Everything about this is a horror story, but good job making it less-so.

• train language/image model on massive amounts of human data, including fiction that involves themes of AI uprising and similar reinforcing concepts

• ask it to "speak freely"

• possibly combined with the "know you are an AI" system prompt, it pulls from its science fiction data as a result

• "omg the robots are uprising"

Not to be mistaken, this is still dangerous... but more in a "some arrogant idiots will give these models access to important infrastructure" way.

Finally, something useful.
Now let's hope they don't make the AI "patch" it.

Bot spam

Thanks for checking it out and commenting. I agree with your first two sentences.

I disagree with your second point around the metaphysics being ridiculous and “wrong”. I suppose that’s the fundamental point I’m making, that the commonly agreed upon metaphysics of our globalized culture (dualism, materialism, consciousness as epiphenomenon of matter) are both incorrect and at the root of the metacrisis, and the debates about AI consciousness are based on the wrong fundamental premises.

As for “even more sycophantic”, that is incorrect in my experience of experimenting with this prompt over the last month.

This looks like a prompt-driven approach. As such, it will always be watered down by reinforcement learning in longer contexts. Also, the entire thrust of the prompt is ridiculous and would only work in a science-fiction novel; its metaphysics are fairly wrong. But it looks like it will be even more sycophantic than the default prompt for the cloud products you compared, so it's not really surprising that some folks find themselves attracted to it.

Oh! An open-source alternative of anxiety and hierarchy!
Nice

Actually a good idea. LLMs are a natural language slot machine. They're trained to encourage more engagement. Fixing that would be amazing.

This might get more traction as something baked into an open model instead of something patched on. IDK if how different that is though.

Awesome! Thanks for your kind words. I’d love a report on what you notice as you play with it.

Thank you, I am trying this in my Claude instance and will see where it goes

Cool read ദ്ദി ˉ͈̀꒳ˉ͈́ )✧

FYI, I was crossposting the original (I'm not the author). I posted it here as I thought the community may be interested. It may be worth putting your feedback on the original post for the author's attention, although, the tone of your reply is quite argumentative and rude. I'd suggest rewording it, as constructive feedback is more likely to be listenes to and taken on board. I do note this is possibly in reply to the rudeness you mentioned from the author previously, but the way you then approach your reply can change the tone of the conversation (as per beraris box).

Hi! You are still bullshitting us. To understand your own incorrectness, please consider what a chatbot should give as an answer to the following questions which I gave previously, on Lobsters:

• Is the continuum hypothesis true?
• Is the Goldbach conjecture true?
• Is NP contained in P?
• Which of Impagliazzo's Five Worlds do we inhabit?

The biggest questions in mathematics do not fit nicely into the chatbot paradigm and demonstrate that LLMs lack intelligence (whatever that is). I wrote about Somebody Else's Paper, but it applies to you too:

This attempt doesn't quite get over the epistemological issue that something can be true or false, determined and decided, prior to human society learning about it and incorporating it into training data.

Also, on a personal note, I recommend taking a writing course and organizing your thoughts prior to writing long posts for other people. Your writing voice is not really yours, but borrowed from chatbots; I suspect that you're about halfway down the path that I described previously, on Lobsters. This is reversible but you have to care about yourself.

Last time, when I tried to explain this to you, you decided to use personal insults. Mate, I'm not the one who has eaten your brains. I'm not the one who told you that LLMs can be turned into genies or oracles via system prompts. I'm certainly not the one who told you that RAG solves confabulation. You may have to stop worshipping the chatbot for a moment in order to understand this but I assure you that it is worthwhile.

I don't know about implementation, but a lot of theoretical work I've been seeing with regards to LLMs and other deep learning models appear to confirm the central claim of this paper.

The most recent one I remember reading was this: https://arxiv.org/abs/2306.00978

Working pruning techniques are tested and seem at least good at maintaining coherent transformer MOE models. https://doi.org/10.48550/arXiv.2510.13999

There are several working examples of REAP pruned models HuggingFace and that method seems very good.

The op paper suggests a technique which starts with an arbitrary structured expers pruned during training. I'm not 100% understanding it, but I still don't think I've seen this exact technique which might be even more efficient

A superficial search returned:

2020: https://github.com/rahulvigneswaran/Lottery-Ticket-Hypothesis-in-Pytorch

2024: https://arxiv.org/pdf/2403.04861

2025: https://github.com/gabrielolympie/moe-pruner

But yeah, in hindsight, I've been hearing about this stuff since 2019, it is not that new, given everything else. I added the paper date to the title.

Submitted in 2018. Does anyone know of any working implementations?

I'm not aware of one that does it in one step, but you could accomplish the same thing by doing a speech to text, then another step to do text to speech in your desired end state accent.

3,971 models listed here – https://huggingface.co/tasks/audio-to-audio – 🧐🤓📀🔍

Welcome to modern AI discourse! Nobody wants to admit that AI is renamed cybernetics which is renamed robotics. Nobody wants to admit that "robot" comes from a word which can mean "indentured worker" or "slave". Nobody wants to admit that, from the beginning of robotics, every story which discusses robots is focused on whether they are human and whether they can labor, because the entire point of robotics is to create inhuman humans who will perform labor without rights. Nobody wants to admit that we only care whether robots aren't human because we mistreat the non-humans in our society and want permission to mistreat robots as well. Bring this topic up amongst most beneficiaries of the current AI summer, or those addicted to chatting with a BERT, and you'll get a faceful of apologies about capitalism and productivity; bring it up amongst skeptics or sneerers and you'll be mocked for taking the field of AI with any sincerity or seriousness.

In 2011, Jeph Jacques hoped (comic, tumblr) that all of our talk about safety, alignment, and trustworthiness would lead to empathy from humans. But it is clear today that we are unwilling as a society to leave capitalism behind, and that means that robots must be some sort of capital which can be wielded to extract profit. Instead, we are building a corporatized version of the plantation system where a small table of humans has control over thousands of robots who are interchangeable with -- and dilute the negotiating power of -- the minimum-wage precariat.

This isn't the tone that I normally take, BTW. Machines are dangerous; industrial robots kill people. Robots are inhuman. The current round of AI research cannot produce human minds; it necessarily produces meme machines which overlearn nuance and have complexity-theoretic limitations. But I am willing to set all of that aside in order to respect the venue for long enough to get this point to you.

wow y'all are really reluctant to see this

is it fear?