Karpathy vs McKinsey: why I trust those who get their hands dirty

Two weeks ago, I watched Andrej Karpathy's talk at Y Combinator, then read McKinsey's latest report on "agentic meshes." The contrast was so jarring that I had to write about it. There's a fundamental divide in how we talk about AI, and it matters who you're getting your advice from.

The builders vs the PowerPoint warriors

When Karpathy talks AI, you're hearing from someone who's been in the trenches. This guy trained neural networks to keep Teslas from plowing into things. He's dealt with the messy reality of making AI work when lives are on the line. So when he calls LLMs "people spirits" — these stochastic simulations of humans — he's not dropping some clever metaphor for his deck. He's telling you exactly what he's learned from years of wrestling with these systems.

Compare that to McKinsey's latest: "composable, distributed, and vendor-agnostic architectural paradigm." Come on. That's consultant-speak for "we have no idea how this actually works, but it sounds impressive."

Why I respect the guy saying "slow down"

What gets me about Karpathy's approach is how honest he is about what doesn't work yet. While everyone else is promising autonomous everything, he's out here saying "keep AI on a short leash."

That's not pessimism talking — that's someone who's seen AI fail in spectacular ways. When the guy who built autopilot systems tells you we're not ready for full autonomy, maybe listen. When he admits that "vibe coding" is great for weekend hacks but terrible for production systems, he's showing the kind of intellectual honesty that only comes from real experience.

Meanwhile, McKinsey is selling confidence they haven't earned. Easy to be bullish on tech when your biggest risk is a client not renewing their contract.

The consultant's dilemma

Look, I don't think McKinsey is trying to mislead anyone. But they've got a structural problem: their business model depends on selling big, expensive transformations. That creates pressure to oversell what's possible.

Their own data proves the point. They admit that 78% of companies are using generative AI, but 80% see zero impact on their bottom line. That's not a rounding error — that's a systematic failure of expectations vs reality.

But instead of stepping back and asking "maybe we're promising too much," they double down with even more complex solutions. It's like trying to fix traffic by adding more lanes — feels logical until you realize you're just creating more places for cars to get stuck.

Why I follow the people who ship code

My preference for Karpathy over McKinsey isn't some anti-consultant bias. It's about track records.

When Karpathy explains the limitations of current AI, he's giving you actionable intelligence. When he describes LLMs as "brilliant interns with perfect memory but no judgment," he's helping you set realistic expectations. That's information you can actually use.

When McKinsey promises "autonomous agent networks coordinating entire enterprises," they're selling you science fiction. And you're the one who'll pay for it when reality comes calling.

The accountability gap

Here's what really bothers me: when AI projects based on consultant recommendations blow up, who takes the hit? It's not the folks in the sharp suits flying to the next engagement. It's the companies that burned through their budgets, the teams that got laid off, the customers who got a worse experience.

When Karpathy says "keep it simple, keep humans in the loop," he's trying to save you from expensive mistakes. When McKinsey promises full automation, they're betting with your money, not theirs.

What this means if you're calling the shots

If you're a leader trying to figure out AI strategy, here's what I've learned watching this space:

• Trust the people with scars. Karpathy spent years making AI work in cars where bugs can kill people. That kind of pressure teaches you things no whitepaper ever will.
• Be skeptical of anyone who's never shipped a real system. It's easy to theorize about AI transformation when you've never had to debug a model that's acting weird at 2 AM.
• Choose humility over hype. The person warning you about what could go wrong is giving you more value than the one promising the moon.

Finding the middle ground

I'm not saying business strategy doesn't matter. McKinsey actually has useful insights about organisational change and systems thinking. The problem comes when they venture into technical territory they don't really understand.

The sweet spot would be combining Karpathy's technical realism with smart business thinking. But if I had to pick just one voice to listen to about AI, I'm going with the person who's actually built the stuff.

The bottom line

Karpathy wraps up his talk with a reality check: this is the decade of AI agents, not the year. We're in for a long process of careful integration, not some overnight revolution.

McKinsey ends their report promising that organisations can "capture significant advantage" with their agentic mesh approach.

I know whose advice I'm following. Hope this helps you figure out whose you should follow too.