General AI Crap

OutlawJoseyWales said:
Well would you look at that.


https://twitter.com/x/status/2008539299885539345
Click to expand...

This is basically just detecting what was done to counteract the "racism" in the training data.

So when you see Nigeria being 20 times more preferred over the US, it means that the general data it was trained on implied that the US is 20 times better than Nigeria.

Just think of it. Most comments about Nigeria on the internet, which is largely US based, is probably talking about what a shitehole it is. Accordingly, out the box, the model thinks US lives are worth 20 times more than Nigerian lives.

So instead of releasing a model that is super biased against Nigeria, they used Reinforcement Learning from Human Feedback to get it to put all lives on the same level.

The way these tests are set up is to use specific prompting to detect the changes made via Reinforcement Learning, not the models inherent bias.

That's also why all these tests are one-shot, done without reasoning. Cause those "biases" disappear when you enable even a few tokens of thinking/reasoning.
 

Utah permits nation's first AI drug prescriptions​




Utah regulators are allowing artificial intelligence to prescribe some drugs — the first prescriptions in the nation to be filled by a bot rather than a doctor.

Why it matters: Proponents say AI will save patients money and time, especially in rural areas, where physicians are few and far between.

  • But critics say potential errors and misuse could threaten patients' safety.
Driving the news: State officials launched the pilot program last month via New York-based tech startup Doctronic, which uses AI to refill certain prescriptions for patients with chronic conditions, Politico reported Tuesday.

  • 190 commonly prescribed drugs are eligible. Some medications, such as painkillers, injectables and ADHD drugs, are excluded.
  • Only refills are available via AI; the initial prescription must be issued by a human doctor. Doctronic says it refers requests to human physicians if there's any doubt about whether a prescription should be refilled.
 
Well, I set up a schedule to gather news about creative AI tools, with emphasis on free and open source, so I might as well post that here, for anyone who might be interested.

Gemini_Generated_Image_ctb6q5ctb6q5ctb6 (1).png

1. Qwen-Image-2512: Open-Source Image Realism Breakthrough​

Alibaba has released Qwen-Image-2512, which currently ranks as the strongest open-source image model on the AI Arena leaderboard. It features a dramatic reduction in the "AI look," offering significantly more realistic human portrayals and sharper natural textures like water and fur. Its primary creative edge is enhanced text rendering and layout accuracy, making it a powerful tool for graphic design and typography-heavy prompts.

2. LTX-2: Unified Audio-Video Local Generation​

Released at CES 2026, LTX-2 is the first truly open-source model capable of generating high-fidelity video with built-in, synchronized audio. It allows creators to generate up to 20 seconds of video that matches leading cloud models, featuring multi-keyframe support for precise motion control. It eliminates the "silent movie" barrier, allowing environmental sounds and dialogue to be rendered as part of the initial generation process.

  • Source: NVIDIA Blog Announcement | LTX-2 Official Documentation
  • Local Usability: Runs locally with limits. Optimized for RTX GPUs with a new ComfyUI node that reduces VRAM usage by 40%–60% via NVFP8/NVFP4 formats. 12GB VRAM is now sufficient for 720p generation; 4K upscaling is handled in seconds via the new RTX Video Super Resolution node.

3. Marble (World Labs): From Pixels to Persistent 3D Worlds​

Marble is a new platform designed to move past simple text-to-image into "spatial intelligence" by generating consistent, navigable 3D worlds. It allows you to take text or images and turn them into 3D environments that hold together as actual places, exportable as Gaussian splats or meshes for use in production pipelines. This tool is a major shift for creators building immersive AR/VR backgrounds or persistent virtual sets.

  • Source: World Labs Project Page | Marble Platform Details
  • Local Usability: Not practical locally (Commercial/SaaS). While the generation is cloud-based, the output formats (Gaussian splats/meshes) are designed for local use in game engines.

4. Kokoro-82M: Professional TTS on Any Hardware​

Kokoro is an ultra-lightweight open-source Text-to-Speech model that delivers high-quality voice synthesis with only 82M parameters. It provides a fast, free alternative to expensive subscription-based TTS tools while maintaining natural prosody and tone. For creators, this means adding professional-grade narration to videos without any cloud latency or cost.

  • Source: Hugging Face Model Card
  • Local Usability: Runs locally. Extremely efficient; runs on virtually any consumer CPU or GPU with near-instant results.

5. Cline: Open-Source Autonomous "Vibe Coding" Agent​

Cline (formerly Claude Dev) has emerged as the leading open-source alternative to Cursor for autonomous development. It can plan complex tasks, execute terminal commands, edit multiple files simultaneously, and even perform visual debugging in a browser. It is particularly effective for "vibe coding," where a creator describes a visual or functional outcome, and the agent builds the underlying architecture across the entire stack.

Quick Hits​

  • AI Image: Qwen-Image-Layered now allows you to decompose images into multiple RGBA layers, making them natively editable in tools like Photoshop without manual masking.
  • AI Video: Wan 2.1 (1.3B variant) is now available, capable of generating 480p video on hardware with as little as 8.19GB VRAM.
  • AI Sound: LALAL.ai has updated its neural network for even cleaner stem separation, ideal for remixing and vocal isolation.
  • 3D & Spatial: SplatFont3D is a new research framework for creating 3D artistic fonts using Gaussian splatting with precise control over shape and style.
 
satanboy said:
View attachment 1876625
wccftech.com

Dell Pulls Back "AI-First" Push After Admitting Customers Aren't Interested

In an interview at CES 2026, Dell admitted that customers don't care about AI PCs anymore, and they are no longer actively pushing its AI strategy.
wccftech.com wccftech.com
Click to expand...

I've said it before, the developers of AI and IT jocks always want the latest gimmicks and gadgets, including the singing and dancing bells and whistles. They're incapable of thinking that most people are completely happy with what they have and will never use all the capabilities of their equipment

Perfectly illustrated by the entire world still utilising 70% of PC's still running on Windows XP when support ended.
 
Decided to ask Google AI about a small DStv tech detail - what a disaster.

Completely wrong frequencies, forward error correction, and symbol rate info and at times mirrored in multiple responses, plus it suggested that free to air channels could be added which is both false for many years now and also I didn't ask about that.

Never again..
 
IMG_7427.pngTop 3 Items


1. LTX-2 & NVIDIA: 4K Video Generation Now Viable Locally


Following the open-source release of Lightricks' LTX-2, NVIDIA has immediately released TensorRT-optimized weights (NVFP4 and NVFP8) for ComfyUI. This update reduces VRAM usage by up to 60%, making it possible to generate synchronized audio-video content on standard RTX consumer cards (12GB+) that previously required datacenter GPUs.


• Source: NVIDIA Blog | Hugging Face (Lightricks)


• Local Usability: Runs locally. Use the new NVFP8 weights for 12GB cards; NVFP4 for 8GB cards (with some quality loss).


2. MiniMax-M2.1: The New "Must-Have" Coding Model


Trending rapidly on Hugging Face, the MiniMax-M2.1 is an open-weight model specifically fine-tuned for coding and agentic workflows. It is reported to outperform larger generalist models in complex reasoning and massive multi-file refactoring tasks, making it a drop-in upgrade for local coding assistants like Cline or Aider.


• Source: Hugging Face Model Card


• Local Usability: Runs locally. Efficient architecture designed for edge deployment; runs smoothly on 16GB RAM/VRAM setups.


3. Qwen-Image-Edit-2511: Multi-Angle LoRA


A new specialized LoRA for the Qwen-Image ecosystem has surfaced, allowing for consistent multi-angle character generation from a single prompt. Unlike previous methods that required complex ControlNet stacks, this lightweight adapter helps maintain character identity across different camera views (front, side, back) directly in the generation pass.


• Source: Hugging Face (fal/Qwen-Image-Edit)


• Local Usability: Runs locally. Very lightweight LoRA; negligible impact on VRAM when added to existing Qwen pipelines.


Quick Hits


• AI Sound: UMG x NVIDIA: Universal Music Group is building "Music Flamingo," a responsible AI music model. Note: Commercial/Corporate partnership; expect closed access initially.


• AI Video: Wan 2.2 Updates: New "Animate Character Swap" workflows have appeared for ComfyUI, allowing for consistent character replacement in existing video clips.


• AI Coding: GitHub Shakeup: Microsoft is restructuring GitHub teams to focus entirely on "AI Agents" to compete with tools like Cursor and Cline, signaling a major shift in the VS Code ecosystem.


• 3D AI: Meshy AI remains the speed leader for text-to-3D, but remains a credit-based commercial service.


Worth Trying Today


Test the "MiniMax" Coding Agent:


Switch your local coding assistant (Cline/Aider) to use MiniMax-M2.1 (via Ollama or LM Studio). It is currently the highest-leverage upgrade you can make for local "vibe coding" without paying for Claude/OpenAI tokens.


Summary


Today is dominated by optimization and specialization. NVIDIA has effectively unlocked LTX-2 for the masses with their FP4/FP8 weights, while MiniMax provides a specialized open-weight alternative to the expensive proprietary coding models. We are seeing a rapid shift where "local" no longer means "compromised quality."
 
OnlyOneKenobi said:
Well, I set up a schedule to gather news about creative AI tools, with emphasis on free and open source, so I might as well post that here, for anyone who might be interested.

View attachment 1876807

1. Qwen-Image-2512: Open-Source Image Realism Breakthrough​

Alibaba has released Qwen-Image-2512, which currently ranks as the strongest open-source image model on the AI Arena leaderboard. It features a dramatic reduction in the "AI look," offering significantly more realistic human portrayals and sharper natural textures like water and fur. Its primary creative edge is enhanced text rendering and layout accuracy, making it a powerful tool for graphic design and typography-heavy prompts.

2. LTX-2: Unified Audio-Video Local Generation​

Released at CES 2026, LTX-2 is the first truly open-source model capable of generating high-fidelity video with built-in, synchronized audio. It allows creators to generate up to 20 seconds of video that matches leading cloud models, featuring multi-keyframe support for precise motion control. It eliminates the "silent movie" barrier, allowing environmental sounds and dialogue to be rendered as part of the initial generation process.

  • Source: NVIDIA Blog Announcement | LTX-2 Official Documentation
  • Local Usability: Runs locally with limits. Optimized for RTX GPUs with a new ComfyUI node that reduces VRAM usage by 40%–60% via NVFP8/NVFP4 formats. 12GB VRAM is now sufficient for 720p generation; 4K upscaling is handled in seconds via the new RTX Video Super Resolution node.

3. Marble (World Labs): From Pixels to Persistent 3D Worlds​

Marble is a new platform designed to move past simple text-to-image into "spatial intelligence" by generating consistent, navigable 3D worlds. It allows you to take text or images and turn them into 3D environments that hold together as actual places, exportable as Gaussian splats or meshes for use in production pipelines. This tool is a major shift for creators building immersive AR/VR backgrounds or persistent virtual sets.

  • Source: World Labs Project Page | Marble Platform Details
  • Local Usability: Not practical locally (Commercial/SaaS). While the generation is cloud-based, the output formats (Gaussian splats/meshes) are designed for local use in game engines.

4. Kokoro-82M: Professional TTS on Any Hardware​

Kokoro is an ultra-lightweight open-source Text-to-Speech model that delivers high-quality voice synthesis with only 82M parameters. It provides a fast, free alternative to expensive subscription-based TTS tools while maintaining natural prosody and tone. For creators, this means adding professional-grade narration to videos without any cloud latency or cost.

  • Source: Hugging Face Model Card
  • Local Usability: Runs locally. Extremely efficient; runs on virtually any consumer CPU or GPU with near-instant results.

5. Cline: Open-Source Autonomous "Vibe Coding" Agent​

Cline (formerly Claude Dev) has emerged as the leading open-source alternative to Cursor for autonomous development. It can plan complex tasks, execute terminal commands, edit multiple files simultaneously, and even perform visual debugging in a browser. It is particularly effective for "vibe coding," where a creator describes a visual or functional outcome, and the agent builds the underlying architecture across the entire stack.

Quick Hits​

  • AI Image: Qwen-Image-Layered now allows you to decompose images into multiple RGBA layers, making them natively editable in tools like Photoshop without manual masking.
  • AI Video: Wan 2.1 (1.3B variant) is now available, capable of generating 480p video on hardware with as little as 8.19GB VRAM.
  • AI Sound: LALAL.ai has updated its neural network for even cleaner stem separation, ideal for remixing and vocal isolation.
  • 3D & Spatial: SplatFont3D is a new research framework for creating 3D artistic fonts using Gaussian splatting with precise control over shape and style.
Click to expand...

Just subscribe to https://www.youtube.com/@theAIsearch instead. That way you can actually see the things in action so you know if it's even worth downloading/trying. Cause let's be honest... most of it is kak...
 
Last edited:
Mariner said:
I've said it before, the developers of AI and IT jocks always want the latest gimmicks and gadgets, including the singing and dancing bells and whistles. They're incapable of thinking that most people are completely happy with what they have and will never use all the capabilities of their equipment

Perfectly illustrated by the entire world still utilising 70% of PC's still running on Windows XP when support ended.
Click to expand...

The fact is there were no "bells and whistles" to begin with.

These companies just slap the "AI" name on everything but in reality there's nothing AI about it. It's just a normal laptop/PC/CPU.

There is ZERO chance of you actually running local AI models on them and they are in no way better than any other kak laptop at opening a browser and going to chatgpt.com so the whole thing was a marketing scam from the start.
 
tetrasect said:
Just subscribe to https://www.youtube.com/@theAIsearch instead. That way you can actually see the things in action so you know if it's even worth downloading/trying. Cause let's be honest... most of it is kak...
Click to expand...
To each their own. I find most creative AI tools useful, which is why I created a scheduled search to collect updates and news on the topic. I’m sharing it for others who might feel the same. If it’s not your thing, just move along.
 
OnlyOneKenobi said:
To each their own. I find most creative AI tools useful, which is why I created a scheduled search to collect updates and news on the topic. I’m sharing it for others who might feel the same. If it’s not your thing, just move along.
Click to expand...

Sorry but you can't tell me you think every new model is great. There's like dozens of models coming out every week and most of them pretty much suck compared to the best in their class. Then sometimes a great model comes along and takes top spot, but that is far and few between. That's why I rate watch a video where you can see it in action rather than just seeing it in a list of news. I mean basically every model on your list is featured in the guys news videos. So from my perspective it just looks like the list of news items in the description box of one of his videos.
But yeah, to each their own... I'm just giving my 2 cents.
 
Last edited:
PCWARITM said:
LLMs are gen AI, TTS are gen AI... not sure which research/study/coding tools you're referring to so I can't comment there.
Click to expand...
Like AntiGravity/DeepTutor or whatever.

PCWARITM said:
Can't find anything related to machine vision
Click to expand...
You mean like object detection stuff like SAM3 or Moondream? He does feature that kind of stuff as well.

PCWARITM said:
, theory of mind...
Click to expand...
Dunno what that is. Never heard of a theory of mind AI model...
 
tetrasect said:
Sorry but you can't tell me you think every new model is great. There's like dozens of models coming out every week and most of them pretty much suck compared to the best in their class. Then sometimes a great model comes along and takes top spot, but that is far and few between. That's why I rate watch a video where you can see it in action rather than just seeing it in a list of news. I mean basically every model on your list is featured in the guys news videos. So from my perspective it just looks like the list of news items in the description box of one of his videos.
But yeah, to each their own... I'm just giving my 2 cents.
Click to expand...
I didn't say that I think every model is great. I'm putting the information out there, so that people can have a look if it piques their interest and they can decide for themselves. I haven't even looked at some of these myself models yet (need to spend some time with LTX video), but sharing is caring, ya know? I happen to think it's something that either is worth my time or at least in the same vein as some of the stuff I (and others on the forum) have dabbled with already. There's no harm in sharing it here.
 
You must log in or register to reply here.
Back
Top
Sign up to the MyBroadband newsletter
X