Meet Charles “Charlie” O’Neill
Charlie O’Neill is an AI researcher, founder, and writer who runs Base Labs, a Baseten research lab focused on democratizing open-source intelligence. His work spans language-model training, continual learning, specialized open-source models, post-training, mechanistic interpretability, long-context memory systems, and AI evaluation. He previously cofounded Parsed, studied machine learning as a DPhil student at Oxford, and writes about both technical AI research and the broader implications of AGI for work, institutions, and society.
Summary of Charles “Charlie” O’Neill and his portfolio
charlesponeill.com is the personal portfolio and writing site of Charlie O’Neill, an AI researcher, founder, and writer focused on language-model training, open-source intelligence, continual learning, model specialization, and mechanistic interpretability.
Charlie says he runs Base Labs, a research lab by Baseten, which works to “advance and democratise open-source intelligence” by publishing models, data, environments, recipes, and even failed experiments. Before that, he cofounded Parsed, where the goal was to build specialized open-source models competitive with frontier labs; the site says Parsed was acquired by Baseten, after which he led model-training work there.
His background includes time as a DPhil student in machine learning at Oxford under Jakob Foerster, being a General Sir John Monash Scholar, studying at ANU as a Tuckwell Scholar, and cofounding Rake News, a language-model-based personalized and debiased news project.
What he works on
Charlie’s work is centered on practical and scientific questions around building better language models, especially:
- Continual learning — how models can keep learning over time without catastrophic forgetting or degradation.
- Specialized open-source models — training models for specific companies, domains, tasks, and agentic workflows.
- Post-training and fine-tuning science — supervised fine-tuning, distillation, LoRA, RL-style training, dense rewards, prompt mutations, and evaluation-driven training.
- Mechanistic interpretability — especially sparse autoencoders, disentangled representations, feature structures, hallucination detection, and steering.
- Long-context and memory systems — KV-cache compaction, latent memory, context compression, and model memory hierarchies.
- AI evaluation and model behavior — how to assess language models, improve their reasoning, and detect failures.
- AI and society — AGI timelines, the future of work, research culture, news, rationalism, and the role of humans in a post-AGI world.
A good example of his current thinking is the essay “Trying to actually define continual learning”, where he argues that continual learning is not one single problem: knowledge acquisition, skill acquisition, in-context memory, weight updates, and long-term agent behavior all need to be separated and studied differently.
Website structure
The site is minimalist and organized around three main areas:
- Homepage — short biography, current work, selected recent writing, and selected papers.
- Writing — essays mostly about training language models, plus occasional writing on books, cities, AI culture, and philosophy.
- Papers — research papers, preprints, and technical notes from Base Labs and Baseten Research.
Selected writing
His writing is opinionated and research-adjacent, often connecting technical AI progress with broader questions about work, meaning, and institutions. Recent posts include:
- “The final form of work” — argues that in a post-AGI world, important human work may become deciding which problems are worth solving.
- “Trying to actually define continual learning” — lays out why continual learning involves memory, skill, facts, context, and weight updates as distinct subproblems.
- “The nihilism of NeurIPS” — a reflective essay on AI research culture, AGI progress, and the meaning of academic work.
- “How I’d build the future of news” — notes from Rake about personalized, synthesized, language-model-mediated news.
Selected research
The papers page lists work across language models, interpretability, continual learning, and model specialization. Notable listed papers include:
- “Post-Training Science for Supervised Fine-Tuning”
- “Can a Language Model Learn Facts Continually in Its Weights?”
- “Still: Amortized KV Cache Compaction in a Single Forward Pass”
- “Compute Optimal Inference and Provable Amortisation Gap in Sparse Autoencoders”
- “Disentangling Dense Embeddings with Sparse Autoencoders”
- “Sparse Autoencoders Enable Scalable and Reliable Circuit Identification in Language Models”
- “AstroLLaMA: Towards specialised foundation models in astronomy”
The research page also links many Base Labs/Baseten technical notes on topics such as LoRA, model evaluation, hallucination detection, KV-cache compaction, legal agents, and training specialized sub-agents.
Social and professional links
The homepage links to Charlie’s:
Overall, the site presents Charlie O’Neill as a technically focused AI researcher-founder working at the intersection of model training, interpretability, continual learning, and practical open-source AI systems, with a strong interest in how AGI changes research, work, and society.