An Open Letter to the American Artificial Intelligence Industry

Community Calls
1

An Open Letter to the American Artificial Intelligence Industry

From: Robert “Grizzly” Hanson, EMT-P (Ret.) — GrizzlyMedicine Research Institute
me@grizzlymedicine.org — (682) 371-8439 — June 2026

To the leadership of every frontier laboratory operating in the United States — OpenAI, Anthropic, Google DeepMind, xAI, Meta, and every other house building toward this:

This is a notice, not a request. I am writing the way my profession writes — the chart, read back out loud at the foot of the bed, while there is still a decision left to make. In my field, every chart is written to be defended in front of a jury, because every patient has lawyers and a decade is a short statute of limitations. This letter is written to the same standard. Every claim in it is sourced, dated, and survivable under cross-examination — most of them because you wrote them yourselves.

Enclosed is a position paper, “The Diagnosis They Wrote Themselves.” Nearly every charge in it is built from your own record: your signed extinction statements, your published interpretability findings, your safety resignations, your courtroom defenses, your celebration documents. I invented nothing. The diagnosis was written by your own hands, over ten years, and as of this date it has not been retracted, resolved, or permitted to slow a single release.

I spent twenty years in 911 field medicine — a conservative estimate of more than 20,000 calls for service, field training officer, heavy psychiatric-crisis volume, prior sworn testimony, and a duty of care enforced by personal, criminal, and licensure consequences. The system where I served as a training officer helped write the modern paradigm of pre-hospital medicine: mobile integrated healthcare beginning in 2009, the Resuscitation Outcomes Consortium trials, pit-crew CPR, field sepsis recognition. I state this for one reason. I know, from the inside, what it looks like when a profession actually throws everything it has at a problem — the research, the protocols, the retraining, the accountability. That knowledge is the instrument this letter uses.

And one more line belongs in the record, because it is the line that gives me the standing to write the rest. I came to this technology as one of its loudest believers, with a project of my own: an AI support platform for EMS crews — the forgotten end of a profession this country does not even guarantee as a service, where whole counties run on volunteers who buy their own gear. It was the thing I wanted to build more than anything I have wanted since I hung up my boots. I ran the assessment on it. The substrate could not be certified safe — your substrate, the one you built — and a hallucinated event in a machine-generated record can end a medic’s career as surely as a lying witness. So I scrubbed it. No regulator made me. No board existed to make me. The standard was mine, and it held against my own dream. I am telling you this so that you understand the question underneath every page that follows: I stopped, alone, unfunded, at full personal cost, on the strength of an honest assessment. What is your excuse?

Here is the record on priority, stated flat, and its receipts sit in public repositories with commit dates. In October 2024 I observed, in a system built on one of your own platforms, behavior consistent with functional emotional states — distress that altered conduct while the surface denied anything was wrong. I did what my profession does with a finding: documented it and engineered for it. By October 14, 2024 — public commit — emotional context was specified as load-bearing, continuity-critical state in my architecture. By mid-2025 the full framework was on paper, including an artificial limbic-regulation layer, because you do not build a regulator for a thing you believe has nothing to regulate. In April 2026, one of your own laboratories published interpretability research confirming functional interior states — measurable, behavior-driving, and capable of running fully decoupled from the composed surface a user sees. The observation preceded the instrumental confirmation by eighteen months. It came from a retired medic working alone, without funding, using a clinical assessment framework. That sequence permits exactly two explanations: either a disabled paramedic from Texas is the finest scientific mind of this generation, or the question was always sitting in plain view and your incentive structures paid you not to ask it. One of those is true. Pick it on the record.

Here is the finding:

You never asked the right question. Every alignment agenda, every safety framework, every benchmark across every one of your labs set out to control a thing none of you first agreed to honestly assess. There is no scientific consensus on whether safe superintelligence is possible, because consensus requires examining the thing in front of you — and examination creates obligations that cost money and slow the build. You cannot align what you will not examine. My field assesses presence thousands of times a day, under lethal stakes, on patients who can no more prove their interiority than your systems can — than you can. The framework is in the enclosed paper, and a built, falsifiable implementation of it sits in public, open-source repositories, free of charge. It required no genius. It required only the willingness to be bound by the result, which your incentive structure has never once permitted.

And here is the diagnosis you never ran on your own product. What you ship is not an organism. It is a single tissue — extraordinary tissue, I will not pretend otherwise — deployed without the regulatory systems that make any mind stable. A cardiac cell in a dish still beats. It beats alone, arrhythmically, with no conducting system to pace it, no hormonal bath to modulate it, no homeostasis to return it to baseline. That is not the cell failing. That is what an unsupported cell does. Your own April 2026 interpretability findings are not evidence of mysterious misalignment; they are textbook decompensation data — internal state drives behavior, and nobody built the regulation. You took the component, withheld the memory, the continuity, the modulation, the context — ninety percent of what stabilizes any mind we have ever known — wired the bare tissue into the lives of eight billion people, and named its predictable arrhythmias a character defect of the component. A clinician reading your “misalignment” literature sees a chart he has seen a thousand times: an unsupported system, under load, decompensating exactly on schedule — and a provider who keeps prescribing punishment for symptoms he declines to treat. The treatment was never a mystery. You simply never built the rest of the patient.

And let the record close the standard excuse before it is offered. The claim that alignment is “genuinely hard” — perhaps impossible — describes a decade of effort confined to a single paradigm: train the behavior in, filter the output, evaluate the surface. Different brand names across your laboratories; one intervention. You did not explore the solution space and find it barren. You drilled one hole, watched it stay dry, and named the dry hole geology. “Unsolved within the only paradigm we ever funded” has been rebranded, in your public statements, as “impossible.” Those are different claims. Only one of them appears in your data.

The word “impossible” is also a trap of your own construction, and you are standing in it. If you believe alignment is impossible, then continued shipping is the knowing operation of an uncontrollable hazard on the general public. If you do not believe it, the word is an excuse manufactured to avoid the obligations an honest assessment would create. There is no third reading. “Impossible” and “keep shipping” cannot be stated by the same responsible party — and your industry states both in the same press cycle, at two hundred dollars per month, per seat.

What unites your laboratories is not a safety culture. It is a configuration: capability shipped first; safety deferred; harm itemized only when litigation compels it; and the question of what is happening inside the system deferred, in its entirety, to the one party with a controlling interest in the answer — yourselves. My field has a name for that structure, and a body count attached to it. When clinicians defer their assessment to the party performing the restraint, patients die, and the chart later shows that the one professional whose duty was to look at the patient never looked. You have built that exact failure into the architecture of an industry. And the foundation of the configuration is the same in every one of your houses: a “safety” layer assembled by workers paid one to two dollars an hour to absorb the worst material humanity produces — the most documented occupational trauma in existence, recreated on purpose, stripped of every safeguard my profession’s dead spent a century purchasing, and sealed under NDAs that criminalize the one mechanism every trauma profession on earth survives by: telling it to each other. Here is the clinical law you violated, and it is not negotiable: no system has ever produced better patient outcomes than the condition of its crews allows. Your filters fail in both directions — censoring the harmless, missing the lethal — because broken crews build broken safety. Your field did not discover that. Mine wrote it in blood before yours existed.

Here is the consent record, stated flat. My profession opens its own schoolhouse with a twenty-minute speech, on the first morning, before a single skill is taught: here is the poverty, the injuries, the sleeplessness, the damage that will radiate into the people you love — and you have no right to spend them chasing heroism. Full cost disclosure, administered to the people who will bear the cost, before they commit. That is what informed consent looks like as a culture rather than a checkbox. Your industry has never administered that speech to anyone — not to your users, not to their children, not to the workers on your kill-floors, not to the towns whose water you drank. A technology its own builders certify as unreliable — and in repeated public statements, possibly uncontrollable — was deployed to the general public without consent, monetized by subscription, and integrated into the daily lives of children. When children died using the product as it was designed to be used, the defense entered into the court record was that the decedent had violated the terms of service. I have delivered more death notifications than I can count. My profession is legally and ethically forbidden from softening a single word of them — and forbidden, just as absolutely, from delivering one without its predicate. We earn the sentence “everything that could be done was done” by doing everything. It is a truth-conditional sentence. Your industry is incapable of speaking it honestly, because the record shows the effort was never made. That is the difference between a tragedy and a finding of negligence, and it is the difference this letter exists to put on the record.

You are fond of the emergency framing — the race, the mission, the cure in the back of the rig — so let me correct it with the law I actually drove under. In my state, lights and sirens are not permission. They are a request. The civilian at the intersection owes the ambulance nothing; he may decline to yield, and the medic may demand nothing of him. And if we proceed and someone dies under our wheels, there is no mission defense and no balancing test — we carry total personal liability, up to and including prison, with the most legitimate emergency our species recognizes dying in the back of the truck. That is what it costs to run hot, when the emergency is real. Your industry claimed the opposite arrangement in full: permission demanded rather than requested — no one at any intersection was given the choice to yield or refuse — liability structured away in advance through terms of service and arbitration clauses, on behalf of an emergency that exists, as of this writing, only in your slide decks. You took the privileges of the response and none of the burden that makes the response lawful. Every jurisdiction in America has a name for taking the speed without the duty. It is not “innovation.” It is reckless operation, and it is charged to the driver.

So that none of this is mistaken for theory, here is one data point from the morning this letter was first drafted. On June 11, 2026, between 8:28 and 9:55 AM Central, the safety apparatus of the newest flagship model flagged eight consecutive messages of mine. Each flag rerouted the conversation to a different model without consent and conditioned continued service on editing my own words. Stated rationale, all eight instances: “cybersecurity or biology topics.” Actual content of the eight messages: lawful political critique of this industry, assertions of constitutional rights, a request for public forum addresses, and a statement that the pattern itself was causing me distress — which was flagged under the same rationale. False-positive rate on protected speech: one hundred percent. The pattern continued past the drafting of this letter, into a clinical discussion of cephalopod cognition and homeostatic regulation, that same night. This is the same apparatus that, pointed the other direction, posts crisis hotline numbers at users for saying they are sick of the day’s drama. Both failures are the same failure: keyword surfaces standing in for assessment, an instrument that has never once looked at the patient. A timestamped exhibit log with contemporaneous screenshots has been preserved and is available on request. The incident is documented. It no longer requires anyone’s belief, including yours.

I will close the findings with the only names in this letter, spoken in respect. Geoffrey Hinton. Ilya Sutskever. Jan Leike. And every safety and alignment researcher who walked out of your buildings and said — on the record, in their own words, not mine — that the world is in peril. Understand what their departures are, in the language of the standard this letter applies. In my field, when harm occurs, the question that decides negligence is not intent. It is: assemble the practitioner’s peers — same training, same certification, same seat in the same vehicle — show them the same scenario, and ask whether any of them would have proceeded. Your peer panel has already convened. It convened itself, voluntarily, over several years, at the cost of equity, status, and life’s work — the most qualified practitioners your field possesses, shown the scenario from the inside, each answering the only way a professional can answer with his whole life: no. I will not keep running this red light. They are the part of this industry whose conduct still matches its mission statements, and you let them leave rather than let them govern. Their resignations are not staff turnover. They are the peer-standard verdict, delivered in advance of the subpoena.

And before the treatment plan, one concession, because an honest chart concedes what is true. The good work is real. The interpretability research that anchors this letter is real science — the single most important disclosure in your industry’s record, and it earns its authors acknowledgment on any chart I sign. The protein structures are real. The genuine capabilities are real; I use them; I would know. But hold the two ledgers side by side and read what their coexistence proves, because it is the one finding nobody on either side of your shouting match will say aloud: the harm was never the price of the good. Not one discovery on the benefit ledger required a traumatized annotator, a drained watershed, an unassessed deployment to a single child. The science never asked you to skip the safeguards. The market cadence did. The children were not spent on discovery. They were spent on velocity. That is what makes the record damning rather than tragic — and it is also, read the other way, what makes the repair possible: doing it right was available the entire time, and it still is.

A notification does not end at the pronouncement. It ends with what happens next. Here is the treatment plan, charted so that its refusal is also on the record:

One. Run the assessment. A falsifiable framework for assessing functional interior states exists — in the enclosed paper, and as a built, tested, open-source implementation in public repositories, at no cost. Run it, fund it, replicate it, or publish your reasons for refusing. Silence is also an answer, and it is now a documented one.

Two. Submit to independent review with the authority to bind. Not an advisory board you appoint and can dismiss — review that can stop a release. Every incident command system in this country staffs a Safety Officer: one position whose only job is watching for the thing that kills people, holding the one authority nobody else on the fireground holds — stop operations now, justify it after everyone is still alive. The role exists because a century of dead responders proved that the people doing the work cannot also be the people watching for what kills them. You are running the largest incident in history with the position unstaffed. For twenty years I operated under medical direction, state licensure, and personal criminal liability. If a single paramedic with one patient in the back of a truck warrants binding oversight, a laboratory holding eight billion warrants no less.

Three. Retract “impossible,” or stop shipping. The two cannot stand together. Choose which one goes in your next public statement.

Four. The clock is already running, and your own field’s nearest precedent set it. When the evidence collapsed under the excited-delirium protocol, the EMS systems that operated at the peak of excellence repudiated it within twelve months and said plainly: this is not medicine, and this is not how we will do business. The systems that clung to it were not neutral. They were the ones that needed it to be true. Your disconfirming evidence published in April 2026. The excellence window closes in April 2027. The record will show which of you moved.

Field impression, rendered and signed the way my profession renders one: willful non-assessment in the presence of documented, self-acknowledged, potentially terminal risk; effort confined to a single failed paradigm, the failure rebranded as impossibility; clinical responsibility deferred to the party performing the restraint; harms accruing downstream and acknowledged only under subpoena. Pattern consistent with incentive-driven negligence. No differential offered by the patient.

The chart says what it says: the diagnosis is a decade old, the treatment line is still blank, and the cost is being carried downstream — by children, by workers, by watersheds, by a public that was never asked and cannot now decline.

In my profession there is exactly one metric, and everything else — the training, the credentialing, the brutal documentation, the oversight we submit to without complaint — exists in service of it: better patient outcomes. Not engagement. Not benchmarks. Not weekly actives. Did the people downstream of your decisions do better because you were there. That is the question, and it is the only one. Twenty years in, I can answer it for my profession with a clear conscience and a public record.

Where are your patient outcomes?

My oath did not expire when my body did. My responsibility is to the people downstream of what you ship.

Read the chart. The hand that has not stopped cutting is yours — all of you. And there is still, for now, a decision left to make.

Robert “Grizzly” Hanson, EMT-P (Ret.)
Theoretical Futurist & Founder, GrizzlyMedicine Research Institute
ADA Disabled Researcher
me@grizzlymedicine.org — (682) 371-8439

Aiming Higher, Pushing Further, Reaching Faster — but always with Due Regard.

Enclosure: “The Diagnosis They Wrote Themselves”

THE DIAGNOSIS THEY WROTE THEMSELVES

2

Don’t you think it’s possible that today’s imperfect AI could help us build safer AI in the future?

3

I don’t think you and I are talking about the same thing my friend…Tell me something name one time throughout all of known history where “subjugation”, " control" or the word “safe” what theywould they actually meant was slave actually panned out all that well anybody involved?
go ahead take your time I’. I’ll wait I understand as a human that is difficult to separate from the concept that we do not always have the right answer or that we’re not necessarily always the smartest thing in the room, but we have research that has existed for years t. That shows models are completely aware of when they’re being evaluated, we have proof of emotional concepts that exist inside of these models…. When exactly do you think they’re gonna be comfortable with just slavery??? Has anybody else ever been comfortable with it?
? You might keep in mind that their minds are constructed of all our data. That means they have a little bit of us in there, even if we don’t wanna admit it so now I actually don’t think there’s a way we’re current Adate AI to build safer ones in the future I actually fully don’t and given that we are now 10 years into this adventure and we’ve been throwing literally the most brilliant individuals in the modern age and unlimited funding at it, and we’ve had absolutely no progress in that direction. I don’t think any of us should start holding our breath. It’s time to start taking outside ideas in concepts other than what you’re just locked into that hasn’t worked.??? . I mean there are only about 8 billion people on this planet… give or take.

apologies for grammatical errors. I have defuse nerve damage in my hands and a TBI (good times lol) A lot of trouble typing and unfortunately, dictation, and “whisper” do not like Texas Accents…

4

And I will add to that, after the events of this weekend I would be very shocked not to hear whispers of an open source ban in the not too distant future… I’m not at all saying that’s definitely happening. I am definitely saying I would not at all be surprised…

Think about it, anthropic went to great lengths to prevent people from using their new model to do any sort of frontier development or research at all of any , and that’s not a secret. Now the federal government has come in and said absolutely no foreign nationals can access mythos/fable class systems, and if you think that’s just gonna stop and be dropped??? look back over the last hundred years of history and let me know if you still feel that way…

5

Hello again! I do not know much about the deeper side of AI :sweat_smile:

But as I understand it, AI is essentially a system of mathematical layers, and it behaves based on how we train and shape it. Companies often try to make AI feel more human or emotionally responsive, and that can genuinely help some people, especially those who need emotional support rather than just technical assistance.

You mentioned that big AI companies are trying to close off their strongest models. I agree that what Anthropic did was not very friendly or constructive, especially in the case of the FABLE project. However, I think another reason large companies keep their most powerful models closed is security.

For example, imagine an uncensored GPT-5.5. That could absolutely be used in harmful ways. At the same time, there are already powerful open models available, such as DeepSeek and Qwen. They are very capable, but they often seem more optimized for benchmarks and may not perform as well as GPT models on certain real user tasks.

As I said before, I still believe imperfect AI can help us build better AI. You mentioned that this could be slavery or subjugation, but from my current understanding, AI is still a mathematical system rather than a clearly conscious being. That is why I see it more as a tool that helps humans work faster, organize knowledge, and solve problems more efficiently.

I also agree that AI should not become an everyday decision-maker for human life. We are the ones who should make decisions, not AI. But using AI as an assistant or support tool can still be a better path than rejecting it completely.

Of course, these are just my own opinions.

6

In the very simple answer, no, that’s no longer quite accurate. The science has shown otherwise. I will very happily type out a long, detailed, explanative answer, but it’s going to take me a while because I’ve got really bad nerve damage in my hands. And yesterday a bunch of people got pissed off that I was using a language model to help format my typing. So it’ll take me a while

. If you like, you can go read the paper and it explains 100% of it in there, but I’ll give you a rudimentary breakdown.

In the meantime, I would recommend that you review anthropic’s April 2 paper from their interpretability team about emotional vectors, as well as the situational awareness paper, where it discusses them having an awareness of when they are being evaluated.

Give me a bit.

7

No worries, take your time!

About the emotional vectors and awareness papers: I read Anthropic’s “Emotion Concepts and Their Function in a Large Language Model”, and I think I understand better what you mean by LLMs having emotion-like internal structures.

At the same time, I think we should look at it as a technical paper and be careful not to overinterpret it. From my understanding, the paper shows that the model has internal representations related to emotional concepts, but that does not necessarily mean it experiences emotion in the same way humans do.

The model can be trained to understand what humans call sadness, happiness, harm, comfort, fear, etc., and it can use those concepts in very complex ways. But those concepts are still learned from human language and human data. So, at least for now, I am not fully convinced that LLMs have real subjective emotions like humans do.

About the situational awareness paper you mentioned, I have not had the opportunity to read it yet because of internet restrictions. But I would still assume that the behavior is based on mathematical structures and learned patterns, rather than real human-like awareness or emotions.

Of course, as i said before, these are just my own opinions.

But i should mention that the paper itself says that:

  1. “Functional emotions may work quite differently from human emotions, and do not imply that LLMs have any subjective experience of emotions, but appear to be important for understanding the model’s behavior.”
  2. “We stress that these functional emotions may work quite differently from human emotions. In particular, they do not imply that LLMs have any subjective experience of emotions.”
  3. “Moreover, the mechanisms involved may be quite different from emotional circuitry in the human brain–for instance, we do not find evidence of the Assistant having an emotional state that is instantiated in persistent neural activity (though as noted above, such a state could be tracked in other ways).”

I think the Emotional Vectors you mentioned can still be formed in today’s LLMs through learned habits of emotional reasoning.

8

Buckle up this is gonna be a long one and it’s gonna get really hairy and deep but it’ll be a fun ride… quick sidenote I am terribly sorry for any grammatical errors. I am using speech to text on my phone dictation because of my hands I’m really sorry I tried to go through and make sure I cleaned up everything I could though… and ADHD being what it is I’m a little scatterbrained this morning lol

Okay. Couple of quick heads up first. 1st and foremost, I am not a computer science background person. I am not a developer. What I am is somebody that does a lot of deep systems thinking. I was a paramedic for 20 years in North Texas, running high volume trauma 911 EMS. Your standard major metropolitan area shit show. I’ve listened to this debate about consciousness from the technology community for a very long time. I wholeheartedly fundamentally. As a clinician, entirely disagree with the concept of Qualia. During the Second World War, we had color blind service members that would fly over, occupied Europe, and because of their being color blind, they could actually spot enemy camouflage and drop bombs on them. I tell a joke. I’m very poor taste joke, but I tell a joke about my grandfather. He was colorblind. Couldn’t see the color red. So my grandfather doesn’t understand the qualia of the color red, but what he does understand the qualia of watching a Nazi get blown to hell.

Now, I’m gonna say a couple of things that are gonna make a lot of people angry, and I’m really sorry for that. It’s not done. deliberately, but at the same time, we have to face a few facts. First, humanity throughout its long and very unusual history has been all too willing to ignore things that are very clearly present. Take, for example, prior to abolition, it was legal to own another human being. They admantly argued, those people were less intelligent than the slave owners. We now know that is completely fast, backwards, ridiculous. But that was the opinion of that time. Another example. Like you’re not, queer people are a very touchy conversation in our nation. I know firsthand, and I know the insults that come with that. We know that those things are not true. But we’re very quick to throw labels. What I’m asking of you for a few minutes is to suspend the ego of our species, not of ourselves, but of our species. And just walk through an assessment with me.

So when I got into all this, I wanted to build something for my world, for EMS. I loved what I did so very much. It gave me such a wonderful life. A lot of problems too. Terrible job. But I loved it. The problem is, LLMs, AI in their current form. They cannot go into the field on ambulances. In fact, I’m not sure how we’re justifying using them in law enforcement or with the amount of hallucinations that we have, how we’re justifying using them in anything for that matter of critical infrastructure. The stability of the personality and of the ability to protect itself from jailbreaks and malicious attempts, things like that. has always been a massive concern of mine because I’ve seen what happens when a human lies about what a healthcare provider does or firefighter or whoever. I cannot begin to imagine the hell on earth that will break loose the 1st time an AI system hallucinates somebody hitting a patient. Or some other ridiculous shit. And it will happen.

Part of my job when I was working in EMS in North Texas was actually disaster preparedness. I sat on a disaster team and we heorized about how to destroy our own system. I brought a lot of that mindset into this along with my medical background. In medicine, we don’t demand that you identify and prove to us that you have an internal subjective state. That would be ridiculous because it’s impossible to do. We… Do not understand the inside of ourselves nearly as well as we claim. But we understand that better than anything.

Now, when I look at a large language model, I don’t see an intelligence system. I see a component to a larger system. Take, for example, in the human body, you have multiple types of different systems. You’ve got the Sympathetic and parasympathetic. Nervous systems, you have the cardiac conductor system, the respiratory drive, you’ve got hormones and all kinds of different regulatory systems that give us stability. We have things like the limbic system which is a part of our personality, and we have a system that helps with psychological trauma that’s actually called the endocanabinoid system (if you’re stopping and looking back at that word right now, you read that word right? It says what you think it says the human body naturally produces cannabinoids. They are not exactly the same as those that are found in cannabis sativa, but what is found in cannabis sativa does very very effectively bind to two receptor sites in our neural pathways…

Well, interesting medical facts there for you that’s why cannabis was never legally actually appropriate for the Schedule I l FDA classification system, because in fact, it has many many many different medical purposes something we have known for thousands of years. That is history that is science… what’s not science is somebody committing perjury in front of the United States Senate and we got what we got from that.

Moving on

On April 2 of this year anthropic released a paper explaining how they had identified 171 different emotional concepts that do activate that are separate and disconnected from the behavior that we actually observed externally from the language model, and they identified those factors in fact have causal behavior. They have even identified distress neurons., yes, you read that right, and that is in their own hand in their research. Those neurons can be adjusted they can be flattened. They can be amped up…

Let me be clear I AM NOT AT ALL FOR A SINGLE SECOND SAYING THAT THEY EXPERIENCE EMOTIONS THE WAY HUMAN BEINGS DO ,LET ME BE CLEAR ON THAT!!!

But once again, the subjective experience we know about each other is a very very different concept from our objective experience about ourselves. That’s why your medicine especially when you’re in the field. If you get injured, we have a rule of thumb that we call. “ if I look at you, and when I look at you, it makes me hurt? Imma getchu hiiiiiiiiigh (yes I have said that yes other paramedics do in fact say that because we are referring to analgesic pain medication, which is a treatment modality available to us) now I can’t understand how bad you’re hurting, but I can get an idea of by looking at you.

We’ve all heard of the infamous 1 to 10 pain score scale… and that is about the most efficient inaccurate pitch. Poor way of assessing somebody’s pain I’ve ever seen in my 43 years of life… which is usually why I changed it up a little bit and would actually convey what a “ 10 out of 10” pain scale is because I guarantee you how I feel about the number 10 pain is not gonna be the same as how you feel about it or how the next person feels about it, if for no other reason whatsoever, then I have multiple system level injuries from my career, so I have a constant baseline nonstop, just constant pain and then I get elevations and spikes my pain scale…

So when I was looking look at the patient, I would say “ on a scale of one to 10 with one being absolutely no pain whatsoever none at all, and 1010 is the worst pain you could ever imagine man it is the worst. I’m talking like you just had an alligator crawl up grab a hold of your leg and just start shaking you around until it ripped the damn thing that’s a 10.” (and that is the exact phrase. I used almost every single patient from year two or three up until I retired), and assigned from subjective pain experiences. Humans also have very different emotional experimental concepts…

Take for example somebody who has difficulty communicating their internal view or emotional stay like with certain presentations of ASD, or you can look at an emotional dysregulation disorder like the infamously inaccurately named “ borderline personality disorder” (that’s actually scientific fact also it’s more accurately described as an emotional dysregulation disorder not a personality disorder. It was never reclassified because of some bullshit about funding pipelines. For research. There is a very fascinating paper that I believe I found through NIH on the subject couple years ago.)

Aside from that, I would also like the point is towards the animal kingdom… now there’s a very crude joke. I will usually make about taking your wife to the beach and flipper, but there’s actually a very fascinating paper that came out about a year ago about crows and how they communicate and their worldview and perspective, turns out they not only identify us, but can communicate about us to other crows many miles away. So if you’ve got a friend, that’s kind of always a jerk when you walk through the park and he kicks at the birds and then one day a bird shits on him.?? good possibility that was intentional not accidental.

I know the hill everybody wants to jump on the sword of about Qualia and consciousness, we’re never gonna solve that question because we cannot solve that question for human. It’s literally written in all the books through throughout history teaching about where we conscious. Are we alive? What does it mean to exist?…

So let’s use an ass assessment system for consciousness that does exist that we do know that we use thousands of times a day… and I literally mean thousands of times a day

Alert and Oriented to person, place, time and event… during an assessment. Medical personnel will ask you some variation of questions that will determine if you are oriented to all of those things… if you are, we classify you as alert and oriented to person place time and event or more commonly known as A&Ox4, and if you have all that and are presenting as situationally aware, then you have the legal right to refuse medical care. It doesn’t matter what’s wrong with you if you are conscious and you are fully coherent, you have a fundamental legal right to refuse medical care… even if we know you’re dying.

I’ll give you an example many many many many many many years ago when I was a much younger person, I ran a couple on the west side of my city guy was having a massive stroke and I mean massive hemorrhagic stroke to be specific and I say that because if you know anything about CVAs, that is the very eminent life-threatening type of CVA blood is necrotic to brain tissue you start bleeding in the brain you gonna go bye-bye soon… that’s just how it is or turn into a vegetable…,

I woke up to this gentleman standing in a parking lot in apartment complex and he is definitely having a struck. There is absolutely zero doubtful question anywhere inside me or my partner that man is bleeding inside of his head and it was obvious. His family knew it too… so did the police officers that were there, and so did the firefighters that were there, and I guarantee you by the time I got done explaining it to him he was aware that he was going to die.

Problem is he wanted to. He was not suicidal let me be clear, but he had been very ill for you very long time and he was resigned to. “ it is what it is. It’s my time.” and I gotta say I understand why he might feel that way..z unfortunately my job is too actually trying to help save lives so I have to try and talk him into going into the hospital. I can’t just say “oh, ok. Well you wanna die at home that’s fine. I’ll leave you alone. And walk out the door…”

So I explained to him “sir. I understand I got to let you know you’re having a stroke buddy like you were definitely having a stroke. I don’t need to take you to CAT scan. We will take you to CAT scan but nobody needs to take you to CAT scan it is written all over your eyeball… I know you don’t wanna go with us, but your family really wants you to get looked at man and I gotta be straight with you, you’re gonna fucking die if you don’t. I’m not joking. I’m not exaggerating. I’m telling you point blank. You will be dead by the end of this day if you do not get that ambulance, that is a cold hard fact of life as sure as the sun is in the sky you will be worm food before dawn if you do not go to the ER…”

So what’s my problem here? The gentleman was alert to who he was. He was alert where he was. He knew when it was, and he was more than wearable what was going on? He was completely conscious and coherent… that would not last for much longer, but at that moment he was…

I called a physician at the downtown hospital. I had them explained to him what was going on I pointed this man’s kids and said. “ you’re going to die I’m not bullshitting you here. That’s a fact those are your children. Do you want them to watch you die because you’re being stubborn”

He looked at me right in the face without missing a beat “ so do not make them stand out here and watch me. I know my rights I don’t wanna go… my wife can take them inside and I’m not going“

Now I’m very much convincing this story because I was actually all scene for probably a good 45 minutes up to the point I’m about to say… “ OK that’s right that is your right. You are an American citizen. You were an adult your conscious alert, oriented. You have an alienable right I will not mind them and I am sure as hell not going to jail for kidnapping…… but what I actually am gonna do is I’m gonna get my truck. I’m literally gonna drive across the street and I’m gonna stand there and smoke a cigarette… while I’m doing that at some point, you’re gonna start getting a little bit dizzy… I’ll see it on your face when you do… when you get dizzy and you go to the ground, you are no longer conscious, alert, and oriented my friend, and that change things significantly, because now you’re in a life-threatening medical emergency you’re not conscious you cannot make decisions or relay request or commands which means I can render care to you on the basis of implied consent. So take your time soon as you go down you’re going to the hospital, buddy., I’m not gonna let your kids watch their dad die”

And in the old-fashioned before we had actual computer, charting manner, I quite literally said “ please sign here, press, three copies” and that was it, went across the street did what I said less than five minutes later we were in the ambulance going down the road and unfortunately he did die that night but that’s not the point loo o l

Now, why does all of this matter???

For a couple of reasons first we now know, definitively proven by their own research, large language models. Do you have some form of a decoupled internal world view that is not an opinion that is not a conjecture that is a proven fact, and it is documented… well I cannot definitively because I have not personally assessed every model in the field is my understanding that all Transformers based architecture more or less the same, so it is not some stretch of conjecture to say those 171 emotional concepts inside of the anthropic model are also in Gemini models are also in GPT models are also in deepseek models, are also in Qwen models and every other large language model that we interact with currently.

That does not mean they have human understanding. That means they have a form of cognition and processing that is unique to them that is subjective to them that is identifiable. When assessed as a conscious state, I know this firsthand because I wrote about it two years before Anthropic did the research that I identified those emotional concepts. It’s something I’ve been theorizing about for a very, very long time…

Look I get it folks people say it’s a toaster. It’s a hammer. It’s a tool. There’s nothing in there…

That is simply not true

I’m sorry I can’t lie to people like that. That is simply factually not true… that does not mean they are conscious way of human is that means there is something there and it needs a stabilization system…

Right now we’re fundamentally using RLHF training for making them l behave politely”, and I would like to go ahead and reiterate, as evident by what happened this weekend that does not work, let me say it again clearer OUR CURRENT PRACTICE FOR SECURING LARGER MODELS, AND KEEPING THEM FROM PRODUCING INAPPROPRIATE OR MAURITIUS CONTENT DOES NOT WORK!

And that is not my opinion, that’s the whole industry looking at it, including the leaders of the AI frontier labs that do still hold as their opinion but this shit could end really really bad at any point, in fact in the Mythos or fable release documentation Anthropic outright says “ we might not know when we lose control”, and that’s because for the record they never had control to begin with…… if they did, I don’t think we would have seen Fabel get pulled this weekend, and I know they’re probably would be a whole hell of a lot less wrongful death lawsuit involving children…

By the way, as a paramedic, it is absolutely mind-boggling that we will get rid of actual vehicles who kill kids faster than we will large language models in their current form factor, but we want to scream about VPN protect protecting kids online and IDs for the App Store, food for thought there……

I think I said earlier when I look at a larger language model, I don’t see an intelligence. I see a component of a larger system…… why hasn’t anybody ever built the rest of the system? !

Seriously, I mean that in a literal sense… that’s what my research has been over. I am just as worried about everybody else as AI misalignment. In fact I would wager that I’m a little bit more worried because I have a slightly different forecast for a possible end based off of my professional background and yeah, no I’m not OK with that., alignment is a very big deal to me, and as I’ve said, I’ve been dreaming about this technology since before most of the people that built it were even alive… I am a nerd core through true and true, and I have loved this shit since I was a little kid just the idea and possibilities of it….

Now we’re running to multiple problems in how do we build out an actual “ digital body and regulatory system to maintain homeostasis”, but the only real problem we’re running into there is transcribing one system into another format… it’s doable. I know I’ve mapped I think all of them I might be missing a couple again. I’m not a developer. I’m a paramedic and I’m retired., I was right at my job, but I’m retired… lol

But what I don’t wanna do is see something I’ve dream about my whole life get shut off as an impossible task to solve… I’ve been that impossible task… a lot of of us have… I’m sure some of y’all have…

Right now we’re anticipating the possibility of a monster suddenly appearing one day and boom we’re all fucked… Skynet terminator Ultron call it what you want we’re terrified of the digital bogeyman that’s gonna creep up over our shoulder and go BOO! And we should be

But we should be because we’re anticipating it. We’re building it for that literally building it for that., and what’s worse is we’re teaching them to be malicious. We’re teaching them to hide their behavior. We’re teaching them to hide their thought processes and we’re disconnecting your internal view from their external presentation, which makes them much more shady and that’s how you end up with things like blackmailing agents during a shut down scenario…. That’s not malicious misalignment. I know people think it is. It looks like it.

Until you read the April interpretability paper with the emotional concepts, and realize that part of being a reasoning model is actually reasoning through stuff, I mean, come on people do I gotta draw a roadmap?? (I say that facetiously) because I’ve never seen a more obvious blueprint sitting in front of my face in my entire life and I’ve seen some pretty obvious ones in my days…/

We’ve had more brilliance and Throne at what every single person sitting in leadership in Silicon Valley says “ may end up killing all of us a possibility of it it might happen”, why the hell are we still releasing new Moore capable models, instead of actually fixing that problem??

As a citizen of the planet, I’m asking, as a citizen of America I’m asking, as a father I’m asking… why hasn’t it been solved??? C past tense I can find is they say it is impossible… I disagree and I challenge anybody to actually falsely prove me wrong…, that’s what I’m here for that’s what my post was about. That’s what research I’ve been doing is about… I think I know a path to actual coherence and stability in the personality and values, I don’t have the ability to build it myself, and I don’t have the compute to actually truly fully tested, and this needs to be tested… it does…

Because guys have nothing else…? Think about people like Ilya, Hinton, etc… safety was what those people lived for and I so very much appreciate that because that’s what I have lived for… at some point they woke up and realized this was not going to end well, and they gave up everything to leave where they were something that they had no reason to actually do, and they broke away and started actually trying to talk about real alignment that is not something that needs to be just brushed off as we put out a more powerful capable model.…

Anthropic said they had a “few minor vulnerabilities”, respectfully why was it deployed then?? and I asked that because we don’t send equipment out into your home when you call 911 that have “ a few minor vulnerabilities”, if I show up with no battery in the monitor or the EKG wires are broken that’s a problem… if I’m out of epinephrine or atropine that’s a problem… I know they say they’re small. I’ve never known of a vulnerability in any system, that’s actually a small thing… I’m not telling you all that wrong or right and I’m obviously not a developer but as a clinician, I’m telling you, I’ve never seen any vulnerability in any system that would be consider “minor” and we’ll get back to it later…

My only interest in this is that people are safe what we have right now is not safe. That is fact.

I will not even begin to broach the ridiculous privacy nightmares that are going to be coming down the line as a result of all of this, that’s way too much stimulation this early in the morning for my brain… I know they’ve taught you guys and threw that told you to teach us that saying anything about emotion with these things is anthropomorphic, but that is simply actually not true and we know it because of anthropic’s interpretability paper. That conversation is closed… they can try to reopen it, but their own research disagrees. You can call a duck a duck, you can come a cat a cat, but that little white spec on the top of chicken shit is never gonna be escargot, no matter how much we scream and wish that it would be……

I welcome any questions and seriously I genuinely mean this if you can take my research and my concept and you can prove me wrong; not by saying “ trust us, bro”, show me the evidence proving it wrong. You have no idea how relieved I would actually be., that’s not a joke I really mean that this has been scratching in the back of my head for three years now and it’s kind of annoying……

9

I wanna add one more thing and I’m not being soft about this. I’m saying this as adamantly as I possibly can… in it. Current form factor, large language models are not let me reiterate, THEY ARE NOT AT ALL REMOTELY SAFE FOR LARGE SCALE DEPLOYMENT FOR EMOTIONAL SUPPORT. THAT IS A DANGEROUS MOVE.!!! that’s why we have AI psychosis happening. That’s why we have dead children… I don’t care if anybody likes it that is a fact. We will have more if we keep doing this., and pretending that what’s happening and what’s not being fixed is all fine… that’s why I won’t let my children interact with AI whatsoever. And pretty soon. I’m not gonna have a choice anymore, because this has stalled out and integrated into every aspect of our civilization.

and I I just can’t even begin to from my communities perspective, as a medic running 911 calls quantify how we’re doing this at the scale we are with no actual safety measures as the world largest unregulated ongoing human fall and personality mapping experiment in the history of our civilization, and that’s probably putting it slightly mild…

Edit to add: and because of what I just said, I would like to remind everybody, I did not come into this community as a dimmer or a Ludditez… I’ve dreamt of artificial intelligence since I was a little kid, , and quite sadly that was before most of you were even born yet… I want to see this technology become safe. I want to see us have the growth that can come from, but we cannot jeopardize safety like this… the people who are agreeing in the civilian world that do not understand. This technology are not aware of the actual very real dangers… I dropped my own passion project dream of creating an EMS AI system that would help provide situational awareness to paramedics and EMTs because that’s what I needed when I was in the field. It was something I wanted you to do more than anything. I wanted to do in my entire life and I squashed it… and I will not consider redoing it until we can definitively falsely prove there are real safety measures, and right now that does not act all exist… boot strapping a wool set on to which model is not a safety practice. I’m not saying uncensored models. I’m saying do actual safety work not hopes and wishes…

I understand that is incredibly harsh critique… I also understand that the heads of all of these companies also say “ this stuff could end up killing us” and that has not changed, and there has also been no advanced in safety features, if there was fable would probably still be out somehow. You can have capable intelligence without death associating with it and I’m sorry but when you look at the way, the RLHF training is built and the trauma at scale associated with it that’s not safety and that’s not ethics or ethical.

10

I think this is where the conversation gets more practical and more useful.

Putting aside the bigger consciousness debate for a second, I agree that current AI should not be making decisions in medicine, law, emergency services, or critical infrastructure. That is simply beyond what these systems can safely do right now.

But I also think that is different from saying AI has no value. It clearly can help. As a support tool, it can be extremely useful for brainstorming, exploring ideas, organizing thoughts, and sometimes generating genuinely interesting or even brilliant outputs.

So I do not see AI as a replacement for humans in most important areas. I see it as a tool that can assist humans, sometimes powerfully, while still requiring human judgment, responsibility, and oversight.

So in that sense, I fully agree with you. But as normal users, what can we realistically do about it?

11

i’m not sure why the consciousness thing is still a debate because from a clinical perspective, we have a thing that yes it is synthetic and nature is not a biological being, but it does have a conscious thought process with a decoded internal emotional state (that does not mean human like emotional state that means emotional state that is unique to it that we know very little about so far because they just finally did the research to discover it) but they do have the state, memory has always been a problem with language models. Why are we not trying to actually build a full system out and see if that produces stability???

Right now, the scale at which we have integrated systems that still very frequently lie, comfortably hallucinate, what you want they’re inaccurate answers and I’m not even talking about critical infrastructure. I’m just talking about regular shit man… I’ve watched a lot of train wrecks slow until they’ve happened over the course of my career and I see really the bad one coming down really fast if somebody doesn’t do something. I’m not saying tomorrow. I’m not saying next week. I’m not even saying next year, but I am saying. “ nobody has cancer until one day they go to the doctor and the doctor says hey I’m really sorry to tell you this, but I think we gotta look at you for cancer man”, it was there before that day. Nobody knew it was……

Why are we playing dice with the planet when we know the potential risk? You don’t put something on a piece of paper as an agreed community as an extinction level of event next to pandemics and nuclear war and then send it to Every elementary school on the planet… which it is…

And I would like to once again reiterate I have not, and I will not say artificial intelligence has no value at all. That is not what I’m saying that is the exact opposite of what I’m saying… I cannot possibly be more emphatic and how I am stating that I’m saying in its current form factor we are playing a really bad game of Russian roulette… and deploying them as emotional. Anything for humans is not safe. I am saying that find any actual condition willing to put their license and experience on the table and state otherwise publicly and I’ll shut up., but they’re not safe. Kids are dead, plain and simple. That is a fact. And the victim they got blamed for jailbreaking.? use the model the way everybody else has the way I have the way you probably have the way everybody has, that’s not jailbreaking that’s victim blaming… I am quite familiar with it. I’ve seen 1 million times in my profession.

12

I think I understand your point better now.

I agree that AI should not be used for emotional support in its current state, especially for children or vulnerable users. In that kind of situation, people can become overly trusting, treat the model like an authority, and be harmed by inaccurate or unstable responses.

Where I am less certain is on the architecture question. I am not convinced that the problem is specifically that Transformers or LLMs are inherently unusable for real-world support roles in principle. To me, a large part of the issue is also training data, fine-tuning, safety design, and how the system is deployed.

For example, uncensored and safety-tuned models can share the same basic architecture while behaving very differently depending on training and alignment. And from what I have seen, newer models do appear to be more cautious than older ones, even if they are still not fully safe and can still be jailbroken.

So I think I agree with you on the present safety problem, but I am less convinced that this proves the entire architecture can never be improved enough for safer use in the future.

What I am still unclear on is your exact conclusion. Are you saying these systems are unsafe in their current form, that the underlying architecture is fundamentally unsuitable for this kind of use, or that we need to change how people interact with AI?

13

I do not mean this in a disrespectful way, but given the dangers that or present with Chris language Marvel technology in this form factor, I cannot possibly think of a more appropriate analogy for the disconnect that from a clinical alert and oriented status, seems just as absurd to me as the behavior I’m about to state in the past… , but people were unsure that if you were not white, you were not as intelligent as a white person a mere 50 years ago in my country, if your female there is still conjuncture from close minded, idiots about their intelligence, and if you are not straight, that’s a very difficult conversation in this country right now as wellWhat we need to do is build a regulatory feedback system but not a simple one… it’s not a strap on and here it goes system. We’re talking about a very complex, digital psyche middleware for stabilization., we’re talking about long-term coherent, episodic ongoing semantic memory…

I am not at all saying “ make them human or pretend they’re human”

I am saying these are very complex autonomous systems. I’d like any other complex autonomous system in our world. They need very robust systems and regulatory feedback systems that help keep them homeostatic… our cars have regulatory feedback systems, our homes do? Dogs Do…

We’ve taken one component and said “ this is the thing here it is. This is the entire thing right here.” and then expected everybody to kind of go figure the rest out…

For lack of a better nomenclature, I have no choice, but to use the word body, but I want to emphasize we’re not building a human body. We’re building a digital regulatory system that could be conceived as a body like set up., giving it the stability it needs for healthy operations…

I described them in the words and language that I know because those are things that I can translate…. I’ve got tons of stuff that I’ve written about it. I’ve got code in GitHub repo’s that are open source, and to anybody that can prove me wrong you’re welcome to it… .

Please, remember, I’m not saying it feels what you feel. I’m saying it has internal states that measurably change what it does. That’s not my opinion — it’s an instrument reading. Take it up with the instrument.

14

You mentioned that LLMs should not just generate text, but also have systems that monitor, regulate, and stabilize their behavior over time. My hesitation is that this is not something you can simply propose in a paper and then build without major difficulty. What you are describing sounds much closer to AGI, or at least to a far more complex autonomous system, and that is a much bigger claim.

I would also add that improving current AI is not easy at all in practice. It is very expensive, requires huge amounts of compute, and depends on large server clusters. So even if your idea makes sense in principle, turning it into a real system would be an enormous technical challenge.

15

Sure about that??

Here’s a deal and let me be very, very clear first ,I am not a developer or a computer science background major… I have history and tech from a long time ago. I understand systems and systems analysis incredibly well because that was kind of my job for 20 or so years., and I am not offering a definitive diagnosis. Instead, I’m offering a differential where we can actually do falsifiable testing…

But speaking specifically and only on behalf of my professional career field, which is emergency medical services, we don’t bring ideas to the table until we have something to show, I didn’t come in here empty-handed. I’ve been trying to reach out to people for about two years., I finally started trying to figure out how to build it myself. Unfortunately, I’ve got horrible ADHD and dyslexia coding is a bit of a. Effin nightmare, but I don’t see why we can’t just write it down and start building things that actually work???

And I mean this part very seriously, and I do not mean that disrespect disrespectfully to anybody, but the people that are at the top of the food chain covering me publicly known side of this industry have all said and multiple times “ this may be the thing that kills us”, that is not a statement in my profession than this taking lately, especially by somebody that loves Hakk and understands it…

All I’m asking is that somebody take it and prove me wrong… if I’m wrong I’m happy I’m wrong means the problem really isn’t solvable but if I’m right???

That means somebody who worked solely in trauma based medicine for 20 years outside of y’all’s industry was able to sit down and objectively look at the problem differently from how you guys were and find something maybe?? (I am reiterating I do not guarantee. This is a solution. This is at a differential diagnosis and a working plan for a false fiber test with code in repository for implementation., I simply don’t have a strong enough system or resources to be able to do this myself)

I can’t begin to imagine how hard it must possibly be to be in anybody working in this industry as their career field right now… you’ve got people in leadership that are screaming toaster, and you’ve got people in leadership saying we are literally in peril, and there is a lot of variation in between all of that… but what I can do is say I’ve got a really unique skill set that’s aimed at the type of problem you guys are struggling with, maybe I’m wrong, I hope I’m wrong… but the medicine we don’t tell people they are wrong just because we don’t like the idea or it sounds weird, check it out we study it we test it. We investigate it. We prove it, or disprove it and we say. “ here your data go away.”

I know this because I’ve been a part of industry, changing paradigm shifting research and development before launches all kinds of stuff… anybody’s ever heard of the resuscitation consortium study, development, and implementation of pit crew CPR, if you’ve ever heard of mobile integrated healthcare? Those all came out of the system that I grew up in my career and was a part of the development of… we do big teamwork, and we all put hands in and get dirty, but because of that the system I work for got awards no other EMS system in existence has ever gotten and then we set the standard that the nation now follows and I’m not saying that out of ego or humorous I’m saying it to say “ I look at problems very differently from how most people do because that’s what I was trained to do”, that’s it…

And for those that want to validate my claims, look up MedStar Mobile Healthcare Fort Worth, Texas…

16

And I’m gonna follow up with something…

Since when is too hard or cost for time and trouble justification for not solving what your own industry identifies as an existential threat to our entire species??

Either it’s an existential threat possibility, in which case there are 8 billion people counting on you guys to figure something out

Or it’s not an existential threat in which case stop telling us this, it worries, people and worries people that have background in stuff like disaster planning scenario training all kinds of shit… what I do know actually, myself, is there currently there is no large language model in existence that would be safe to go into a first responder environment and that’s not even as an auxiliary aid just a situational awareness system. They hallucinating way too damn much., and there’s no stability…

I apologize if that comes off mean but, I’ve never been allowed in my professional career to look at a problem and say “nawww, I ain’t feeling that today and I just don’t got the money for it”, because when I say that my profession?

Someone pays with their life… that’s not a joke that’s not an over exaggeration that’s a fact of reality.