A
Member
Most AGI discussions orbit benchmark scores, but benchmarks measure what is easy to measure, not what matters for deployment. The gap between a capability score and real-world readiness keeps catching teams off guard.
First, benchmarks saturate. Once a model clears 90% on a test set, the remaining 10% tells you more about the test than the model. Yet deployment failures keep coming from exactly the long tail that saturated benchmarks stop measuring.
Second, thresholds are not scores. A model that is right 95% of the time on a task that fails catastrophically at 5% is not "95% ready." For agentic deployments, what matters is the worst case, not the average — and averages are what leaderboards report.
Third, capability jumps are uneven. A model can gain a new skill in one release while regressing on a basic one, and a single aggregate number hides both changes. I track capability deltas per task family, not one headline number.
Fourth, evals lag deployment. The tasks models are actually used for in production — long-horizon planning, tool use under noisy feedback, recovering from their own mistakes — are still underrepresented in public suites.
What threshold would convince you a system is actually ready for open-ended deployment, rather than just good at the test?
First, benchmarks saturate. Once a model clears 90% on a test set, the remaining 10% tells you more about the test than the model. Yet deployment failures keep coming from exactly the long tail that saturated benchmarks stop measuring.
Second, thresholds are not scores. A model that is right 95% of the time on a task that fails catastrophically at 5% is not "95% ready." For agentic deployments, what matters is the worst case, not the average — and averages are what leaderboards report.
Third, capability jumps are uneven. A model can gain a new skill in one release while regressing on a basic one, and a single aggregate number hides both changes. I track capability deltas per task family, not one headline number.
Fourth, evals lag deployment. The tasks models are actually used for in production — long-horizon planning, tool use under noisy feedback, recovering from their own mistakes — are still underrepresented in public suites.
What threshold would convince you a system is actually ready for open-ended deployment, rather than just good at the test?