Mohit

conv model : so we get the sematics out from a model transformer model : so that the attention is spread and the relevance is maintained in continous latent space

best correlation between Build time vs Compile time is thinking docker build and docker run

@nengjiali this looks amazing from first glance, can you create a blog post / video for same would be amazing to get all things at one place

I spent 4 brutal months building a full Stereo Visual SLAM system from scratch in C++17 + CUDA. NO pre-made libraries. NO black-box magic. I just wanted to understand the underlying mechanics behind SLAM. Here’s the intuitive breakdown (the full SLAM pipeline, the math that almost broke me, and the real KITTI footage)👇 cc @aelluswamy your work in this space is a massive inspiration for tackling this from absolute scratch

@solairaiapp @spiritbuun 😂

@spiritbuun 0bit incoming

Had a huge breakthrough in quantizing weights today. Holy fug. You have no idea how small we're going and how high quality we're going. Soon.

@pupposandro and what about unsloth ? were those guys also not having better kernels ? @UnslothAI

@alexandr_wang hm still gpt 5.4 xhigh looks good .. now sure on pricing though

1/ today we're releasing muse spark, the first model from MSL. nine months ago we rebuilt our ai stack from scratch. new infrastructure, new architecture, new data pipelines. muse spark is the result of that work, and now it powers meta ai. 🧵

so apparently unsloth wrote better trt kernels for LLM that the official nvidia team generalised .. Guess we should all shift to using unsloth rather than using nvidia and we need same for diffusion models also

@pupposandro so tensorrt also doesnt have fused kernel for this deltanet + full attention layers ?

hm qwen3tts model

one thing that top models still struggle at is quantization .. seems simple at first but gets complex with added complexity of trt

its interesting to see how top tier models are also not that efficient in fields where data is a sparse ..

so claude is still figuring out with me on how to do PTQ reliably and how to bypass onnx protobuf limits and achieve quantization on all possible layers with min conversions back to original quant ..

qwen 3 tts explained ! @Alibaba_Qwen

@MarioNawfal now real delta is in making cheap interceptors , one main interceptor guiding ultra cheap small ones

🇮🇳 India just unveiled an AI-powered kamikaze drone with a 2,000km range and 12-hour endurance. Every major power is now racing to mass-produce cheap one-way attack drones, the weapon that's rewriting the rules of modern war. Expensive interceptor missiles vs. cheap drones. The math isn't pretty. Source: Zero Hedge

@anirudhbv_ce @GoogleResearch you are late to the party ..

I implemented @GoogleResearch's TurboQuant as a CUDA-native compression engine on Blackwell B200. 5x KV cache compression on Qwen 2.5-1.5B, near-loseless attention scores, generating live from compressed memory. 5 custom cuTile CUDA kernels ft: - fused attention (with QJL corrections) - online softmax -on-chip cache decompression - pipelined TMA loads Try it out: devtechjr.github.io/turboquant_cut… s/o @blelbach and the cuTile team at @nvidia for lending me Blackwell GPU access :) cc @sundeep @GavinSherry

@cursor_ai composer2 is shit

Omni 3.5 is close source .. really @Alibaba_Qwen maybe that why the founder left ..

shortcut learning in CNN , give same pixel to a class and model will learn to hack it to use it to its favour .. Classic fuckup

so today I realised, the gap between attention and convolution is not different and anyone that was in the deeply in the field could have come come up with it

@ViralOps_ When you read this article and realize its a shit article

when you realize you could have cloned any influencer in seconds. but you wasted hours doing everything manually.