Ashish Kumar

Hey, I'm Ashish, an undergrad (class of 2029) who got hooked on machine learning and never really recovered. What I enjoy most is building AI systems from the ground up. Give me a paper and an empty folder and I'll happily spend a week turning its equations into working layers, then another week finding out why the loss curve looks wrong. Most of that energy goes into large language models: tokenizers, pretraining runs, fine-tuning, evaluation, and the serving code that finally lets you chat with the thing you trained. I like knowing every part of the machine instead of trusting a black box, and I like showing people the numbers, including the unflattering ones. It's slower than stitching existing libraries together, but it's the only way I've found to actually understand what's going on inside. Honestly, watching a model I built from nothing string its first coherent sentence together is still the best feeling I know.

Language models sit in the middle, but they're not the whole picture. One eye stays on generative media: diffusion and flow models for images and video, speech and text-to-speech, and multimodal models that look and read at the same time. The other stays on world models, agents and reasoning, the pieces that feel closest to the big question of how we get to general intelligence. Underneath all of it is the machinery. I write CUDA kernels to see where the milliseconds go, I think about memory far more than is healthy, and I'm teaching myself digital design so that one day the model and the chip can be designed together. Robots, embodied AI and AR/VR are where I'd love all of this to land, somewhere a model has to deal with a body, a camera and the messy real world. It's a long list, and I have no intention of making it shorter.