Descript's AI audio tool worsened after they trained it on extremely bad audio (e.g., vacuum cleaners). They learned the model that best fixes terrible audio is different from the one that best improves merely "okay" audio—the more common user scenario. You must train for your primary user's reality, not the worst possible edge case.
Related Insights
Instead of waiting for AI models to be perfect, design your application from the start to allow for human correction. This pragmatic approach acknowledges AI's inherent uncertainty and allows you to deliver value sooner by leveraging human oversight to handle edge cases.
Users mistakenly evaluate AI tools based on the quality of the first output. However, since 90% of the work is iterative, the superior tool is the one that handles a high volume of refinement prompts most effectively, not the one with the best initial result.
A 'GenAI solves everything' mindset is flawed. High-latency models are unsuitable for real-time operational needs, like optimizing a warehouse worker's scanning path, which requires millisecond responses. The key is to apply the right tool—be it an optimizer, machine learning, or GenAI—to the specific business problem.
Features designed for delight, like AI summaries, can become deeply upsetting in sensitive situations such as breakups or grief. Product teams must rigorously test for these emotional corner cases to avoid causing significant user harm and brand damage, as seen with Apple and WhatsApp.
The effectiveness of an AI system isn't solely dependent on the model's sophistication. It's a collaboration between high-quality training data, the model itself, and the contextual understanding of how to apply both to solve a real-world problem. Neglecting data or context leads to poor outcomes.
Developers often test AI systems with well-formed, correctly spelled questions. However, real users submit vague, typo-ridden, and ambiguous prompts. Directly analyzing these raw logs is the most crucial first step to understanding how your product fails in the real world and where to focus quality improvements.
A critical learning at LinkedIn was that pointing an AI at an entire company drive for context results in poor performance and hallucinations. The team had to manually curate "golden examples" and specific knowledge bases to train agents effectively, as the AI couldn't discern quality on its own.
The most fundamental challenge in AI today is not scale or architecture, but the fact that models generalize dramatically worse than humans. Solving this sample efficiency and robustness problem is the true key to unlocking the next level of AI capabilities and real-world impact.
Fine-tuning an AI model is most effective when you use high-signal data. The best source for this is the set of difficult examples where your system consistently fails. The processes of error analysis and evaluation naturally curate this valuable dataset, making fine-tuning a logical and powerful next step after prompt engineering.
The central challenge for current AI is not merely sample efficiency but a more profound failure to generalize. Models generalize 'dramatically worse than people,' which is the root cause of their brittleness, inability to learn from nuanced instruction, and unreliability compared to human intelligence. Solving this is the key to the next paradigm.
