Conference Presentation, Fireside Chat
How Harvey Built a Research Lab on a Budget | Gabe Pereyra
- Application layer companies are expected to compete with frontier labs by leveraging the frontier ecosystem to build frontier intelligence, with the anticipated outcome that every company must adopt an AI-focused playbook.
- The organization plans to operationalize a system enabling individual law firms and enterprise customers to customize models for specific work while protecting client data, viewing this as the end game rather than building a single best legal model.
- Product development priorities are projected to shift from individual productivity tools to organizational-level productivity and resource allocation within firms and enterprises.
- The team anticipates winning with a limited budget and current staff, likening the strategy to a "Moneyball" scenario, while growing to a size that avoids direct competition with frontier talent for high-salary researchers.
- A growing pool of PhD candidates不愿 working at large labs is expected to increase available talent for research and infrastructure, alongside new infrastructure and talent making post-training and serving models easier.
- Future datasets are expected to be future-facing to address the gap between realistic synthetic data and actual production distributions for tasks like email drafting, while synthetic data helps general improvement alongside private data training.
- Open-source models are predicted to become competitive enough for post-training to specific frontier intelligence levels, though potentially not for general frontier intelligence, despite current skepticism regarding this playbook.
- The team expects to face a performance gap regarding current models' ability to manage large context sizes, specifically citing data rooms of 80 million tokens.