Aww y’all still mad your color revolution failed?

It seemed pretty clear to me. If you have any clue on the subject then you presumably know about the interconnect bottleneck in traditional large models. The data moving between layers often consumes more energy and time than the actual compute operations, and the surface area for data communication explodes as models grow to billions parameters. The mHC paper introduces a new way to link neural pathways by constraining hyper-connections to a low-dimensional manifold.

In a standard transformer architecture, every neuron in layer N potentially connects to every neuron in layer N+1. This is mathematically exhaustive making it computationally inefficient. Manifold constrained connections operate on the premise that most of this high-dimensional space is noise. DeepSeek basically found a way to significantly reduce networking bandwidth for a model by using manifolds to route communication.

Not really sure what you think the made up nonsense is. 🤷

I’m personally against copyrights as a concept and absolutely don’t care about this aspect, especially when it comes to open models. The way I look at is that the model is unlocking this content and making this knowledge available to humanity.

Ah yes, they must be stealing IP from the future when they publish novel papers on things nobody’s done before!

drill baby, drill

I do indeed hate living under capitalism

The fact of the matter is that this is a perfect example of LLM actually doing something useful and helping researchers figure things out. I also love how you’re now playing an expert on deciphering ancient scripts. You should go let the researchers know asap what a bunch of dummies they are for not being able to figure it out on their own.

Maybe find a new hobby other than sealioning into threads to screech about how much you hate LLMs. It’s frankly tiring of watching people perseverate over it.

I’m sorry you felt the need to argue a point nobody was bringing up, and which added absolutely nothing of value to the discussion.

This whole thread was just you trying to make a straw man.

You may not realize how scientific process works, but it’s not based on trust. What actually happens is that researchers publish a paper that explains their thesis, and provides supporting evidence. Then you have this thing called peer review where other experts in the field examine the findings and make their own assessments. The reality is that hallucinations and fabrications aren’t exclusive to LLMs, humans do this stuff all the time all on their own. This is why we have the scientific method in the first place.

And I’m reminding you over and over that it’s completely beside the point. I’m sure when they publish the research they will provide the reasoning for their hypothesis, and how they tested it. Then other researchers will examine their findings, and point out problems with the research if they exist. That’s how scientific process actually works.

You literally just made up a baseless argument that the researchers aren’t doing due diligence. I’m skeptical of your thesis and I’m not seeing any attempt on your part to provide any supporting evidence for it.

I get the impression that you don’t understand how science actually works. Science is about examining the evidence, then making hypothesis, and testing them to see if they’re viable. Proof is never guaranteed in the scientific process, and it’s rarely definitive. Seems to me like you just wanted to bray about AI here without actually having anything to say.

Nobody was claiming a proof, that’s just the straw man the two of you have been using. What the article and the original post from researchers says is that it helped them come up with a plausible explanation. Maybe actually try to engage with the content you’re discussing?

In other words, you’re saying neither of you could be arsed to click through to the actual discussion on the project page before making vapid comments? https://blog.gdeltproject.org/gemini-as-indiana-jones-how-gemini-3-0-deciphered-the-mystery-of-a-nuremberg-chronicle-leafs-500-year-old-roundels/

I’m actually building LoRAs for a project right now, and found that qwen3-8b-base is the most flexible model for that. The instruct is already biased for prompting and agreeing, but the base model is where it’s at.

I’m not sure what you’re claiming I blew past here. I simply pointed out that nobody is expecting LLMs to validate the solutions it comes up with on its own, or to trust it to come up with a correct solution independently. Ironic that you’re the one who actually decided to blow past what I wrote to make a personal attack.

That’s where the human comes in though. The value of genAI is that it can generate outputs that can trigger ideas in your head which you can then go and evaluate. A lot of the time the trick is in finding the right thread to pull on. That’s why it’s often helpful to talk through a problem with somebody or to start writing things down. The process of going through the steps often triggers a memory or lets you build a connection with another concept. LLMs serve a similar role where they can stimulate a particular thought or memory that you can then apply to the problem you’re solving.

the media doesn’t exactly bring it to people’s attention, so you really have to go digging on your own to find out

Yup, and this is precisely why it was such a monumental mistake to move away from GPL style copyleft to permissive licenses. All that achieved was to allow corporations to freeload.