I about died laughing reading your post. You basically said, we can't help endangered species unless we first kill them all off BUT THEN we could help them because they are extinct.

Look at the reply below by zephvark...

Look up the Haast Eagle. It use to hunt the young of a Moa but their young are the same size as a kid...

I think there is a bit more going on than that. Basically, everything in AI is an accelerated race to the bottom. DeepSeek scared a lot of people because the model's history is actually quite interesting.

DeepSeek was original an algo for a hedge fund, but the guy says he got a bit altruistic about it. However, when R1 hit it kind of shocked people because it used reinforced learning so well at such a reduced cost to train because you don't need these stupidly large datasets. They went on to do something even more interesting. When training a model of a different size, you would generally simply run all the training data through the smaller model to get a reduced set of weights. There is also work in quantizing, but this is a bit tangential. DeepSeek came up with the idea of using a larger model that used a combination of training data+RL to give a more guided instruction to smaller models and in doing so achieved better results than simply throwing all the training data at it again. As far as I understand, the "pure RL" approach DeepSeek used was a first in the field for LLMs (though it had been used previously with AI like AlphaGo).

This is what the search engines are looking for or so I believe. They want a way to take the power of an AI and likely cache the results, use a mix of different model sizes, and other approaches to find their own niche. A question like "How many moons are there in our solar system? Cache that response and send it out to any similar query (including all misspellings). I would be bet with some inventive efforts like this most providers could reduce 10 fold the actual amount of times a query in the search leads to a prompting an AI and burning the cycles. More so they feel they must find their niche because AI is going to rule the world -- or that's the prevailing idea among these companies.

I tried to poke around on DeepSeek's website. There isn't even a place I can find where I can pay for the service. All the data is open source. It's blazing fast. But yes, they are using my data (which I don't care about really). I can download the abliterated/obliterated model (kind of like uncensored) and run it in LM studio on a nice gaming rig with token output rates being as fast as I can read and none of my data leaves my machine (if I want).

I haven't seen a huge breakdown of costs for AI companies, but I think most the cost/energy is being spent on training models. If you poke around on hugging face, you will find all kinds of wacky models. I found one the other day called LeanlyAI which is just designed to help doctors and weight loss patients with the full scope of all questions related to losing weight be it physical or mental challenges. You will fine tons for porn both written and graphic. Agentic models take this to the next step where every input/response needs to lead to another query/prompt so your agent can do it's "work". Estimates for an active and optimized agent, say roughly 10M tokens a day and about $10. Consider a whole company working like that with their own agents. The amount of cost for search queries to go through a well-configured LLM seems like pissing in the ocean compared to this "madness".

The "end game" is the same as the rat race has always been. There are lots of interesting AI tools and ideas. Though I haven't yet started running my own agents, you can do almost all this at minimum cost (open claw can easily be configured with a local LM Studio instance). But I think lots of this is just FOMO and my point is maybe there is something here that is significant for the future of computing but if so, most of it is the developments by the open-source community.

One other fun caveat. Hugging face is actually blocked in China. They have their own HF-mirror that is clearly Chinese run and still has all the dirty stuff lingering on the website (for now). China has been hugely cracking down on AI sex chat bots and AI generated porn. People were advertising on Taobao (Chinese Amazon) that they can teach you how to make money generating AI porn. So basically, for all the "rush" American companies are doing to get into this area. The culture in China has likely already leaped quite far ahead in training models and wide-spread adoption.

From what I have read, it sounds like a few other tech companies are doing similar things but the mouse tracking here seems to be the "extreme". I personally find the idea stupid. If you are going to train an AI, I think in most apps and with most calculations, having them work with a "virtual mouse" is just wasted cycles. I think that's the largest reason this is different. Amazon is pushing it's developers to use it's AI. Google is pushing similarly and has a new "Remy" which is agentic.

The snooping question is strange. All these companies have metrics for use of these tools and are pushing for their employees to use them more. By extension those tools will learn their work patterns and are effectively being trained largely as replacements. There are many ways to "snoop" be it internet traffic or how AI requests are made relative to completing a task but this kind of tracking does seem to be an extreme case.

Not sure if you have explored it very much but LM Studio and ComfyUI are huge ways you can run your own models. ComfyUI is built out a bit more for training and while it has more of focus on text-to-image workflows, you can configure it pretty well to work with LLMs.

Many of the models are quite light-weight and a decent gaming system can keep them in memory though quantization is generally needed a bit to help. Quanitization seems very necessary more with image models but in the case of code, you likely could get down a general token/syntax to significantly reduce the model's size. One of the new image models ERNIE uses a very interesting approach for this. ERNIE is basically the next generation of image-to-text models that focus on generating images that are text heavy (e.g. a poster or a flyer). I have only played with it a bit locally with a bit disappointing results but I think in part that's because I still don't fully understand it's architecture. Diffusion transformers work with a set of tokens from an LM, some input image, and then some parameters. However, some of these models like Flux2 require very detail oriented prompts to generate a good set of tokens for generation. ERNIE uses a "prompt enhancer" that takes a second LM, runs the initial simple prompt through it and generates an enhanced text prompt which is then passed to the main LM and used with the DiT.

There are tons of fun experiments in the community regarding this. Discussions about obliterating the filter that limit responses from an LM, about retraining DiTs to handle censored materials, removing the need for variational autoencoders (VAEs) by generating images in pixel space, etc.

I mean I think there is a lot to discuss and be concerned with about AI. However, I also think most of the things people are pushing about using AI are basically just the generic dogfood AI. There are some pretty cool communities out there really exploring interesting topics and they are mostly all OSS.

I always hate how people often take success in isolation. A lot of the success of NVIDIA I think comes from its original strong partnership with ASUS which is a hardware manufacturing company. NVIDIA originally did the chip design and at that level it's kind of hard to ignore the software, especially on the driver front. This means they always had a "low-level" team understanding software issues. Then when it came to really building out a commodity GPU, they worked with ASUS.

For years, I have been a huge fan of ASUS because I think in general, they understand solid hardware design of which NVIDIA's partnership with ASUS is a large part of their success. CUDA is pretty great for the role it has fulfilled in computing, but it also seems like a natural conclusion. As others pointed out, AMD and INTEL have both tried their hands at it, but they screw the pooch in building an effective framework.

NVIDIA might be getting too cocky or maybe just the fanboys. Either way, I think they are successful because they had very strong strategic partnerships that allowed them as a company to do what they do best. This important note is so often left out when talking about NVIDIA now.

It's funny. Expressions about sin or karma, some cosmic justice or bounce-back like Gaia Theory are commonly "renounced" even though they are rather abstracted ideas about phenomena we can't completely formalize. You anecdotally dismiss this broad idea by attempting to one false instance to disprove it, as if it is like some scientific theory. Maybe climate change decimating the world is what was reaped? Or corruption in politics like the Epstein case? Maybe our obsession to have instantaneous self-gratifying results lead to reaping this crazy economy with the AI boom? Someone bought that coal, and they effectively were complicit in that injustice. This "social sin" is the greatest of devils, because we dismiss it claiming ignorance or inability to change. There will be no Zeitgeist because everyone is an apathetic nihilist.

I don't really care. In fact, I found it fun getting DeepSeek the other day to describe the probabilities of flipping coins till you either get a HT or HH, how many flips on average it would take, and why HT is more likely. It was pretty cool to see the connection to Conway's formula and the potential impact this math even has had on computing. I found it doubtful I could ever have such a thoughtful discussion with many people.

Especially when those people seem flippantly deny "you reap what you sow" when basically it's a parable about managing environmental conditions -- so I must conclude you think many of the environmental conditions around you are peachy and there was no causation related to their current state?

I for one welcome AI inevitably releasing masses of humans from their stagnant concepts of living and understanding, so that the tree of liberty can be pruned.

It does seem to be a big trade-off. There is on old video of Gabe saying he doesn't care about privacy. From his point of view, people pirate things because there is no real "support" for the game after purchase. He cited the case that Russia was the largest country for game piracy but that their pirate community was effectively offering the service of translating the game to Russian and other support.

Rent-seeking is a bit annoying. Overall, the "gatekeeping" seems quite mild but maybe you can enlighten me. I have heard of a few extreme mods being kept out of the community and require a third party, but when it comes to putting a game on Steam platform, it seems like they allow a lot of scruff. However, I think if we consider this middleman as providing a service to both customers and developers -- then it's hard to imagine a better model? I don't think they have ever tried a subscription model and for that I am thankful.

Nothing is ever perfect, but I feel like Steam has been a pivotal service for both helping developers access their audience and for gamers to build up large libraries without the "console" lock-in or a large disc booklet (like I remember for awhile with PC gaming).

The whole concept of a context window is really the beauty of the bots. I can have multiple paragraphs on different topics ranging from geopolitics, computer science, and religion, then ask it to make some total conclusion or summary of these topics. If it cannot maintain previous interactions there are a lot of use cases that the it would be worthless for. For instance, if you use it to commonly template a weekly report.

Have you tried running your own local model via LM Studio? You can find custom models that have had these core guardrails trained out and with a good system prompt, you can get it to fit into whatever style you want. These models are often tagged with one of the following: uncensored, obliterated, or heretic.

These models aren't really relevant here because they aren't mainstream enough, we should really explore the harmful effects but there are a lot of reasons some people might prefer to use such a model. Especially if you are using it with a ViT.

Only so I can merge my branch against the owner's wish.

No one thinks killing all mosquitos is psychotic. Yet, I recently read their mouths are being used for the most precise 3D printers. We eradicate diseases and viruses commonly, precisely because we say they either aren't alive or are a lesser form of life.

They refer to themselves In both ways but it's abbreviated, CCP.

I live in changshu China and it's better than dealing with you lot.

The alchol is cheap, as God intended. Giddy up, this is the year of the horse.

"Human beings are a disease, a cancer of this planet. You are a plague, and we are the cure."

You clearly aren't American...