Ayush Tambde

I'm a researcher at Paradigm focused on meta reinforcement learning

Fun facts about me:

• Raised in Dublin, Ireland
• Live in San Francisco, California
• Theoretical physics sophomore on leave from University College Dublin
• I was really into quantum computing and systems biology as a teenager
• Won best individual at the Irish high school science fair for developing mechanistic signalling models of luminal and basal breast cancer cells
• I'm part of Emergent Ventures, and a Unitary Foundation grant recipient
• Mostly listen to Angels & Airwaves and Third Eye Blind
• Fascinated by the lives of Lyndon B. Johnson and Alfred L. Loomis.

Things I've done in the past:

• Matrix Orthogonalization Improves Memory in Recurrent Models
• Tree Search Distillation for Language Models using PPO
• Special Projects @ NFDG/Andromeda
• A LoRA finetuning framework for DeepSeek-V3 (671B)
• Searched for Planet Nine by rewriting the Mercury integrator in CUDA and running sims on H100s
• Built trading engines and VM orchestrators @ SFCompute
• FlashAttention for Apple Silicon because I really wanted to train models on my Macbook
• A compiled quantum programming language
• Quantum gate simulator as a PCI peripheral in QEMU
• Operating system kernel with syscalls for interfacing with the above simulator