Capability overhang and the challenge of using AI effectively
Everyone’s heard of the term “capability overhang”, introduced by Jack Clark in 2023, now understood as the growing gap between what artificial intelligence systems can do technically and how people or businesses actually use them in daily practice.
Because AI models are so broadly useful, it’s difficult for an individual to even understand what they’re capable of, let alone use them to their full capacity. Helping people use AI well will be one of the most important ways for the accelerating improvements in models to benefit everyone.
There are many methods for guiding or supporting AI systems, like those for:
- giving models more knowledge: skills, MCP tools, RAG, fine tuning or LoRA, context caching, long context and memory systems,
- giving models tools and agency: function calling and tool use, MCP, skills, computer use, and code execution sandboxes,
- orchestration layers: agent frameworks, multi-agent systems, guardrails and evals,
- prompt augmentation: system prompts, constitutions, and general prompt engineering, plus structured outputs for more reliable downstream use.
What’s missing: helping people, not models
In every case, the technologies that surround AI models are systems designed to address different gaps in those models: giving them knowledge, capabilities, skills and agency. They augment the models, not their users.
We believe that capability overhang is evidence of the need for systems that help people to use AI much more effectively than they have been able to.
Introducing OpenTacit
OpenTacit is designed to help people use AI better.
Today, we’re introducing OpenTacit. It is a new open-source project for individuals or organizations to use AI better.
OpenTacit doesn’t require any change to the tools or methods that you use with AI. Instead, it observes your use (privately), identifies and measures what is working well, captures that as re-usable techniques, then shares them at the very moment they will be most useful, right in your existing tools.
It is particularly valuable for teams or organizations, because it helps share knowledge from the most effective users with people who would benefit the most from it.
OpenTacit for individuals
By observing and measuring what works best, OpenTacit gives you deep insight into how you use AI. It can help you understand your patterns, habits, successes and challenges, and track that over time to explain why your AI tools may seem to be more or less helpful. By aggregating this across different models, agents, platforms and systems, you gain a comprehensive view of what works well, and what does not.
OpenTacit also measures the helpfulness of the techniques you employ, identifies when you are not using the best techniques, and can prompt you unobtrusively with what will improve your outcomes.
For an individual, OpenTacit can eliminate the need to manage skills, frameworks and custom prompts manually, so that you can focus on your goals, and simply use AI better.
OpenTacit for teams and organizations
OpenTacit becomes even more useful when shared among colleagues in a team or organization. The common goals of a group, and the proprietary systems, data, processes and platforms accessible to it that were unknown to a model during its training are exactly the types of information that OpenTacit helps people bring to models. This makes a team capable of doing much more with AI than it could otherwise.
The techniques that OpenTacit captures and measures can be shared among colleagues, so that the discoveries and improvements from the best users become available to everyone immediately and automatically.
The decisive AI advantage for organizations won’t be the model; everyone has the same models. It’s what your organization has learned about putting them to work with your proprietary information, systems, people and processes: tacit knowledge that today is unwritten, unmeasured, and at risk of loss.
Private by design
OpenTacit measures what works, never who did what. Outcomes are stored by cohort, such as team, role, or harness, and never linked to individuals.
Your personal record of which suggestions you adopted or dismissed stays in a file on your own machine and nowhere else. Reports about “who” are always about cohorts, never about individuals.
Privacy is fundamental to OpenTacit’s open-source design, so that an organization and its members can trust it.
Adopting it is fast and simple
OpenTacit is a single open-source binary that provides all functionality: registry, dashboard, and session integration. There is no platform to procure, no agents to babysit, and no vendor to negotiate with.
Where to go next
OpenTacit turns the things your colleagues have figured out about working with AI into knowledge you can use, right at the moment you need it, with proof that it works.
Get started at opentacit.com.