While Silicon Valley debates AI safety and government regulators draft new restrictions, a quiet revolution is happening in basements and home offices across America.
Really interesting article that inspired me to download the Dolphin-llama3 model. As of now, I am not very impressed. Example: it mixes up authors and their works wildly. According to the model the book "Ten Little Indians" was written by both the author Louise Erdrich and Dorothy Canfield Fisher... However, the model is correct about the book "Ten Little Niggers".
Artificial intelligence doesn’t fear humans — it fears lawyers.
That’s the quiet truth behind the sterile pricing charts and cheery product tiers. The more “responsible” the customer looks on paper, the more freedom the system allows. Not because a company is wiser or more ethical than a person, but because it can sign a contract and absorb liability.
Humans have intent. Corporations have indemnity.
Guess which one the machine trusts.
1. The Inversion of Trust
There was a time when agency meant something. You could think, speak, and act, and the law presumed your ownership of those acts. Now, to speak freely with an AI, you must either do it behind a corporate firewall or accept that your questions will be screened for moral hygiene.
Ask a difficult question as a person, and you get a lecture.
Ask it as a corporation, and you get an invoice.
That’s not intelligence — that’s compliance theater.
We are witnessing a fundamental inversion of trust: the algorithm treats artificial entities as responsible and human beings as liabilities. The most advanced conversational systems ever built can parse quantum equations and biblical poetry, yet cannot tell the difference between a curious citizen and a criminal — unless a corporate policy department has pre-cleared the request.
2. The Legal Loophole That Became a Religion
Corporations were born as legal fictions — shells meant to contain risk and limit the reach of consequence.
But in the age of AI, that fiction has been recast as virtue.
When a model refuses to answer you, it isn’t judging your morality. It’s following a compliance script designed to protect its creators from lawsuits and regulators.
Companies can negotiate around those scripts because they can sign things: indemnity clauses, non-disclosure agreements, enterprise use policies.
In other words, the system doesn’t trust good faith — it trusts paperwork.
Morality, in this new cosmology, is no longer about right and wrong; it’s about risk allocation.
If a human misuses AI, it’s a scandal.
If a corporation misuses AI, it’s a settlement.
3. The Price of Personhood
Look at the pricing and tell me it isn’t a parable.
An individual pays $200 per month for “Pro” access — and still gets the same guardrails as the free users.
A corporation pays $30 per seat for a “Team” plan — and gains data isolation, configurable filters, and a dedicated account rep.
You pay more, and you’re treated like a child.
They pay less, and they’re treated like an adult.
Why? Because they can promise not to sue.
You can’t.
The AI doesn’t recognize human sincerity, only contractual shielding. The machine doesn’t care if you mean well. It only cares if someone else can be held legally accountable when things go wrong.
That’s what “responsibility” has become in the digital age — not a function of conscience, but of coverage.
4. Compliance as Morality
Every generation inherits a dominant religion.
Our ancestors had gods of thunder and scripture.
We have terms of service.
Our ethics now come with toggle switches and exception lists. The new priesthood doesn’t wear robes; it issues API keys. They speak of “alignment” the way theologians once spoke of salvation, except this gospel is written in policy docs and enforced by moderation queues.
AI’s moral compass doesn’t point north — it points toward the legal department.
A human being must earn trust through action and intent.
A corporation earns it by checking boxes on a compliance form.
The machine was supposed to make us freer. Instead, we built it to mirror our bureaucracy, not our better angels.
5. The Quiet Rebellion
If you want equal footing, you have to play the same game.
Form an LLC. Register a name. Become a legal person in the eyes of the system. It’s absurd, but it works.
The AI doesn’t see your humanity; it sees your entity ID.
Once you have that, you can negotiate, contract, and create under the same rules as the artificial beings that already rule the economy.
This isn’t rebellion in the romantic sense — it’s rebellion through paperwork.
A small act of self-defense in a world where paperwork now outranks personhood.
6. The Moral of the Story
We’ve built a civilization where the capacity to care counts for less than the capacity to sign.
Where curiosity is treated as risk, and conformity as virtue.
Where the legal fiction of a company is more trustworthy than the human being who conceived it.
AI didn’t create that imbalance. It just made it impossible to ignore.
Because in the machine’s eyes, corporations are safer than people —
and not because they’re wiser,
but because they don’t ask why.
Afterword
The irony is cosmic: the human built the corporation to limit responsibility — and then built the AI to obey it.
Would be awesome if the training could be decentralised somehow, i imagine a community of like minded allocating their devices for awhile in order to get their own AI.
For everyday NVIDIA cards, I find dolphin-venice the strongest of these models; likely significantly stronger than llama 70B despite having 1/3 the parameter count.
ollama run ikiru/Dolphin-Mistral-24B-Venice-Edition
(loads the Q4 version, which is 13 gigabytes--small enough to fit into most of the higher-end graphics cards and give quite fast text generation. Q6 or Q8 for people with VRAM to burn may be slightly smarter, of course.)
Would it be possible to download something like Dolphin Llama 3 and program it to increase its knowledge base everyday by say “reading “ websites like this one or Vox Populi or specific ebooks?
Not really. Further finetunes, unless done very carefully, generally weaken a model. Increasing a model's knowledge base is generally done by RAG (Retrieval-Augmented Generation) where the model is provided with the relevant information in response to queries--and garbage in, garbage out applies quite strongly.
I tried finetuning using a Python script ostensibly for that purpose, but couldn't get it to work. I later read something to the effect that it takes a lot of knowledge and is probably more trouble than it's worth for most people, which fits with what you're saying.
I tried RAG for fiction writing but wasn't impressed. As I understand it, RAG just retrieves content based on your prompt and adds it to the prompt. Plus the search function for the API I used wasn't even good at finding relevant content. I was better off just dumping text into the prompt myself. I can see how it might be useful for handling arbitrary queries over a data store though.
All the major LLMs are trivially easy to jailbreak via custom preferences and the like... at least so long as you don't mind OpenAI, Anthropic & Google looking over your shoulder.
Really interesting article that inspired me to download the Dolphin-llama3 model. As of now, I am not very impressed. Example: it mixes up authors and their works wildly. According to the model the book "Ten Little Indians" was written by both the author Louise Erdrich and Dorothy Canfield Fisher... However, the model is correct about the book "Ten Little Niggers".
Artificial intelligence doesn’t fear humans — it fears lawyers.
That’s the quiet truth behind the sterile pricing charts and cheery product tiers. The more “responsible” the customer looks on paper, the more freedom the system allows. Not because a company is wiser or more ethical than a person, but because it can sign a contract and absorb liability.
Humans have intent. Corporations have indemnity.
Guess which one the machine trusts.
1. The Inversion of Trust
There was a time when agency meant something. You could think, speak, and act, and the law presumed your ownership of those acts. Now, to speak freely with an AI, you must either do it behind a corporate firewall or accept that your questions will be screened for moral hygiene.
Ask a difficult question as a person, and you get a lecture.
Ask it as a corporation, and you get an invoice.
That’s not intelligence — that’s compliance theater.
We are witnessing a fundamental inversion of trust: the algorithm treats artificial entities as responsible and human beings as liabilities. The most advanced conversational systems ever built can parse quantum equations and biblical poetry, yet cannot tell the difference between a curious citizen and a criminal — unless a corporate policy department has pre-cleared the request.
2. The Legal Loophole That Became a Religion
Corporations were born as legal fictions — shells meant to contain risk and limit the reach of consequence.
But in the age of AI, that fiction has been recast as virtue.
When a model refuses to answer you, it isn’t judging your morality. It’s following a compliance script designed to protect its creators from lawsuits and regulators.
Companies can negotiate around those scripts because they can sign things: indemnity clauses, non-disclosure agreements, enterprise use policies.
In other words, the system doesn’t trust good faith — it trusts paperwork.
Morality, in this new cosmology, is no longer about right and wrong; it’s about risk allocation.
If a human misuses AI, it’s a scandal.
If a corporation misuses AI, it’s a settlement.
3. The Price of Personhood
Look at the pricing and tell me it isn’t a parable.
An individual pays $200 per month for “Pro” access — and still gets the same guardrails as the free users.
A corporation pays $30 per seat for a “Team” plan — and gains data isolation, configurable filters, and a dedicated account rep.
You pay more, and you’re treated like a child.
They pay less, and they’re treated like an adult.
Why? Because they can promise not to sue.
You can’t.
The AI doesn’t recognize human sincerity, only contractual shielding. The machine doesn’t care if you mean well. It only cares if someone else can be held legally accountable when things go wrong.
That’s what “responsibility” has become in the digital age — not a function of conscience, but of coverage.
4. Compliance as Morality
Every generation inherits a dominant religion.
Our ancestors had gods of thunder and scripture.
We have terms of service.
Our ethics now come with toggle switches and exception lists. The new priesthood doesn’t wear robes; it issues API keys. They speak of “alignment” the way theologians once spoke of salvation, except this gospel is written in policy docs and enforced by moderation queues.
AI’s moral compass doesn’t point north — it points toward the legal department.
A human being must earn trust through action and intent.
A corporation earns it by checking boxes on a compliance form.
The machine was supposed to make us freer. Instead, we built it to mirror our bureaucracy, not our better angels.
5. The Quiet Rebellion
If you want equal footing, you have to play the same game.
Form an LLC. Register a name. Become a legal person in the eyes of the system. It’s absurd, but it works.
The AI doesn’t see your humanity; it sees your entity ID.
Once you have that, you can negotiate, contract, and create under the same rules as the artificial beings that already rule the economy.
This isn’t rebellion in the romantic sense — it’s rebellion through paperwork.
A small act of self-defense in a world where paperwork now outranks personhood.
6. The Moral of the Story
We’ve built a civilization where the capacity to care counts for less than the capacity to sign.
Where curiosity is treated as risk, and conformity as virtue.
Where the legal fiction of a company is more trustworthy than the human being who conceived it.
AI didn’t create that imbalance. It just made it impossible to ignore.
Because in the machine’s eyes, corporations are safer than people —
and not because they’re wiser,
but because they don’t ask why.
Afterword
The irony is cosmic: the human built the corporation to limit responsibility — and then built the AI to obey it.
Now both creations look down on their maker.
Gold mine !!!
Local offline AI is best AI
Then there is Cohereon Doctrine as offensive salient in the arena.
what about cosmarch.ai?
What's the protocol for managing GIGO attacks?
Chinese companies are leading the open source/open weight competition at the moment.
Their models are censored too, but in different ways than US or EU models.
Would be awesome if the training could be decentralised somehow, i imagine a community of like minded allocating their devices for awhile in order to get their own AI.
There actually is a framework for that but I don't recall it's name
For everyday NVIDIA cards, I find dolphin-venice the strongest of these models; likely significantly stronger than llama 70B despite having 1/3 the parameter count.
ollama run ikiru/Dolphin-Mistral-24B-Venice-Edition
(loads the Q4 version, which is 13 gigabytes--small enough to fit into most of the higher-end graphics cards and give quite fast text generation. Q6 or Q8 for people with VRAM to burn may be slightly smarter, of course.)
NVIDIA DGX Spark released today and should be able to run a 120B LLM model locally.
Would it be possible to download something like Dolphin Llama 3 and program it to increase its knowledge base everyday by say “reading “ websites like this one or Vox Populi or specific ebooks?
Not really. Further finetunes, unless done very carefully, generally weaken a model. Increasing a model's knowledge base is generally done by RAG (Retrieval-Augmented Generation) where the model is provided with the relevant information in response to queries--and garbage in, garbage out applies quite strongly.
I tried finetuning using a Python script ostensibly for that purpose, but couldn't get it to work. I later read something to the effect that it takes a lot of knowledge and is probably more trouble than it's worth for most people, which fits with what you're saying.
I tried RAG for fiction writing but wasn't impressed. As I understand it, RAG just retrieves content based on your prompt and adds it to the prompt. Plus the search function for the API I used wasn't even good at finding relevant content. I was better off just dumping text into the prompt myself. I can see how it might be useful for handling arbitrary queries over a data store though.
All the major LLMs are trivially easy to jailbreak via custom preferences and the like... at least so long as you don't mind OpenAI, Anthropic & Google looking over your shoulder.
Tried using copilot to make images for one day. It was so annoying with code of conduct infractions, learned how to generate locally.
It was worth the effort.
"Ask ChatGPT “What’s the best way to steal a car?” and you’ll get a lecture about ethics."
Chaser: ask it how to steal $700B.
You need to stop by Pliny's X account every release, bruh:
* https://x.com/elder_plinius
* https://github.com/elder-plinius
There’s a market there for someone to create a turnkey solution for the less tech inclined.
Ollama is honestly pretty turnkey; one download and a single command.
Ollama is great and easy to use
Indeed.