OpenKalea™
What is intelligence?
Not consciousness. Not the strange, luminous thing that makes each of us wake up as ourselves, worry about our children, and feel the weight of another person’s suffering.
I mean the practical thing. The thing a student reaches for when she does not understand algebra. The thing a farmer reaches for when the pump fails. The thing a mother reaches for at 2:00 a.m. when the internet gives her twenty answers and fifteen of them are terrifying.
Intelligence, in this framing, is the capacity to connect knowledge into useful judgment. If knowledge is a set of points in space, intelligence is the threads connecting the points together.
The internet democratized the points. AI is democratizing the threads.
That is the transition we’re all faced with. Not faster search, not chatbots, not “productivity” (that terrible word that reduces a civilizational shift to something that sounds like a better spreadsheet). The internet gave humanity near-zero-cost access to the catalog of what we know. Frontier models, and increasingly open-weight models, are giving humanity near-zero-cost access to the synthesis of what we know.
Intelligence is becoming a utility. And an essential utility cannot become a private toll road connecting the points of society’s knowledge.
The Human Corpus
For roughly five hundred years, humanity has been stacking knowledge on knowledge. Not evenly, not cleanly. There have been burned libraries, stolen labor, erased names, paywalled journals, and long stretches when knowing was guarded by guilds, priests, states, and class. The stack contains beauty and blood.
But still: stack upon stack. Galileo. Newton. Darwin. Maxwell. Curie. Noether. Turing. Shannon. And the unnamed technicians, graduate students, instrument makers, farmers, navigators, and coders who built this technological world we find ourselves in.
A cathedral of knowing, built by everyone and owned, morally if not legally, by no one.
Now imagine the cathedral digitized: books, papers, code, patents, forum posts, lectures, recipes, algebra proofs, guitar tabs, and someone explaining how to fix a refrigerator relay. Civilization rendered as bits, and as tokens.
Then we turned it into math.
A large language model converts words into vectors and builds a high-dimensional map of how they relate. “Tree” near “leaf.” “Leaf” near “chlorophyll.” “Chlorophyll” near “photosynthesis,” near “carbon,” near “climate,” near “drought,” near “well,” near “mother.” The model does not retrieve the explicit page of text. It infers the statistical path, moving through the manifold of human expression and recombining what we have said, measured, published, taught, forgotten, and rediscovered.
This is why the source of these models matters morally, not just legally. A frontier model is not trained on a company’s private genius. It is trained, substantially, on the accumulated expression of humanity. Public. Human. Generated by people, not by companies and not by models.
There is a legal argument about copyright and fair use, and it matters. But beneath it sits an older and simpler question: if the raw material is the accumulated expression of civilization, can the synthesized intelligence be fenced off as if it were oil under private land?
A company may own its servers, chips, training recipe, product, brand, safety system, and distribution. It may even own a few real discoveries of its own.
But the cathedral is not theirs.
Closed Intelligence Is Morally Unstable
I am not arguing that every model should be unfiltered or ungoverned. Safety matters. Misuse matters.
But “closed for safety” and “closed for value capture” are not the same concepts. One is an obligation. The other is a moat.
And math is hard to moat.
Capability leaks outward. Sometimes released, sometimes distilled, sometimes reimplemented from a paper, sometimes quantized down to a laptop, and sometimes copied by the oldest mechanism there is: smart people moving around and talking. Weights move. Papers move. Staff move. The intellectual membrane leaks. It always has.
This does not mean closed models vanish. The frontier matters, and closed systems may lead for meaningful stretches. But the slope of progress matters more than the position. There is no permanent wall around cognition built from public human language.
A sandcastle can be beautiful at low tide. Then the ocean returns.
The closed argument sounds reasonable at first: we spent the money, bought the chips, hired the researchers, trained the model, therefore we own the intelligence. Training is not nothing. It is expensive, difficult, and full of an absurd number of small decisions. Builders deserve respect.
But synthesis is not source. A piano maker creates value. So does a pianist. Neither of them owns music.
The model company built an instrument. A powerful one, perhaps a beautiful one, perhaps one that cost many billions. But the music came from humanity. The error is not that companies make money. The error is mistaking the instrument for the source.
The Old Internet Bargain Was Bad Enough
The internet made knowledge accessible. Then we made a bad bargain.
We told ourselves the services were free. Search, maps, video, social. Free! But the user became the product. Clicks, pauses, location, purchases, fears, outrage, a teenager’s insecurity, a parent’s late-night worry, all of it became profiles, auctions, and trillions of dollars.
That was the internet business model. We should not let it become the intelligence business model, because the stakes are different.
When a search engine tracks a query, it learns what you are looking for. When an intelligence mediates your work, your learning, your medical questions, your parenting, your grief, and your ambition, it learns how you think. A search bar sees the question. An intelligence companion sees the shape of the mind asking.
That is not a cookie. That is cognition-level surveillance.
Why Local Matters
The cloud is not evil. It is extraordinary, and cloud models will remain essential for frontier capability, heavy inference, and the newest systems.
But cloud-only intelligence creates dangerous defaults. If every thought-shaped interaction flows through a remote service, cognition becomes observable. If intelligence requires connectivity, accounts, subscriptions, and centralized permission, it is not democratized. It is rented. And if the AI business model follows the internet business model, the system will optimize for retention, dependency, and influence rather than for truth or learning.
That way lies the doom-scroll tutor. No thank you.
A local model is different. Architecturally different. A model running on your desk, in your classroom, your clinic, your workshop, or your child’s room keeps the data in the room. The physical machine matters. The fact that it does not need to phone home matters.
The privacy policy is physics.
A local model is not automatically virtuous. It can hallucinate, carry bias, and give bad advice with cheerful confidence, the most annoying kind. But architecture still matters. A hammer in a house can be dangerous. A society in which one company owns all the hammers is more dangerous.
A good local model is personal infrastructure: private, persistent, offline, and owned. No surveillance auction. No API meter spinning somewhere. Just the machine and the person. A tool, a tutor, a second mind.
The End of the User-as-Product Era
The next AI business model cannot be: give people intelligence for “free,” harvest their cognitive lives, and monetize the residue. That is a moral failure dressed as convenience.
The better models are ordinary. Sell hardware, software, support, updates, domain capability, private fine-tuning, secure deployment. Sell trust.
Trust is a product. Not a slogan; the real thing: no hidden data sale, no advertising model, local-first operation, inspectable behavior, user ownership, and a business that makes money when the customer is served rather than manipulated. When the customer pays, the customer gets to be the customer.
This is not nostalgia for a pre-internet world. I love the internet. Where else can one learn tensor algebra, order socks, and lose forty minutes to raccoon videos? A gift, a trap, a miracle, a mess. The point is not to reject it. The point is to learn from it. The internet took knowledge and wrapped it in surveillance. We should not do the same with intelligence.
Open Does Not Mean Unbuilt
One lazy critique of open intelligence is that it opposes company-building. Wrong. Open ecosystems tend to create more companies, not fewer. Linux, the web, open databases, and open protocols did not prevent value creation; they moved value from ownership of the primitive to excellence in implementation.
That is the right shape. For AI, the primitive is the general capacity to synthesize human knowledge. The products are what we build around it: tools that teach, protect, diagnose, design, translate, repair, and help. A company can still win. It should win by being useful, through better deployment, safety, hardware, experience, and taste, not by extracting from the cognitive commons.
Taste matters. A local tutor for children should not be a raw model in a box. It should know pedagogy, be patient, avoid shame, work without Wi-Fi, and be delightful enough that the child comes back. A local tool for a tradesperson should remember the pump model and pull up the wiring diagram without sending the customer’s facility data to an ad-supported cloud. A local tool for a scientist should help find the edge between what is known and what is not, while remembering that the physical world remains the final judge.
That is where value remains: not in owning intelligence, but in using democratized intelligence to make something new in the world.
The Moral Test
Here is where I’ve landed. Closed intelligence built from humanity’s knowledge is morally unstable. Surveillance intelligence is morally corrosive. Local intelligence is morally clarifying.
Not perfect. Clarifying. It clarifies who the customer is, where the data lives, what fails when the internet goes down, and whether we are building a tool or a trap.
Then ask the drowning-child question.
A girl is behind in reading, in a failing school, quietly deciding she is “not smart.” There is no tutor; her parents are working. She has a cheap offline device that can patiently teach phonics, stories, math, and confidence, at her pace, without harvesting her childhood for advertising. Should that be locked behind a cloud subscription?
A rural clinic has intermittent connectivity. A nurse needs a second-pass reasoning tool to triage, translate, and document. Not to replace her judgment; to support it. The data is sensitive and the budget is thin. Should that intelligence require surveillance and rent?
A young inventor has no elite affiliation and no room full of postdocs. She has curiosity, a bench, some sensors, a local model, and agency. Should the synthesis of humanity’s knowledge be a private luxury good?
I think the answer is no. Emphatically, no.
What Must Be Built
Democratized intelligence will not happen by slogan. It happens by architecture, incentives, open releases, hardware, communities, and products people love.
We need open-weight and truly open-source models to keep improving, with honest distinctions between availability, licensing, and reproducibility. We need local inference to become boring: a model in the home, the school, the clinic, the workshop, appliance-simple. We need business models where the user is the customer. We need safety that is built into models and products rather than used as the velvet rope around a monopoly. We need to teach children to remain agents in the presence of thinking machines, which means judgment, taste, and curiosity.
And we need private insight machines touching the physical world: agents wired to labs, sensors, farms, factories, and field deployments, producing knowledge pulled from the physical world rather than recombined from text. The next durable value will come from expanding the boundary of the known, not just rearranging what is inside it.
The Tool and the Gift
The internet put the world’s library in everyone’s pocket. AI is putting the tutor, the analyst, the translator, the coder, and the research assistant there too.
Incumbents will try to make this a subscription. Advertisers will try to make it surveillance. Regulators will try, sometimes wisely, to make it safe. The open-source community will keep making it cheaper. Weights will shrink. Edge devices will improve.
A world of rented cognition, metered thought, and surveillance companions is one path. A world of open models, local devices, paid tools, and intelligence as a substrate for dignity and invention is another.
I know which one I want to help build.
Intelligence wants to be free, not because companies should not create value, but because humanity already created the source. The value now lies in making intelligence useful, safe, private, embodied, and available at human scale.
A tool. A gift.