Pierce Freeman and Richard Diehl Martinez review the June announcement of GPT-5.6 Sol, Terra and Luna while access was limited to trusted partners. They compare the tiers and discuss training uncertainty, but their guesses about parameter counts, checkpoints and undisclosed training runs are not confirmed technical details.
Pierce Freeman and Richard Diehl Martinez distinguish model weights from the harness: tools, orchestration, context handling and permissions surrounding a modelAn AI agent harness is the software framework that packages a model with tools, instructions, context management, execution controls, and user interaction.. They argue that a changed harness can unlock different capabilities, which makes headline model comparisons incomplete without knowing the evaluation setup.
Pierce Freeman and Richard Diehl Martinez compare laboratory cyber testsAI safety is the field and practice of reducing harmful failures, misuse, loss of control, and unintended consequences from AI systems., simulated networks and hardened software. Their central caution is that success on one task does not establish universal real-world compromiseEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria.. They discuss Terminal-Bench and incentives to optimize for public scores while acknowledging that many reported results are not independently auditable.
Pierce Freeman and Richard Diehl Martinez trace the shift from consumer protections and risk frameworks toward national-security oversightGovernance defines decision-making authority, rules and accountability within a system or organization.. The June order establishes voluntary access and classified cyber testing, not blanket mandatory model licensing. They debate whether future controls will target models or harnesses and how secrecy affects public scrutiny.
Pierce Freeman and Richard Diehl Martinez react to Anthropic's allegations of large-scale distillationKnowledge distillation trains a smaller or different AI model to reproduce useful behavior learned from a stronger teacher model. and debate the distinction between training on public material and extracting a competitor's outputs. They examine data-labeling costs, contracts and fairness without settling copyright law, attributing wrongdoing to an entire nationality or treating payment as permission for every use.
Watch on YouTube




