#31 Everyone's Reading Karpathy's 2025 AI Review. Here's Your Plain-language Translation

A restaurant metaphor that explains RLVR, reasoning models, and AI's biggest year before you tackle the technical original.

Dec 23, 2025

Hey, welcome back! I’m Laura. Just a regular human trying to make sense of the AI era without losing my mind.

This week, we’re unpacking Andrej Karpathy’s 2025 AI Year in Review. It’s essential reading if you want to stay ahead as an AI-savvy marketer, but let’s be honest, it’s a little bit dense.

Why this review matters for marketers

We don’t need to become engineers. But we do need to understand what’s actually changing under the hood, because those shifts directly affect what we can build, promise, and deliver to customers.

If you’re new to Karpathy, here’s the short version: he’s one of the most respected voices in AI. Former Director of AI at Tesla, founding member of OpenAI, and someone who can explain deeply technical ideas with rare clarity. When he writes about where AI is heading, the industry listens.

His latest review is both highly relevant and technically dense. It covers concepts like RLVR, reasoning traces, and inference-time compute. All crucial, but not exactly light reading for most marketers.

So here’s what I’m going to do.

I’ll tell the same story Karpathy tells, but through a metaphor that actually sticks. When you’re done, his original post will feel clearer and more meaningful. You’ll understand why 2025 was pivotal and how to explain these ideas to your team, your clients, or that exec who keeps asking, “But what does this actually mean for us?”

Think of this as your primer. Then go read Karpathy.

Let’s go.

The restaurant metaphor: training ghost-chefs

Picture a gigantic restaurant.

Not a cozy neighborhood place with twelve tables. A sprawling kitchen-city serving millions of customers at once. Breakfast orders, lunch rushes, late-night meals, allergy lists, vegan substitutions, and someone insisting on pancakes at 2 a.m.

Your job is to train the newest chef.

Except it’s not really a chef. It’s a ghost-chef.

No taste buds. No real understanding of food. But it can read every cookbook ever written, overhear a million dinner conversations, and perfectly imitate how humans talk about flavor.

Act 1: The old training recipe (and why everyone hit a wall)

At the start of the year, every top kitchen trained its ghost-chef the same way.

Pretraining meant reading everything. Cookbooks, food blogs, menus, Reddit arguments about carbonara, grandma’s handwritten recipe cards, Michelin plating guides. All of it.

This gave the ghost fluency. It knew how people talk about food.

Supervised fine-tuning came next. You showed examples of how to respond to customers. Be clear. Be polite. Structure your answers. Don’t be weird.

Better manners, sure. Still not great at thinking.

Then came RLHF, Reinforcement Learning from Human Feedback. Human taste-testers sampled outputs and voted. “That was helpful.” “That was strange.”

And this worked. For a while.

Then every kitchen hit the same wall.

Reasoning is slippery. You can’t just tell a ghost-chef to “think better.” Does that mean step-by-step reasoning? Trying multiple approaches? Checking mistakes and backtracking?

Even humans disagree on this. So how do you train it?

Act 2: RLVR, or when everything changed

In 2025, a new method took over: Reinforcement Learning from Verifiable Rewards, or RLVR.

This is the breakthrough Karpathy focuses on, and it explains most of the progress we saw this year.

Instead of asking humans, “Did this taste good?”, you give the ghost-chef tests that score themselves.

Did the soufflé rise to exactly the right height? ✅ or ❌
Does this code actually compile? ✅ or ❌
Does the math check out? ✅ or ❌

No “almost. or “pretty close,” Just right or wrong.

The surprising part

Run enough of these drills and something strange happens. The ghost-chef starts inventing behaviors that look like reasoning.

It breaks problems into sub-steps.
Checks its work.
Backtracks and tries again.
Do small calculations before committing.

You didn’t teach it how to think. You taught it how to win, and thinking emerged as a strategy.

Why this changed everything

Because these tests are automatic, the kitchen can run far more reps. No human bottlenecks. Just vastly more practice.

Then came another realization. You could let the ghost think longer during service. Pause. Try multiple approaches. Check its work before serving.

For marketers, this matters. The AI tools you’ll use in early 2026 are fundamentally different from those in early 2025. Not because the models are bigger, but because they’ve been trained to reason instead of just pattern-match.

That’s why they feel so much more capable.

Act 3: Animals vs. ghosts, and why this feels uncanny

People often describe AI using animal metaphors. It’s growing. It’s evolving. It learns like a child.

That metaphor breaks quickly.

Humans are trained by life. Survival, social consequences, physical reality, embarrassment, risk.

Ghost-chefs are trained by text imitation, puzzle rewards, approval clicks, and leaderboards.

So they develop jagged intelligence.

A ghost-chef might plate a flawless seven-course tasting menu, then confidently serve soup in a wine glass because the question was phrased oddly.

It’s not broken. It’s optimized.

Sharp spikes where rewards exist. Flat spots everywhere else.

Like someone brilliant at their job who still doesn’t know how to cook a fried egg. (You know someone, like me… 😅)

The benchmark problem

Benchmarks are meant to be neutral tests, like standardized cooking competitions. But when RLVR (Reinforcement Learning from Verifiable Rewards) trains on anything that can be auto-scored, benchmarks quietly become training manuals.

No one is cheating. The rules simply define what gets practiced.

You can dominate competitions and still collapse during real dinner service when everything goes sideways.

The 2025 shift was simple: benchmarks still matter, but they’re no longer enough.

Marketing translation: treat benchmark scores with caution. What matters is how a tool performs on your use cases, not how it scored on a standardized test.

Act 4: Cursor and the rise of the kitchen manager

Now the story shifts from better chefs to better restaurants.

Cursor isn’t just a chef. It’s a kitchen manager.

It gathers context, breaks work into tasks, calls the ghost-chef multiple times, balances cost and quality, and gives humans a clean workstation to supervise from. It even includes an autonomy slider, from “just suggest” to “run the whole kitchen.”

That’s why people started saying “Cursor for X.”

Not one prompt. A full workbench.

Think of it this way. LLM Labs graduate capable culinary students. Apps like Cursor turn them into working professionals.

For marketers, this signals the next wave of tools. Not “ChatGPT for marketing,” but full workflow assistants that understand context, remember preferences, and execute multi-step projects with minimal hand-holding.

Act 5: Claude Code, the ghost that moves in

Then came something stranger.

Not a ghost you visit on a website. A ghost that lives with you.

Claude Code works inside your environment. It opens your files, reads your codebase, uses your tools, and responds instantly.

That matters because real work isn’t abstract intelligence. Its context, setup, private data, and fast feedback loops.

It feels less like searching Google and more like working alongside a colleague.

A little creepy. Extremely useful.

Marketing angle: This is where AI shifts from a tool you use to a collaborator you work with. That changes how we delegate, collaborate, and structure work.

Act 6: Vibe coding and losing the knife skills barrier

In 2025, people started building real software by describing what they wanted.

“Make me a menu planner that uses my pantry, calculates macros, generates a shopping list, and looks good.”

And it happens.

No knife skills. No six-month detour to learn a new language.

That’s vibe coding.

For professionals, code becomes cheap, temporary, and disposable. Like cooking a quick snack just to test the oven.

For marketers, this is huge. You can prototype landing pages, dashboards, and interactive campaigns without waiting on a dev queue. The gap between idea and execution collapsed.

Act 7: Nano banana and the GUI moment

One last shift.

Text works well for computers. Not always for humans.

Chatting with LLMs today feels like using a 1980s terminal. Functional, but clunky.

People prefer visuals. Diagrams. Slides. Whiteboards. Interactive tools.

“Nano banana” hints at the GUI era of LLMs (Graphical User Interface). Not just images, but text, visuals, and world knowledge fused.

Like a chef who can explain a recipe, sketch the plating, and understand what you meant by “make it feel cozy.”

Not perfect. But clearly a new interface.

Why this matters: the next generation of AI tools won’t be chat boxes. They’ll be multimodal, visual, and spatial, closer to creative partners than search engines.

The punchline: what 2025 actually taught us

2025 showed us ghost-chefs that are smarter than expected and dumber than expected. Incredibly useful. Still far from their ceiling.

The next breakthroughs won’t be about making the chef smarter.

They’ll be about:

better kitchen managers
better co-living assistants
more people who can suddenly cook
better ways for ghosts to communicate with us

The real question was never whether the ghost understands food.

It’s whether we’re learning to build kitchens that make its weird, jagged talents useful.

Now go read Karpathy

Go read the original article. Seriously.

With this framework in mind, his technical explanations will click. When he talks about reasoning traces, inference-time compute, process supervision, or test-time training, you’ll recognize the ghost-chef patterns.

As marketers, we don’t need the math. We need the concepts. We need to know what’s changing, what’s possible, and where the opportunities are.

2025 wasn’t incremental. It was the year training methodology shifted, vibe coding became real, and AI tools stopped being novelties and started becoming infrastructure.

And 2026 will move faster. Strap in.

Your Turn

If this helped the ideas click, or if you think I oversimplified, missed something important, or got a detail wrong, tell me.
I learn faster in public, especially when people disagree thoughtfully.

Discussion about this post

Ready for more?