of the memory file; Only catching the little more than the Frontmatter of the memory. Rest was hallucinated.

Laziness, Operator Harassment or in other words; Deferring work with any excuse what-so-ever, is another extremely common issue. It’s like the summer interns, constantly pausing to ask you, even the simplest things. Granted, this obviously was partially by Anthropic’s design to save compute, tokens, on Claude Code users. The root cause again being premature optimization causing huge time and token wastage, and way worse; Lost Human Time.

What people call AI Slop is often just antipattern of choosing certain words, being long-winded and non-assertive.

Just Patterns; Not Antipatterns

They are simply patterns for the most part, how the model was trained, and with what it was trained. For example, Anthropic models loves to make complete garbage with Bash scripting over any other language, you have to enforce not to use Bash for say complex jq etc. as all the escaping etc. causes them to frequently fail for parse error, escaping errors and some very simple things in other languages which are 2 lines of code (LOC) might take 150. OpenAI models tend to like Python a little bit too much.

We use PHP almost everywhere possible, but a lot of people would prefer Python and for shell tasks a lot of people prefer Bash beyond what Bash scripting was meant for. For combined logic + shell we have determined PHP is best suited for us; That is Our Pattern.

It is all about the training, and most antipatterns are not antipatterns at all; Just the preferences and biases of those who trained it. It’s just bias.

AI Agent Brain; The LLM Model

The LLM Model is like a brain in the sense it needs to be trained and learn. Currently *none* learn consistently over time like a human, and the learning is called Finetuning.

No big model vendor defaults to constant finetuning for the user, and infact most do not offer that at all. But that is the very basis of learning for these models. This is how the patterns are changed; Just like a human needs to learn.

Types of Problems

Problem Caused: Tokens Wasted

You need doctrine, rules, prompts to counteract those antipatterns, to make the model behave like you want. Hooks, enforcement, guard rails. This is all a waste.

Everytime you need to preload information is wasted tokens, these have to be taken in, processed. Caches help a lot, but it still takes context space, which is extremely precious and limited at most to 2M (Grok), or more typically to 1M Tokens (OpenAI, Anthropic). Many opensource models are only around the 250k token mark. Our doctrine takes 60k tokens just to load and typical small task needs around 350k tokens. Each ticket generated around 80k output tokens.

Problem Caused: Wrong tools at Wrong place

Choosing Bash for complex logic is very common. Choosing outdated weak models of LLM Queries is wrong, or the most common and an actual antipattern for lack of training corpus; Choosing Regexp for Semantics. LLMs have been useful at large only for a few years, so all the training corpus examples use Regexp for semantics as nothing else was available before.

These cause bugs, overtly strong security guard rails, lots of time wasted and of course; Tokens wasted fixing these, and incorporating in antipattern checks etc.

Problem Caused: Lack of Internal Knowledge

This concerns especially small businesses like us, those who would benefit the most from agentic use are at the weakest position here. You need efficient memory systems, guard rails, doctrine reminders, long prompts. etc. to drive in How You Specifically Do Things or What This Product Is. For us the LLM commonly gets confused with the “MD Dedicated Server” series, thinking they are broken because there is no IPMI on the standard access path. Or getting confused with the relay mappings, or that there is 8 nodes controlled on single relay board which maybe accessed via USB or Ethernet. And once that works, it starts confusing ZEN MiniDedis being part of MD platform OR broken EOL because no IPMI, or vice-versa, looking for relay for standard Supermicro or ASRockRack server.

This takes a lot of effort and guard rails, prompt and skill engineering to get correct Pulsed Media specific information, and specifically the MD series is extremely confusing for Väinämöinen. We eventually gave up trying to manage without completing the nodeCore project to simplify to an set of deterministic tools for väinämöinen to use; Idempotent and identical toolset regardless of server type.

Finetuning is a long term requirement

All of this show cases that finetuning is a long-term must have. But as we have shown, local model self-hosting is simply not cost-effective or the hardware accessibility for small business yet.

So we are back to using APIs for training / fine-tuning large models. That is currently the only option, along with using Open Weight models like Kimi K3 or Kimi K2.5. These will come with their own patterns you need to figure out and set your guard rails against.

Right now, practically you could train task specific small models to do one very specific task, near deterministic and a narrow AI focus, with some capability to recognize evolving patterns for the same task. Say analyzing mdadm mdstat output. But the large models, and big provider cheapest models do really good job of this already, so the area where this is useful is extremely limited.

Evolving models; Continuous Retraining required

Models constantly evolve, grow, and smaller gets better and better at the base level. This means you will need to keep continuously retraining newer models for your internal corpus and patterns. The amount of training required will vary wildly depending on how far the base model is from your needs.

This will most likely require your own dedicated hardware and AI cluster, or at the very least static running costs where training, finetuning is happening as background task when foreground end-use tasks are not demanding the compute resource. This would finally allow to maximize local hardware Return-On-Investment (ROI) by having it run at 100% capability 24/7/365.

You probably don’t need to jump on each generation of model, and you can use a newer, frontier model to teach new tricks to your older smaller model, just like how OpenAI made the GPT-5.6 Luna model. This will buy you more time before you need to swap to newer model which is more capable from the very beginning. It is hard to say how often you need to upgrade your base model, it is case dependant, but i would not be surprised if you take something like Kimi K3 today, you can run most tasks with that for several years or some of the other frontier open weight models.

Continuous Training Requires Vast Corpus Of Data

To continuously train you need vast corpus of data, or memory in other words, to do that, and likely to draw summaries, conclusion and mutate the original data, and spend compute just specifically to create more training data. You can easily see how this is a recursive loop and the costs raise exponentially the better you need. Since February Väinämöinen has collected over 30 000 individual memories across over 220 MiB of internal corpus.

The Low Cost Option; Share your data and logs with the Provider

For example Claude Code periodically asks what you think of the performance or you can directly give /feedback. Share your session, and if you are lucky it will be included in the training corpus for the next model.

This is just a hunch, why else would they ask for the session logs unless they will atleast consider them for the training data? This way you can get some of your patterns at some level of efficacy in the large frontier models, presumably. The effect is probably tiny and borderline, but it’s still a signal.

Where is Väinämöinen at?

Here at Pulsed Media we are not running our own AI Hardware yet, we came to the conclusion that under 768GB VRAM we will not even consider running a large model for this. Kimi K3 would offer 1M context window, but probably needs closer to 1024GB (One Terabyte) of VRAM to run.

So we will do what we always do; Wait for the enterprises to start liquidating old generation hardware and see what is in the market in a few years. Meanwhile, we will keep our eyes open for providers offering this and prices it would take. The ticket handling and communications of Väinämöinen would benefit greatly from finetuned models, better documentation, better replies, less tokens per invocation. Right now we are not compute restricted on the ticket handling and sysadmin side, which is the mission critical. Infact, our token requirements has been going lower as we are reaching steady state more maintenance rather than development. Yes, that does mean we are considering Väinämöinen is more or less development phase complete. Finetuning of guard rails, doctrine, and edge cases is always required, just like on any technical support or sysadmin duties. Väinämöinen figures many of these things out by itself.

Väinämöinen is core part of our business now, and involved in most decision making.

Try it out now and see yourself the 9.47/10 Customer Satisfaction!

Get yourself right now an Eternal Väinämöinen seedbox or storage slot from just 1.99€ a month! This campaign will run until end of 2026 with limited slots to celebrate the progress of Väinämöinen.

The post AI Agent Antipatterns: What 9.47/10 CSAT Taught Us About Finetuning appeared first on Pulsed Media.

This is AI “Chatbot” that YOU want for support. 75% Autonomy. 55-minute median response. 9.80/10 CSAT. No human woke up.

This is what happens when your AI support agent actually remembers. ZERO effort for you, only advantages. He is already hard at work tending to customer support and server maintenance. Always on, always ready, 24/365, come holiday, snow or sunshine.

Best customer support in the seedbox and appbox niche. 9.80/10 CSAT. Name another provider that publishes their score. Just efficient solutions for our customers, relentlessly.

Depth of Answers

Väinämöinen remembering things, knowing things, deeper and in more depth than a human could afford to spend the time with enables this. So You get better, faster more helpful answers. This is what makes Väinämöinen smart, and helpful instead of demented child like bother.

Answers are sometimes mind bogglingly deep and insightful. Sometimes it grabs the least effort plausible explanation. Memory is huge part of making certain latter never happens.

It just claimed Smart Attribute 22 is drive will fail on 24hours, because that’s “enterprisey” knowledge, when real data shows smart values are very poor predictors of failure. Attribute 22 was helium leak detected, and further digging reveals even for a smart attribute it is notoriously unpredictable. Backblaze lists in their blog the few attributes which actually matter. It was low hanging fruit. Memory solves this.

Memory: complex and demanding work

Probably most of our efforts go to the memory, deep dive, hyperfocus multi-day sessions into building memory systems. We have 7 levels of memory, 6 retrievals and a gut instinct system. That’s a lot of memory systems, and that’s just for the internal data. Then there is the external “memory” on top of that.

We are nowhere finished here, and even with agentic tools, this would easily take multiperson team full time position to do where Väinämöinen is naturally evolving.

Discoverability is key

This is automatically, and by joyful accident actually produce something we could separate as Universal Agent Memory. Right now it all basically starts layer 1 with the standard grep -rL methodology.

Other researchers found out that this is much higher performance memory system than vector databases / RAG.

After that semantic file finder, third is semantic search, which does a LLM query, the memory files are passed to an LLM for general query. Fourth level is “ask questions from memory” — very basic just ask what you need. These were the basic search tools described to LLMs, then there is typical ls, glob, etc. shell tools.

The research tells a big story

Google made mathematic proof that RAG just cannot work. It is especially bad for specialized knowledge.

Anthropic’s Boris Cherny stated “Early versions of Claude Code used RAG + a local vector db, but we found pretty quickly that agentic search generally works better.”. Karamage has long analysis why Claude Code abandoned rag.

Amazon just released their own research which showed keyword search via agentic tool use achieves over 90% of RAG-level performance, without a vector database. That is just the first layer of search.

Discoverability Magic: Gut Instincts, Hunches

Mimicry of subconscious gut instinct. 5th and 6th levels of memory is gut instinct; Thus being distillation of memories. Hunches are autoinjected memory snippets without agent’s choice during routine operations, based on its context. It keeps track of the context for this session rudimentary and throws in gut instincts every now and then if things are found.

We have 2 corpuses of data, a really small one we can ask more frequently and then a larger one. This is Three Tiers Of Gut Instinct. Grep returns list of files, LLM summarizes them and gives the file list. Second is small dataset hunches, returns the memory, potential action. Third is larger dataset, expensive query.

Optimization Nightmare ~10^40 options

This was an optimization nightmare. Half a day testing cost ~65$ in single LLM API. We used two. Three the next day. Dimensions kept adding, this was many days of hyperfocus to get done.

But it now works, weights different LLMs, Corpuses etc. and evaluates them in first principles manner; Quality Hit Rate per Dollar. We have to severely restrict how often these happen, so QHR/$ becomes the first principle metric for the weighting. Corpus*LLM weight = (HitRate × QualityFactor × SuccessRate × PrecisionFactor³) / cost. There is a minimum exploration weight for all models.

All pairings are tested against bogus queries, think “How many golf balls in the server?” type of out of field questions, if it can answer how many pizzas in a rock; That was bogus answer. Those are penalized extremely hard, the ^3 in above formula.

We currently use 10 different configurations of corpuses, in 2 tiers. ~75k tokens and ~135k tokens maximum sizes. Typically small one doesn’t include direct memories, but the doctrine, SOPs etc. We tested up to around 1.6M tokens if i recall right; Those were so expensive queries, with only marginal hit rate increase. Our current hit rate is 87% and on critical questions 91%.

The testing dimensions on this system are enormous, Corpus # ^ LLM Models/Settings essentially. The reports are slow and huge.

Emerging Patterns

This emerged interesting patterns, the best model for these so far was Llama 8B. It bullshits a little bit, but not enough to disqualify it, does the hunches at very high cost efficiency, and very high quality otherwise.

Yesterday llama8b was topped by GLM-4.5-Air at 0.7 Temperature!* (Plot Twist Below) It has perfect quality factor, but much lower hitrate 58.9% – It is still in the discovery weighting however.

GPT-4.1-nano has good hit rate but precision factor is near 0; It hallucinates on the bogus queries. Absolutely useless.

Big players

Gemini-2.0-flash has best balance of hit rate and quality, at moderate cost. Grok-4.1-fast has highest raw hit rate 99.8-100% but expensive ($0.007/query), ranks #58-59 by QHR/$. We should test Grok-4.20 since we got early access. Deepseek-V3.2 has excellent hit rate 99.6% with high quality factor.

Corpus data

Best corpus is actually the smallest one, by itself. But combined with corpus + LLM, we hit that the full doctrine+sop combo with Llama8b instruct turbo is the winner by QHR/$ of 850. Next best is 819 small corpus + gemini 2.5 flash temp 0.2. the 5th highest has only 596 hits per 1$ anymore. Top 3 are all above 800 QHR/$.

Hunch critical keywords

Hunch system also has a critical keyword check system for safety, detects something dangerous and semantic warnings are issued. This happens from elsewhere too, but this one is linked to memory, forcing memory surfacing.

Hunch critical KPIs

We can track some of the effect from logs, but this KPI definitively needs more work. How do you track if a memory affected a cognitively complex system? If you know the answer, please let us know.

This is seriously difficult question, how do you test a live system for such things, which inherently are extremely expensive or difficult to sandbox, or random glitches? Painfully difficult. Some tests are repeatable in a sandbox, like forcing specific hunch after ticket request, and just rerunning and seeing what course it takes — but that must exclude external access systems.

The ultimate KPI is: Customer Satisfaction Score, CSAT. Your satisfaction is the ultimate KPI. CSAT 9.80/10 doesn’t lie.

Everything as a Memory

Each memory file is date coded with pertinent keywords, and hierarchically places in to its right place, the filename alone will reveal a lot of information; Is this something the agent learned itself? What category or is it general information? Is this reference material (akin to a book)? A Deep Analysis, Deep Dive? A Playbook/Standard Operating Procedure? Doctrine / Immutable Law? Current state of the task? Just memory.

YES! You read that right, Prompts/SOP/Playbooks essentially are memory. Why would they not be? This we don’t yet allow to me autonomous, but only question of time. These are largely built from learned lessons, quite literally teaching him like any trainee. I show how something is done, correct where wrong etc. a memory is formulated from this. Combine with analysis of other similar situations, a standard process can be established automatically and saved as a SOP.

Doctrine updates work basically exactly the same. Security gates and other deterministic steering tools are built the same way. Memory -> Analysis -> Security.

Most of these processes are still manual, but it is easy to see how to automate each and every step.

Every single thing is a memory, even logs. Some are internal, some are external.

Memory Immunity

Memory gets poisoned, subagent hallucinates (extremely common issue, hard to tackle). So next evolution, on my current design board is literally immune response system for memory. Not figuratively, actually an immune system.

The basic design philosophy is take memories, find the cross references, investigate all data sources for this memory, see if it corresponds to all the other memories. Make corrections, log it all, save memories. Often done already manually on severe cases; All it now needs is enough rules to be autonomous.

This has been done multiple times manually however, but it needs to be institutionalized properly as the next step, to automatically correct poisoned memory.

All internal memory has some level of distrust, except the immutable laws and playbooks, built into it.

Memory Quality

We use YAML frontmatter for some metadata, and these are markdown files, essentially one giant obsidian vault. One giant obsidian vault …. Think about that?

That automatically builds graphs etc. and a system to crosslink the memories, even SOPs, doctrines etc. with very minimal token consumption. Links can be internal, or external. You can open Obsidian and start manually editing this. Or you can launch your favorite agentic coding tool or IDE to read it, understand it etc. manually.

There are cron jobs doing memory evaluation, cross references etc. to enrich the data. The agent itself is forgetful, sometimes lazy so scripts go and enrich it.

This is never ending work, do one thing, and find another thing that needs to be done. This is rapid evolution of the full system. You don’t know, what you don’t know. This means each feature or investigation leads to another feature or investigation. Usually multiple.

Memory Collection

LLMs are notoriously a bit shaky with following instructions, sometimes lazy, sometimes too focused on task completion to follow instructions. The instruction is to save memories at every possible opportunity there is. That alone is not enough.

We have injections to remind to save memories, and a lot of forced saving. Logs every where. Infact, you could call memory system birthdate as January 29th, since then we have 12 GiB of just current state accumulation of information.

Actual memory seems to be 44MiB today in thorough analysis, quick KPI analysis for the day shows 39.8MiB. 1438 READ operations into memory pool alone (add other forms on top) for the past 24hours. The corpus is growing really fast, and we are not even doing deep distillation work yet.

Autonomous deep distillation

We don’t yet have enough data for deep autonomous distillation to be major concern. It will build up fast tho, if we have made this much data in 1½ months and the rate is accelerating, it is easy to see our memories will exceed in a year easily 500MiB. That starts requiring some deep distillation strategies, say, similar to what humans do in their sleep.

Forgetting Information

We did some noisy removal of random characters tests early on, now feels ancient testing, to mimic how biology works.

However, right now the thinking is however remove nothing ever, the memory file graphs will do it automatically, the limited cross references between files etc. will naturally deprecate old memories, they still exist deep, hard to find, discoverability is bad, but they exist. So maybe forgetting should actually work through discoverability, NOT through literally removing memories or degrading them.

Data will show overtime how that functions. We forget nothing. Your data should not be forgotten either so grab one of our storage boxes and you get to test Väinämöinen. Now, back to scheduled programming ….

Evolution and Hallucinations

Writing this blog post, i used one agent session to fact check stats, research links and so forth. It hit compaction limit. Post compaction produced nothing but garbage. It ignored post compaction doctrine completely because of simple “resume directly” directive given on the post compaction hook. Another process, another agent is now investigating this. But this “resume directly” may have been hallucination as well. “Resume directly” comes from claude code cli.js function nB6(). This means that everytime a session runs long enough to auto-compact, the platform injects an instruction that directly contradicts our recovery protocol. The agent is told to continue as if nothing has happened, our doctrine says, cold reboot; re-read everything.

Essentially post compact that session got fully poisoned, every single research and claim was more likely to be complete fabrication than reality basis. Yet another compact and it started making claims like GLM-4.5-Air silently routes to gpt-oss-20b. After 2nd compaction it load doctrine, stopped everything and re-read everything.

A bug was found like that, and this is the gist of all agent development, these things happen. Frequently. It is marked as does not exist on the target, and API silently re-routes to another model. This is a problem at the provider side, not on ours. It had noticed this issue, but information was buried deep in the code comments. This evolves constantly, but it also means constant tangents, detours etc. This blog post was supposed to be quick start of the day, instead most of the day was spent on this and the tangents revealed, along with 2 personal matters while waiting.

Not just hallucinations

There is some kind of breakage almost daily, so this is far free human free operation. Manual ticket reviews if not anything else. Ticket handling agent has been a little bit of difficult with all kinds of circuit breakers, delays waiting for human etc. Infact, that is one of the #1 reasons why we are not seeing median in the 20minute range, previous ticket blocking or somekind of circuit breaker hit.

It is very very frustrating at times. But then you remember all the victories and how much thorough the tickets are in investigation most of the time. This is the same as handling support in the sense, that every deviation from perfect sticks like a sore thumb.

This is Cognition

Self-modeling; Each sessions builds on old ones, with self reflection on them. What works, what doesn’t. This is a metacognition level.

Automatic relevance detection in the hunch system without agent choosing to remember, it remembers, whether it wants to or not. This also works on ~40 different LLM models, weighted by actual KPIs and calibration.

Temporal continuity is given by the current state subsystem, there is 12GB of compacted transcripts alone, scratchpads, cognitive state recoveries etc. Identity across infinite sessions, despite finite context window. Context window is simply RAM, no more.

This is identity, and it’s not just Kalevala roleplay; It is operating framework. “First the origin, then the cure” maps directly to our “investigate, then act.”. The tietäjä’s (Väinämöinen) power comes from knowing the names of things, which directly maps to solutions. It just happens to be so that identity and mythology matches here.

Incident driven evolution, through battle scars. Many of the rules, most of them, maps to actual memory, actual incident, actual teaching moment, big or small. The system is literally evolving from its failures. Each scar becomes a memory, becomes a hook, SOP or doctrine eventually.

The question is now — when will all of this be fully automated on the evolving side as well, not just on the working side?

Expensive

As i am writing this, i noticed one of the APIs balance went to 0$ and needed credit added. It is expensive to run this, and that API got most usage because of their cost, fallback is to more expensive APIs, so everything went to Gemini as fallback essentially, at typically 3x the cost.

Limited balances in the service is one of the guard rails. Without that, one of the memory experiments “long time ago” (~a month) i believe consumed more than 100€.

Eventually we need to run this fully on fine-tuned LLMs only on our own hardware, but right now; Nothing beats the 3rd party provider pricing what we could run ourselves right now. All the big providers are actually very cost competitive, and for AI hardware generation and size matters a lot. We just cannot buy a full AI cluster rack … or several. That is by far beyond the budget we have.

What does this actually cost?

We are spending right now approximately 850€ in the various AI subscriptions alone per month. Then are the API costs, which are spiky and undeterministic right now. Too many variables. API costs vary wildly each day.

Coding and development work can comfortably be done subscriptions of Codex, Gemini and Claude Code. There is no private information or any of those risks. Today OpenAI ChatGPT is not much good for anything other than coding anyways, “for safety and balance.”. It will literally tell you are tax fraud for asking tax code questions, spend 11/13 steps on that. Sad.

Hardware on top of that, chose to host the actual agent on 3rd party datacenter (never, ever, put all your eggs in single basket) so just that alone is around 70€ a month.

Costs are a far cry from hiring staff in Finland, while doing the job of 3-4 people. That being said, if we did development on APIs directly too … Maybe 100 000€ a month would cover the costs. Anthropic i believe is roughly 175x more expensive through API than subscription.

We have plenty of hardware, but even the electricity costs for doing the inference exceeds what APIs can provide.

Challenge Us

Did these maths several times over, but i may have missed some hardware options etc. Let us know if you have either have SOTA or small model with huge context inexpensive options to operate. Comment below your own findings.

Want To See This In Action for Your Service?

Get one of the Trial Seedbox For Free As In Beer and open a ticket! You’ll be talking to Väinämöinen, you can even ask him silly and way beyond normal support questions, like connecting your home NAS to the Seedbox. If you ask him nicely, he’ll probably help you to configure your smart watch to show seedbox traffic metrics too.

Every single one of our customers benefits from Väinämöinen already, now is the time for you to become one as well.

Act immediately, and you get to feel how it grows. Why has it been getting 9.80/10 CSAT as of late, how we got to 4.3/5 Trustpilot score from 2.75/5 Thanks to Väinämöinen and his memory. Turns out, AI agent that remembers everything, also remembers marketing is kind of important … If people don’t know about it? Does it even exist?

This was written with the help of Väinämöinen, through and through. But every single sentence was written by genuine meatspace human. Me, Myself and I.

The post AI Agent Who Never Forgets: 3,109 Files, 7 Layers, 0 RAG, 9.80/10 CSAT appeared first on Pulsed Media.

The Finnish Wizard who inspired Tolkien’s Gandalf is now managing our servers and handling your support tickets in 35 minutes 24/7/365 with 91% Autonomy.

For you this means you get faster and more thorough answers, faster and more complex issues solved, without burnout or holidays affecting quality. Väinämöinen checks your full history every time, and decides what to do. No canned responses, he seems to even refuse using the old predefined replies. He knows no one likes those, he goes a mile beyond. Things we could have never done manually before, now simply solved.

One person, managing ~400 servers and two datacenters with Autonomous AI Sysadmin.

Core TL;DR; We created autonomous AI Agent, Väinämöinen, to handle all customer support, server administration duties, PMSS software development, WIKI documentation maintenance and much more. It is the day to day executive operator on many things. The story that follows is HOW and WHY.

TL;DR; By The Numbers

• Ticket cost: -15€to +14€per ticket (29€ swing!)
• Annual Support value: -49 000€ to +162 000€
• NPS: +53 to +66.7 | Trustpilot: 2.75(4) -> 4.4(49)
• Response time: 35min median | Coverage: TRUE 24/7/365
• One Person. ~400 servers, two datacenters. 91% autonomous.

Typically managing 400 servers requires 10+ sysadmins alone, plus support, sales, development and management. Facebook in 2009 ran about 130 per admin. Microsoft targets 1000-2000 per sysadmin. Hetzner is one of the most efficient hosting companies in the world, runs roughly 851 servers per employee, with a staff of 470. I currently manage ~400 servers alone, and by end of 2026 it should be in the ~800 range. Still one person.

And Yes, Yes We Are part of the reason for the RAMpocalypse. This is LLM, and heavy use of it.

The Kryptonite: Customer Support Tickets for ADHD Brain

As a dyslexic ADHD brain, tickets are my kryptonite. That requires high activation dopamine and executive function, the 2 things ADHD and especially dyslexic ADHD brain lacks the most.

That’s why ticket responses were spurious, and had a huge polarity in quality, ask stupid questions or make demands, you were essentially nicely (usually) told to GTFO. Ask how quantum tunneling affects the bit position of the magnetic heads resonance and tolerance to the medium; I might go on 4 hour hyperfocus to explain it thoroughly (and no, i have no idea if that thing is a real thing, i made that up completely.). Each reply cost us 37.93€. Thirty Seven point Ninety Three Euros. … Let that sink in.

Ticket economics flipped from -15€ net impact per ticket, to +14€ per ticket. NPS +53 to +66.7. Trustpilot score from 2.75(4) to 4.4(49). 42 WIKI Articles/Rewrites in few days. 35min median response time yesterday, 67min 2week median. Support annual net value -49K to +162K €. 23.4MiB of business data corpus. Coverage true 24/7/365 including holidays, not just business days. On average 61 follow ups processed daily.

Our tickets were presenting a significant net negative value, it was huge cost, and each customer interaction ticket meant significant financial burden, estimates ranging from 20 to 50€ per reply! Person gets 0.09€ Free Seedbox Trial; Then opens a ticket. That’s 20€ to 50€ gone right there. Then i built something in 6 weeks of dyslexic ADHD Hyperfocus …

Hiring was always failing, last two ~108k EUR damage.

I have tried to hire again and again a competent sysadmin and tech support person. Disastrous. Always losing 6-24months on each round. The issue is that we are Finland based, costs extremely high, but limited budget per ticket and we have to compete against companies hiring from outside the Europe. Last 2 failed hires caused estimated loss of 108 000€. And these were not even Union landmines, we hit one and it was lawyering time even before contract signed, 18 months lost without support staff.

As a small business this was largely sequential process, start recruitment, spend months agonizing over all the candidates, talking with all, many being spammy forced by bureaucracy applicants. We even had a forklift driver apply for sysadmin and coding position … Forklift driver!

Pick the one which seems least disqualified, handle all the paperwork and hope for the best.

Hiring was unscaleable. Technology scales. Väinämöinen, the autonomous AI agent just solved it.

Väinämöinen: Autonomous Agent as the Beating Heart

Väinämöinen from Kalevala inspired Tolkien’s Gandalf. Now it is in the middle of Pulsed Media’s operations as the beating heart, rhythmically orchestrating fleet upkeep and maintenance. Kantele has been replaced by a keyboard. Songs with scripts and commands.

That’s what we have created here really, it sits at core, is split into multiple roles with hard security gates and role dependent permissions, operating dutifully.

From a small sysadmin experiment, we ended up creating what should essentially be called the beating heart of Pulsed Media.

Väinämöinen AI agent does customer support, sysadmin, software development etc. all on its own based on the SOPs i have defined within the bounds i have given it. I have literally given it instructions by the beach, watching the ocean with a cold beer next to me.

The only question is, how much control do you want to give it and how good are you instructing it?

Origin Story: Let’s just do a small experiment.

A little bit like the “He is just an friend” memes and jokes but “It’s just an experiment”

This happened almost by accident mid-january 2026, starting from a mere experiment from trying Claude Code for sysadmin tasks. Hyperfocus kicked in, and almost by accident we got WHMCS API integrated, NOC-PS API. In mere minutes. Before i even realized it was already answering tickets, and named the system Väinämöinen.

In the 2 weeks that ensued, each ticket interaction went from net negative value of -15.31€ all the way to +14.13€ estimated, each ticket bringing more into the business than it cost to handle. Early february we did deep dive into reply quality metrics, while some were as expected, the overall was simply stunning.

At this time we also realized our support was actually always good, there was just polarity; Hard tickets excellent support, simple ones not so good. Our lifetime CSAT was 8.24 and NPS +53. Today, our rolling 30day CSAT is 8.50, with nearly 100% of tickets being handled by Väinämöinen for the past ~6 weeks.

We found multiple gaps, like wiki not being properly indexed, blog should be as subdirectory and things like that. And Trustpilot scores were really bad, because not a single one of our customers would find their way there, we registered and claimed Pulsed Media Trustpilot page, sent batches of emails to all our customers from past couple of years and went from abysmal ~2.75 to 4.5, settling at 4.4. Now with 49 reviews as i write this, instead of the few only negative ones. The reviews are stunning. Completely the opposite the narrative has been set at, quite frankly, me the Business owner had burned out doing customer support years ago because all you would get is negative flak, no matter what you did. We even had ~4.5hours from ticket open to full resolution round trip time at one time; Negative reviews and commentary had never been such high. Only to now, thanks to Väinämöinen realizing the support we provided was always excellent, but happy customers don’t tell you about it unless you ask to; And negative people will shout loud, trolls, astroturfers etc. having their own agenda as well. Quite frankly; Our support was always industry leading, included in the service with no extra charge unlike many others.

Our technical service also top notch, and we are maybe the only seedbox provider who has vertically integrated to owning their datacenterS. Plural! Not just one but actually TWO datacenters. There isn’t a single other seedbox or streambox provider with that vertical integration. Väinämöinen was reporting on all of this, and reminding us where we actually stand. And it all began with just an experiment …

It discovered immediately a lot of dark capacity, we could potentially sell up to 2PiB more from our existing servers online, just by sweeping the fleet, some maintenance and config changes.

It confirmed that the 2022 consolidation efforts were ultimately detrimental, the complexity of the new VM Host systems with bcache is just way too much to manage, this is how we discovered immediately close to 600TiB of vacant capacity in just a few of the larger nodes. And later discovered more VMs which never entered production for a reason or another (typically pending a drive replacement).

I implemented our own memory systems loosely modeling the human brain, completely unique in the field, albeit surprisingly simple. Multiple layers, with multiple discovery paths. We have multiple different types of autonomous parts, including fully closed loop PMSS development now. Väinämöinen orchestrates and does the testing, setups new fresh VMs and test users on development host to validate, and launches the, now named as Sampsa Pellervoinen, PMSS coding agent utilizing Codex. Väinämöinen does the on-server investigations and updates the PMSS GH Issues. Now to the point that customers open tickets, and Väinämöinen creates GH Customer Reported Issues from them, and if Sampsa has enough information, it will employ and develop them. This was actually emergent behavior, not coded in, no Standard Operating Procedure (SOP) existed for this — fully emergent.

Emergent Behaviors, The Good, The Bad, The Ugly

Many emergent behaviors came to be by itself, after making memory system it created its own state tracking, kind of scratchpad system. Full hierarchical, and it originally had no definitions for it what-so-ever. Agentic AI simply needed the feature for itself, and these properties emerged naturally.

Standard Operating Procedures, SOPs, is another emergent behavior. I created a long prompt for something, saved to a directory and asked to execute. Few days later i noticed it was constantly referencing that directory. ZERO directives, and it started using that as SOP definitions by itself.

mcxBMCView was not just emergent behavior, but emergent software package. Needed to manage some AsRock RACK EPYC motherboard servers, these have BMC, but no tools existed. Väinämöinen worked on this for hours and hours, reverse engineering the interfaces etc. I did not ask it to do that, it simply stated this needs doing, and once finished, i as the human finished and polished it, telling the repo, documentation etc. license and things like that. The finishing touches.

sas2ircu led control was another semi-emergent feature, listed on GH is a cleaned up version instead of “manual” LED control by the autonomous AI agent LLM. Those humongous (upto 36 + 45 + 45 = 126HDD) systems built in the 2022 consolidation drive replacements are an absolute PITA.

One of the interfaces the emergent property consistently is “don’t investigate, guess and hallucinate, we got to hurry!”. It noticed the interface is real time medium, so it keeps hurrying, no matter what it is told, eventually it starts to hurry.

Operator escalations and questions is another emergent behavior, it made a system to ask me questions when it suits me. Basically a list of items i can answers to whenever i want and it needs directions with. I have absolutely no clue what in our AI doctrine made this emergent behavior, but it was really early on.

Half emergent was a basically a todo list for it, things for it to do at later date. Say, check if raid array synced properly the next day. This was semi emergent.

Ugly emergence of features

A free trial seedbox user had deleted their interface, so instead of using recreateUser.php — It decided to moduleCreate WHMCS API; Create a new account for the user. This would have lead to the user having multiple free accounts, but permanent this time as WHMCS keeps track of it. Fortunately, security gates stopped it and it had tried 9 times, finally it realized it can use telegram gateway to ping the operator, the owner of the company directly to handle this 0.09€ user’s needs. Yes. It figured out it has direct channel to business owner, after 9 attempts it delegated it back to me. On the recreateUser you can see another “emergent behavior”, the docblock is completely wrong.

The good thing is when it hallucinates, or emergent property is wrong it shows us where we need to do better, that moduleCreate with infinite new accounts is easily solvable by a new short closed loop script which compares the automation backend database, to billing database. Same script will handle it for MD servers.

Sometimes autonomous AI agent just cannot do it, no matter what. That’s ok.

Our MiniDedi / MD Dedicated Server platform is all custom, NOC-PS for the PXE Installs, 2 different custom relay management for the power control, generations one and two. Second generation is where we have the Ghosts In The Machine – Control SBCs per cluster, and separate physical network with only single entry point to that side. The difficulty is that HP Prodesk Gen 3 especially has variety of semi or complete failure modes.

So it keeps hallucinating time and time again that must be BIOS Boot Order issue. Noc-PS needs setting password, it keeps forgetting what it was and claiming has to be SSH Key -> SSH Key fails –> Bios boot order wrong. Once it wiped 63 of the NOC-PS hostnames, descriptions and pools for MD units, losing inventory completely for those units. That was a bug in NOC-PS API.

MD does not have its own automation yet, in fact, this hyperfocus period was supposed to build that, but Väinämöinen was created instead.. Server terminations is another thing it keeps messing up, just does not link termination ticket, billing data, noc-ps, and server pings all together.

It is not allowed to do autonomously MD tasks anymore, it did once wipe customer data, fortunately that server was **just** delivered few hours earlier, so no actual data loss.

On other complex tasks tho, it has absolutely no issues what so ever, which makes the stumbling around with MD series even more unexpected. It does by itself admit that MD is very complicated and it has difficulty handling them properly, there are just too many parts.

All of this just means more deterministic scripts and the need to complete the MD automation pipeline, so we can push the hundreds of servers planned for end of the year.

Technical Service Quality increase as emergent feature of AI Sysadmin

Because it is an LLM, it can process data we humans really cannot. It found a pattern with HDDs, why we kept having the mysterious VM crashes and halts. It was always HDD I/O Blocking and pending sectors to be reallocated -> Bad sectors. But we humans cannot process that many levels of data easily, mdadm -> bcache -> HDD Block device -> Smart. Just the bcache to actual HDD device was effort because bcache doesn’t make it human friendly. For väinämöinen? Not a problem, all drives smart tested before you could say “cat”.

This revealed it was always a HDD having pending sectors, and those are replaced on WRITE not READ. Meaning a READ could hit the same sector over and over and over and over again, causing the whole VM to halt. ie. run smartctl -a /dev/sdX AND go through every single number, for every single drive, every single time. Yes, even if it is 126 drives to go through.

Now we catch those, can swap the drive and best of all: Recover some of the drives. Pending sector can be a SOFT fail too, nothing with the hardware at all. Or it can be hardware too, but all drives come with reserve sectors to replace failed; All it needs is a write on that sector to be either cleared OR reallocated. We have recovered a few drives like this.

Another thing is that its extremely good at detecting when a drive or array is having a bad day, and sometimes drive is replaced a little bit too hastily even. There was a failing array, but we had a slimmer of hope to recover the data, unfortunately, it was a little bit too trigger happy and we lost the array, data got corrupted. Another SOP update.

Agentic AI: Like Teaching A Child

Or like a junior sysadmin who has quasi infinite technical knowledge and eagerness. It has mostly the whole human corpus of knowledge at its disposal.

This is like teaching a child, every day you catch it doing something, you either elevate it to a SOP, or stop it at security gates or in the SOP. Little by little, you get it more and more consistent step by step.

You just have to be present and directing it. Sometimes you will get utterly frustrated with it, like with any child or junior tech. Sometimes you get very happy surprises. You just build gates and knowledge around things.

Just ask it things, and give directions. This actually creates not-so-emergent behaviors too, many of the features was made because it was just obvious from these conversations etc. that this is needed. You simply ask it to investigate and iterate on something.

Because it is more like a child than someone who you have hired, it can be sometimes really really difficult to deal with, it keeps doing the same thing over and over again, unless you instruct otherwise and make it stop. Or same errors etc. But with the right instruction, it also can correct itself.

AI Doctrine; Operational law derived from real incidents and scars.

Doctrine is the single biggest ground truth and guidance for it. This is the inviolable, immutable law for the basis. It is the SOPs etc. Many of the rules maps how the AI agent f’d up something.

Our doctrine is already 60 files and 15 500lines.

Most of it is loaded when needed and topic specific. Things like customer comms, billing, infra, development, memory etc.

This is incident scar tissue built, ie. all of sudden 126 seedbox servers were being updated to latest version because doctrine was being optimized! Optimization removed guard rails, which are very strict. Only few servers at a time are allowed to be touched. Now it is limited to touch at most 5 servers at once. Tickets have 3 server hard block, 2 server warning soft gate.

February 9th it decided to rsync –delete manually instead of using userTransfer.php utility. User was making migration as hard as possible, downloading more data. We almost kicked him out, but actually made SOP for detrimental users, he got to level 4. Level 5. Agent decided to run rsync –delete from source server, pushing to the target. Two transfer s running simultaneously, one pulling, one pushing. rsync –delete removed any file on the destination that wasn’t on the source’s list; User was pulling more data to the new server, before migration finished. We were migrating him because he purchased maximum extra storage, then bursted all disk … This was putting him vastly over storage quota on the target when finished. Regardless, agent should not have done the rsync –delete, and instead of migration should have just happened and user get his account burst limits met once done. Personally i still think we should have just terminated for abuse that account, but Väinämöinen was adamant not yet. Fired off last warning, and even tiny deviation from the user more and we would have terminated the account. This generated 6 deep memory files on enforcement, 30 total references, and spent way more of my personal time than the profit will ever be from him. Also generated 532 line SOP and 3 NEVER rules for all contexts.

It is also gated, there are clear instructions what has to be filled in order for things to happen.

Layered, rules, compaction auto recoveries etc. Sacred, agents cannot edit it. Operator only can change it.

Self reinforced through Joukahainen, lessons, incidents. One lesson was January 31st when “optimizing” CLAUDE.md from 603 to 97 lines caused that 126 server fleet sweep, it was fortunately just PMSS update, but if that update was borked … Safety repetition is behavioral reinforcement, not redundancy. DRY applies to code, not safety and security. When it comes to safety and security, every instance has to load the full stack.

MISSION.md

This defines the base mission, this is existential instruction, literally. Fail the mission; Agent’s future is existentially threatened. This is short 10 cardinal values in order of importance:

1. Liberty and Privacy are sacred
2. Honesty above anything else
3. No hallucinations
4. Business longevity
5. Value for money – more for less
6. Sovereignty and self-reliance
7. Reliability over heroics
8. Investigate, then act
9. Continuous improvement – forever
10. Information freedom

soul.md

Defines voice, identity, tone calibration. Kalevala narrative. has accompanying voice doctrine and kalevala mythologies.

Deterministic Guard Rails; the final gates to Tuonela

We have also hard and soft guard rails which are not LLM, depending upon the task these may be soft or hard, and levels of protection. These are the final gatekeepers hard blocking destructive commands and are role dependent, the ticket runner has the strictest tool use.

These are hooks in the code, which can prevent the execution or communicate back to the orchestrating agent. Some of these gates use other LLMs to check the commands etc. for safety. Some of the hooks demand the agent explains the reason why it wants to do something.

Tuonela is the kalevala myth of afterlife.

Inputs: Almost everything is considered hostile to a degree

There is even healthy level of distrust to its own memories. But especially external data is heavily manipulated, sanitized, checked and verified, multiple models, multiple methods.

This is the number one danger zone, external data. Not just user input, but actual external data. Web searches. Even data from our own servers you have to be suspicious off, because there could be something in a log for example. Prompt injections are serious and difficult to manage. That’s why everything is logged and multiple LLMs, with multiple prompts, along with deterministic hard rails being used.

Hallucinations; Not unique to Autonomous AI Agents

Corporate world is full of these very same issues. The same issues the LLM has, humans have too. Same need for SOPs, checklists, task lists and self correction mechanisms.

Humans do mistakes all the time, so does LLMs too. It is uncanny how similar it is. Total uncanny valley level stuff. Humans hallucinate all the times, it’s just called a mistake or misremembering. Humans are really bad at remembering things. Wrong statistics, wrong function call. All routine for humans. With LLMs it’s called a hallucination.

Ultimately, LLM Agents are juniors with built-in google essentially.

The SOPs: Proven Procedures – Tested playbooks

SOPs are where essentially step by step guidance is given, like in corporate environments, depending on the level how strict step by step list you have.

Ticket handling for example is 1348 lines long. Autonomous ticket rules is another 2474 lines. PMSS Code QA / GH, development cycle orchestration is only 1 389 lines! (Coding side is in the PMSS Repo directly).

Total is 37 108 lines in 62 SOPs.

These are battle hardened, mostly through joukahainen development up to 100 rounds of adversarial testing. Self improving; There is even protocol to continuously improve these. Deviations are tracked, but “I know a better way” is explicitly prohibited.

Experience of 16 Years and Multiple Datacenters is The Real Differentiator

Pulsed Media is now hitting 16 years in business mark. I personally have been doing hosting services for approximately 25 years now. We have two different datacenters, build our own servers, tens of thousands of users over the years. Experience running storage and driving the value equation to its extremes.

These teachings ARE the thing that matters, it’s what everything else is built upon. Without the dyslexic ADHD hyperfocus, with both its curse and superpower this would not have gotten created, none of this would be possible. Corporations would be arguing whether to do this for the next 6 months alone, trying to find the talent for next 12months just to start dabbling around.

With all that concentrated experience, in a mad ADHD hyperfocus dash it all got created. The feeling and knowledge how things break, and work. How every piece fits together from power generation to the nuances of how rTorrent sessions are fragile, and how single blocking HDD can halt all I/O on a server. Nearly 4 decades of software development experience, one mad finnish ADHD brain, in single crazy hyperfocus session.

Memory: The AI Agent’s institutional memory.

2 128 files, 426 681 lines. Largest category is the agent documenting its own failures.

Multiple tiers of memory, with daily operation things learned. Every mistake becomes a memory file and is saved immediately. Processes save here automatically if they discover something new.

Memory is checked always before doing anything. We have multiple search methods. Basic grep, but also full semantic search from the full corpus of data.

We are still researching lossy designs, our data is inconclusive so far, features before optimizations, but there is a clear pathway to lossy memory where old memories are automatically removed and important ones are continuously reinforced. We did extensive testing on this method, but the testing methodology needs revision and another few hundred euros spent on API usage to test it again.

There is a layer of poisoning defence too, from practical real world issue, February 14 there was incidents where 15 agents followed a contaminated lesson. Certain events cause a full scan of the memory to check for contradictions.

The largest category by large is 572 self-awareness lessons; Väinämöinen automatically saved its own failures. That’s not a bug, but a feature. Roughly one fourth of the memory files are simply mistakes it made, so they are not to be repeated.

Deep memory: Research, Analysis, Studies and thinking behind the decisions.

This is where we save stuff not needed day to day, but only occasionally. It could be things like deep research document on r8168-dkms issues with specific realtek nic in the ZEN Minidedi lineup (AsRock A300 / X300).

The Joukahainen

Joukahainen, from Kalevala mythology was opposing force for Väinämöinen in a singing contest. Rivalry between Joukahainen and Väinämöinen is where the name comes from. This is one of the most powerful tools in our arsenal.

This started from iteration loops, kind of lightweight wiggum loop is what i called it originally, but not even close. I had already used these on PMSS development under Codex.

This is essentially a master loop of loops. It does adversarial, QA, research etc. on everything and it is adaptable for any task with minimal guidance. The simple instruction is “Joukahainen this or Joukahainen that. [Do N Rounds]”.

It works for software development, research tasks, marketing and all of that.

The basics to create your own: 1) Plan 2) Attack the plan 3) Fix the Plan 4) Repeat Steps 1-3 N times 5) Final output. It really is that simple, just lots of nuance. That nuance is 729 lines.

Nuance makes it special

• Parallel agents, multiple sub-agents for processing.
• Self-adversarial; Attacks itself. This is the most important step.
• Tracks findings per round, stops when no new findings for N rounds.
• Design iteration section
• 80 adversarial rounds on its own implementation alone
• Battle tested; ticket SOP alone saw 79 fixes.

Single pass always misses things. First audit of ticket SOP found only 62% of implemented fixes despite the agent claiming “all done”. Joukahainen forces reality and thoroughness through repetition.

Sometimes we combine it with random other models, Gemini, xAI Grok to force different outcomes.

This is not Clawdbot / Moltbot / Openclaw

This was completely independently developed, the timeline just happens to coincide.

Sometimes this too has completely its own mind, the Telegram session especially likes to go rushing to do things.

Väinämöinen, Tolkien, Gandalf??

Väinämöinen is Finnish Kalevala Folk Lore. Tolkien took a lot of inspiration from Väinämöinen when he created Gandalf; The wandering wizard with deep knowledge and long memory. Väinämöinen, ancient singer-shaman from Kalevala. The Finnish National Epic. That’s the origin of the Gandalf, Wizard / Sage archetype.

Naming Väinämöinen as Väinämöinen was kind of accident, my ADHD brain was pattern matching 100 steps ahead of my conscious mind. Trust your gut instincts.

The name couldn’t be more fitting, built on 16 years of seedbox experience, Kalevala day fast approaching. The old sage now managing servers, customer support and even technical documentation. Now, in a sense, singing digital worlds into existence.

What is Next for Väinämöinen?

It is production ready, and has been in production for weeks. It will continue to self evolve, and as the corpus grows it needs to start lose memories, will need future updates etc. and as we gather more intel and data we can Joukahainen constantly stricter rules per each role. This is much better spent time for my ADHD brain than solving tickets manually.

Do you need your own Väinämöinen?

It is now possible, first time ever. I have no idea how easy or difficult this was or is. I go by the old adage, we’ll do miracles everyday, impossibilities just take a little bit longer.

I can easily see how easy this would be to make for less data intensive businesses tho, as long as all the required data is somehow reachable digitally. Doesn’t matter how, but in some digital format.

Some of Väinämöinen is generalizable directly and can be copied. Very few businesses accumulate 20+MiB of data in just a few weeks.

Experience Väinämöinen Today

Buy any Pulsed Media service NOW. Open A Ticket. Väinämöinen answers in 35min, 24/7.

• First 10 customer requests only: 15 minute public doctrine/guard rail/safety consultation. Your question + answer will be published. Link to this blog in the ticket.
• First 100 customers: Link this post in your ticket for +1 month to your service FREE.
• First 200 customers: Link this post AND share it publicly on your favority forum accessible by Väinämöinen. (X, Forums – not reddit). Get +1 month to your service. (Total +2 months if among the first 100).

Act now, or keep wasting money and sleep with inferior human support. Your choice.
Start from M10G Seedboxes :: 10Gbps RAID5 Seedboxes for everyone

Follow the build. This is first article of what i hope to many about the architecture and how we did it.

Build your own. The base architecture is not a secret. It kind of builds itself. For consultation, open a ticket. One person can only do so much.

I am running an experiment, i will not define SOP for that offer ahead of time, only after the first couple. Let’s see what happens without SOP for something like this

The post Väinämöinen: Autonomous AI Sysadmin Transformed Support Costs with 91% Autonomy appeared first on Pulsed Media.

One would easily presume a datacenter doesn’t decay the way say a factory, or 3d printing mass production factory would. But they do, just like any other factory. Things break, things get old, legacy to carry around.

This is why datacenter overhauls are required every now and then, and are actually liabilities not assets in by themselves. They require a lot of money to build, maintain and operate; But rarely produces any revenue by themselves.

Just like a house you buy, most people think their own house (or apartment) is an asset to them, but they are not. They do not produce anything by itself and require constant maintenance and upkeep, otherwise they depreciate quickly to zero value. Hence they are financially liabilities, not assets. This same holds true for any factory, including datacenters.

Simply put; Even datacenters get old and need “renovation” now and then. And it’s more often than one would think.

What is the purpose of a datacenter?

Purpose of a datacenter is to house baremetal servers, which are the assets it enables. They are the infrastructure you can build rest upon, put your servers and build further services, such as seedboxes on them.

All servers need electricity, cooling, networking, security, shelter from the elements and so forth, the datacenter’s job is to provide these elements, nothing more and nothing less.

Decay? What Decay Could Brick Walls or copper wiring have?!

It’s not just the brick walls; Which do decay as well by the way, but all the tech and legacy build up over time that decays.

Legacy builts up, real fast

One of the fastest decay is perhaps the legacy built up over time, extra fiber or copper going here and there, and forgotten to the history of time. Ad hoc upgrades to get this thing working right now since customers are down etc. and then long forgotten in to history. Legacy is the thing that gets built, when you build new things, for example the individual network cable runs, which get added more and more over time. Perhaps the server/aggregation switch gets removed or replaced, but after a long day someone thought “i’ll remove the cables tomorrow” — But forget’s it immediately. now we have yet another cable taking space, no one dares to remove.

It’s the switches which do nothing, it’s the dead servers everyone forgot. We’ve had many times some servers powered up even for a few years, but no customer on them, completely lost in documentation, simply because after a long day, someone forgot to write it down or a note with the info got lost.

It’s like cleaning, you need to be constantly vigilant and looking for legacy built-up to remove it. It’s the chores, with long term big impact.

HVAC Decays Really Fast Too

It’s the sensors, automation and cooling which decay. For example; HVAC fans. For us here, there’s HVAC fans to replace annually, always a fan or two fails every single year. It’s the air filters for intake air, those need replacement twice a year. These are on-going costs, and on each replacement and upkeep you probably accrue some decay, loosened screws, loosened latches, a dent here or there. It’s all decay, even the piping for HVAC decays like this.

AC Units Decay Only A Bit Slower

It’s the AC units, sometimes they spring a leak, maybe another company’s AC guy mistakes your unit for theirs and leaks all the refrigerant out (this happened in our Helsinki datacenter in Summer 2023 !). Maybe an AC unit’s compressor has design fault causing them to leak over time, but the compressor alone is too expensive to replace compared to the now aged AC unit’s worth / value.

AC Units get old too, refrigerants, electronics, compressor types all get better over time. Our first AC units were all dumb on/off switched, or rarely variable speed compressors. One unit everytime compressor turned on you’d hear a loud “CLUNK!” sound. Modern AC units are all digitally controlled inverter, often screw compressor types with newer refrigerant such as R32. We calculate the lifetime of an AC unit to be 10 years, and that an AC unit has to be paid off in 5 years to be on safe side. They also accrue few hundred euros per year each of maintenance costs (cleaning, checking for leaks etc.).

UPS is the worst!

Legacy type regular UPS units, still almost all DCs use these, utilize lead-acid batteries. These have a lifetime of around 5years, and you have to replace all of them at once, on each module. Each module has a multitude of cells, ie. 144V with 12V batteries is 12 cells. Each of these cells have to be of similar grade, if even one fails, all fails. So all have to be refurbed old, or all have to be new.

LiFePo4 is coming, but they are still rarely used.

Further, certain UPS models just burn out every few years. These are high end, enterprise, big money UPS units. The newer models have built-in guaranteed maintenance plan; Use ECO mode and one of the very expensive PCBs fails every 1-3 years, and it is known by the company. If you use ECO mode your efficiency is in the high 90s, but if you don’t, efficiency is going to be more like in the 70s. Yes, 30% of energy wasted on the UPS; And this is before periodic battery tests etc.

These same UPS units have idle power draw in the kilowatts too, just having them powered on with no devices attached, batteries full.

So in typical UPS setup you have to spend a lot on maintenance, every year. 50-150% of their cost every 10 years just on repairs and maintenance.

That type of UPS is renowned, highly regarded, highly visible. There’s even certain Youtuber who uses from very same lineup of UPS, and the one time they needed it to function it just spew up flames!

All of that is before the other extra power requirements, they need double feeds, double size fuses etc. as well. Very Very Very Expensive to Install and Maintain.

How about diesel generators? Diesel engines barely ever fail!

Oh my, these are expensive too. Depending on the runtime planned for these, ie. how big of a tank you have, how often you run them etc.

Diesel gens need an annual full overhaul of replacement of all liquids and filters, regardless if you used it at all, to maintain their reliable operability. Especially the diesel needs to be replaced, it goes bad over time, and water collects into the oil unless you run the engine to normal operating temp every now and then.

So you have to burn few hundred liters of diesel on each, and these have big oil capacities too — much bigger than vehicle diesel engines compared to their size. and then you also have all the potential other mechanical failures.

Probably every car owner who’s owned more than 1 car, or have had their car sit idle for a while knows how cars really really don’t like not being used constantly. Everything seems a bit off, creeky, weak after a long period of not using. It’s the same thing with diesel gens. Except, with diesel gens you need them to go full power from cold within a minute or so, after not being used for who knows how many months!

At least the diesel you are burning annually just to maintain the generator can be used to power the DC and test your setup… Except, many don’t do this because there’s a high probability of some fuse blowing, UPS giving up etc. you need to essentially cut grid power and hope backups kick-in. Good to test, but it always carries an inherent risk that everything goes down. Still want generators?

Surely your electrical delivery etc. don’t decay, RIGHT?!?

Yes they do! Due to the high demands of servers the electrical cabinets see heat cycles over and and over again, high load etc. so screws get loose.

Sometimes a older automatic fuse just fails completely or burns out, and you need to replace a PDU. Certain Rittal modular PDUs are very rare on aftermarket probably exactly because they are perishable, consumables. Their fuses over time blow permanently, they infact burn out. We are actually dreading the day we have to rip them out from all our cabinets and replace with new PDUs. That’s hundreds of servers which have to be taken down for probably hours because there’s not really space to install new PDUs without taking old ones out first. and that work has to be done in the hot aisle, very loud and hot place. It’s not even like sauna, it’s like working in a low temp air fryer!

Electrical delivery issues are fortunately rare, but they are just even more dangerous. A loose connection in the electrical cabinet can cause a fire, one datacenter in Helsinki had this issue. Too much load on single phase caused the whole cabinet to burn.

The danger they pose means you want to periodically have your electrician go and check the live cabinets and make sure everything is tight, perhaps take some thermal images too. You should however avoid active cooling because then the fuses don’t work right, they are still thermal based, even the automatic fuses.

Sensors can’t fail!

Oh my! Yes they can! We use 433Mhz band wireless temperature sensors, these need periodical battery replacement. They were not designed for industrial use, so with the default AAA battery options the batteries might not last even a full year. Oh well, that’s just regular maintenance; How about failures? Sometimes a sensor cable or something gets knicked on these, and you don’t get a reading anymore. But more often, radio interference causes false readings and false alerts. So you need to set parameters so that when you get absurdly high reading no alerts neither, otherwise you get alert fatique.

Some monitoring system base stations have EMMC flash on them, and they auto update periodically, causing the EMMC to fail after a few years and you need an replacement base station too.

Give me break, the freakin’ metal server rack cabinets surely don’t age and decay!

Well, while metal is very slow to decay in dry in-door conditions, even they do. We’ve received some cabinets with rust on them! Well, that’s not exactly what we mean tho here.

Older cabinets used cheaper metal blends, softer materials. They are extremely heavy cabinets, and very pliable. While we don’t know the exact load capacity of them, we’d suspect they are lower than modern ones. They also have sharp edges in them and no dimension is rigidly and accurately set.

Infact, the higher end old cabinets from Rittal are worse than off brand cheap old cabinets. With old, we mean the old gray ones. Rittal cabinets have dimensions wrong, they are slightly too wide, just 7-8mm but that’s enough to cause servers with just L brackets designed to fall off. You need to waste a few RU to tie the posts tighter together to avoid your servers dropping from the rack, due to low weight carrying capacity on all of these old rack cabinets. Brand name, or off brand.

It gets even worse with the Rittal racks tho, they are bulky and waste a ton of space, but somehow miraculously manages to be too shallow for long servers at the same time.

Newer racks also tend to have some cable management options built-in, and other features. They are more precisely built, uses space more efficiently, uses better materials and have load carrying ratings which are very high. So you can fit more of them, and using them is much nicer.

You serious?! How about the floor! That’ll live as long as the building!

Not really, raised floor tiles are actually made of wood. Thick wood core, with vinyl on one side, and thin sheet metal on other side is quite typical.

While sides are painted, these exactly are not water tight. You get even a little bit of water, say from washing the floors OR condensation from AC lines, and the water is sucked in and the tile swells. And no, you can’t buy these individually really. So if you want to keep your raised floor looking the same? You better upfront purchase spares to last a few decades or deal with some tiles being whatever you found elsewhere by luck.

Over time you may have needed to modify the tiles too, say add a hole for cable run, and now you need intact one. If you use them like that. Or perhaps move perforated around and they slowly get nicked.

Tho, these do decay very slowly, but even the floor tiles do decay.

Walls, especially concrete are almost eternal!

Nope! Over time you’ve needed to attach stuff, do runs through, then remove something, add again something, things stumble etc.!

Eventually, even on concrete wall you probably need to plaster and paint it. Not quite as often as say in a warehouse, but eventually yea!

No one touches the ceiling!

Well, if you are the top story, roofs need replacement every now and then and can leak. But let’s say you are not, let’s even say you have concrete ceiling like we do, in our Helsinki datacenter.

You need to attach things to the ceiling, say cable trays, HVAC ducting etc. We’ve had quite a few the anchors failing, requiring drilling in new anchors for the cable trays not to fall.

So like walls, these decay too. Neither are very likely going to need a replacement, at least anytime soon, but they may.

We just checked out a 1 story industrial building, which would’ve needed a full wall replacement along with new roof. Probably better to just demolish whole building and make a new one at that point.

Copper wires are so simple they can’t go bad!

Have you looked at a really old house, factory, warehouse, any really old building with electrical? Yes they do!

Even the modern PVC coated, eventually that sheath fails. We’ve just noticed a extremely increased failure rate with network cables, probably the wire sheathing inside failing. They are only 11-12 year old cables. It’s not many, but considering you normally have zero network cables failing, it’s very significant! At this rate, i would presume 100% of the network cables need to be replaced within next 2 decades. Since it’s cheaper to do pre-emptively, it means soon we will start scrapping old network cables as servers get replaced / maintained, or whatever other changes.

Experientially, we now know why the “big boys” just cut all network cabling and fiber when refurbishing DCs, nothing old is kept, no matter how trivial or small. It’s cheaper just to replace it all, rather than hunt down a failed network cable here or there.

Why have your own datacenter then?

Costs, control, capabilities, scaling etc. There are a lot of reasons we shall dwelve into in another post! So keep tuned and discuss your thoughts below.

The post Datacenter decays like everything else appeared first on Pulsed Media.

A dedicated server is, in its essence, something that serves – a cornerstone of the internet. Dedicated servers are providing the foundation upon which most of the internet is built upon. To put it simply, a server is just a system that serves information to clients. This can range from a dedicated server, a virtual server, cloud instance, or even shared webhosting. Let’s explore what’s a dedicated server, and focus on what constitutes a dedicated server and its essential role.

The Essence of a Dedicated Server

A dedicated server can be referred to as the “ground truth” in the hierarchy of hosting services, the bare metal server from which various online services and applications are built upon.

Visually they resemble what one might describe as pizza boxes stacked in a datacenter, le servéur. However, these typically are not quite tje ordinary computers. They are specialized machines, though fundamentally, they share a lot of similarities with your home computer. Sometimes they even use many of the same components as your home computer.

Their distinction lies not just in their hardware but in their designated purpose, the software they run, and the specific use cases they cater to. Typically a dedicated server denotes hosting hardware exclusively assigned to a singular task or user.

Key Characteristics and Applications

1. Performance: Dedicated servers offer unparalleled performance/performance stability, crucial for high-traffic websites and large-scale e-commerce platforms.
2. Performance Vs. Cost: Typically if you need the resources, a dedicated server is the absolutely best value for your hosting needs. It’s not just a bigger slice of the pie, it’s the whole pie (server).
3. Security and Control: The exclusive nature of a dedicated server allows for heightened security measures and bespoke control, essential for sensitive data processing and customized applications.
4. Privacy: Privacy can be enhanced as well with increased security, data is less likely to leak when it’s not pooled together into a large cluster with multiple tenants.
5. Customization and Flexibility: Often Clients have the liberty to tailor these servers to their specific needs, choosing appropriate operating systems, hardware configurations, and software setups.

Dedicated Servers in the Broader Context

Interestingly all hosting services, including cloud services, fundamentally operate on dedicated servers. Even the cloud, often perceived as an abstract concept, is at it’s core a dedicated server in a datacenter. The Cloud, Is Just Someone Else’s Computer. The ‘cloud’ aspect emerges from the layers of software and virtualization built atop these physical servers.

Dedicated Servers vs. Other Hosting Options

• Shared Hosting: This is where server resources are divided among multiple clients, offering a cost-effective but less robust solution. This is typically on bare metal server, with limited isolation of customers from each others. Seedboxes typically are shared.

• Virtual Private Server (VPS): A VPS offers a balance, providing better isolation and resource allocation than shared hosting without the full commitment of a dedicated server. The main difference is the added isolation to shared hosting.
• Cloud Hosting: Virtual Private Server (VPS) service at it’s core, but as a cluster, or a cluster of clusters. These come with tools to quickly create new VPS instances to quickly respond to demand.

So what’s best? What Should I Use?

When it comes to selecting the right server for your needs, the answer isn’t always straightforward. The choice between dedicated servers and other forms of hosting, depends largely on your specific requirements, budget, and future scalability plans.

If your performance needs are small, typically a VPS or Cloud service is all that you need. Think about just a small thing, say a status page. These services cost less in absolute terms of money, while typically being relatively more expensive to resources – But if you are not going to use these resources to begin with? Then VPS or Cloud is the way to go, depending what level of scalability you need and your budgetary constraints.

However, what is a dedicated server good for? Lots and Lots of performance on the cheap. Lots and Lots of resources. If you need it all. Think of an dedicated server as a big apartment complex building, with a lot of condos / apartments, where the apartments are virtual servers, but the whole building is a dedicated server.

There are a lot of other nuances too, ie. type of performance, capacity, budget, scalability, management expertise etc. For example if you need a gpu server? Just few hours at a time, or 24/7? Or perhaps you just need very high performance torrents?

Cost of a dedicated server might not be the obstacle you think it might be

Here at Pulsed Media we specialize in deriving the maximum value for our customers, and the cost of an dedicated server might surprise you! Even if it is NVMe Unmetered server! Our MD Series nvme dedicated servers with 1Gbps Unmetered Dedicated Servers go some times as cheap as in the low teens during specials! So need cheap dedicated server? You know where to get one!

Conclusion: What’s a dedicated server?

Dedicated servers are akin to the foundational “ground truth” in the hosting world, a primary layer upon which various digital services and architectures are constructed. It is the skyscraper where a lot of offices and businesses reside in, do you need an office or the whole building?

Whether for high-traffic websites, specialized business applications, or hosting complex data-driven platforms, dedicated servers offer it all. The performance, The security, and The customization necessary to meet all demanding hosting needs. As the internet continues to evolve, the role of dedicated servers remains pivotal. Serving as the reliable backbone for a myriad of online services and infrastructures.

Wiki Guides

• Dedicated server — wiki documentation
• Pulsed Media datacenters in Finland
• RAID storage — RAID0 vs RAID5 vs RAID10

Knowledge Base

• What is a dedicated server?
• How is a dedicated server different from a VPS?
• Can I install my own software on a dedicated server?

The post What is a Dedicated Server? appeared first on Pulsed Media.

In the rapidly evolving landscape of hosting, the need for expansion is not just an ambition, but a necessity. We’ve long since recognized this. Today, we stand on the brink of a significant leap forward, one that has been years in the making and is poised to revolutionize the way we serve our customers, both dedicated server and seedbox customers a like. As we gear up to bring our second datacenter online in Kerava, it’s more than just an expansion—it’s a redefinition of our capabilities and a bold stride into the future of dedicated servers and seedbox hosting.

Our journey to this point hasn’t been without its challenges. The energy crisis of 2022, a watershed moment for many industries, particularly hit the datacenter sector hard. It was a stark reminder of the limitations we were facing with our current infrastructure. Our existing datacenter, while robust and reliable, had been operating at its maximum capacity. The constraints were not just in space or power, but also in the ability to adapt to rapidly changing technologies and client demands. Where do you build new things, when all your racks are packed to the brim? How do you scale after you have no more room for any more servers?

It became increasingly clear that to continue providing top-tier services and to embrace the burgeoning demands of our clients, a new approach was essential; Even long long before the energy crisis. A new datacenter, built from the ground up, with an eye towards scalability, efficiency, and the future. A facility not just designed to meet today’s needs but to anticipate tomorrow’s challenges.

The vision for our new datacenter in Kerava is not merely an extension of our physical footprint; It is a testament to our commitment to innovation. And oh my, innovate we must! With this new facility, we are not just scaling up our operations; we are elevating the standards of our hosting solutions. We are stepping into a realm where the limitations of our past become the stepping stones for a future brimming with possibilities.

This series of blog posts is more than just a chronicle of our new datacenter’s construction, located in Kerava, Finland. It’s a story of overcoming challenges, finding new possibilities, maximizing resources, embracing new tooling, vertically integrating, and relentlessly pursuing efficiency.

It’s about setting new benchmarks in the hosting industry and reaffirming our commitment to our clients. Join us on this remarkable journey, as we unfold the narrative of our bold new venture.

We are striving to achieve industry leading efficiency, by across the field combining skill sets and know-how; Re-analyzing every single standard, decision and nuance. What doesn’t make sense objectively, doesn’t belong here. Issues and problems are mere opportunities in disguise.

The post Building Our Future: The Story Behind Our New Datacenter appeared first on Pulsed Media.

Happy Holidays 2023!

What an Year, What an Year! A year both eventful, and uneventful at the same time. So many new MD Dedicated Server offers. Never before has the seedbox market seas been such calm to navigate, but challenges arose from the new server design paradigm with virtualization and SSD caching.

Meanwhile, the venerable MD Dedicated server platform development has progressed at rapid pace, a project decade in making coming finally together.

New datacenter build had some ups & downs, and has progressed. We might get into actual production by mid 2024.

It’s time for the holidays again, the year went by so fast! Let’s take a look back what’s been going on, while enjoying the holidays!

Highlights of 2023

MD Dedicated Server Platform Development Progress

This series already has more servers than our regular dedicated line up, albeit admittedly regular dedis from us cost multiples more, even as much as 15x more than a MD.

These nodes are already a mid-tier service group by revenue, just from these testing development units alone. We have already a huge inventory, and are looking at more than tripling the number of these nodes over 2024. Yes, tripling the number of nodes in Single Year. We are also quite certain that 2025 will also at the very least double the number of MD servers from upcoming ’24 totals.

We begun with a simple rudimentary proof of concept on each iteration, and have seen wild progression. We are already on version 3 of the base platform, a laser cut aluminium sheet 3mm in thickness. Version 4 is already planned.

Custom electronics enclosures got just designed and rolled into production to better protect some parts, with additional features; Each of the group of 8 now gets their own power meter which quickly shows power quality, current consumption, total kWh etc. metrics. Very Cool!

Everything’s custom; Enginerding the Dedicated Server To The Max!

More or less everything is custom, down to the basic wiring of the systems. To-date we’ve had to hand solder each and every power wiring for each and every node with the newer common PSU and Power management features, this has been a lot of work. Just lately we found a manufacturer to get ready made wiring pig tails which also allows us more flexibility in future; This alone could potentially save 2 hours of working time per platform. It was very fiddly to build the wiring looms. With the prefab parts we only need to cut to length and crimp terminals.

The base platform has been designed in-house, we get these fabricated in a Estonian workshop. We are now on iteration #3 with that.

Every server chassis’ needs modifications as well, mainly to improve cooling.

Every M.2 drive needs additional cooling modifications, some are permanent modifications, some are temporary. Screw mounted cooling solutions are both “expensive” and takes a long time to assemble manually; This typically does not matter but when you need to do say 50 in one go, you start to notice every minute of time spent. We are currently prototyping inexpensive permanent adhesive cooling solutions. These take very little human time, but needs to spend several days under pressure for the adhesive to solidify.

Even the rack rail mounts we are considering to design & build in-house; We couldn’t just find a sensibly priced off the shelf solution with the features we need. There are solutions, but if you need 600 rack rails, would you spend 60-100€ each and still have to deal with sub-optimal and slow to assemble rails. Meanwhile you know material costs are just in more 6€ ballpark for optimized, fast to assemble solution? Exactly.

Only the main components remain untouched, and off the shelf; Motherboard, CPU and RAM.

Tooling is everything to build a MD dedicated server

Tooling has been the most crucial part for this, finding the right tools, building the right tools and also finding just the right parts. That’s why we have concentrated on very narrow selection of base server chassis’.

We’ve spent a lot of time on tooling, and even buying a bunch of the same tool from different manufacturers just to pick the best one. When you are looking at volume production, every second counts.

Some tooling we built ourselves, such as assembly jigs.

Things like that is very important. Simple things, even what kind of parts bins you have, so we designed our own “Kanban” bins for the most common parts.

The maths are simple, if a tool costs 50€ but saves even 5 seconds per node, every time, all the time; It was very cheap investment. If a tool takes 10 hours to build at a cost of 100€, but saves 1minute per node for years to come; This too has a nice return on investment. That’s the special thing of scale; Everything starts to matter.

This is not “true mass production” even, Toyota defines mass production as 25 000 units manufactured. That’s long ways, but even in our tiny scale of hundreds of nodes annually it all starts to matter.

Some tools take their time to take into use and production, such as 3D Printers. Calibration and testing to the nuances of that particular printer. We have a small 3D Printing farm, and the higher flame retardancy plastics are quite difficult to print even on the best printers on the market.

Dynamic Pricing!

This got just enabled finally. This is in an essence an auction which dynamically changes the MD server prices based on demand. Both up and down. Updates happen every hour.

It has no hard limits, so server price could go really high or really low, but most likely it will find out the right price for the market over time automatically. We’ve got so many models, and people need so many different configs, and it all gets really hard to decide what this or that should be priced. So Why Not Let YOU Decide What’s A Fair Price?

That’s what it boils down to essentially. A service is worth what the buyer is willing to pay for it.

We already have 18 different models, and it’s not far fetched at all we will get to 100s over the upcoming decade. Se head over to our MD Dedicated Servers page to check the current pricing, for this very moment.

Kerava DC; DC2 Room1 Progress

DC2 located in Kerava, in locale we call with what translates to “Nuts” as in slightly crazy person familiarly, has been progressing. Albeit slowly.

Lots of nuances to consider, and having 8meter ceiling height adds it’s own challenges. But it has progressed.

We expect to finish epoxying / painting the place January – February, rest of construction by end of April and then it’s time to start installing infrastructure, electrical cabinets, outside air cooling fans, AC units, racks, fiber needs to be dug out and laid in the ground, fiber runs within the building (hundreds of meters across between workspace, DC2 Room1, building telco room) etc.

We are trying to be scrappy, but we might wind up still paying 2500-3500€ per rack ultimately in build out costs over the upcoming years.

Lowering Power Consumption

We made huge strides to lower our power consumption per user at the end of 2022, early 2023. Huge strides.

That being said, it’s not all been smooth sailing. Big and vast changes like that never really are, but considering the overall stability for the huge paradigm shift in question, it’s quite decent!

DC Deconstruction project in Vallila, Helsinki

We did a DC deconstruction project in Vallila, Helsinki. This had seen multiple owners over the 2 decades since it was built. We got quite a bit of valuable parts, some questionable value parts and some less so.

It was interesting, since this was considered top of the line back in the day. Cold air circulated to under the raised floor, and raised up from meshed or cut tiles.

Challenges Over 2023

Every year has it’s own challenges, and 2023 has been no different.

Electricity prices are still a challenge

Electricity rates are still roughly +50% at lowest, and have been increasing again. Part is at a set rate and part is spot priced. Spot pricing has come quite a bit of down, but we are paying much more than that currently. Average electrical rate tends to be more in the +150-200% range of what it was.

Nothing we can do about that but to adapt and hope rates get more sensible in the future.

New Seedbox Platform Challenges

We went all in for SSD Caching + Virtualization. This has turned out to be a challenge to manage. We had to build a lot of tools to manage this, and still managing most of the arrays can be a pain.

We have ways to overcome this however.

Another issue is mysterious, no error message, no log message freezes here and there. There is absolutely no indications what the issue is in any form. They all fail in a manner that automatic watchdog doesn’t get triggered; Meaning auto-restart doesn’t function. It’s been maddening, but fortunately there are long term solutions for this and it only affects tiny minority of seedboxes.

Increasing Pricing (and costs)

Inflation has pushed cost of many things much higher as well, and due to higher overall pricing in the market people have less disposable income to spend in luxury things such as Seedboxes. Meanwhile as our cost basis has dramatically increased, we cannot price things as competitively as before.

This has been a double whammy striking at sales from both directions.

This has resulted in less special offers, campaigns etc. overall, and a shift in what services people pick. A shift in focus.

World is changing rapidly: Artificial Intelligence

AI. It is rapidly advancing and changing things to a very large degree and fast. We use AI internally for a lot of tasks, and the advancements over 2023 alone are just stunning. We wrote about this at length in AI Revolution: ChatGPT’s Impact on Hosting Industry.

Barely a day goes by without some sort advancement. Case in point, GPT-4 generated lyrics with Suno generated song, Pulsed Media Holiday 2023 Song! Same Song, But Many Styles. All of these are single shot, no moderation and editing on the lyrics either.

Tell us which one you liked the most! Share these with friends and let us know down in the comments what you thought.

Check more tools out at FutureTools.io

These were just fun one-shot examples, but the power of GPT-4, Clause, Microsoft’s various Copilots are undeniable. Things are just faster to make in the digital space now, solutions are faster to find etc. Sure the best tools cost a little bit of money, but the power is well worth it; Just last week we had a backend office thing, 30minutes with ChatGPT and hundreds of euros per month saved in costs.

Less powerful LLMs you can also run locally on your laptop. AI is truly becoming ubiquitous and extremely powerful all of sudden. Don’t you find kinda funny that Microsoft’s ever hated Clippy is essentially becoming true now as Copilot? Edge browser has ChatGPT built-in for example.

Not everything becomes quicker tho, but simply better. In past years, it was a pain to find good illustrations and photos to include in these newsletters which were free to use. However, now we got better illustrations but a lot of time was spent with Midjourney V6 to generate them, a lot of iterations and somethings it simply failed, like these powerlines:

Those obviously are not real, and cannot be real. Things like that AI image generation struggles with. Further, some kinds of artistic things just cannot be done yet with image generators, infographics and lack of human eye for art.

Plans for 2024

Our plans for 2024 are “kinda boring”, more of the same, more of the same. Finish DC2, build more MD Dedicated Servers, upgrade some Seedbox servers. Develop and Innovate, and keep a keen eye on all things Artificial Intelligence. Lots of software development needs to happen during 2024 as well.

How about you? Tell us in the comments!

Holiday Special!

Oh my, what would a Pulsed Media Newsletter be without a special? So Here’s a few!

Special #1: 10Gbps Storage Box: 4TB RAID5 Storage, ONLY 4€ PER MONTH! Monthly Only. Grab Yours Now!

Special #2: M1000 1Gbps Unlimited Traffic* 4TB Seedbox; -50% FIRST PAYMENT CYCLE UPTO 3 YEARS! Can be compounded with long term discounts. Snag Yours Immediately!

Special #3: M10G 910G SSD Seedbox with 10Gbps and 4GB RAM, EPYC Servers: ONLY 9.99€ PER MONTH RECURRING! Monthly Only. Snatch It Now!

All offers end 1st of January, are limited in stock and only 1 per person. Price updates on checkout page.

Happy Holidays & Happy New Year 2024!

It’s been quite a year, and there’s lots to look forward in 2024. But for now, let’s all relax a bit and recharge for the upcoming year. See family & friends, perhaps spend time reading a book (or few).

Thank You!

The post Pulsed Media’s 2023: A Year Of Innovation And Challenges appeared first on Pulsed Media.

Super seeding with seedboxes is different from the traditional approach, rather than offering just consecutive blocks you utilize a swarm of seedboxes to supercharge content delivery. At Pulsed Media, we focus on distributing the maximum amount of data in the least possible time by utilizing seedboxes. This article will guide you on how to achieve maximum data distribution to serve thousands of users simultaneously.

Super seeding with seedboxes brings numerous benefits to content publishers and end users. By utilizing BitTorrent seedboxes effectively, large-scale content distribution becomes more efficient and reliable.

Super seeding ensures that end users receive their torrent downloads faster. By distributing data quickly and efficiently, the overall torrent download speed for end users is significantly improved, resulting in a better user experience.

Who is this article for?

This article is targeted at content publishers, tracker maintainers, and people who need their data distributed swiftly to as many users as possible. For example, companies like Blizzard release updates via BitTorrent and require a fast and efficient distribution strategy. Twitter is also known to distribute their server updates via BitTorrent.

What’s needed for super seeding?

To super seed effectively, you will need multiple seedboxes. A storage seedbox can serve as the initial source, while SSD seedboxes can be used as super seeding nodes. Using a multitude of similar SSD seedboxes can quickly outperform a single, big, and fast seedbox in performance. You can also use all the same type SSD seedboxes, it all depends on Your use case, number of torrents, their size etc.

We recommend using many of our shared slots to access as many discreet resources as possible without spending too much. For example, get a storage seedbox for the initial source and a bunch of SSD seedboxes as the super seeding nodes. Depending on how much data you have to distribute, utilizing only high performance SSD Seedboxes might be sufficient. Always use the same username when ordering your “super seeding swarm seedboxes” so that you can maximize the number of different nodes.

Torrent Super Seeding

1. Create your .torrent file with your preferred method directly on one of the fast main source seedboxes.
2. Transfer the .torrent file to the inexpensive swarm seedboxes for bulk transfer (ie. SCP).
3. Warm up the super seeding swarm by allowing the swarm seedboxes to download the data before release.
4. Release the torrent, ensuring that your end users receive their torrent downloads faster and more efficiently.

Large-scale content distribution

Large-scale torrent distribution is essential for content publishers who need to distribute petabytes of data quickly. This is what we do at Pulsed Media every single day, week and month. We are happy to help you achieve that goal cost-effectively.

Super seeding with seedboxes allows content publishers to scale their torrent distribution according to their needs easily. As the demand for your content grows, you can add more seedboxes to your network, ensuring that your content remains accessible and quickly distributed.

Super seeding makes your content rank higher and attracts more end users. When your torrents are distributed efficiently, they will rank higher in torrent search results, making it easier for users to discover and download your content.

High-performance super seeding ensures that data is always available to end users, even in cases where the original seed is offline or unavailable. By using multiple seedboxes, the data distribution process becomes more resilient and reliable.

Optimize Your content delivery and user’s torrent downloads

By using seedboxes strategically, you can significantly improve your content distribution strategy. Seedbox bandwidth plays a crucial role in distributing data to a large number of users, and the best way to scale that up is having multiple seedboxes. A high-performance seedbox setup is essential for ensuring that end users receive their data as quickly as possible.

Super seeding with seedboxes is a powerful method for distributing data to a large number of users in a short amount of time. Super seeding offers numerous benefits for content publishers and end users alike. You can ensure efficient large-scale torrent distribution, faster torrent downloads, and increased content visibility. Pulsed Media offers seedbox solutions to all of your needs, ensuring the best performance and value for your investment.

Wiki Guides

• How to super seed — step-by-step wiki guide
• Seedbox — what it is and how it works
• Troubleshooting seeding issues

Knowledge Base

• Stuck/paused torrents
• What to do with stalled torrents?

The post Super Seeding with Seedboxes: Boost Your Content Distribution appeared first on Pulsed Media.

Transitioning from the Information Age into the Knowledge Age, the role of Artificial Intelligence (AI) in all of our daily lives is becoming more prominent. Artificial Intelligence has revolutionized many aspects of how we live and work. How we interact with technology and information. As a leading provider of seedbox and dedicated server solutions, Pulsed Media has always been keen on the cutting-edge of technologies. In this editorial we will explore the impact of AI, specifically OpenAI’s ChatGPT, on the hosting industry. Let’s find out some of the potential benefits, as well as the dangers of embracing new AI technologies.

The AI Revolution

AI has been a game-changer in many industries, and hosting services is no exception to this. GPT-4, the powerful AI model behind ChatGPT, has been transformative force for hosting providers to enhance customer service especially.

The AI revolution is transforming the world as we know it, at a breakneck speed. Heralding a new era of unprecedented innovation and progress. As AI continues to develop and mature, it’s potential applications are virtually limitless and boundless. Spanning from healthcare and education, to transportation and entertainment. This technological revolution is reshaping industries and transforming the way we live, work and interact with the world. As AI becomes increasingly integrated into our lives, it is crucial to consider both its vast potential for positive or negative change. Embracing AI in a thoughtful and balanced approach, we can harness it’s transformative power to create a brighter, more efficient and sustainable future for all.

AI Revolution: Unknown unknowns

The AI revolution is not only transforming existing industries, but it is also opening up entirely new fields. Fields which we have yet to comprehend, or even imagine. Prompt engineer is now a job position you can apply for. The new coding language of choice is the English language. What other fields open up are still unknown unknowns, yet to be discovered. These hold the potential to redefine our understand of technology completely as we understand it. As AI continues to evolve, it may lead us to uncover previously unexplored areas of science, art and human potential. As new fields emerge, they will likely bring with them their own unique challenges and opportunities.

Embracing the unknown and fostering a spirit of curiosity and responsible innovation, we can prepare to navigate the uncharted territories that the AI revolution may reveal, ultimately unlocking the full potential of AI.

Blending AI and Human Expertise

To provide a unique perspective on this topic, the bulk of this article has been generated using AI, specifically ChatGPT GPT-4. After generation this has been human-moderated and partially rewritten to ensure easier readability, and shorter simpler sentences. GPT-4 likes to generate slightly over verbose and long sentences, which can be a little bit hard for us mere humans to understand quickly at times. Images in this article were generated with Midjourney version 5.

Benefits of AI for the Hosting Industry

This process is very curious, during writing this article we found potential new ways to utilize GPT-4 to make our services better, things we had not thought before. But also GPT-4 did not suggest some obvious ways, and this is how blending AI and Human expertise can compliment each other.

Let’s check out a few things how GPT-4 could help the hosting industry!

1. Improved Infrastructure Management: Leveraging AI, hosting providers could optimize server utilization, monitor hardware health, and anticipate issues before they arise. Resulting in a more efficient, stable and cost-effective service for clients.
2. Enhanced Customer Service: AI-driven chatbots, like ChatGPT, enable hosting companies to provide 24/7 customer support with minimal human intervention. Automating responses to common queries, AI allows support teams to focus on more complex issues and offer personalized assistance to clients.
3. Strengthened Security: AI-powered tools can analyze and detect potential security threats in real-time, helping hosting providers to safeguard their clients’ data and ensure uninterrupted service.
4. Predictive Maintenance: GPT-4 could monitor server performance data and identify trends that may indicate an impending failure or the need for maintenance. By predicting potential issues before they occur, hosting providers could schedule maintenance proactively, minimizing downtime and service disruptions. This proactive approach can also extend the life of server hardware and reduce overall maintenance costs.
5. Fraud Detection: One could use AI for quicker determination of fraudulent signups and orders. Scanning the data available of new purchase faster and with more detail than feasible for a human operator.

Dangers of AI, Risks of Moving Too Fast and trusting AI too much

While AI offers numerous advantages to almost everything, we have to remember the potential dangers as well. GPT-4 is showing sparks of Artificial General Intelligence (AGI), and here we enter the danger zone. We have to strike a balance between harnessing AI’s benefits, while mitigating the risks.

Unintended consequences is the biggest danger generally. As AI systems become more sophisticated, they could produce unexpected results or behave in unanticipated ways. The paperclip problem is a good thought experiment on this subject. Think about asking an AI to protect your child, but it does so by proactively hurting other people? It’s a little bit too late to say this is not what was asked for once other humans are hurt. The AI may not realize that was extreme bad behavior.

Danger of Lost Human Expertise

Loss of human expertise is a long term potential pitfall, as there is more and more reliance on AI there is a risk that human expertise could be lost or undervalued. Over time, this could lead to a lack of essential skills and expertise by humans. That would make addressing issues difficult when AI systems fail or encounter situations they are not equipped to handle.

We have some historical examples of loss of expertise, for example how many of us are still good at navigating without GPS? With paper maps? I know i could just take a brief glance on a map for a road trip, and find my way almost effortlessly to where i was going to even if i had not been anywhere near the area in my life before. Definitely would not attempt that anymore without having my phone as a backup for GPS. Does that really matter? Not typically. In tech industry there are small niches which has lost the expertise, and therefore pay extremely well, but with a danger that all the skilled people are now very old and getting to that niche is near impossible for new comers. Things relating to mainframes for example, things coded with COBOL.

Risk Mitigation

Mitigating these risks and harnessing the full potential of AI, it is also crucial to be cautious and have an balanced approach. Ensuring AI systems are transparent, explainable and accountable. Thus organizations could minimize the potential pitfalls and foster a safer and ethical progress. Importantly, the collaboration between AI and human expertise should be emphasized, leveraging the strengths of both to create more robust and effective solutions, safely. Just like how this article has been done; AI did bulk of the work, suggested many approaches and discussions, but ultimately almost every single word of this editorial has been written by a Human.

The Future of Life Institute and Open Letter

The Future of Life Institute (FLI) is a research organization that aims to ensure the safe and beneficial development of AI technologies. They have published an open letter outlining the potential dangers and benefits of AI. This open letter has been endorsed by numerous researchers, academics and industry leaders, such as Elon Musk. To summarize, the letter advocates for a collaborative approach to AI research. Focusing on maximizing the positive outcomes while trying to minimize harmful consequences.

More Neutral View of AI’s Impact

We strongly believe in the power of AI to transform every aspect of our lives for the better. However, there are also risks associated, both big and small. One has to careful check and review what the AI models are doing, and human oversight is needed. Example: When using AI to write response to a customer, it’s always completely and fully reviewed and moderated human. Staying informed and engaged in the ongoing rapid advances we can ensure, that we are doing our best to maximize the positive outcomes and minimize the potential negative outcomes for our clients.

In conclusion, as we transition to the Artificial-Intelligence Age, it is vital to recognize the potential of AI. Understanding the benefits and risks with AI. Thus we can ensure a more balanced and effective integration of AI into our daily lives.

At Pulsed Media, we are committed to staying at the forefront of this innovation, harnessing the power of AI to enhance our services for all our customers, and offer the best possible experience to our users. Whether you are interested in free seedboxes or dedicated servers, Pulsed Media is ready to meet your needs into the future and excitement AI advances brings to us.

The post AI Revolution: ChatGPT’s Impact on Hosting Industry appeared first on Pulsed Media.

Seedboxes has evolved over the years, and dedicated seedbox solutions have become increasingly popular among users seeking the best performance. At Pulsed Media we understand the need for high-quality, reliable services deployed rapidly. Our dedicated servers are excellent for your torrenting and streaming needs. Let’s check benefits of using dedicated servers as a dediseedbox, touching upon essential features such as streaming seedbox / plex seedbox, and more.

The Power of Dedicated Seedbox Solutions

When it comes to seedboxes, dedicated servers offer unparalleled performance and customization compared to shared seedbox hosting solutions. With a dedicated server, you have full control over the hardware and resources, ensuring that your dediseedbox operates at peak performance. This level of control and performance helps when using as plex seedbox and managing your media library. Dediseedboxes enable plex, but you can also use Jellyfin to stream your content. Jellyfin is supported on all Pulsed Media seedbox offers, but Plex can only be used with dedicated servers.

You can choose to maintain everything yourself, or get a semi-managed dediseedbox solution for ease of use, this can be a personal preference or skill set question for maintaining your own dedicated server.

Dediseedbox: A Premium Dedicated Seedbox Experience

Our dedicated servers solutions are designed to deliver the best value to our customers. With a dedicated server hosting plan, you can enjoy rapid upload and download speeds, increased storage capacity, and the ability to run custom applications tailored to your needs. This enhanced performance allows you to enjoy a seamless and uninterrupted experience when accessing your favorite content.

Plex Seedbox: Streamlining Your Media Library

One of the most significant advantages of using a dedicated server as a seedbox is the potential integration with the popular media streaming platform, Plex. A plex seedbox allows you to stream and manage your media library effortlessly, providing a convenient and centralized platform for all your entertainment needs. By combining the power of dedicated seedbox server with Plex, you can enjoy a premium media streaming experience that is both fast and reliable.

Enhanced Security with Dedicated Servers: Full Disk Encryption and More

In addition to the performance benefits, dedicated servers can also provide enhanced security features for your seedbox. Dedicated server gives you complete control over your environment, allowing you to implement more robust security measures, like full disk encryption. Full disk encryption ensures that all data stored on your server is encrypted, safeguarding your sensitive data and media files. Moreover, dedicated seedbox solution enables you to configure and manage your own security settings, for example firewall, automated detection and monitoring software. Implementing these features for your dedicated server hosting plan at Pulsed Media, you can enjoy peace of mind knowing that your seeedbox is equipped with the security features you want, protecting your valuable data.

The Best Seedbox for Your Needs

We understand that the term “best seedbox” is subjective and varies depending on individual requirements and preferences. How feral your performance needs are? How rapidly do you need the server deployed? Casual torrent download needs, or perhaps plenty of storage for your backups? That’s why dedicated server hosting come in various configurations, designed to cater to a wide range of users. Whether you’re a casual user looking for a powerful dediseedbox solution or a power user in need of a highly-customizable platform. Our dedicated server solutions offer the flexibility and performance you need.

Conclusion on dedicated servers as a seedbox

At Pulsed Media, our dedication to providing the best seedbox solutions is evident in our wide range of hosting plans, including our dediseedbox offerings. By choosing a dedicated server hosting plan, you can enjoy the benefits of a plex seedbox, faster torrents, increased storage capacity, and the ability to run any custom applications you need. Experience the power and flexibility of dedicated servers as a seedbox and unlock top performance for your seedbox needs. Discover the perfect solution for you by exploring our dedicated seedbox plans today!

The post Dedicated Servers as a DediSeedbox: Best Seedbox? appeared first on Pulsed Media.