Share

In a recent Dwarkesh podcast, Andrej Karpathy argued that AI progress slows dramatically with each added “9” of reliability—99%, 99.9%, 99.99%. True for self-driving cars. But the wrong framework for law.

When he says AGI is ten years away, he means infrastructure: AIs that can consistently operate software, blend modalities, remember tasks for days, and never make unsafe mistakes. Those are engineering hurdles — not barriers to legal value.

In patent work, perfection isn’t the threshold. Augmentation is.

We don’t need agents that run error-free for days to gain many AI benefits. We need systems that produce strong work a patent attorney can verify and refine. That’s a different kind of reliability.

Every good attorney already works human-in-the-loop — reviewing claims, citations, drawings, and scope. AI fits naturally into that model: it analyzes, summarizes, drafts; the attorney applies judgment.

The threshold for value isn’t flawless autonomy. It’s faster, better starting points.
We engineer around the missing nines by showing sources, flagging uncertainty, and keeping humans in the loop at optimal points. Good AI design makes errors visible and fixable — and that’s the difference between autonomy and augmentation.

Source: LinkedIn Post by Michael Drapkin

More Insights

View all insights
View all insights

Try our tech

Can your invention go the distance?

Our Discovery Agent will work with you to explore whether your invention has what it takes, and you’ll get a detailed report to help you with the next steps. Request a demo to experience this transformational tool.

Discovery Agent