We scan new podcasts and send you the top 5 insights daily.
Rob Ferguson cites reinforcement learning finding 4–6% better video compression after PhD-level work plateaued, because the outcome was easy to score. He advises breaking work into small verifiable parts; if you can measure it, you can improve it with AI. That is the great secret to AI, he says.
Unlike traditional software that optimizes for time-in-app, the most successful AI products will be measured by their ability to save users time. The new benchmark for value will be how much cognitive load or manual work is automated "behind the scenes," fundamentally changing the definition of a successful product.
The most effective AI strategy focuses on 'micro workflows'—small, discrete tasks like summarizing patient data. By optimizing these countless small steps, AI can make decision-makers 'a hundred-fold more productive,' delivering massive cumulative value without relying on a single, high-risk autonomous solution.
Businesses should prioritize AI projects that can completely automate a recurring workflow. Transforming a multi-week manual process into an instantaneous one delivers transformative value, far exceeding the gains from projects that only offer partial assistance to a human user.
Don't assume AI can effectively perform a task that doesn't already have a well-defined standard operating procedure (SOP). The best use of AI is to infuse efficiency into individual steps of an existing, successful manual process, rather than expecting it to complete the entire process on its own.
Instead of asking an AI for a one-off task, identify recurring workflows and have the AI turn them into a "skill." This creates a reusable asset that dramatically improves efficiency and output quality over time, turning the user into a system builder.
The most effective AI companies don't try to automate everything. They ask which specific, repetitive task creates the most value when partially automated. This pragmatic approach delivers measurable results by using AI to augment human workers, not replace them.
AI models improve dramatically in domains with objective feedback, like coding (unit tests) or science (lab results). Progress is slower in subjective fields like creative writing where feedback is opinion-based, explaining the uneven impact of AI across different types of knowledge work.
Unlike more abstract domains, AI research is particularly suited for automation by AIs. The tasks are verifiable, allow for iterative improvement, and can be broken down into containerized environments for reinforcement learning.
Instead of merely outsourcing tasks to AI, frame its use as a tool to compound your learning. AI can shorten feedback loops and help you practice and refine a craft—like messaging or video editing—exponentially faster than traditional methods, deepening your expertise.
Major advances like RL on Chain of Thought could have occurred earlier on less powerful models. The real bottleneck was often not the core concept but the mundane details of infrastructure, implementation, and hyperparameter tuning—tasks that AI labor can massively accelerate.