Rakesh🐊 (@rozeappletree) header
Rakesh🐊 (@rozeappletree) profile picture

❤️ FOSS 📊 Math 🧠 Intelligence

More →
Joined August 2019
Photo by Rakesh🐊 (@rozeappletree)Photo by Rakesh🐊 (@rozeappletree)Photo by Rakesh🐊 (@rozeappletree)Photo by Rakesh🐊 (@rozeappletree)Photo by Rakesh🐊 (@rozeappletree)Photo by Rakesh🐊 (@rozeappletree)Photo by Rakesh🐊 (@rozeappletree)Photo by Rakesh🐊 (@rozeappletree)Photo by Rakesh🐊 (@rozeappletree)Photo by Rakesh🐊 (@rozeappletree)
Are you proud of what you are building?
2
1
1
880
Rakesh🐊 retweeted
Fabel 5.2 is insane 😳 Built this amazing world and riding dragon This looks so beautiful 😍 ❤️
53
24
4
734
43,048
Rakesh🐊 retweeted
this 1.7B parameter model is magical they say ✨
This is a 1.7B param model btw, using the browser, making html files, and managing itself in Pi. This model hits 4o levels on most benchmarks lol Great for workflow automations.
1
7
122
8,913
Rakesh🐊 retweeted
space bunny has been the top model on opencode for 5 days straight 40T tokens
93
37
38
2,802
227,007
Rakesh🐊 retweeted
Sharing my slides on "Getting started with Local AI", presented at a recent hackathon in Qatar! Learn more about: * benefits of local AI * where to find the best models * quantization * dense vs. MoE models * using @Llama_app
9
21
177
16,169
Rakesh🐊 retweeted
GRPO problem explained very simply
New blog! This one covers my experiments on fine-grained credit assignment with an agentic judge, and how it affects the models' performance and reasoning quality and some thoughts on RL science. Link below, please do check it out, and RT/QT if you like it:)
12
27
1
512
30,600
Rakesh🐊 retweeted
Animation by Claude Opus 5.5 Paintings and paper textures generated with GPT Image 2.6 via @fal All motion written in code
56
32
5
715
31,491
Rakesh🐊 retweeted
Cinema app design that shows you the view from your seat before you book.
619
3,298
999
46,062
4,281,607
Rakesh🐊 retweeted
The desktop: every run gets a fresh HF Sandbox (an isolated VM that boots in seconds). Learn more at huggingface.co/docs/huggingf… The image (Xvfb + XFCE + Firefox + xdotool) is built by a Docker Space and pulled straight into the sandbox via huggingface.co/docs/hub/en/j…. The desktop is ready in ~18s.
1
1
6
463
Rakesh🐊 retweeted
Im tired of infinite canvas startups. Im tired of generative video snake oil providers. Im tired of hype for hyperframes, as html and playwright is a dumb long term solution for video rendering. So I decided to build Raea. Open source. Free for everyone. Faster and more capable than hyperframes or remotion. (10-20x faster with real footage and complex scenes). And it renders without a browser…. Capable of more than just video. It’s built on @BevyEngine so it’s highly capable at producing 3D art, shaders and games. It’s also built on Blitz Dom @dioxuslabs , and a bridge between the page and the 3d world behind it allows you and your agent to write everything in typescript and react… without a browser. This video below was made entirely with Claude, Raea, and a weekend. I never had more fun playing and toying with creative work than I did making this. But to be honest, that’s not where I’m stopping. I want Raea to be an engine that correctly empowers creatives and agents, while not cognitively offloading the parts that make your work unique. I’m excited to continue working on this project as long as I can, and always open to feedback!
164
168
34
2,414
100,968
Rakesh🐊 retweeted
We're in a computer graphics renaissance Drawings, photos, renders are all something world models can "see into", world vertigo
73
369
27
5,659
164,537
Introducing the Together Compiler: continual compilation for AI inference! I’m super proud to finally share this work from our research team! ✨ Today’s inference is “Polymorphic” or shape-shifting, it doesn’t sit still. On OpenRouter, 80% of traffic goes to models released in the last three months and the top ten most popular open models each use a different architecture. Our customers constantly change their usage behaviors. Together Compiler leverages agents to continually optimize our deployments immediately as these changes occur. Together Compiler research highlights include: 1. 252x better throughput: TKCC lets agents try 252x more optimization experiments per hour compared to prior agent inference optimization systems. 2. State-of-the-art TPS across InferenceX workloads we tried: TKCC outperforms the best InferenceX engine across all tested settings in just 24 hours & 2 nodes per compile. Including 1111 TPS on MiniMax M3.1 B200 and 1205 TPS on DeepSeek v4.1 Flash GB300 at batch size 1. NVIDIA, AMD, Trainium2. 3. Continuous recompilation: TKCC currently serves many trillions of production tokens per day and adapts to traffic shifts within just a few hours of recompilation. 4. Bitter lessons: TKCC’s design builds on 2+ years of building frameworks to accelerate AI Inference, starting from ThunderKittens and Megakernels. Compilers have always used abstractions and IRs, but we show that agentic compilers now reach higher performance ceilings without abstractions. From an incredible team @reyna_abhyankar @dylan__lim @realDanFu @togethercompute!!
94
142
45
968
173,676
Rakesh🐊 retweeted
This is actually a common misconception. Hidden states in LLMs are used to predict all future tokens not just the next one! And in fact we see representations that correspond to plans, or the broader situation transformer-circuits.pub/202… anthropic.com/research/globa…
Remember one thing when you read JEPA vs LLMs in the context of the platonic representation hypothesis. Hidden state of an LLM does not represent the full representation of the LLM, it carries only what is needed for the next token. It's not an encoder.
14
30
8
380
19,788
Rakesh🐊 retweeted
There is so much alpha in this blog post for anyone learning model optimization. Check it out! llamaindex.ai/blog/the-infer…
We got OCRia-1 generating tokens nearly 10× faster on a single B200. Getting there meant making much better use of the same GPU. We tuned how OCRia-1 runs for the demands of document parsing, testing every change for both speed and accuracy. Some improvements were only 2–5%. Seventeen smaller optimizations later, “just a few percent” starts adding up. Our infrastructure team wrote up what worked, how they tested the changes, and what they learned about making document models run faster. We also plan to contribute some of these optimizations back to the open-source vLLM project. Read the full breakdown → llamaindex.ai/blog/the-infer… Watch the video below to see OCRia-1’s token generation before and after the optimizations.
2
7
41
4,296
Everyone’s excited about decision models. When should you buy one, or train your own? We fine-tuned a 2B model for a production classifier: frontier-level quality, 35ms, ~1% of the cost. We’re sharing when we think it’s worth it, and when it isn’t: airbnb.tech/ai-ml/frontier-q…
29
46
12
606
49,291
Rakesh🐊 retweeted
Agents can now control phones. Introducing Phone Harness (YC F26) Let your agent use your personal phone, or its own iPhone/Android in the cloud > Test and use real mobile apps > Message and call people > Grab verification codes for 2FA Open source. Connect through MCP, CLI, or API. Get started at phone-harness.com
157
114
57
1,662
224,568
Rakesh🐊 retweeted
Microsoft released a new 4B model called Mage-VL-Cowork that sets a new SOTA for on-device computer use 🔥 I built a demo that lets you submit a task like "Go to Wikipedia and find out when the Eiffel Tower was completed." Watch it click, type, scroll, and use a terminal on a real Linux desktop. The whole thing runs on @huggingface: the model in a Space, the desktop in an HF Sandbox. No VMs to manage. 🧵
9
16
6
156
12,486
Rakesh🐊 retweeted
Building Voice AI for India means building for the acoustic reality of Bharat - and doing it at a cost that India won't mind paying for. ​Please give it a listen. The Jio moment of Voice AI is here! ​Super excited and motivated 🙏
35
9
2
153
11,005
Rakesh🐊 retweeted
First a CPU, then a GPU. A computer isn’t complete without storage. Meet qwenFS: Qwen3.5-2B with a filesystem compiled directly into its weights. Nothing runs outside the model. It adds 8 layers in front of Qwen’s own 24, using attention to hold the files and Gated DeltaNet state to hold the tables. Every save and read runs in the model’s own forward pass. A write stores its text in the copy layer’s attention log, keyed by file, version and offset. A read uses those same keys, so attention copies the text back exactly token by token.
Meet qwenGPU: Qwen3.5-2B with a SIMT GPU inside its weights, 45 compiled layers in front of Qwen’s own, using its Gated DeltaNet layers as registers and memory. It runs real CUDA in its forward passes. The host compiles the kernel to PTX and loads it, then the model writes every step of the launch itself: every register, branch and memory write. The GPU has 4-lane warps, a 4-stage pipeline, 15 registers per lane and 126 words of global memory, all held in DeltaNet state. Qwen generates each launch token by token, each record running one machine block for one warp.
84
153
46
1,831
136,675
Rakesh🐊 retweeted
We @huggingface reproduced this paper!! It's built with ML Intern, an open-source harness to do AI research, train models, and more Model: huggingface.co/ML-Intern-lab… Demo: huggingface.co/spaces/ML-Int…
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
64
181
26
1,640
96,723