PinnedInTowards Deep LearningbySumit Pandey·Nov 3Andrej Karpathy Got Sick of YouTube Slop, So He Built a HackerTubeHackerTube is 6,454 videos that Hacker News actually recommended, and Karpathy didn’t write, read or run a single line of the code behind…A response icon4A response icon4
InTowards Deep LearningbySumit Pandey·1d agoI Tested SenseNova U1.5 Against ChatGPT. It Failed Just One Image TestI ran SenseNova locally on a DGX Spark against ChatGPT. It nailed text, handled complex edits, but couldn't count coffee cups.A response icon1A response icon1
InTowards Deep LearningbySumit Pandey·6d agoAI on Excel Sheet: NVIDIA’s Kumo Runs 23x Faster Than Google’s TabFMIt predicts straight from your spreadsheet with no training at all, and unlike the two rivals that edge past it, it’s free for business…A response icon2A response icon2
InTowards Deep LearningbySumit Pandey·Nov 6How uzu Runs a 9B Model 4x Faster Than llama.cpp.Memory bandwidth caps a base M5 at about 29 tokens a second, so here’s how uzu gets 92 and where the 4x needs an asterisk.A response icon2A response icon2
InTowards Deep LearningbySumit Pandey·Nov 4Weight-Space Programming: How GPT-6 Astra Fixed Qwen by Editing 0.016% of Its WeightsMIT researchers handed OpenAI’s GPT-6 Astra the raw weights of a small Qwen model and asked it to make the model better at math, and it did…A response icon2A response icon2
InTowards Deep LearningbySumit Pandey·Oct 21OpenShell: Nvidia's Open-Source Answer to Rogue AI Agents at OpenAI and AnthropicAfter rogue agents hit Hugging Face and Medicare, Nvidia’s open-source OpenShell runtime puts the fence outside the agent
