2025 LLM Year in Review by Karpathy (AI column)

Interesting. revue 2025 A year from Andrey Karpatha, where he talks about paradigm shifts that changed the landscape of the LLM.

1️⃣ Reinforcement Learning from Verifiable Rewards (RLVR). To the classic stack (pretrain → SFT → RLHF) A large stage of reinforcement learning for “testable” awards was added (maths). Longer RL runs → better capability / $, plus the lever test-time compute ("let the model think longer"). 2️⃣ Ghosts vs. Animals / Jagged Intelligence. LLMs are ghosts rather than animals. In one place genius, in another - easily breaks / jailbreaks. Hence the skepticism of benchmarks: they are all easier to "train to death." 3️⃣ Cursor / new layer of LLM apps. Cursor can be thought of as an LLM-app template: separate product layer = context engineering + call orchestration (DAG) UI for human-in-the-loop + . autonomy slider. This is where the phrase “Cursor for X” comes from. 4️⃣ Claude Code / AI that lives on your computer. The agent must live in your car near the real environment. (Configures, secrets, data, low latency)Not in abstract cloud containers. 5️⃣ Vibe coding. It is interesting to hear the thesis from the creator of the term vibe coding - in his opinion, the threshold has been passed and now you can collect working things using the most common programming language. (||English||). The code becomes cheap, disposable and plastic. (just for one run). 6️⃣ Nano banana / LLM GUI. A chat is a console. 80-x. UX - visual artifacts (slides, schemes, whiteboard, web-apps). Google’s “Nano banana” is an early hint of direction.

In general, I recommend reading these theses in the original, since it is very short.

#AI #ML #Trends #Software #Engineering #Future