Aashi Dutt

Here is my take "Attention Residuals Explained: Rethinking Transformer Depth" ⚡️Depth is finally getting its Attention moment. Read more: datacamp.com/blog/attention…

Full tutorial 👇 datacamp.com/tutorial/qwen-…

📍Stack: Qwen 3.5 (9B) via Ollama OpenCV (frame extraction) Pydantic (structured spec) Streamlit (UI + preview)

I just built a system that turns a gameplay video into a playable browser game 🤯 Using @Alibaba_Qwen 3.5 small models with @ollama Here is a abridged version of the workflow:

Introducing 𝑨𝒕𝒕𝒆𝒏𝒕𝒊𝒐𝒏 𝑹𝒆𝒔𝒊𝒅𝒖𝒂𝒍𝒔: Rethinking depth-wise aggregation. Residual connections have long relied on fixed, uniform accumulation. Inspired by the duality of time and depth, we introduce Attention Residuals, replacing standard depth-wise recurrence with learned, input-dependent attention over preceding layers. 🔹 Enables networks to selectively retrieve past representations, naturally mitigating dilution and hidden-state growth. 🔹 Introduces Block AttnRes, partitioning layers into compressed blocks to make cross-layer attention practical at scale. 🔹 Serves as an efficient drop-in replacement, demonstrating a 1.25x compute advantage with negligible (<2%) inference latency overhead. 🔹 Validated on the Kimi Linear architecture (48B total, 3B activated parameters), delivering consistent downstream performance gains. 🔗Full report: github.com/MoonshotAI/Att…

Built a complete person detection system with real-time video streaming on embedded hardware. Without writing a single line of code. 🤯 My only instruction: "Create a person detection demo with a real-time video dashboard" Here's how @claudeai Code + a custom skill automated the entire embedded dev workflow: 🧵

Fully AI-Generated CUDA Course 🚀 - Remotion for cool slides - Gemini 3.0 for video understanding to make scripts more human-friendly - ElevenLabs v3 for natural text-to-speech - LTX-2 for avatar video generation

Mercury 2 is live 🚀🚀 The world’s first reasoning diffusion LLM, delivering 5x faster performance than leading speed-optimized LLMs. Watching the team turn years of research into a real product never gets old, and I’m incredibly proud of what we’ve built. We’re just getting started on what diffusion can do for language.

@crystalsssup Waiting on "Restarting now..." 😢

@AashiDutt you can ask kimi claw to debug it or reconnect:)

Introducing ZUNA, a 380M-parameter BCI foundation model for EEG data, a significant milestone in the development of noninvasive thought-to-text. Fully open source, Apache 2.0.

👉Read it here: datacamp.com/tutorial/glm-i…

⚡️Learn about: - Hybrid autoregressive and diffusion design for stronger layout and sharper details - Text rendering (the hardest part of slide images - can mis-spell/ mis print certain words😅 ) - Strong on knowledge-dense, information-heavy visuals

🌸A model made these slides for me in under 5 minutes.👇 In this quick tutorial on building a prompt to infographic deck generator with #GLM-Image (from @Zai_org) wrapped with a @streamlit app.

👷‍♀️Built a demo for testing @deepseek_ai OCR-2’s vision-first document reading with: - Visual Causal Flow to learn a better reading order (instead of naive top-left → bottom-right scanning) - Stronger handling of layout-heavy PDFs 👉datacamp.com/tutorial/deep-…

🤌Tried something with @openclaw and @ollama for building a Local Data Analyst 🧠 💪No cloud APIs. and No data leaving your machine. 👇Links for you: Tutorial: datacamp.com/tutorial/openc… Code: github.com/AashiDutt/Open…

🚀I built a Generate-and-Edit image app using @BlackForestLabs's FLUX.2 Klein 4B and @Gradio . Here’s what makes this setup interesting 🧵👇 📍Video playback is sped up for demonstration purposes.

👉Full tutorial (code + explanations): datacamp.com/tutorial/flux-…

💥The underrated part: session history. Every generated or edited image is stored. You can reload any previous version and continue editing from there. This enables: • Branching edits • Easy comparisons • No lost “good” intermediate results