In this video, the speaker discusses their personal experience and philosophy around choosing AI models for different types of work, emphasizing that the best model for someone depends on their unique workflow and thinking process rather than benchmark scores alone. They highlight their frequent use of Chad GPT 5.6 Sol, a model they consider “dumber” in some respects but highly effective for their complex knowledge work, particularly due to its ability to handle lengthy, detailed prompts and maintain persistence in delivering results. The speaker contrasts this with other models like Fable 5, which excels in handling high-level ambiguity and conceptual thinking, reflecting different strengths suited to different user needs.
The speaker explains that OpenAI’s models, including Chad GPT 5.6 Sol, focus heavily on reinforcement learning to improve task-specific performance, especially in coding and knowledge work, while Anthropic’s models like Fable 5 invest more in pre-training on large datasets to create more general-purpose, philosophically nuanced assistants. This distinction leads to different “model families” with unique characteristics: OpenAI’s models are precise and explicit, ideal for detailed agentic workflows, whereas Anthropic’s models are better at interpreting ambiguous, high-level intent and conceptual tasks.
A key insight shared is that AI models should be thought of like family members—each with distinct personalities and strengths—rather than simply ranked by benchmarks. The speaker argues that benchmark scores do not fully capture the nuanced ways models support different types of work. Instead, users should reflect on their own work habits and cognitive styles to select models that best accelerate their personal workflows. For example, those who prefer detailed, technical prompting might favor OpenAI’s models, while those who thrive on abstract reasoning might lean towards Anthropic’s offerings.
The speaker also touches on the current limitations in AI tools for knowledge workers compared to engineers, noting that many AI developments have been driven by engineering needs, resulting in tools optimized for coding rather than broader knowledge work. They express a desire for more sophisticated AI harnesses tailored to non-technical knowledge work, which involves iterative thinking and process rather than code verification. The speaker is actively developing a tool to help users articulate and share their work styles and passions, aiming to facilitate better model selection and usage.
Finally, the speaker encourages viewers to experiment with different models and find the one that feels most comfortable and effective for their hardest work. They emphasize that the AI model landscape is evolving rapidly and becoming more complex, but maintaining a clear understanding of one’s own workflow and needs can help users navigate this complexity. For those interested, the speaker offers access to detailed benchmark results and a continuously updated tool to assist in choosing the right model mix, underscoring the importance of personalized AI adoption in the ongoing model race.
