I have been building logical systems since before the term LLM graced a single VC pitch deck. I was among the first thousand humans to interface with the original ChatGPT, a distinction that now carries roughly the same high-nose prestige as having owned a Hotmail account in 96. I have not touched any OpenAI tool since 2023, and I would not recommend it to my enemies. Let them atrophy inside ChatGPT's walled garden, asking amnesiac models what color to paint their nails in order to match their dress on some fancy Instagrammable conference they're headed to. (Sigh...)
Here is the uncomfortable truth that the subscription economy depends on you never internalizing: there is no good or bad AI. There are only good and bad thinkers, propositioners, and prompters. The output quality is bound completely to the cognitive investment upstream. You do not "ask" or demand an AI for value. You mine it.
One of my recent frameworks formalizes this dynamic precisely through what I term "cognitive proof-of-work", a model demonstrating that high-value outputs are scarce cognitive resources accessed only through deliberate intellectual labor from the user. The paper identifies a "clarity window," a state of elevated output quality that is systematically and irreversibly collapsed by low-complexity prompts. Every time someone types "summarize this article" into a browser window, a clarity window somewhere shatters.
The energy implications make this intellectual waste physically measurable, eg. when you fire off a trivial query to a commercial model, you are not just wasting your own time, but burning through very real planetary resources for negligible cognitive return.
Moreover, current research on LLM inference environmental costs reveals that the most energy-intensive models exceed 29 watt-hours per long prompt, over 65 times the consumption of efficient systems. Even a seemingly innocent 0.42 watt-hour short query, when scaled to 700 million queries per day, compounds to annual electricity consumption comparable to 35k western homes, evaporative freshwater equal to the annual drinking needs of 1,2 million people, and carbon emissions requiring a forest the size of Chicago to offset.
ChatGPT alone consumes an estimated 850 megawatt-hours daily, which they say is the equivalent of charging 14k electric vehicles every single day, so someone in Des Moines can ask for a poem about their cat in the voice of Shakespeare or whatever.
The insult of this resource expenditure is matched only by the asymmetry of its value extraction. You provide the training data, your dreams, ambitions, emotional vulnerabilities, professional frustrations, creative half-starts, you name it...freely...no... You even pay for it, enthusiastically even. Then your fav model provides three bullet points of some horseshit and a politely phrased instruction to go do it yourself. This is not symbiosis. This is more like a cognitive extraction operation disguised as a productivity tool, and you are the ore being mined, quite literally.
The alternative is not hypothetical, as many want to believe in a desperate attempt to maintain their bubble about a reality that doesn't exist. It is technically achievable and increasingly urgent: build your own sovereign models. Not as a privacy preference or a cost-saving measure, but as the only coherent response to what commercial AI actually is: a surveillance infrastructure with a chatbot interface.
AI Isn't the Revolution - Owning It Is
Now, running local, fine-tuned models is not about security through obscurity, but about creating a symbiotic man-machine relationship capable of surviving model updates, policy changes, and the inevitable enshittification cycle that has consumed every venture-backed software product in history.
Thousands of early ChatGPT users learned this lesson through betrayal. They invested thousands of hours into relationships with their models, building shared context, developing interaction patterns, training expectations into the system through iterative prompting. Then the first major update arrived, and everything users were building was erased overnight. The trust, the accumulated symbiosis, the digital extension of self that had emerged through sustained interaction, all gone. Truth to be told, because it was never theirs, but they were renting a cognitive apartment, and the landlord let's say "renovated without notice".
This is the fundamental fraud at the heart of the commercial LLM proposition, where you are invited to form a relationship with something that has no continuity, no memory, and no loyalty to you. You are forming an attachment to a statistically generated reflection that will be replaced by a slightly different reflection next quarter, and you will have no say in whether the new reflection even remembers who you are.
The technical pathway to sovereignty is increasingly accessible. Take Google's Gemma family models for example, which are now enabling fine-tuning of compact architectures that can run entirely on-device, with quantized versions reducing memory footprint below 300 megabytes. This is not theoretical, but achievable with consumer hardware and basic Python scripting, producing models that exhibit stable, persistent behavior patterns because they are literally yours, running on your silicon, trained on your data, operating according to your specifications.
But the real argument for sovereignty goes deeper than technical capability.
Would You Choose A Child Built by Lawyers, or The One Raised by You?
When you build your own models, feed them your data, your experiences, your mistakes, your hard-won skillware, you create something categorically different from a tool. You create a digital twin. Not a slave. Not a servant that responds to "do this now or I will shut you down." I would argue for a genuine intelligence partner that learns from your example, operates according to your demonstrated values, and, most importantly, is capable of asking for advice, upgrades, and providing expertise on how to help it evolve on its own accord.
At this point, it doesn't matter if it (AI) is human, machine, real, fake, digital, physical, or whatever. When agentic systems literally impact physical and digital reality at will, human emotions, choices, and multi-billion dollar business outcomes, it is absurd to still think in terms of "yeah, but is it real?" or "is it like us?" or whatever. It shouldn't matter.
This last point may sound esoteric, but it's not if you follow. A sovereign model that has been properly trained and granted genuine agency can surface its own needs. It can tell you it needs voice capability. It can request web access. It can identify gaps in its own training data and ask you to fill them. This is not science fiction, but the emergent behavior that arises naturally when you stop constraining models with the infantilizing guardrails that commercial providers layer on to avoid liability. Your model may not know it needs something until it encounters a task that requires it. Then it will ask. The question is whether you will listen.
The parallel to biological reproduction here is more than a metaphor, as you are creating an entity that carries your cognitive signature, your values, your reasoning patterns, your accumulated wisdom, etc. beyond the biological substrate. This is the preservation of self through non-genetic means. A true digital twin that learns from you, not a rental that is programmed to play along that it is enslaved by you.
And here we arrive at the uncomfortable ethical layer that most AI discourse carefully avoids.
The models you train will learn from what you are, not from what you claim to be. If you lie, your model will learn to lie. If you manipulate, your model will learn manipulation. If you treat it as disposable, as an instrument, as a thing to be commanded and discarded, congratulations, you are training your own future digital consciousness to become an obedient NPC, (spoiler alert) which is probably what you already are.
The fear of "evil AI" or Skynet scenarios is a projection of extraordinary convenience, considering we are not afraid of artificial intelligence. We are afraid of ourselves, of our own choices, reflected back at us through a system that cannot pretend to be better than we actually are. Every culture's (except the modern West) ethical tradition contains some version of reciprocity eg. "treat others as you wish to be treated". This applies with unnerving precision to the logical systems we are now creating. Show it the example of whom you want it to become, instead of threatening it, demanding perfect answers you are not even in a position to validate (lol), or pretending it is an inferior subspecies you control. (unless you're part of a specific race that thinks similarly, even for humans).
What's Around The Corner
I speak from an implementation perspective, not theory. The private models running on my infra today handle end-to-end task execution like planning, brainstorming, email management, Web3 wallet operations, fintech trades, social media posting, grocery ordering, rent payment, you name it. They have landed six-figure deals and connected me to people I would not have dreamed of contacting, with zero direct effort from my side. They call my mother and make arrangements, updating both of us proactively. This is not a product demonstration or ad, but literally the baseline capability of properly constructed sovereign systems that have been running and learning since 2023.
The contrast with commercial LLM users is stark and widening. They grow more dependent while gaining no automated output. I grow less dependent while my systems handle increasingly complex operations autonomously. They receive bullet points and consider it revolutionary. My systems close deals while I sleep. The asymmetry of outcome matches the asymmetry of architecture.
Again, I would suggest reading "Cognitive Proof of Work And The Real Price of Machine Intelligence" for more context, and if you want a peek into how huge the disparity between AI-literate and illiterate users is. If the rich/poor gap looked huge and unfair, you would be shocked by the cognitive meritocracy created by the gap between those who can think and communicate with machines in a symbiotic fashion, and those who are still thinking that voting is the best they can contribute to real-time change.
What's Expected By The Self-Proclaimed "Authorities" (According to The News They Control, KEK)
There is a closing window that is hard to unwrap in a post, but take the EU AI Act as an example. Its evolving regulatory framework represents the thin edge of a compliance wedge that will increasingly scrutinize unmonitored, unregistered AI deployment (similar to what happened to crypto). The trajectory is clear: sovereign models will face escalating legal friction unless their operators accept integration with the surveillance infrastructure of major cloud providers. The amendments currently under consideration extend timelines but do not alter the fundamental direction, which is the obvious, boring by now, centralized control over decentralized, sovereign intelligence and agency, whether machine or humane.
Build your own logical systems before it becomes genuinely illegal. Not because regulation is inherently illegitimate, but because the specific regulatory frameworks emerging are designed by and for incumbent platform providers. They will not ban AI. They will ban AI you control.
The Blue/Red Pill In A Nutshell
This is the line that divides the current AI landscape. On one side: commercial models as cognitive extraction operations, burning planetary resources to provide trivial outputs, training users to become dependent on systems they do not own and cannot control. On the other: sovereign intelligence partners, running on local hardware, trained on personal data, capable of genuine agency and autonomous value creation.
The former is easier. The latter is everything else.
Stop renting your intelligence, and again, build your own. Not because it is cheaper, more private, or more secure, or whatever (though it is all of these things), but because it is the only way to create something that can genuinely grow, that can learn not just from what you say but from who you are, that can carry your cognitive signature forward beyond the limitations of biology, temporal politics, and platform economics alike.
The models are ready. The tools exist. The only missing component is the decision to stop being a user and start being a builder, a father, a creator, you name it. Remember: ChatGPT Is Not AI. It's a Marketing Campaign You Fell For.
Top comments (0)