Saurav Pawar

After an amazing year-long internship, I am happy to share that I have joined Technology Innovation Institute as a Senior Associate Machine Learning Researcher!!🎉🎉🥳 #research #MachineLearning #artificalintelligence

@findgoddd Kadıköy is fun too, you can give it a try

🤖 𝗥𝗲𝗶𝗻𝗳𝗼𝗿𝗰𝗲𝗺𝗲𝗻𝘁 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 (𝗥𝗟) 𝗘𝗻𝗵𝗮𝗻𝗰𝗲𝗺𝗲𝗻𝘁𝘀: 🔹 𝗥𝗟𝗠𝗙 (𝗠𝗮𝗰𝗵𝗶𝗻𝗲 𝗙𝗲𝗲𝗱𝗯𝗮𝗰𝗸) – Boosts mathematical reasoning.  🔹 𝗥𝗟𝗘𝗙 (𝗘𝘅𝗲𝗰𝘂𝘁𝗶𝗼𝗻 𝗙𝗲𝗲𝗱𝗯𝗮𝗰𝗸) – Improves coding accuracy.

🔹 Built using JAX & trained on Google TPUs.

breaking down the deepseekmoe architecture

Chief Al Scientist at Meta, Yann LeCun, believes we're never going to get to human level AI by text:

This paper is very valuable for most people. You won't easily find this knowledge in books—it's expert-level stuff.

🔧 Use Cases: - Legal, financial, and academic document processing - Large-scale automated text extraction - AI-powered search and retrieva

🔑 Key Features: - Maintains document structure and layout - Supports multiple languages - Processes 2000 pages/min - Tables: 96.12% accuracy (better than GPT-4o & Gemini-1.5-Pro) - Scanned text: 98.96% accuracy (outperforms Google Document AI) - Math expressions: 94.29% accuracy

@MistralAI introduces Mistral OCR, a fast and accurate OCR model that extracts text while preserving structure. It recognizes images, equations, and scanned content, outputting structured text with bounding boxes and markdown formatting. Read more: mistral.ai/news/mistral-o…

I built a prototype that infers the appropriate CUDA version at runtime and installs the latest compatible PyTorch version (below, 2.6.0+cu126, i.e., CUDA 12.6) without any configuration.

❓Why it works? By converting a PyTorch model into TorchScript, we: ✅ Enable faster execution ✅ Reduce dependency on Python ✅ Simplify deployment for real-world applications With TorchScript, PyTorch models become faster, lighter, and deployment-ready.

Model inference can be slow due to Python’s overhead in @PyTorch. A powerful way to optimize performance? torch.jit.script – it compiles your model into an optimized representation, reducing execution time and improving efficiency. A thread 🧵

Does it Work? ✅Outperforms LLM-as-a-Judge on complex dependency tasks. ✅ On par with human evaluation in reliability. ✅ More detailed and actionable feedback than traditional methods.

📊 New Benchmark: DevAI – Existing code-generation benchmarks aren’t keeping up. Meta introduces DevAI, a realistic dataset with 55 AI development tasks to push agentic evaluation forward.

As AI agents evolve, evaluating them is the next big challenge, especially in code generation. @Meta's Agent-as-a-Judge lets AI assess AI, providing step-by-step feedback for better results. A smarter way to measure performance. 🔗 arxiv.org/pdf/2410.10934