Karpathy’s 2025 retrospective is the clearest articulation I’ve seen of what foundational AI labs are actually building.
We’re not “evolving animals,” we’re “summoning ghosts.”
LLMs have completely different optimization pressures than biological intelligence. Humans evolved for tribal survival. LLMs optimize for imitating text, solving puzzles, and winning upvotes on LM Arena. Different pressures, different shapes in the intelligence space.
This framing finally explains what confuses everyone about AI capability.
GPT-5 aces the bar exam but gets tricked by simple jailbreaks. Claude writes PhD-level philosophy but hallucinates citations. Gemini solves competition math that stumps IMO medalists but fumbles spatial reasoning.
Capability spikes near verifiable domains where RLVR concentrates optimization pressure. Everywhere else, you get a different entity entirely.
I’ve been thinking about what this means for AI product builders.
The teams struggling with AI deployment are treating capability as uniform. They ask “can AI do this task?” and expect a yes/no answer. But ghost intelligence doesn’t work that way.
The teams winning are asking a different question: “Does this task live near a verifiable domain?”
If yes, the ghost might be superhuman. Build for autonomy. If no, the ghost needs guardrails. Build for human-in-the-loop.
This is why Cursor works. This is why Claude Code runs on localhost instead of the cloud. The best AI products in 2025 mapped the jagged edges and designed around them.
The companies that internalize Karpathy’s ghost framing will build better products than the ones still thinking in terms of “smarter or dumber than humans.”
There’s no single axis. Just different shapes.
HOW TO WIN IN THE AI AGE
You are competing against people who copy what exists and let AI remake it.
Most design, writing and product work comes from AI copying existing references in generic mode, which is why the internet feels like an endless conveyor belt of bland.
The moment you push past that baseline and force the model to do something uncomfortable, you separate yourself from the entire curve. Giving it original ideas + references outside your niche.
The people who win are the ones who insist on originality in a world that keeps asking for “good enough.”
The reason this matters is because default effort produces default outcomes, and default outcomes disappear instantly in a feed that never stops moving.
I bet these default startups have more and more friction standing out going further as AI makes startup making more like a machine.
People can FEEL the difference between something generated and something directed, between something assembled and something authored. AI will happily produce the average case forever, but it will never choose to take the creative risk that changes how something feels. And that's where the alpha is!!!
Anyone can generate. Almost no one curates, edits, reworks, reshapes, interrogates, or pushes until the thing becomes unmistakably theirs.
The edge is in the taste you bring to the model, the choices you make that the average user will never even see, and the willingness to throw away nine versions to get to the tenth that feels alive.
In a world where everyone has access to the same tools, the only real advantage is the person who refuses to behave like a default setting.
@boltdotnew the new update looks great for the UI however because I’m someone that uses discussion mode heavily to design my apps, it’s not that efficient to keep clicking the plus sign and doing it repeatedly after each prompt entry. Having it stay on would be great!
The Journey of an Ai Business Trainer! At AI Consultant Pros we’re developing a system to train solopreneurs and business owners to become AI Power Users.