Question Can LLM's theories be banned?

[marcin]

Simple question + the poll.

 

[COLGeek]

Use of LLMs is not necessarily bad. They can be used to help clarify a point or to make something easier to read/understand.

The vast majority of LLM users don't really know how they work, nor understand the content they can produce. Those users assume the AI aspects are actually knowledgeable, they are not.

LLMs are only as good as the data used for training. In addition, highly technical content is often just wrong when viewed through the eyes of an actual human expert (I see this very often on tech sites).

Users of forums should not post LLM/AI generated content that does not have some actual independent input by the user themself. Something the user actually wrote and understands. Using AI/LLMs to post something that the user knows nothing about should be avoided.

Asking a LLM random questions and posing random prompts to develop a theory is a way to formulate a feasible sounding hallucination as the LLM tries to produce an output to match the given prompts. This is not science and epitomizes the term GIGO.

 

[marcin]

@COLGeek Could you give an example of a physical theory created with LLM that you consider reasonable?

 

[COLGeek]

@COLGeek Could you give an example of a physical theory created with LLM that you consider reasonable?

None that I am aware of. Do you have an example?

Now, I will say that an LLM can be used to draft a reasonable summary or comparison of some technical information (including an explainer for a theory). However, I would suggest such output be taken with a grain of salt. As noted earlier, I have seen many LLM well-written outputs that were simply wrong.

Last year, I conducted a personal study of the major consumer LLMs for a topic I am considered an expert on. I know the information regarding the topic is readily available and accessible. NONE of the LLMs ever got the "right answers". I would prompt (teach) the correct information back to the LLMs to see if they learned. They did not. Subsequent queries for several months produced the same results.

LLMs have a place, but not for theory generation that can taken at face value. Skepticism is very much needed when looking at these LLM generated results.

 

[marcin]

None that I am aware of. Do you have an example?

No. Thx for the honest answer.

Now, I will say that an LLM can be used to draft a reasonable summary or comparison of some technical information (including an explainer for a theory). However, I would suggest such output be taken with a grain of salt. As noted earlier, I have seen many LLM well-written outputs that were simply wrong.

I agree.

LLMs have a place, but not for theory generation that can taken at face value. Skepticism is very much needed when looking at these LLM generated results.

I agree. Why then is this forum a place for them?

 

[COLGeek]

LLMs themselves should not be banned. As previously noted, they can be helpful.

We moderate content as needed. While you may not see it, we do remove such content on occasion.

 

[marcin]

Thank you for it.

They can be helpful to summarize, compare or explain existing theories, but they are mostly used for GIGO, as you rightly called it. I don't want to ban LLM's explanations, but their hallucinations, which is also your expression.

 

[Astraeus]

I respectfully — but firmly — disagree with the sentiment expressed in this thread.

LLMs are not perfect, but they are tools. To dismiss or disqualify any content simply because it shows signs of LLM assistance is short-sighted. These systems are here. They are powerful. And — perhaps most importantly — they democratize access to structured thinking.

I understand where some of the frustration comes from. You've probably seen your share of poorly formed theories, and I get that fatigue. But to discount all ideas assisted by AI is not just unfair — it's intellectually dangerous.

When I introduced my UMT post, I started with a conversation. That wasn't a gimmick — it was intentional. AI didn’t write UMT. It helped me test it, structure it, and verify internal consistency. The theory is mine. The refinement? That’s where the tool mattered.

We shouldn't be afraid of these tools. We should be using them to make our thinking sharper.

 

[marcin]

@Astraeus thank you for speaking up. Truly disrespectfully, your Pastry-Baked Heresy: A Manifesto is the perfect example of LLM's garbage. It would be master class trolling, if you didn't praise your clarity so much. I think that you think that you figured out a way to avoid criticism with somewhat funny standup. One of your definitions particularly caught my attention.

Sprinkles
Local quantum deviations from expected geometry.
Commonly dismissed as “instrument noise,” but actually the cosmic equivalent of sarcasm.
Function: Decorative. Disruptive. Possibly intentional.
Cannot be predicted, only savored.

Your whole post is intentionally disruptive. I would probably keep my mouth shut if you didn't mess with the Einstein field equations by adding your own term like all the people who want to change or ridicule the basics of physics without the basic knowledge. You also seriously praise yourself

We don’t simplify.
We sharpen.
We coat the raw edge of revelation in metaphor, not to sweeten it, but to make it palatable.
Because the universe is not here to be clean. It is here to be real.

and write about the invented data that you'll be pointing to.

We will laugh in the halls of orthodoxy and point to the data when they call us fools.

Please, point me to your data now and also answer my last question in your thread: https://forums.space.com/threads/pastry-baked-heresy-a-manifesto.70914/post-616472

If anyone wonders how much trolling there is in your trolling, you're seriously putting your donut on top of your Universal Motion Theory.

An asymmetrical donut.

UMT had been pointing there all along. Bounded curved motion doesn’t collapse inward — it wraps. It folds into itself without vanishing. And as we worked through the math, it became clear: the toroidal structure wasn’t just a metaphor. It was required. It sits directly in UMT’s falsifiability table — the theory lives or dies on whether these shapes can be seen.

I wouldn't take it is so seriously if it weren't a bunch of people who were truly enlightened by a donut without your sweeteners.

I am not a physicist or a cosmologist, so this theory wasn’t conceived by studying data or devising complex formulas. I am an explorer of consciousness, and the idea came as a flash of inspiration while I was meditating and inquiring with a friend, Valerie. A few days beforehand, I saw the following image on the Wikipedia…
https://evolvingsouls.com/wp-content/uploads/Curved-Universe.jpg
And we had been experiencing torus-shaped energy fields (like the one pictured below) during our meditations and self inquiries for several weeks…
https://evolvingsouls.com/wp-content/uploads/Animated-Torus.gif
Then, in a flash of inspiration, the two images came together (as pictured below) and we realised that the universe had toroidal geometry. This is not a new idea, but it was new to us, and we knew it was significant.
(...)
https://evolvingsouls.com/blog/toroidal-universe/

 

[COLGeek]

We shouldn't be afraid of these tools. We should be using them to make our thinking sharper.

We actually agree, a bit. An LLM can be useful as you stated and as previously discussed.

However, it is critical to understand how LLMs actually work. Think of them a bit like search engines on steroids. They will try to formulate a response to the questions asked. Those responses can be true, or not.

There are endless cases of well written (by LLMs) piffle and nonsense. They are well reasoned and thoroughly wrong at the same time.

 

[Astraeus]

If anyone wonders how much trolling there is in your trolling, you're seriously putting your donut on top of your Universal Motion Theory.

I am. I won't deny the data, just because it isn't perfectly shaped.

I find it curious that you’re offended on Einstein’s behalf, when he himself expressed clear discomfort with singularities.

Example quote (Einstein & Rosen, 1935 paper on the Einstein-Rosen bridge):“The solution given by Schwarzschild... suggests that the singularity is not real.”

In 1947, Einstein wrote in a letter to Max Born:“I do not believe in singularities because I do not believe in mathematical infinities as representing something real.”

If half the energy spent dismissing ideas were spent asking better questions, imagine where it might lead. These tools are free. Ask your own. See what unfolds.

I understand — we spend years learning the universe through familiar patterns. Then along comes someone questioning it — with equations, and yes, a little flair of confection. I know what I’m proposing is disruptive. But are you suggesting I should stay quiet? That attitude doesn’t look like skepticism — it looks like suppression.

We actually agree, a bit. An LLM can be useful as you stated and as previously discussed.

However, it is critical to understand how LLMs actually work. Think of them a bit like search engines on steroids. They will try to formulate a response to the questions asked. Those responses can be true, or not.

There are endless cases of well written (by LLMs) piffle and nonsense. They are well reasoned and thoroughly wrong at the same time.

We might agree more than you think. You’ve said LLMs can be useful. I’ve used them not as sources of truth, but as instruments of inquiry. They’re not search engines on steroids. They’re cognitive amplifiers. They help me test, iterate, and explore edge cases — far faster than human discourse alone allows.

Yes, they can echo nonsense if misused. So can people. The key is the question. If the query is thoughtful, the tool sharpens it. If it’s hollow, the tool reflects that back.

I’m not here to evangelize or troll. I’m here to explore. And I’m comfortable with friction — that’s often where discovery begins.

 

[marcin]

I am. I won't deny the data, just because it isn't perfectly shaped.

I find it curious that you’re offended on Einstein’s behalf, when he himself expressed clear discomfort with singularities.

Example quote (Einstein & Rosen, 1935 paper on the Einstein-Rosen bridge):“The solution given by Schwarzschild... suggests that the singularity is not real.”

In 1947, Einstein wrote in a letter to Max Born:“I do not believe in singularities because I do not believe in mathematical infinities as representing something real.”

Is this the data confirming your UMT or a donut shaped black holes or a donut shaped universe?

If half the energy spent dismissing ideas were spent asking better questions, imagine where it might lead. These tools are free. Ask your own. See what unfolds.

I understand — we spend years learning the universe through familiar patterns. Then along comes someone questioning it — with equations, and yes, a little flair of confection. I know what I’m proposing is disruptive. But are you suggesting I should stay quiet? That attitude doesn’t look like skepticism — it looks like suppression.

Oh no, please, I'm dying to know what coordinates you used in your donut's D_μν tensor. Don't be quiet and answer my question please.

We might agree more than you think. You’ve said LLMs can be useful. I’ve used them not as sources of truth, but as instruments of inquiry.

Compare this statement with your previous statements:

We begin with a truth so soft, so obvious, that it slips through the fingers of high theory:
(...)
But the glaze of truth is unmistakable.
(...)
It is a sign they have reached the edge of the Crumb Horizon and glimpsed the frosted core of truth.

 

[COLGeek]

Yes, they can echo nonsense if misused. So can people. The key is the question. If the query is thoughtful, the tool sharpens it. If it’s hollow, the tool reflects that back.

While the questions posed are important, so is the training data. Results must be vetted before considered as factual. Many leap to incorrect conclusions, when they don't perform this step. An assumption of accuracy is flawed.

Writing poems, song lyrics, recommendation letters, limericks, etc are perfectly fine outputs of LLMs.

Well-founded, researched, and vetted theories, not so much.

We'll see how well received your UMT is.

 

[marcin]

@Astraeus the more I look into it, the more intentional joke it becomes.

Donut Tensor (D_μν)
A corrective term to Einstein’s field equations accounting for observable asymmetries in gravitational topology.
Predicted by those who stared into the ngEHT and saw pastry instead of perfection.
Units: curvature per bite
Behavior: grows unstable under spherical assumptions.

and

Note: All components are understood in pastry-coherent units (PCUs), normalized against a reference cruller.

That definitely does not answer my question about your coordinate system, but it tells me, that you are mixing your fairy units with the actual units in the Einstein field equations. That makes your tensor a joke, and since it's the only mathematical formulation in your theory, it makes your whole theory a joke.

 

[Astraeus]

UMT is guided solely by observation. Its foundational principle is simple: it must never contradict what is observed.

Still skeptical? Visit Universal Motion Theory. Copy the text. Paste it into an LLM. Ask it, “Is this guy crazy?”

As we've already discussed, LLMs are excellent summarizers. Let the system tell you what it sees.

 

[marcin]

@Astraeus you're adding the inverse square meter to the curvature per bite. It's like adding unicorns to electric charge.

 

[Geoff Dann]

LLMs are like every other computer system humans have ever devised in one key respect: garbage in ---> garbage out.

Everything depends on the quality of the prompt.

 

[COLGeek]

LLMs are like every other computer system humans have ever devised in one key respect: garbage in ---> garbage out.

Everything depends on the quality of the prompt.

....and the data the prompt queries.

Even then, an LLM will only provide a likely response based on similar data. It is a guess, possibly correlated with something else, possibly pure fiction.

They will attempt to respond to nearly any question/prompt, no matter if well founded or not.

 

[Geoff Dann]

....and the data the prompt queries.

Even then, an LLM will only provide a likely response based on similar data. It is a guess, possibly correlated with something else, possibly pure fiction.

They will attempt to respond to nearly any question/prompt, no matter if well founded or not.

Yes. Although in most cases the data is pretty good. Exceptions come were the data was distorted for commercial, political or legal reasons, and in cases where there is entrenched dogma in academia (e.g. materialism in physics, anti-realism in a lot of postmodern political philosophy). There are ways to compensate for at least some of this. The LLM will even tell you where it thinks its own data might have been distorted.

 

[billslugg]

I'm sure most of the data is pretty good. Unfortunately, this is what gets us into problems. One of the prerequisites for trusting an AI output is that you know the input is true. Give us a way to tell this and we'll believe it. Until then I consider all AI output as false and in need of verification.

 

[Geoff Dann]

I'm sure most of the data is pretty good. Unfortunately, this is what gets us into problems. One of the prerequisites for trusting an AI output is that you know the input is true. Give us a way to tell this and we'll believe it. Until then I consider all AI output as false and in need of verification.

It all needs to be checked for mistakes, yes. AI is capable of indescribable brilliance, but also regularly makes very silly mistakes. In general it is pretty good though. It is exceptionally good at explaining the relationships between different sets of concepts.

 

[COLGeek]

This was from earlier in the thread:

Question - Can LLM's theories be banned?

Simple question + the poll.

forums.space.com

Recently, I re-ran the experiment over multiple platforms. Zero correct answers.

I'll not share the specifics due to privacy, but the technical topic's information is readily available and has been for years, with years of subsequent publication and discussion via multiple public venues. The types of resources used to train LLMs.

 

[billslugg]

We understand AI is capable of brilliance and silliness. The problem is sorting one from the other. No one seems to be able to do that.

 

[COLGeek]

We understand AI is capable of brilliance and silliness. The problem is sorting one from the other. No one seems to be able to do that.

What is generally recognizable is LLM written content. Especially on forums. They tend to stand out for their style and syntax.

 

[Astraeus]

We understand AI is capable of brilliance and silliness. The problem is sorting one from the other. No one seems to be able to do that.

You raise an important point, AI output does range from insightful to flawed, and distinguishing between the two is critical. But that’s precisely why engagement matters. If we’re concerned about quality, we need to evaluate content directly rather than disqualify it based solely on its origin.

Blanket dismissal makes it impossible to ever identify the ‘brilliance’ you acknowledge AI might produce. The act of scrutiny, of assessing reasoning, evidence, and internal consistency, is what separates skepticism from evasion. I'm fully open to critique of any specific claims I’ve presented, and I welcome it. But rejecting them without examination doesn’t protect standards—it prevents discourse.

Much of the conversation here appears focused on discrediting LLM-generated content rather than evaluating the ideas themselves. The thread title, “Can LLM’s theories be banned?”, reflects that tendency.