October 27, 2025 · JustSayAI

RL Is Broken, AGI Is a Decade Away: Karpathy’s Cold Shower for the Industry

90 % of the AI agents we rave about are just pretending to be useful.

Andrej Karpathy’s new two-hour interview is so dense it maxes out the context window—and that’s the point. He dumps ice water on the sector: AGI is at least ten years away.

His yardstick isn’t theory; it’s the decade-old self-driving promise. Waymo looked “two years out” in 2014. Today, true driverless still isn’t here.

The industry’s biggest hallucination? Mistaking a slick demo for a shipping product.

99.9 % despair: the movie you’ll never finish downloading

Karpathy frames it as the “problem of nines.” 99 % reliability and 99.99999 % live in different universes. To replace a human you need nine nines—99.9999999 % uptime. A model that’s only 99 % is the torrent stuck at 99.9 %: looks almost done, unwatchable forever.

He’d rather unplug models from the web. LLMs spit out the highest-probability answer and have no clue what they don’t know. One high-confidence garbage source and you’re toast.

Eating garbage: today’s AI is brute-force stupid

Two counter-intuitive truths:

First, garbage in, garbage model. We cheer trillion-parameter monsters, but Karpathy says they’re bloated because they’ve been force-fed the open web’s sewage. A truly smart model might only need 10 B parameters.

Second, the straw. Reinforcement learning isn’t a miracle—it’s “actually terrible.” We use it because everything else is worse. Picture sipping the Pacific through a cocktail straw: RL tries to suck a supervision signal through a bandwidth-starved pipe.

That’s why today’s coding assistants are useless to him. His code is “over-abstract and concise”; the tools regurgitate StackOverflow boilerplate and clutter it.

Intelligence is forgetting, not memorizing

We’ve defined smart backwards. Memorizing the internet isn’t genius—it’s a party trick. Human minds compress and forget; that compression is the algorithm. You don’t recall every grade-school lunch, just the taste of your first crush. The details vanish, distilled pattern remains.

Karpathy praises DeepSeek OCR for the same reason: it compresses 1,000 text tokens into a handful of image tokens while keeping 97 % accuracy. Maybe future models train on pixels, not prose.

We’re summoning ghosts, not breeding animals

An “animal” would evolve on its own; a “ghost” is a mirror of us. Right now we’re scraping the internet, jamming it into transformers, and chanting emerge. The hope is that while reflecting humanity, the ghost eventually shows us something we never knew we knew.

Irony: a system engineered to output the most likely answer can’t flag when it’s clueless. Maybe Karpathy’s right—pull the Ethernet cable and start over.

Real intelligence is compression and amnesia.

【Subscribe to AI Daily Brief】

🌈 New here? Grab the AI Daily Brief – two concise emails a day, listen with one click.