• Discovered LLM tasks with a new type of scaling law by running the Inverse Scaling Prize ($250K pool). Executed everything from designing task formats to evaluating 20+ language models on 99 dataset submissions and analyzing the causes of inverse scaling.• Created the KTS technique for targeted LLM steering, leveraging my code for distributed fine-tuning and synthetic data generation, and using Huggingface, LangChain and OpenAI; KTS reduces Llama 2 jailbreak success rate by 44% with… Show more • Discovered LLM tasks with a new type of scaling law by running the Inverse Scaling Prize ($250K pool). Executed everything from designing task formats to evaluating 20+ language models on 99 dataset submissions and analyzing the causes of inverse scaling.• Created the KTS technique for targeted LLM steering, leveraging my code for distributed fine-tuning and synthetic data generation, and using Huggingface, LangChain and OpenAI; KTS reduces Llama 2 jailbreak success rate by 44% with negligible impact on helpfulness. Show less

• Helped establish the field of Cooperative AI with a publication highlighting problems with reinforcement learning algorithms in multi-agent bargaining, using GCP, Docker, Ray, PyTorch, Wandb for RL experiments.

• Implemented ensembling and model sparsity in the internal Samsung deep learning framework, used by Samsung R&D offices worldwide for efficient neural network training and inference.• Created novel methods of ensembling and data augmentation in computer vision, performing 1000s of CNN training experiments using PyTorch, TensorFlow, Docker, SLURM, Ray. My methods (GPS) and metrics (DEE) later saw adoption by external deep learning and medical imaging research groups across the world.

• Built a library for protein search—a crucial component of the in-house end-to-end drug discovery pipeline—using Python, GitLab CI/CD, pandas, RDKit, Open Babel.• Addressed data scarcity in modeling protein-ligand affinities by implementing GNNs and Neural Processes for transfer learning and boosting accuracy on rare proteins, working cross-functionally with chemists.

• Built a U-Net-based segmentation system for image background removal from prototype to production for an e-commerce client in an Agile team, using Python, AWS, PyTorch, OpenCV, PIL, pandas, numpy.• Developed OCR for passport images using YOLO, Faster R-CNN, and LSTM-CTC, from zero to 97% document-level accuracy in weeks; the startup later zeroed in on it as the sole long-term product.