AI's Next Frontier: Capturing Human Process, Not Just Output
Models have ingested nearly everything we've ever produced – but how we produced it remains largely undocumented
Today's leading AI thinkers agree on a compelling possibility for unlocking major advances in foundation models and LLM capabilities: capturing the largely invisible processes behind human thought and creation. While the internet contains vast amounts of information equivalent to countless Libraries of Alexandria, almost all of this data represents end products rather than the underlying processes that created them.
Consider anything you've ever produced – whether a tweet, an email, a photograph, a symphony, or an event. The final output represents only a fraction of the actual work: the thinking, planning, consideration, reconsideration, comparison, and self-editing that went into its creation. Much of this crucial process occurs either below conscious awareness or never makes it into any documented form.
Current generative AI achievements have largely relied on reverse engineering human-quality outputs by analyzing massive collections of finished products. However, this approach misses the rich landscape of how humans actually arrive at these conclusions.
Learning from Past Approaches
The next evolution in AI training might draw inspiration from an unexpected source: Robotic Process Automation (RPA). RPA providers like UiPath pioneered a more basic form of task automation by recording users completing simple computer interface tasks and then replaying those interactions. While rudimentary, this approach highlighted the value of capturing step-by-step human processes.
At the Cerebral Valley AI Summit this week put on by Eric Newcomer, both Scale founder Alexander Wang and Anthropic CEO Dario Amodei emphasized this gap in AI training data. Wang pointed to the lack of detailed, step-by-step documentation of even routine tasks like booking flights – a frequently cited use case where AI agents still consistently struggle. Similarly, Cohere's Aidan Gomez, speaking on the No Priors podcast, discussed the potential of building reasoning models that break down tasks into discrete steps, solving each challenge individually to tackle larger problems.
Three Paths Forward
There are three potential approaches to training models on human processes rather than just their end products:
Manual Process Documentation
Systematically recording human task completion steps
Likely implemented through low-cost, developing market labor participation (mechanical turk style)
Particularly valuable for Scale given its business model
Simulation-Based Learning
Generating high-quality, cost-effective process simulations using a small subset of actual recorded ones for ground truth
Creating variant scenarios for comprehensive training
Combining with real-world data for robust learning
Direct Neural Interface
Brain-computer interfaces like Neuralink aim to capture thought processes directly
Could potentially eliminate the ‘lossiness’ of keyboard/mouse-based input
Offers the possibility of precise recording and transcription
The gap between human thought and action contains vast amounts of untapped information. While generative AI has made remarkable progress working backward from end results, the key to achieving reliable, human-level or superior performance across domains may lie in understanding and incorporating the complete process of human reasoning and decision-making.