InTowards Deep LearningbySumit Pandey·1d agoHuawei Researchers Turned a Photo Editing Model Into a 3D ScannerMarigold V2 pulls depth, shape, and material out of a single ordinary photo, and it runs on one consumer GPUA response icon2A response icon2
InTowards Deep LearningbySumit Pandey·6d agoDid ChatGPT Steal a $1 Million Math Proof? Even OpenAI Can’t Fully Rule It OutIt wasn’t ChatGPT, the proof leans on a loophole most mathematicians ignore, and no one is collecting the million dollars.A response icon4A response icon4
InTowards Deep LearningbySumit Pandey·Sep 9I Asked GPT-6 and Fable 5.1 to Build an End to End Software, Outcome Shocked MeOne of them won the build and lost the analysis, and daily token limits shaped the outcome more than any benchmarks could.A response icon19A response icon19
InTowards Deep LearningbySumit Pandey·Sep 7They Wanted $79 a Month to Draw Rectangles, So I Built a Free Tool For EveryoneIt runs inside your browser, draws the boxes and traces the masks for you, and never uploads a single image. No account, no bill.A response icon3A response icon3
InTowards Deep LearningbySumit Pandey·Sep 2A 30MB Rust Browser Fixed the Worst Part of My AI News PipelineA small AI news digest that runs itself, the one job it kept failing at, and the tiny open-source browser that finally fixed it.A response icon8A response icon8
InTowards Deep LearningbySumit Pandey·Aug 25I Tried Viral DeepSeek Harness with Qwen 3.8 27B, and It was Beyond My Imagination.How a DGX Spark box, an Qwen 3.8, and the viral Deepseek harness gave me the best coding/personal-assistant experience of my lifeA response icon16A response icon16
InTowards Deep LearningbySumit Pandey·Aug 24The Night I Found Out My Computer Was Smarter Than I ThoughtFreeToken: efficient local MoE serving that runs 35B to 753B models on consumer GPUs faster than existing engines.A response icon2A response icon2
InTowards Deep LearningbySumit Pandey·Aug 20Why Claude is Getting Worse Day by Day ?I think claude (including claude code) is getting worse day by day, I don’t know if it’s only me or others also feel the same.A response icon22A response icon22
InTowards Deep LearningbySumit Pandey·Aug 18Why Qwen3.8 27B Looked Brilliant in Testing but Failed to Ship My AI NewspaperQwen 3.8 27b on a DGX Spark: perfect on my benchmark, and it ran for twenty minutes on the real job, timed out and published nothing…A response icon6A response icon6
InTowards Deep LearningbySumit Pandey·Aug 14Qwen 3.8 Max: The Greatest Open-Source Model of All Time Has Been ReleasedFor the first time, an open model (Qwen 3.8 Max) doesn’t feel like the cheaper alternative. It feels like the frontier.A response icon5A response icon5
