Joseph Cooper's Blog

Rebuilding haiku.joecooper.me for CPU-only inference.

Training a multiturn conversation model on a single GPU using Markov text generation and DeBERTa-based reranking.

Using softmax, A* search can generate diverse candidates for use with test-time compute.

What you really need to know as a user, not developer, of modern AI systems.

I've put an app online to check your resume with a GPT-wrapper simulator.

Retrieval-based in-context learning (RetICL) improves LLM performance by retrieving relevant labeled examples during inference.

Prompt synthesis with activation maximization achieves 95.9% on Yelp Review Polarity with Llama-3.2-1B and a 4-token prompt.

Benchmarking various models to find the best way to deduplicate factoids for AI applications.