AGI Is Late, Bring Snacks — What Andrej Karpathy Actually Said, and Why It Matters More Than Musk’s Duel

There’s a particular sound the tech world makes when it’s over-caffeinated and under-delivered. It’s the hiss of yet another “AGI next spring” prediction escaping into the atmosphere like steam from a cracked kettle.

Last week, Andrej Karpathy sat down for a long, chewy conversation with Dwarkesh Patel and, in one sitting, punctured half the hot air balloons currently floating over Silicon Valley. If you build models, buy GPUs, or write those giddy slide decks promising to fire the entire back office — watch the full interview here: youtu.be/lXUZvyajciY.

Let’s walk through the highlights, stitch them together, and leave with one useful idea: the near future belongs to small skills, not big theatrics.

The Unsexy Headline: AGI Is Decades Away

Yes, decades. Not “two quarters and a miracle.” Decades. Karpathy’s reasoning is boring in the best way:

• Agents aren’t ready. They still fumble basic computer use, hallucinate URLs, and treat “Are you a robot?” as a philosophy question.
• Multimodality is patchy. Real-world tasks need text, vision, audio, and control — together. We’re not there.
• Memory is mush. Long-term continuity — the ability to carry context across days and projects — remains a science fair project taped together with “notes” features.
• Continual learning is fragile. We can fine-tune, sure. But the kind of on-the-job learning you and I do — incremental, robust, safe — still fights with catastrophic forgetting and data hygiene.

You can wave your hands and call today’s agents “junior analysts.” They’re more like enthusiastic interns who keep asking where the stapler is.

Reinforcement Learning: Signal Through a Soda Straw

Karpathy’s verdict on RL is a wonderfully impolite sigh. Yes, reinforcement learning works — but it often works through a straw: millions of actions and a single scalar reward dribbled in at the end. The useful credit gets smeared across a long action chain; noise swamps learning; everyone celebrates a chart that twitches north and calls it science.

That doesn’t mean “no RL.” It means better scaffolding: dense intermediate rewards, curriculum design, and old-fashioned supervised fine-tuning where it fits. If you want a slogan: less hero worship for policy gradients, more craft in how we shape the learning signal.

“Skynet” Won’t Arrive. The Black Box Will.

The apocalypse isn’t a red-eyed AI kicking the door in. It’s computational delegation creeping across your life until nobody remembers how any of it works. Optimizers optimize, dashboards decorate, and humans stop deciding. That’s the drift Karpathy worries about: loss of situational awareness.

Translate that into product terms: make explanations and override levers first-class citizens. Every automated workflow should be pause-able, inspect-able, and boringly auditable. If your agent can book flights, it should also print a receipt of why — data sources, assumptions, and the thresholds that tipped a choice. “Because vibes” is not a reason.

Meanwhile, in the Hype Corner: Musk’s AI Coding Duel

Elon Musk, never one to resist a stadium, floated a duel: Karpathy vs. Grok-5 — think Deep Blue vs. Kasparov, but with code. Karpathy, being sane, joked he’d rather team up with Grok-5 than posture against it.

Get ABV — Applied AI Reviews’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

For dessert, we get the claim that Grok-5 could be AGI and will learn “almost instantly” via some dynamic RL pixie dust. Lovely. Wake me when the unit tests pass and the model can use a spreadsheet without deleting the entire quarter.

The Useful Bit: Teaching a Bee-Brain to Count

Now for the part builders should clip and tape above their desks. Karpathy showed how to add a new skill to a tiny LLM — his nanochat d32, which he cheerfully compares to a bee’s brain. The task: “How many ‘r’s in ‘strawberry’?”

The method is gloriously practical:

1. Synthesize a dataset of prompt–answer pairs (“How many r’s in X?”), with variations in casing, spacing, phrasing.
2. Fine-tune via SFT (supervised fine-tuning) or RL to reinforce success.
3. Bake in step-by-step reasoning so the model doesn’t jump to an answer but counts: one, two.
4. Wire a Python tool so the model can offload arithmetic when logic alone stutters.

Result: even a toy model can pick up a new skill when the examples are clean and the scaffolding is deliberate. You can read his walkthrough here:

“How to add a new skill to nanochat” → github.com/karpathy/nanochat/discussions/164

Two takeaways the size of billboards:

• Tokenization matters. Where you put the space matters. Whether you tokenize “strawberry” as sub-words matters. Pretend this is beneath you and your evals will punish the arrogance.
• Break the task down. “Count r’s” is trivial to humans and non-trivial to a 30-million-parameter network until you teach it a procedure. Procedural thinking is the ladder over the wall.

Agents Won’t Replace You. Good Prompts and Small Wins Might.

If a bee-brain can learn to count letters, your mid-sized production model can learn to reconcile invoices, tag support tickets, or triage IT requests — if you treat skills as shippable features. That means:

• Micro-skills, not nebulous superpowers. “Be my researcher” is a vibe. “Given two PDFs, extract dates, amounts, and payees into CSV” is a SKU.
• Synthetic data with teeth. Cover corners, adversarial cases, and dumb errors humans actually make.
• Tool use that’s deterministic. If the agent controls a browser or a spreadsheet, record what it did and why. Replays aren’t a luxury; they’re your warranty.
• Memory with governance. Persistence helps — until it leaks. Give users obvious controls, retention windows, and a kill switch that actually kills.

Will this get you AGI? No. It will get you software that works — which, last I checked, is still legal.

So, Where Does This Leave the “AGI Tomorrow” Crowd?

With a choice: keep promising or start shipping. The interview’s through-line is restraint, and it’s refreshing. Build the unglamorous things — the credit assignment signals, the long-term memory that doesn’t mutate your model, the multimodal stacks that can read a screenshot and a PDF and an API doc in the same breath. When the smoke clears, the winners won’t be the ones who shouted “AGI” the loudest; they’ll be the ones whose systems didn’t fall over when asked to book a flight, attach a receipt, and email accounting.

Final Word

The next few years won’t be an elegant sprint to omniscience. They’ll be a workshop: soldering better rewards, bolting on safer memory, and teaching models crisp little skills until the pile of small wins looks suspiciously like competence. If that sounds slow, good. Slow is how you build things that keep working after the demo ends.