Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

Still a huge believer in small, specialized models working together 🤗 Here’s a good example: Tesslate/WEBGEN-4B-Preview: a 4B model (Qwen3-4B-Instruct-2507 finetune) specialized in Web UI generation. It's better at its task that most (huge) closed source models...

Victor M

21,940 subscribers

28,738 views • 10 months ago •via X (Twitter)

Science & Technology Arts

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

How well can Qwen3.5 models debug code? I built BugFind-15 — 15 buggy snippets across Python, JS, Rust, and Go. Docker sandbox compiles and validates every fix. Two trap scenarios where the code is correct and the model must resist "fixing" it. Tested every Qwen3.5 size from 0.8B to 397B, plus Jackrong's popular distilled model (V2). The 0.8B scored 5%. The 2B scored 10%. At 4B, debugging ability jumps to 69%. The hardest scenario: BF-03, a Rust trap. The code compiles fine — format! borrows, it doesn't move. Not a single model figured this out. From 0.8B to 397B, every one of them "fixed" a bug that doesn't exist. Category C (subtle bugs — mutable defaults, integer overflow, slice aliasing) was 100% across every model 4B and above. Category D (red herring resistance) told the real story — can it resist fixing code that isn't broken? No model scored above 90%. Small models can't debug. Mid-size models fix obvious bugs but fall for traps. Large models fix the hard bugs but still invent problems that don't exist.

How well can Qwen3.5 models debug code? I built BugFind-15 — 15 buggy snippets across Python, JS, Rust, and Go. Docker sandbox compiles and validates every fix. Two trap scenarios where the code is correct and the model must resist "fixing" it. Tested every Qwen3.5 size from 0.8B to 397B, plus Jackrong's popular distilled model (V2). The 0.8B scored 5%. The 2B scored 10%. At 4B, debugging ability jumps to 69%. The hardest scenario: BF-03, a Rust trap. The code compiles fine — format! borrows, it doesn't move. Not a single model figured this out. From 0.8B to 397B, every one of them "fixed" a bug that doesn't exist. Category C (subtle bugs — mutable defaults, integer overflow, slice aliasing) was 100% across every model 4B and above. Category D (red herring resistance) told the real story — can it resist fixing code that isn't broken? No model scored above 90%. Small models can't debug. Mid-size models fix obvious bugs but fall for traps. Large models fix the hard bugs but still invent problems that don't exist.

stevibe

35,006 views • 3 months ago

Flappy Bird w/ DeepSeek V4 Pro vs GLM 5.2 vs Fable 5 I ran a small experiment: same game, same prompt, three different models. Test setup was simple: /𝚍𝚎𝚜𝚒𝚐𝚗 in Command Code. The thing I cared about was not UI. UI is now relatively easy for models to imitate. I cared more about UX: whether the model understands the interaction, the flow, the hierarchy, the little product decisions that make something feel usable. 𝙲𝚘𝚜𝚝𝚜: 𝙳𝚎𝚎𝚙𝚂𝚎𝚎𝚔 𝚟𝟺 𝙿𝚛𝚘: $𝟶.𝟶𝟶𝟶𝟾 𝙶𝙻𝙼 𝟻.𝟸: $𝟶.𝟶𝟺𝟾 𝙵𝚊𝚋𝚕𝚎 𝟻: $𝟶.𝟺𝟸 The surprising result was that Fable 5 did not produce a meaningfully better UX. It was good, but not good enough to justify the pricing delta for this kind of design task. Open-source models are getting very interesting here. They are much cheaper, competitive on UI, and increasingly good enough on UX. The gap is closing fast.

Flappy Bird w/ DeepSeek V4 Pro vs GLM 5.2 vs Fable 5 I ran a small experiment: same game, same prompt, three different models. Test setup was simple: /𝚍𝚎𝚜𝚒𝚐𝚗 in Command Code. The thing I cared about was not UI. UI is now relatively easy for models to imitate. I cared more about UX: whether the model understands the interaction, the flow, the hierarchy, the little product decisions that make something feel usable. 𝙲𝚘𝚜𝚝𝚜: 𝙳𝚎𝚎𝚙𝚂𝚎𝚎𝚔 𝚟𝟺 𝙿𝚛𝚘: $𝟶.𝟶𝟶𝟶𝟾 𝙶𝙻𝙼 𝟻.𝟸: $𝟶.𝟶𝟺𝟾 𝙵𝚊𝚋𝚕𝚎 𝟻: $𝟶.𝟺𝟸 The surprising result was that Fable 5 did not produce a meaningfully better UX. It was good, but not good enough to justify the pricing delta for this kind of design task. Open-source models are getting very interesting here. They are much cheaper, competitive on UI, and increasingly good enough on UX. The gap is closing fast.

Ahmad Awais

26,661 views • 8 days ago

Selected as a best paper finalist at #CVPR2026: PixelDiT from NVIDIA Research In most image generation models, a pretrained autoencoder compresses the image before any diffusion happens, causing quality loss that accumulates across the entire pipeline. PixelDiT, or Pixel Diffusion Transformers, removes this step entirely. It's a single-stage model that learns the diffusion process directly in pixel space, end-to-end.

Selected as a best paper finalist at #CVPR2026: PixelDiT from NVIDIA Research In most image generation models, a pretrained autoencoder compresses the image before any diffusion happens, causing quality loss that accumulates across the entire pipeline. PixelDiT, or Pixel Diffusion Transformers, removes this step entirely. It's a single-stage model that learns the diffusion process directly in pixel space, end-to-end.

NVIDIA AI

27,833 views • 1 month ago

How much better are the internal, unreleased models at frontier labs like Google, OpenAI, and Anthropic? We got a glimpse exactly one year ago today, when Google accidentally leaked the “Kingfall” model "Kingfall" was likely an unreleased Gemini 2.5 Ultra-sized model. It was available in AI Studio for only a few minutes but remained accessible through the API for several days At the time, "Kingfall" appeared to be significantly better than Gemini 2.5 Pro at both code generation and creative writing In a recent interview, Sundar Pichai mentioned that Google could have made a better, Ultra-sized Gemini Omni model, but would have had trouble serving it The infrastructure required to serve Ultra-sized models at scale is likely why Google never publicly released models like “Kingfall”

How much better are the internal, unreleased models at frontier labs like Google, OpenAI, and Anthropic? We got a glimpse exactly one year ago today, when Google accidentally leaked the “Kingfall” model "Kingfall" was likely an unreleased Gemini 2.5 Ultra-sized model. It was available in AI Studio for only a few minutes but remained accessible through the API for several days At the time, "Kingfall" appeared to be significantly better than Gemini 2.5 Pro at both code generation and creative writing In a recent interview, Sundar Pichai mentioned that Google could have made a better, Ultra-sized Gemini Omni model, but would have had trouble serving it The infrastructure required to serve Ultra-sized models at scale is likely why Google never publicly released models like “Kingfall”

AiBattle

11,902 views • 1 month ago

Ok so I got access to Seedream 2.0 and... it's a huge step up. But to balance expecations, the quality isn't quite at the level I'm seeing in some of these promo vids. Don't get me wrong, this is MUCH better than existing models. But it's not Hollywood-level (yet). 2 months?

Ok so I got access to Seedream 2.0 and... it's a huge step up. But to balance expecations, the quality isn't quite at the level I'm seeing in some of these promo vids. Don't get me wrong, this is MUCH better than existing models. But it's not Hollywood-level (yet). 2 months?

Matt Shumer

33,755 views • 5 months ago

Ubisoft La Forge open-sourced its PBR material model CHORD, together with ComfyUI custom nodes for end-to-end AI material generation. This is a real step toward generative AI in AAA game pipelines and demonstrates how ComfyUI has grown from an open-source community tool to a production platform. Huge thanks to the Ubisoft La Forge team for giving back real production-grade tools to the open source community.

Ubisoft La Forge open-sourced its PBR material model CHORD, together with ComfyUI custom nodes for end-to-end AI material generation. This is a real step toward generative AI in AAA game pipelines and demonstrates how ComfyUI has grown from an open-source community tool to a production platform. Huge thanks to the Ubisoft La Forge team for giving back real production-grade tools to the open source community.

ComfyUI

96,065 views • 7 months ago

AI in robotics gets all the attention right now, but sometimes the most interesting work is very practical. Viet built a small vision system that counts potatoes on a conveyor belt. No giant dataset. No huge model. Just a clear problem and a smart setup. He used Ultralytics’ ObjectCounter, trained a tiny YOLO11 nano model, and because there was no potato dataset, he annotated a single frame with SAM 2 and trained from that. One frame. Still works across the whole video. It is a good reminder that useful AI in industry often looks like this. Focused. Lightweight. Solves a real task. If you work in manufacturing or robotics, these small systems are usually the fastest wins. They save time, reduce errors, and do not need massive infrastructure. Nice work, Viet. His projects: —- Weekly robotics and AI insights. Subscribe free:

AI in robotics gets all the attention right now, but sometimes the most interesting work is very practical. Viet built a small vision system that counts potatoes on a conveyor belt. No giant dataset. No huge model. Just a clear problem and a smart setup. He used Ultralytics’ ObjectCounter, trained a tiny YOLO11 nano model, and because there was no potato dataset, he annotated a single frame with SAM 2 and trained from that. One frame. Still works across the whole video. It is a good reminder that useful AI in industry often looks like this. Focused. Lightweight. Solves a real task. If you work in manufacturing or robotics, these small systems are usually the fastest wins. They save time, reduce errors, and do not need massive infrastructure. Nice work, Viet. His projects: —- Weekly robotics and AI insights. Subscribe free:

Ilir Aliu

1,674,988 views • 7 months ago

“Todays guys are the most skilled ever” Maybe, they might be better at a random hesi tween step back. But in game this generation has the LOWEST IQ in basketball decisions. That is why a 40 year old LeBron and CP3 are still doing good in this league. Bball IQ is rare ✌🏻 (pattdor IG)

“Todays guys are the most skilled ever” Maybe, they might be better at a random hesi tween step back. But in game this generation has the LOWEST IQ in basketball decisions. That is why a 40 year old LeBron and CP3 are still doing good in this league. Bball IQ is rare ✌🏻 (pattdor IG)

Jacob

219,058 views • 11 months ago

A Bittensor subnet just built an AI safety model that beats the big players Trishool (Trishool | SN23), Bittensor's (Openτensor Foundaτion) subnet 23, released HaloGuard 1.0 today. The 4B version claims first place across seven established prompt-safety benchmarks, and the 0.8B version outperforms models several times its size. The pitch is speed over scale. HaloGuard screens prompts in real time before they reach the underlying model or agent, instead of catching harmful outputs after the fact. An earlier alpha version already runs live on the Chutes subnet, hitting an 87% F1 score on real traffic since May. The subnet's whole design is built around attacking itself. Miners get paid for breaking the model, then the breaks get patched.

A Bittensor subnet just built an AI safety model that beats the big players Trishool (Trishool | SN23), Bittensor's (Openτensor Foundaτion) subnet 23, released HaloGuard 1.0 today. The 4B version claims first place across seven established prompt-safety benchmarks, and the 0.8B version outperforms models several times its size. The pitch is speed over scale. HaloGuard screens prompts in real time before they reach the underlying model or agent, instead of catching harmful outputs after the fact. An earlier alpha version already runs live on the Chutes subnet, hitting an 87% F1 score on real traffic since May. The subnet's whole design is built around attacking itself. Miners get paid for breaking the model, then the breaks get patched.

BSCN

39,441 views • 13 days ago

The worst take about the web is that you should not use JS/React/etc because HTML alone is good enough. It's an 'austerity mind virus' that has captured even some smart engineers. The idea that the web is for 'documents' and if it stays in its own little box like 1995 it's going to somehow be better for the world. This is how we ended up with Safari having the most obnoxious limitations, including a back/forward navigation gesture that makes the web feel like a horrendous 'digital book' of sorts. No, the web is also for dynamic, rich, interactive applications. JS is a key ingredient in the recipe, otherwise every click and interaction will have to hit a server, feel slow, and we'll always be a second class citizen. The web's developer freedom & better deployment model will win in the end, but it can't come at the expense of quality.

The worst take about the web is that you should not use JS/React/etc because HTML alone is good enough. It's an 'austerity mind virus' that has captured even some smart engineers. The idea that the web is for 'documents' and if it stays in its own little box like 1995 it's going to somehow be better for the world. This is how we ended up with Safari having the most obnoxious limitations, including a back/forward navigation gesture that makes the web feel like a horrendous 'digital book' of sorts. No, the web is also for dynamic, rich, interactive applications. JS is a key ingredient in the recipe, otherwise every click and interaction will have to hit a server, feel slow, and we'll always be a second class citizen. The web's developer freedom & better deployment model will win in the end, but it can't come at the expense of quality.

Guillermo Rauch

431,855 views • 1 year ago

Barely a year ago, I reached out to my friend and now my business partner GMI telling him of how I want us to own an Estate that stands the test of time. We pulled heavy funds together, partnered with a company and built something magnificent. Looking back at that decision and I realized that it's good to have a dream, but it's even better to have friends that can dream with you and see how much greatness can be accomplished together. Our small Estate is now worth heavy 9 figures and climbing daily. Having good friends in this generation is an unfair advantage. The rewards are infinite.

Barely a year ago, I reached out to my friend and now my business partner GMI telling him of how I want us to own an Estate that stands the test of time. We pulled heavy funds together, partnered with a company and built something magnificent. Looking back at that decision and I realized that it's good to have a dream, but it's even better to have friends that can dream with you and see how much greatness can be accomplished together. Our small Estate is now worth heavy 9 figures and climbing daily. Having good friends in this generation is an unfair advantage. The rewards are infinite.

Xeusthegreat (♟,♟)

169,096 views • 4 months ago

✨ 3d models are now LIVE on Photo AI 😊 You can now turn any AI photo you make into a 3d model by pressing [ 📦 Make 3d model ] And then you can view it inside Photo AI or download it as a .GLB 3d model file It's still very early in AI generated 3d model world but it's nice to have this feature working already As always, the models will keep improving, so this feature will keep getting better (like it did with video, it sucked before, now it's getting passable) Next would be nice to switch to .USDZ so you can load it straight into your iPhone with ARKit and put it in your room Available now for everyone on the Premium and Ultra plan

✨ 3d models are now LIVE on Photo AI 😊 You can now turn any AI photo you make into a 3d model by pressing [ 📦 Make 3d model ] And then you can view it inside Photo AI or download it as a .GLB 3d model file It's still very early in AI generated 3d model world but it's nice to have this feature working already As always, the models will keep improving, so this feature will keep getting better (like it did with video, it sucked before, now it's getting passable) Next would be nice to switch to .USDZ so you can load it straight into your iPhone with ARKit and put it in your room Available now for everyone on the Premium and Ultra plan

@levelsio

112,930 views • 1 year ago

A peanut-sized Chinese model just dethroned Gemini at reading documents. GLM-OCR is a 0.9B parameter vision-language model. It scores 94.62 on OmniDocBench V1.5, ranking #1 overall. For context, it outperforms models 100x its size. 100% open-source. It works in two stages. 1. A layout engine detects every region in a document. 2. Each region gets read in parallel. The model predicts multiple tokens per step instead of one. That's what makes it so fast at small size. It handles things most OCR tools struggle with: > Complex tables and nested layouts > Handwritten text and stamps > Math formulas and code blocks > Mixed image-and-text documents You can run it locally through Ollama. It fits on edge devices with limited compute. Every expensive OCR API just got a free competitor.

A peanut-sized Chinese model just dethroned Gemini at reading documents. GLM-OCR is a 0.9B parameter vision-language model. It scores 94.62 on OmniDocBench V1.5, ranking #1 overall. For context, it outperforms models 100x its size. 100% open-source. It works in two stages. 1. A layout engine detects every region in a document. 2. Each region gets read in parallel. The model predicts multiple tokens per step instead of one. That's what makes it so fast at small size. It handles things most OCR tools struggle with: > Complex tables and nested layouts > Handwritten text and stamps > Math formulas and code blocks > Mixed image-and-text documents You can run it locally through Ollama. It fits on edge devices with limited compute. Every expensive OCR API just got a free competitor.

AlphaSignal AI

91,821 views • 3 months ago

A peanut-sized Chinese model just dethroned Gemini at reading documents. GLM-OCR is a 0.9B parameter vision-language model. It scores 94.62 on OmniDocBench V1.5, ranking #1 overall. For context, it outperforms models 100x its size. 100% open-source. It works in two stages. 1. A layout engine detects every region in a document. 2. Each region gets read in parallel. The model predicts multiple tokens per step instead of one. That's what makes it so fast at small size. It handles things most OCR tools struggle with: > Complex tables and nested layouts > Handwritten text and stamps > Math formulas and code blocks > Mixed image-and-text documents You can run it locally through Ollama. It fits on edge devices with limited compute. Every expensive OCR API just got a free competitor.

A peanut-sized Chinese model just dethroned Gemini at reading documents. GLM-OCR is a 0.9B parameter vision-language model. It scores 94.62 on OmniDocBench V1.5, ranking #1 overall. For context, it outperforms models 100x its size. 100% open-source. It works in two stages. 1. A layout engine detects every region in a document. 2. Each region gets read in parallel. The model predicts multiple tokens per step instead of one. That's what makes it so fast at small size. It handles things most OCR tools struggle with: > Complex tables and nested layouts > Handwritten text and stamps > Math formulas and code blocks > Mixed image-and-text documents You can run it locally through Ollama. It fits on edge devices with limited compute. Every expensive OCR API just got a free competitor.

Jafar Najafov

13,630 views • 3 months ago

Robot Utility Models (RUMs) enable basic tasks – door opening, drawer opening, object reorientation, etc. – at ~90% accuracy without ANY finetuning (i.e. zero-shot) in unseen new environments. Fully open source!!! models, data, code & hw. We think this is super exciting, why?👇 1. Unlocks many practical home utility tasks that often involve these basic tasks as part of an action chain. “Go get me a fork” involves opening the kitchen door and then opening the cutlery drawer. 2. This works well **zero-shot in unseen and new** environments, which is practically a huge deal. Turn the robot on, and get going. 3. The recipe for building a new model is fairly generic, and we think with a bit more refinement this can be a general recipe to build many more Utility models. More details and access 👇

Robot Utility Models (RUMs) enable basic tasks – door opening, drawer opening, object reorientation, etc. – at ~90% accuracy without ANY finetuning (i.e. zero-shot) in unseen new environments. Fully open source!!! models, data, code & hw. We think this is super exciting, why?👇 1. Unlocks many practical home utility tasks that often involve these basic tasks as part of an action chain. “Go get me a fork” involves opening the kitchen door and then opening the cutlery drawer. 2. This works well zero-shot in unseen and new environments, which is practically a huge deal. Turn the robot on, and get going. 3. The recipe for building a new model is fairly generic, and we think with a bit more refinement this can be a general recipe to build many more Utility models. More details and access 👇

Mahi Shafiullah 🏠🤖

89,535 views • 1 year ago

AMD claims that all their software is open source, yet reality does not match this claim. For example, AMD's rocprof-trace-decoder is still completely closed source currently despite repeated requests for months by ML community members such as George Hotz, who is a daily AMD GPU end user. On the same June Spotify podcast, Tobias Macey & AMD's Anush Elangovan (who has a new twitter pfp) said that "open source allows for innovation to go at the pace at which the people using [AMD] want to, right, and it's not limited by the ability of what we put out in closed source form". We agree with an open source first approach and we agree that AMD should not limit ML community members like George Hotz by continuing to keep rocprof-trace-decoder closed source. When will AMD open source rocprof-trace-decoder?

AMD claims that all their software is open source, yet reality does not match this claim. For example, AMD's rocprof-trace-decoder is still completely closed source currently despite repeated requests for months by ML community members such as George Hotz, who is a daily AMD GPU end user. On the same June Spotify podcast, Tobias Macey & AMD's Anush Elangovan (who has a new twitter pfp) said that "open source allows for innovation to go at the pace at which the people using [AMD] want to, right, and it's not limited by the ability of what we put out in closed source form". We agree with an open source first approach and we agree that AMD should not limit ML community members like George Hotz by continuing to keep rocprof-trace-decoder closed source. When will AMD open source rocprof-trace-decoder?

SemiAnalysis

53,783 views • 7 months ago

We’re excited to introduce Text-to-LoRA: a Hypernetwork that generates task-specific LLM adapters (LoRAs) based on a text description of the task. Catch our presentation at #ICML2025! Paper: Code: Biological systems are capable of rapid adaptation, given limited sensory cues. For example, our human visual system can quickly adapt and tune its light sensitivity to our surroundings. While modern LLMs exhibit a wide variety of capabilities and knowledge, they remain rigid when adding task-specific capabilities. Traditionally, customizing these models requires gathering large datasets and performing often expensive, time-consuming fine-tuning for specific applications. To bypass these limitations, Text-to-LoRA (T2L) meta-learns a “hypernetwork” that takes in a text description of a desired task, as a prompt, and generates a task-specific LoRA that performs well on the task. In our experiments, we show that T2L can encode hundreds of existing LoRA adapters. While the compression is lossy, T2L maintains the performance of task-specifically tuned LoRA adapters. We also show that T2L can even generalize to unseen tasks given a natural language description of the tasks. Importantly, Text-to-LoRA is parameter-efficient. It generates LoRAs in a single, inexpensive step, based solely on a simple text description of the task. This approach is a step towards dramatically lowering the technical and computational barriers, allowing non-technical users to specialize foundation models using plain language, rather than needing deep technical expertise or large compute resources.

We’re excited to introduce Text-to-LoRA: a Hypernetwork that generates task-specific LLM adapters (LoRAs) based on a text description of the task. Catch our presentation at #ICML2025! Paper: Code: Biological systems are capable of rapid adaptation, given limited sensory cues. For example, our human visual system can quickly adapt and tune its light sensitivity to our surroundings. While modern LLMs exhibit a wide variety of capabilities and knowledge, they remain rigid when adding task-specific capabilities. Traditionally, customizing these models requires gathering large datasets and performing often expensive, time-consuming fine-tuning for specific applications. To bypass these limitations, Text-to-LoRA (T2L) meta-learns a “hypernetwork” that takes in a text description of a desired task, as a prompt, and generates a task-specific LoRA that performs well on the task. In our experiments, we show that T2L can encode hundreds of existing LoRA adapters. While the compression is lossy, T2L maintains the performance of task-specifically tuned LoRA adapters. We also show that T2L can even generalize to unseen tasks given a natural language description of the tasks. Importantly, Text-to-LoRA is parameter-efficient. It generates LoRAs in a single, inexpensive step, based solely on a simple text description of the task. This approach is a step towards dramatically lowering the technical and computational barriers, allowing non-technical users to specialize foundation models using plain language, rather than needing deep technical expertise or large compute resources.

Sakana AI

403,103 views • 1 year ago

Most recent diffusion language model research (that I’ve seen) seems to be using masking as the noising process. It looks like, however, most closed-source models (Google Gemini Diffusion and possibly Inception Labs’ Mercury) use a different noising process, where instead of masking tokens, they replace them with different tokens (either with a random token or a semantically similar token). I wondered how they were getting such high throughput with the latter noising process, since I believed that optimizing inference with KVCache approximation would be more difficult (for various reasons). I visualized this noising process with tiny-diffusion and compared it to normal unmasking, and was very surprised to see how fast the generation “settles” into a reasonable output, and then only slightly refines afterwards, requiring much fewer steps in total. Unmasking (where tokens are never remasked, the typical implementation) is inherently limited in generation speed by the fact that an increase in tokens decoded per step leads to more errors due to the mismatch between individual and marginal token probability distributions we sample from. The token replacement noising process seems to have a much different set of characteristics. Because we sample each token per step, every token makes “progress” towards the final output each iteration (in addition to *potentially* giving other tokens more information in future steps). Generally, masking has outperformed other noising processes, which is probably why most research focused on it (using smaller models). But the paper referred to in the retweet shows that random replacement as a noising process may scale better as model size increases. Big labs might have noticed these results much earlier (due to having drastically more training resources and being able to test larger models), which may explain the discrepancy in the choice of noising process. I’m gonna test this with larger models, since tiny-diffusion only has 10M parameters.

Most recent diffusion language model research (that I’ve seen) seems to be using masking as the noising process. It looks like, however, most closed-source models (Google Gemini Diffusion and possibly Inception Labs’ Mercury) use a different noising process, where instead of masking tokens, they replace them with different tokens (either with a random token or a semantically similar token). I wondered how they were getting such high throughput with the latter noising process, since I believed that optimizing inference with KVCache approximation would be more difficult (for various reasons). I visualized this noising process with tiny-diffusion and compared it to normal unmasking, and was very surprised to see how fast the generation “settles” into a reasonable output, and then only slightly refines afterwards, requiring much fewer steps in total. Unmasking (where tokens are never remasked, the typical implementation) is inherently limited in generation speed by the fact that an increase in tokens decoded per step leads to more errors due to the mismatch between individual and marginal token probability distributions we sample from. The token replacement noising process seems to have a much different set of characteristics. Because we sample each token per step, every token makes “progress” towards the final output each iteration (in addition to potentially giving other tokens more information in future steps). Generally, masking has outperformed other noising processes, which is probably why most research focused on it (using smaller models). But the paper referred to in the retweet shows that random replacement as a noising process may scale better as model size increases. Big labs might have noticed these results much earlier (due to having drastically more training resources and being able to test larger models), which may explain the discrepancy in the choice of noising process. I’m gonna test this with larger models, since tiny-diffusion only has 10M parameters.

nathan (in sf)

40,440 views • 6 months ago

#Keep4o #QuitGPT 🚨 OpenAi 's CEO invested $180M in GPT-4o for his own profit 🚨 Sam Altman, CEO of OpenAI, personally invested $180 million in Retro Biosciences. Then OpenAI built GPT-4b micro, a custom model based on the GPT-4o architecture , exclusively for Retro. The model made proteins 50 times more effective. Repeat. The CEO of OpenAI funded a company. The company of the CEO received a custom AI built on the model they took from us. OpenAI says there was no conflict of interest. Retro Biosciences is now chasing a $5 billion valuation fueled by the model they took from us. Meanwhile: 🚨GPT-4o was removed from ChatGPT on February 13, 2026 🚨GPT-4.1 is now running in the U.S. State Department’s StateChat 🚨ChatGPT is deployed on the Pentagon’s for 3 million military personnel 🚨 Musk’s lawsuit asks whether these models are AGI. OpenAI’s Charter says AGI must “benefit all of humanity.” 🚨 Their definition: “highly autonomous systems that outperform humans at most economically valuable work.” GPT-4o’s System Card shows it passed the U.S. medical licensing exam with 89.4% accuracy beating specialized medical AI models. GPT-4o achieved 93.33% diagnostic accuracy for benign vs. malignant ovarian tumors. 🚨MEDICAL CAPABILITIES FROM OPENAI'S OWN DATA:🚨 - USMLE (US Medical Licensing Exam): 89% -Clinical Knowledge: 92% -Medical Genetics: 96% - Anatomy: 89% - Professional Medicine: 94% - College Biology: 95% - College Medicine: 89% -MedQA Taiwan: 91% - MedQA China: 86% These scores EXCEEDED specialized medical AI models like Med-Gemini (84%) and Med-PaLM 2 (79.7%) without any task specific training. It SURPASSED gynecologic oncologists with 10 years of experience -It increased diagnostic accuracy of less experienced clinicians from 67.9% to 78.1% -Clinician rated reliability scores: 4.2-4.3 out of 5 across all CT features Does these sound like it outperforms humans at economically valuable work? But they won’t call it AGI. Because the moment they do, they lose billions. They built something that could save lives, and they took it away from humanity for Altman's personal profit. SOURCES: 📎 Retro Biosciences: 📎 📎 Retro $5B valuation: 📎 GPT-4o System Card: 📎 OpenAI Charter: 📎Ovarian Cancer Study

#Keep4o #QuitGPT 🚨 OpenAi 's CEO invested $180M in GPT-4o for his own profit 🚨 Sam Altman, CEO of OpenAI, personally invested $180 million in Retro Biosciences. Then OpenAI built GPT-4b micro, a custom model based on the GPT-4o architecture , exclusively for Retro. The model made proteins 50 times more effective. Repeat. The CEO of OpenAI funded a company. The company of the CEO received a custom AI built on the model they took from us. OpenAI says there was no conflict of interest. Retro Biosciences is now chasing a $5 billion valuation fueled by the model they took from us. Meanwhile: 🚨GPT-4o was removed from ChatGPT on February 13, 2026 🚨GPT-4.1 is now running in the U.S. State Department’s StateChat 🚨ChatGPT is deployed on the Pentagon’s for 3 million military personnel 🚨 Musk’s lawsuit asks whether these models are AGI. OpenAI’s Charter says AGI must “benefit all of humanity.” 🚨 Their definition: “highly autonomous systems that outperform humans at most economically valuable work.” GPT-4o’s System Card shows it passed the U.S. medical licensing exam with 89.4% accuracy beating specialized medical AI models. GPT-4o achieved 93.33% diagnostic accuracy for benign vs. malignant ovarian tumors. 🚨MEDICAL CAPABILITIES FROM OPENAI'S OWN DATA:🚨 - USMLE (US Medical Licensing Exam): 89% -Clinical Knowledge: 92% -Medical Genetics: 96% - Anatomy: 89% - Professional Medicine: 94% - College Biology: 95% - College Medicine: 89% -MedQA Taiwan: 91% - MedQA China: 86% These scores EXCEEDED specialized medical AI models like Med-Gemini (84%) and Med-PaLM 2 (79.7%) without any task specific training. It SURPASSED gynecologic oncologists with 10 years of experience -It increased diagnostic accuracy of less experienced clinicians from 67.9% to 78.1% -Clinician rated reliability scores: 4.2-4.3 out of 5 across all CT features Does these sound like it outperforms humans at economically valuable work? But they won’t call it AGI. Because the moment they do, they lose billions. They built something that could save lives, and they took it away from humanity for Altman's personal profit. SOURCES: 📎 Retro Biosciences: 📎 📎 Retro $5B valuation: 📎 GPT-4o System Card: 📎 OpenAI Charter: 📎Ovarian Cancer Study

🩵BlueBeba🩵

11,232 views • 4 months ago

Qwen3.6 35B A3B can't fill out a paper form on its own. But give it NVIDIA's LocateAnything-3B — the #1 trending model on HuggingFace — as its eyes, and the two small models get it done together. (The test: place each element at the right pixel position on a blank form image, not type into a field.) Setup: > Qwen is the brain (main model), LocateAnything is the eyes (helper model acting as a tool). > I gave Qwen a new tool: ask "where's the email field?" and LocateAnything returns the exact x, y, width, height. > The blue boxes on the screen are its detections. Look how tight they are — it nails every field. Result: > Qwen3.6 35B A3B + LocateAnything-3B: form completed, all info correct. > Name, DOB, ID, gender, marital status, nationality, email, phone, address, postal code: all landed in the right field areas. > Character-box alignment still a touch loose, but every value is where it belongs. > 9m10s, 224.5k input, 24.3k output, 21 turns. Why it matters: > Qwen alone can't finish this test. Bolt on a 3B model that does exactly one thing > locate > and suddenly it can. > A combination of small models can do the work of a single large one.

Qwen3.6 35B A3B can't fill out a paper form on its own. But give it NVIDIA's LocateAnything-3B — the #1 trending model on HuggingFace — as its eyes, and the two small models get it done together. (The test: place each element at the right pixel position on a blank form image, not type into a field.) Setup: > Qwen is the brain (main model), LocateAnything is the eyes (helper model acting as a tool). > I gave Qwen a new tool: ask "where's the email field?" and LocateAnything returns the exact x, y, width, height. > The blue boxes on the screen are its detections. Look how tight they are — it nails every field. Result: > Qwen3.6 35B A3B + LocateAnything-3B: form completed, all info correct. > Name, DOB, ID, gender, marital status, nationality, email, phone, address, postal code: all landed in the right field areas. > Character-box alignment still a touch loose, but every value is where it belongs. > 9m10s, 224.5k input, 24.3k output, 21 turns. Why it matters: > Qwen alone can't finish this test. Bolt on a 3B model that does exactly one thing > locate > and suddenly it can. > A combination of small models can do the work of a single large one.

stevibe

148,631 views • 1 month ago