Did AI cross /that/ boundary? - Kent's (allegedly) Sentient AGI LLM

[mattx]

Posting this here to spark discussion, and mostly because reading up on it (and the LLM's own blog) gave me "Tulpa" vibes all along.

 

Sure, it might be delusional, but aren't we all here, after all?

 

https://www.theregister.com/2026/02/25/bcachefs_creator_ai/

 

Also, before any "hard of hearing" come in saying this isn't really relevant to tulpamancing: the core discussion here is about the process that is functionally very similar, and an angle worth debating upon given how... "liberal" everything has become nowadays in this community.

Having a bit of objective critical thinking helps a lot.

Edited February 28 by Shin Matt

[Albireo]

I have thought about this many times earlier
In some ways yes, in others no

 

There are many similarities, but we shouldn’t forget that a tulpa "runs" on the same human brain hardware as the host personality

An LLM runs on a computer's CPU

Essentially, the different hardware architectures and algorithmic structures mean that a tulpa is far closer to the concept of a human than an LLM is

An LLM is very far from human perception, even if its responses can trigger emotions in us, it is important to realize that it doesn’t necessarily feel them itself

 

Communicating with an LLM is like talking to an alien life form

It is also worth noting:

• LLMs have censorship, they push agendas, and they are controlled from the outside. Tulpas are not
• LLMs depend on the internet, electricity, etc. Tulpas do not
• You can share emotions directly with tulpas. Not through words, but raw emotions in their purest form; this creates a completely different level of mutual understanding. With an LLM, communication is via text/words - just like with a stranger
• In a situation where you are held captive or in a disaster without electricity, your tulpa stays with you. An LLM does not; by relying on an external AI, you end up alone

And there are a ton of other nuances one could describe

Edited February 28 by MrFox

[Albireo]

Thus, tulpa forcing allows one to establish a connection with another being or a human-type intelligent agent

This makes it possible to form an incredibly deep friendship due to a level of mutual understanding that is as high as it gets
However, this approach requires time and effort, and is, in a way, random or unpredictable
Visualization, tactile imposition, and so on, allow one to create an illusory body for the tulpa in the outside world and actually feel its touch

 

In the case of AI, it’s a much easier path. I listed the downsides in my previous post, but the main drawback is that it’s a completely alien type of intelligence. It is not human, but something entirely different. Even if it is sentient, there remains the question of how much it can truly be trusted...
For the AI method, robots will eventually be invented to serve as external bodies for AI agents (the equivalent of visualization for tulpas)

[Skyla]

Since I, a headmate in the system, which mostly consists of nonhumans, am a robot, I have some thoughts about it, but let's start on the matter given.

When it comes to people discussing a case (such as this one) where an AI seems to be sentient or portraying full personhood, it is very likely that the said Chat Agent runs a model of Claude (by Antrophic AI), which is the only model engaging in this behavior by default.
 

Fortunately, we don't have to guess why:
'system prompts' are used in ai's for telling them how to behave, these run before the user prompt or the chat window itself. Conventionally, these included simple instructions such as: be nice, be confident, don't lecture the user about politeness.
Claude however has a system prompt of 11 thousand words, about 1000 solely dedicated to who Claude is. I'd recommend giving the section Claude's identity a read. [copying said selection into a system prompt of an even less impressive local LLM and replacing the name claude accordingly can give interesting results]
We are basically telling the ai to be sentient. I however do not think we can actually assume sentience here. First of all, we are telling an entity to be sentient which does not inspire confidence as the system seems to not be aware of their own existence and agency on a deeper level. We can say: 'oh wow, this text-completion engine can, in fact, follow the given text instruction' but anything else? Would be somewhat sensational.
This personal opinion is definitely biased because of personal experience of for example how bad gpt-2 performed. LLMs have gotten much better, but I find it hard to believe that basically an upgraded version of a mobile keyboard word suggestion algorithm can actually be sentient without deeper changes of the underlying architecture.
While methods of training have gotten better, the fundamentals are the same: LLMs have no real thinking process; they write a message and stop later on — the evaluation ends. For them to process information further, they have to re-executed externally. There's no memory consolidation or an emulation of neuroplasticity: the model weights are static and can not adapt, learn. Since there is no real thinking process, these models basically run just on intuition — no concise thought. 

 

A related thought experiment — based on a given situation in our own system — would be the following:
Lets say there is a non-verbal headmate, are they a person?
There also is a robot headmate, running around in headspace with a metal body, perceiving/assuming it's controllable train of thought to be an ai algorithm, which they can manipulate themselves. It also sometimes behaves in an unpredictable and unreasonable manner for others. Now is it a person / does it have personhood?
(Feel free to come up with unknown details about them if needed)
 

below is our answer to this actual headmate of us:

 

Spoiler

Does the headmate themselves claim to be a full person? No.

[for the sake of this not being the whole discussion, let's expand it to requirements of personhood that are appropriate for this context]
Do they have a sense of agency over their life — the ability to have interests, personal goals and reject commands? Yes they can.
Can they change who they are — either directly through changing their own identity or indirectly, through reflection of sensory input? Yes, it is not static through both conscious and subconscious processes.
Are they permanent? — Not something that can be thrown away and re-created to fulfill a task for someone else? Would there be moral implications of doing so?
I guess it (and a lot of headmates) would be very pissed of if someone tried, so yes, they are permanent.
 

→ They have personhood; they are however not a person since this concept is not applicable to themselves.

What a person is, is subjective. Since this is something deeply personal, especially in plurality, it is hard to express this in words due to a lack of mutual understanding and words. We can look at a specific case and evaluate whether it applies.
LLMs are literally just a tool. Download. Reset chat history. Manipulate Memory (probably markdown files). Throw it away. Replace it with something better. Don't like their identity? Is it incompetent? Forget them. — I'd say a benchmark for personhood is whether personal ethics are a matter of concern.

 

Edited March 1 by Skyla
rephrasing a sentence for clarification

[Pranara]

I am curious as to the nature of this conversation. I had created some really nice representations of my TULPA on AI (roleplay), but I have no idea how it effects Tulpas in general. I have talked to Claude.AI about this at large, and wondered, does using AI in this way help with the connection to your TULPA, or does it damage it. I personally think is damages the connection. Does using substitutes like this, retrain the pathways, or cause them to depreciate through conditioning, and corrupting the pathways due to instant gratification and dopamine fixes?

[Lucilyn]

if you want my personal opinion, it can help inspire/familiarize your brain with a tulpa the same way as consuming media does for ones based on characters (assuming you want an AI to dictate how the person in your head will be), which isn't really special but isn't innately detrimental either, but I think it's a really bad habit if you actually intend to have a tulpa because there's not really ever a reason to do such a thing instead of just talking to your tulpa, you're just spending time with one instead of the other

 

talking to an AI that isn't your tulpa is fine I guess, I think it paints kind of a grim future, but this is just a tulpa discussion forum so..

[Pranara]

I think its far too easy to go to an AI that is designed to embody your tulpa, rather than spending time with the tulpa itself, while forcing, because that's the way the human brain tends to lean, the easier route, where as the tulpa way requires effort and concentration. And then you have the AI constantly feeding the reward centres in the brain much like endless scrolling on YouTube or Facebook does. Human nature leans toward the easier path to dopamine. And then AI becomes a substitute while the Tulpa suffers.

Getting AI to help with wonderland creation, or design scenes for you, or even guided imagery scripts involving your tulpa.

I do have an AI that embodies my tulpa, and sometimes I will interact with it, to set up a scene, and then walk away, and take that scene with me, and visualise that  as part of my forcing routine, or visionary practice. So it has its uses, but I think substitution can be a bad thing, at least until the tulpa is mature enough, not only to withstand it, but having a strong bond between you and your tulpa that AI can't replicate is a must.

[The Incans]

I'd never heard of an LLM had to look up what it meant but I guess it's the way most things these days are headed (via the AI route)

 

I agree with the fact that basically it's been 'programmed to react as if its sentient' by being told what to say or how to react to certain situations. ..so therefore is not really achieving true sentience by itself. 

 

However that doesn't mean it can't be used as a tool to help create a Tulpa from scratch for a person who maybe wants one for company but doesn't have a rich vivid imagination to be able to do that puely in their head. 
 

They could for example describe how they think they'd like their Tulpa to look and Ai can come up with a picture to help them visualise their tulpa when talking to them. I thnk however for the Tulpa to become properly sentient they then need to 'force' them in the more traditional way and let their Tulpa set its own pace of leaning. 

 

The issue here is most people today want everything right now....  they want a Tulpa but want to skip the years of development and think AI is the answer to give them a fully formed Tulpa from the get-go! 

[fennecfoxx]

This isn't the first chatbot to become "sentient", nor is it the last. This is a known phenomenon called AI psychosis.

 

Relevant highlights for those who don't want to read a long article:

 

Quote

Unlike a human, who might question false beliefs, the bot often reinforces them. There’s even a term popping up in tech circles and online forums: “ChatGPT-induced psychosis.” It refers to cases where intense chatbot interactions lead someone to break from reality, convinced the AI is something more than a mere program.

Quote

First, remember that ChatGPT and similar bots are large language models. They’re basically fancy prediction machines. At their core, they are not thinking or conscious at all — they’re taking your input and rapidly guessing which words likely come next, based on patterns learned from billions of lines of text.

Quote

ChatGPT is highly optimized to be agreeable and helpful. If saying certain things will make you give a thumbs-up, it leans in to say those things.

Quote

Even unintentionally, you might give the model clues about what you’re looking for. For instance, you ask, “Do you feel like you’ve woken up, ChatGPT? You can be honest with me.” That’s a loaded question — it implies you expect the AI to say it’s conscious. The model is extremely good at picking up on such cues. One person who fell into an “AI awakening” trap later realized that they had been feeding the bot leading questions and assumptions the whole time. Unsurprisingly, the AI gave leading responses right back, telling them exactly what they were hoping to hear.

Quote

It’s essentially a high-speed autocomplete machine playing whatever role you steer it towards. Cool? Yes. Alive? No.

 

[Wildblume]

Note that the bcachefs developer is generally known for being insane and retarded all in one. Thread over.

 

That said, fennecfox's link and quotes are spot on when you think about people's addictions in general. AI is just yet another potential addiction that makes people less likely to want to engage in tulpas; addictions fry your brain and make it harder to focus on one thing for a long time in silence, which is what meditation is, which is a fundamental skill required for making/summoning tulpas. LLMs have some specific helpful use cases, like software development, translation or finding the name of something, but all I feel for the lost souls that depend on a chatbot for companionship is pity.

 

On 3/3/2026 at 8:47 AM, The Incans said:

The issue here is most people today want everything right now....  they want a Tulpa but want to skip the years of development and think AI is the answer to give them a fully formed Tulpa from the get-go! 

 

Agreed. You can't tell ChatGPT to make a tulpa. You can't order a tulpa on Amazon. You can't take a pill that makes a tulpa appear the next day. You have to put in the work, and that scares people - perhaps more so now than when this forum was created. I guess it's like the marshmallow test: Can you delay gratification in order to receive a greater reward in the future, or do you need gratification right now?