It’s striking to see Dario Amodei and Sam Altman publicly agreeing on something: AI may be moving too fast and we need to slow it down.
Slowing down may be necessary, but is it sufficient? If AI is moving faster than humans and organizations can manage, slowing it down buys us time. It doesn’t solve the underlying problem.
How do humans get ahead of AI? Shouldn’t that be the bigger question?
In my last blog, I wrote about two-way interoperability between humans and AI agents. Humans need to participate in the reasoning—not simply validate an answer after the AI has finished reasoning. I want to advance that thesis here.
Slowing AI down doesn’t get humans ahead
AI models and autonomous agents are producing more reasoning, analysis, decisions, and output than humans can possibly consume. Slowing that output down doesn’t fundamentally change the relationship.
The AI still reasons by itself, with context from humans. The AI still creates the output. The AI still determines what the human sees. The human still reacts.
We need to change who is leading whom.
One way to do that is for humans to participate in the reasoning as peers to AI—not simply as reviewers at the end. But humans can’t do this alone because the volume of AI reasoning and output is already too large to manage manually. You cannot solve that by putting more approval buttons in front of people.
Having worked on so much breakthrough software in the past, I may be biased here, but software has to help.
AI reasoning needs to be organized and structured into understandable steps and phases. Software tools need to help humans identify what matters, contribute their expertise, modify the reasoning, and send that reasoning back to AI.
Reasoning should move back and forth between all three: AI, software, and humans.
Most enterprise software was built for humans operating software—not humans collaborating with autonomous AI systems that can generate enormous amounts of reasoning and work.
Satya Nadella has been a leader in pushing the industry toward a different model with his Learning Loop: AI, humans, and software systems iterating together rather than AI simply producing output for humans.
Rather than going over the cliff a little later, we should put humans and organizations back in a position to decide where we’re going.