Raunak Chowdhuri

🚀 DeepSeek AI just released DeepSeek-OCR 2 The new 3B model achieves state-of-the-art performance in OCR, visual, and document comprehension — especially on complex layouts. What’s new? 👇 🧠 DeepEncoder V2 Instead of reading images in a rigid grid (top-left → bottom-right), the model now learns a human-like reading order: → Builds global context first → Decides what to read next → Follows structure naturally Why this matters: ✅ Better column detection ✅ Stronger table understanding ✅ Accurate label–value linking ✅ Improved mixed text + layout handling ✅ More reliable form parsing 📊 Results: • Outperforms Gemini 3 Pro on benchmarks • >4% improvement over DeepSeek-OCR v1 This shows how much OCR quality depends on how models read — not just how big they are. #AI #OCR #DocumentAI #ComputerVision #LLM #DeepSeek #GenAI Link: https://lnkd.in/gTgicvBh

29

3 Comments

We just shipped Chimera SDK 26.04.01, and I'm excited to announce a feature our customers have been asking for since day one: God Mode. The new --iddqd compiler flag makes any neural network invincible. Models compiled with God Mode enabled are immune to quantization loss, accuracy degradation, and latency spikes. Every operator runs at theoretical peak. Every benchmark hits ceiling. Early results from our internal validation: Quantization Loss: 0% Accuracy Degradation: None Latency: Instantaneous CPU Fallback Ops: Negative twelve Power Consumption: Generates power FPS: Infinite TOPS/W: Yes Happy April 1st. The DOOM post was real though.

52

The critical trade-off between throughput vs. sample off-policyness has always been central in RL, from robotics to video games, I have seen it myself multiple times as a key design driver. But at the current scale of large foundation models, the centrality of its role in RL-based post-training efficiency has become even more evident. The high level architecture in the picture describes the custom system adopted by MiniMax in training their M2.5 model released a few days ago (link in the comments). They claim that this element alone provided an approximate 40x(!) training speedup. There are so many variables at play in these applications, and so much space for improvement, I cannot think of a more exciting space to work in. 

27

1 Comment

For LLMs, variability is a feature. For hardware, determinism and guaranteed accuracy are requirements. Engineering teams don’t build on “pretty good.” They build on results that are reproducible, verifiable, and safe to ship—run after run. Determinism is what turns simulation from an exploratory tool into an operational capability you can trust in production. That’s the bar we set at Vinci4D.ai: FEA/FVM-grade (or better) fidelity with repeatable, production-ready results that don’t drift as models scale. If it can’t be defended in a design review, supplier discussion, or manufacturing sign-off, it doesn’t belong in the workflow. If you’re responsible for thermal integrity, packaging reliability, or PCB warpage at scale, what’s the one requirement you still can’t guarantee end-to-end?

85

3 Comments