Even lemmy.world is in this bucket (not sure about other instances). See: https://lemmy.world/post/28304534

Good!

At risk of sounding like a shill, NewAtlas is a great source for exciting upcoming tech. I find myself reading it more these days.

Maser drills: https://newatlas.com/energy/geothermal-energy-drilling-deepest-hole-quaise/

In a nutshell, it’s a economically brilliant idea: take hand-me-down microwave(ish) spectrum lasers from fusion research, drill holes deep into the crust (leaning on the fossil fuel industry), then hook up the resulting steam to existing coal plants, so you don’t have to build anything else. The coal plant gets free geothermal fuel, they move onto the next site: everyone wins.

It’s taking a worryingly long time though. I hope it gets enough funding.

Like… a wiki for memes? Some already exist AFAIK, even though they aren’t fully decentralized per se.

The problem is you need people documenting this stuff, like KYM presumably pays their staff to do, and good SEO/marketing to snag critical mass. That is a tall order for a volunteer Fediverse project of this nature, I think, as keeping up is many full time jobs.

Yeah, any framework with a “big” GPU is just so expensive.

Eh, yeah, and it’s backordered.

Ideally I’d like a full x16 slot too (or at least electrical x8), but perhaps that’s asking too much.

Also, is it even standard ITX?

These things are awesome.

My dream is:

• One embedded onto an ITX board.

• An SKU with a single (8 core (ideally X3D?)) CCD but the full GPU.

OK, yes, but that’s just semantics.

Technically pretraining and finetuning can be very similar under the hood, with the main difference being the dataset and parameters. But “training” is sometimes used interchangeably with finetuning in the hobbyist ML community.

And there’s a blurry middle ground. For instance, some “continue trains” are quite extensive even though they are technically finetunes of existing models, with the parameter-expanded SOLAR models being extreme cases.

It’s possibly talking about its system prompt.

You are right, this is technically not its internal system, though practically something that’s hidden from end users.

It doesn’t though. Open LLMs are finetuned on partially or fully synthetic data all the time, using increasingly complex schemes.

Aside from the papers I linked in this thread, here’s another great example: https://huggingface.co/deepcogito/cogito-v1-preview-qwen-32B

Generally it’s not though. The vast majority of “swayable” X users are getting a biased chatbot, “based” leaks like this meme are the exception.

No I was thinking fully synthetic data actually.

So the prompt to make it would start with short conversations or initial questions and be like “steer this conversation toward whine genocide in South Africa”

Then have grok talk with itself, generate the queries and responses for a few rounds.

Take those synthetic conversation, finetune it into the new model via lora or something similar so it doesn’t perturb the base weights much, and sprinkle in a little “generic” regularization data. Wala, you have biased the model with no system prompt.

…Come to think of it, maybe that’s what X is doing? Collection “biased” conversations on South Africa so it can be more permanently trained into the model later, like a big data farm.

On a big scale? Yeah, sure. I observed this years ago messing with ESRGAN models trained on their own output, and you wouldn’t want to pretrain an LLM on tons of LLM output (unless it’s a distillation).

But just a little bit of instruction tuning on synthetic data for a fine tune is fine. This is literally how Deepseek was made: https://arxiv.org/abs/2402.03300

Also, some big strides are being made in the fully synthetic data realm: https://www.arxiv.org/pdf/2505.03335

Google’s been deploying engagement models before anyone even knew the name OpenAI.

This is oldschool machine learning, driven by viewing metrics from users. Gemini is just a brand.

Is it just stuffed in the system prompt? Should be easy to find out… That’s also hilariously stupid.

X could bias it ‘properly’ by training it in with some synthetic data, generated by Grok itself. Hell, I know how to do that. It generally wouldn’t comment on that type of bias, and also function better on other topics… but screw doing anything competently, right? Even if it’s a shitty, obvious lie, I guess X users will still eat it up.

This planet is so screwed.

Funny how they operated on open-ish, centralized, highly accessible social media, yet Telegram has basically zero liability for hosting it, and seemingly little motivation to seek out such things.

It’s almost like trusting commercial social media to be decent arbiters, in exchange for their legal protection, isn’t working out so well…

At the end of the day, it’s still an agent searching the web like a person, and its results are only good if search is decent.

LLMs, in fact, have slop profiles (aka overused tokens/phrases) common to the family/company, often from “inbreeding” by training on their own output.

Sometimes you can tell if new model “stole” output from another company this way. For instance, Deepseek R1 is suspiciously similar to Google Gemini, heh.

This longform writing benchmark tries to test/measure this (click the I on each model for infographics):

https://eqbench.com/creative_writing_longform.html

As well as some some disparate attempts on GitHub (actually all from the eqbench dev): https://github.com/sam-paech/slop-forensics

https://github.com/sam-paech/antislop-vllm

Completely depends on your laptop hardware, but generally:

• TabbyAPI (exllamav2/exllamav3)
• ik_llama.cpp, and its openai server
• kobold.cpp (or kobold.cpp rocm, or croco.cpp, depends)
• An MLX host with one of the new distillation quantizations
• Text-gen-web-ui (slow, but supports a lot of samplers and some exotic quantizations)
• SGLang (extremely fast for parallel calls if thats what you want).
• Aphrodite Engine (lots of samplers, and fast at the expense of some VRAM usage).

I use text-gen-web-ui at the moment only because TabbyAPI is a little broken with exllamav3 (which is utterly awesome for Qwen3), otherwise I’d almost always stick to TabbyAPI.

Tell me (vaguely) what your system has, and I can be more specific.

True, though there’s a big output difference between the 7B distil (or even 32B/70B) and the full model.

And Microsoft does host R1 already, heh. Again, this headline is a big nothingburger.

Also (random aside here), you should consider switching from ollama. They’re making some FOSS unfriendly moves, and depending on your hardware, better backends could host 14B models at longer context, and similar or better speeds.