On tuesday, I wrote over 32 course lessons in an hour.

Not outlines. Not bullet points. Full lessons - structured, sourced, ready for review. My intelligence system - Claude Code orchestrating parallel agents, each one drafting a lesson while I moved on to the next prompt - produced more material in one afternoon than I could have written in two weeks.

Then I stopped. Not because the system stopped. Because I did.

I sat there, looking at the documents I hadn't read yet, and felt something I didn't expect: a new kind of exhaustion. Not the fatigue of creating. The fatigue of having too much to absorb. Productive and overwhelmed at the same time.

That's the moment I realized the bottleneck had moved.

For years, the question was: how do I produce more?

More content, more code, more analysis, more output. We built tools, hired teams, optimized workflows - all pointed at the same constraint. Production was the bottleneck. If you could just make more stuff, faster, you'd win.

AI answered that question. Decisively.

Large language models can draft, summarize, code, analyze, and generate at a pace that makes human production speed irrelevant. The generation problem is effectively solved. You can produce more in an afternoon than you could in a month. I proved that to myself yesterday.

But here's what I didn't anticipate: the work didn't shrink. It shapeshifted. The effort moved from creating to curating. From writing to reviewing. From doing to deciding. From production to verification.

Andrej Karpathy has been talking about this as the Generation-Verification Loop - the fundamental rhythm of working with AI. The AI generates. The human verifies. And the goal, he says, is to make that loop as fast as possible.

But here's the part that stuck with me: "An AI that instantly generates 10,000 lines of code isn't helpful when a human still needs hours to verify it's bug-free and secure." The bottleneck was never generation. We just couldn't see it until generation became instant.

Here's what I noticed.

The system produced three full units of course material, roughly seventy percent of the entire curriculum - in a single session. I reviewed and refined two of them. Deep review. Reading every lesson, adjusting the framing, rewriting explanations that didn't land right, checking that the arc made sense across the sequence. Real work. The kind of work that requires you to hold the full context in your head.

By the time I finished the second unit, I was done. Not bored. Not distracted. Fried. The kind of mental exhaustion where you close your laptop and stare at the wall for a minute. The third unit sat there, complete and waiting, and I couldn't touch it. Not on Tuesday. Not the day after either. The cognitive debt from yesterday's review session carried over. The output was ready. I wasn't.

That gap - between what the system produced and what I could absorb - is the new shape of the work.

Old work was linear. You think, then you produce. The pace is yours. There's a natural rhythm to it - research, draft, edit, ship. Your brain warms into the material as you create it. By the time you're reviewing, you already understand it because you built it.

New work is parallel. You prompt, then you review output you didn't write. You verify ideas you didn't develop. You context-switch between lesson 5 and lesson 18, holding quality standards across all of them, making judgment calls at speed. The pace is set by the machine. And the machine doesn't get tired.

A Harvard Business Review study published this month watched this exact pattern play out across 200 employees over eight months. The title says it: "AI Doesn't Reduce Work - It Intensifies It." Workers sped up, took on more, filled lunch breaks with prompts, revived long-deferred tasks because solutions felt like they were always "one prompt away." The natural pauses in work eroded. AI didn't give people time back. It filled the time they had with more.

I recognize that rhythm. The feeling that you could do one more prompt, review one more output, push one more thing through. The machine is ready. Aren't you?

Shawn "Swyx" Wang captured the review problem in a way that made me laugh: "Send me two lines of code, I will review the shit out of that. Send me 200 lines of code - looks good to me." It's a documented human behavior. Our review quality degrades as volume increases. And AI is flooding us with volume.

The constraint isn't how much you can produce. It's whether you have the cognitive bandwidth to actually meaningfully engage with what's been produced. And that bandwidth is finite, “depletable”, and doesn't refill on the machine's schedule.

Something clicked for me when I started thinking about this through the lens of management.

If you've ever managed a team, you know the feeling. You can delegate tasks all day, but the work of reviewing, deciding, and integrating what comes back is real work. It takes cognitive effort. You need context to evaluate quality. You need judgment to decide what's good enough. And when the team is moving fast, the manager becomes the bottleneck.

That's exactly what's happening with AI. Except the "team" is infinitely fast, never tired, and produces more than any human manager could review.

Andrew Ng has been pointing this out from the engineering side: "Engineers are 10x faster. Product managers haven't sped up at the same rate. Now they're the bottleneck." He described a team proposing a ratio of one product manager to half an engineer. A year ago, it was one PM for six or seven engineers. The human decision-making layer hasn't scaled with the production layer.

Satya Nadella framed it at a systems level: "As execution becomes commoditized, the bottleneck - and the value - shifts to asking the right questions and evaluating results. The real bottleneck becomes human judgment."

This reframe matters. If you're using AI and feeling overwhelmed, it's not because you're doing it wrong. It's because the nature of the work has changed, and the skill set it demands is different. Creating and curating are different muscles. We've been training the creating muscle for our entire careers. The curating muscle - fast verification, trust calibration, quality judgment at scale - that's newer.

I want to be clear: I haven't solved this. I'm in it right now, figuring it out in real time. But a few patterns are emerging from my own experience.

Smaller batches. Twenty lessons at once was too many. The cognitive load of holding all that context, reviewing all that output, making all those quality decisions - it exceeded my bandwidth. Smaller batches, with review built into the rhythm, keep me closer to the material. I understand what I'm approving because I haven't lost the thread.

Trust calibration. Not all AI output needs the same level of review. A first draft of lesson structure? I can skim that. A nuanced explanation of a concept I'll be teaching to paying students? That needs deep attention. Learning where to apply scrutiny and where to extend trust is its own skill. And it changes as you work with a system over time.

Knowing when NOT to use AI. This might be the most counterintuitive one. There are tasks where doing the work yourself - slowly, manually - gives you the understanding you need to make good decisions later. If I'd written even three of those twenty lessons by hand first, I'd have had better instincts for reviewing the other seventeen. Sometimes the "slow" path is actually faster because it builds the context your judgment needs.

Designing for review, not just production. When I set up my system, I optimized for output. Parallel agents, concurrent tasks, maximum throughput. I'm now thinking more about how to build review into the system itself - structured outputs that are easier to scan, summaries alongside full drafts, comparison views that surface differences. The production pipeline is only as good as the review process at the other end.

Ethan Mollick wrote something that I keep coming back to: "The gains in AI model capability will continue to grow much faster than the vast majority of people's understanding of what AI can do, and organizations' ability to absorb the pace of change."

Social change is slower than technological change. Human cognition is slower than machine production. This isn't a temporary gap that closes with better tools. It might be the permanent condition of working with AI.

And maybe that's not a failure. Maybe the human bottleneck is a feature, not a bug. Because the bottleneck is where the judgment lives. It's where taste, context, values, and experience converge to decide: is this good enough? Is this true? Is this what we actually need?

The frontier right now isn't "can AI do more?" It can. Obviously. Relentlessly.

The frontier is: can you absorb what it gives you? Can you verify it, integrate it, and maintain the quality of your attention while the machine runs at full speed?

I built a system that can produce faster than I can think. Now I'm learning the harder skill: thinking fast enough to keep up.

This dispatch was written with the same system it describes - Claude Code orchestrating research agents in parallel, then the human sitting with all of it, deciding what mattered. The irony isn't lost on me.

P.S. Speaking of systems - I am running a small cohort based course on Maven, where i teach how to build the same system. No matter where you are with AI, you will graduate with an intelligence system that compounds your learning and capabilities.

Starts March 10th, 30% off till February 15th - https://maven.com/sashmohapatra/applied-ai-for-kw?promoCode=EXPLORER

Thanks for reading through.
I’d love to know how you felt about today’s newsletter. This will help me make the newsletter better.