Yash Patil joins Jacob Effron to examine when enterprises should move beyond prompting and harness optimization toward custom model weightsPost-training adapts a pretrained AI model to improve its behavior, capabilities, or alignment after the main training stage.. Patil argues that proprietary judgment, expert feedback and company-specific evaluationsEvaluation measures how well an AI system performs against defined tasks, criteria and failure conditions using repeatable evidence. can provide useful training signalsFine-tuning continues training a model on selected data so its behavior becomes better suited to a task, domain or operating environment.. He describes reinforcement learningReinforcement learning trains an AI system to choose actions using feedback about the results of its earlier choices. as optimization against a defined objective, stressing that rubric-based rewards can help with non-verifiable tasks but that transfer across distant domains remains limited.
Yash Patil and Jacob Effron explore the economics of combining training and inferenceInference cost is the expense of running a trained AI model to process inputs and produce outputs., including GPU utilization, algorithmic efficiency, token-efficient policies and workload-specific serving. Patil outlines a task, environment and verifier framework for training, and uses a legal-document workflow as an illustrative example. The discussion also covers evaluation confidentiality, reward hacking and engineers retaining responsibility for system design. Company performance claims are treated as Patil's account, not independently established results; subscription appeals and promotional interruptions are omitted.
Watch on YouTube
