- As various “year-in-review” posts roll out recently, Andrej Karpathy, OpenAI’s former co-founder, has also delivered his annual summary on large language models.
Earlier this year, his talk at YC went massively viral online, where he introduced several new ideas:
- Software 3.0 has arrived: from humans writing code (1.0), to feeding data to train models (2.0), we’re now entering the 3.0 era where we directly “cast spells” on models via prompts.
- LLMs are the new operating system: not a basic commodity like running water, but a complex OS that schedules memory (context window) and CPU (inference compute).
- The decade of agents: don’t expect AI agents to mature in one year. To move reliability from 99% to 99.999%, we likely need a decade.
In today’s “2025 Year in Review,” Karpathy again dissects what kind of “brain” AI has grown this year.
Below is a carefully translated version of Karpathy’s annual review. APPSO also provides more accessible interpretations without changing the original meaning.
If you want to read the original, click: https://karpathy.bearblog.dev/year-in-review-2025/
2025 was a year of strong growth and major volatility for LLMs. Below are several “paradigm shifts” that I personally found noteworthy and somewhat surprising—changes that not only reshaped the industry landscape, but also strongly challenged my conceptual understanding.
TL;DR
- 2025 is exciting, but also a bit overwhelming.
- LLMs are emerging as a new kind of intelligence: much smarter than I expected in some ways, and much dumber than I expected in others.
- No matter what, they are extremely useful. Even at today’s capability level, I believe the industry hasn’t mined even 10% of their potential. There are still many ideas to try, and conceptually the field still feels wide open. As I mentioned earlier this year, I simultaneously (and seemingly paradoxically) believe: we will see continued rapid progress, but there is still a lot of hard work ahead.
Buckle up. We’re taking off.
1) RLVR: teaching AI to “think” like doing Olympiad math
Before explaining this complex foundational concept, how were LLMs trained in the past?
In early 2025, the “classic recipe” used by major labs was very stable:
- Pretraining: like GPT-3, let AI read the entire internet and learn to produce language.
- Supervised Fine-Tuning (SFT): humans write “correct answers” to teach the model how to respond.
- RLHF (Reinforcement Learning from Human Feedback): the model generates multiple answers, humans rate them, and the model learns what people prefer.
What changed?
In 2025 we added a powerful new ingredient: RLVR (Reinforcement Learning from Verifiable Rewards).
What does that mean?
Simply put: instead of having humans rate answers (humans are slow and subjective), we let AI work on tasks with objective, checkable answers—like math problems or coding tasks. Right is right, wrong is wrong, and machines can verify it automatically.
Through millions of self-play trials and errors, models spontaneously evolve strategies that look like “reasoning.” They learn to break big problems into steps, and even learn advanced techniques like “checking their work” (see the DeepSeek R1 paper).
Core comparison:
- Old paradigm (RLHF): like teaching kids to write essays. Without a single correct answer, it’s hard for AI to know exactly where it went wrong; it mostly imitates human tone.
- New paradigm (RLVR): like sending AI into an Olympiad training camp. You don’t teach it how to think—give it enough problems and right/wrong feedback, and it discovers patterns on its own.
This approach is so effective that it consumed most compute in 2025. As a result, models didn’t necessarily get bigger, but training ran longer. We also gained a new “knob”: letting AI think longer. OpenAI’s o1 was the beginning, and o3 became the true inflection point.
2) Ghost vs. animal: AI is not an “electronic pet”
In 2025, I (and the whole industry) finally developed an intuitive understanding of the “shape” of LLM intelligence.
A chilling metaphor: we’re not evolving/raising animals like pets—we’re summoning ghosts.
Why?
Because everything about AI differs from biology. The human brain is optimized for survival and reproduction in the wild; the LLM “brain” is optimized to imitate human text, score points on math, write code, and win approval in arenas.
Jagged intelligence:
Because RLVR exists, AI capabilities can suddenly spike in some areas (math, programming), producing a spiky profile. This leads to a ridiculous phenomenon:
- It can be a once-in-a-generation genius (solving advanced calculus instantly),
- yet also behave like a clueless elementary student (falling for simple logic traps).
Karpathy references a meme: human intelligence is a smooth blue circle, while AI intelligence is a red sea-urchin full of spikes—very vivid.
This also explains why I’ve lost trust in today’s benchmarks.
What is “leaderboard gaming” really about?
Since benchmarks are verifiable, you can train specifically for them with RLVR. Labs are effectively doing “test prep,” growing spikes exactly in the directions of exam questions. “Training on the test set” isn’t just cheating anymore—it has become a new art form.
3) Cursor: not just an editor, but a “foreman”
Cursor’s explosion in popularity reveals a new truth: the LLM application layer is thicker than we thought.
People are now talking about “Cursor for healthcare,” “Cursor for law.” What do these apps actually do?
- “Context engineer”: helps organize all background materials and feed them to the AI.
- “Foreman”: quietly coordinates multiple LLMs in the background, breaks down complex tasks, and also tries to save you money.
- “Remote control”: provides an “autonomy” slider so you decide how much to delegate to AI.
Prediction: model labs (e.g., OpenAI) will train “generalist college graduates,” while app developers (e.g., Cursor) will provide private data and tools to assemble them into “specialized construction teams.”
4) Claude Code: a “cyber ghost” living on your computer
Claude Code (CC) impressed me. It’s not only an agent that can write code—more importantly, it lives inside your computer.
In contrast, I think OpenAI took a wrong turn.
OpenAI’s early agents ran in the cloud (ChatGPT), far away from your real environment. While cloud agents sound like the endgame of AGI, in today’s jagged, transitional stage, local-first matters most.
Why does local matter?
Because your code, configs, keys, and messy environment all live locally. Anthropic got the priority right by putting AI into a small CLI (command-line interface).
It’s no longer a webpage in your browser; it becomes a “cyber ghost” residing on your machine, ready to help at any time. This is what future AI interaction should look like.
5) Vibe Coding
What is Vibe Coding?
It’s a term I casually coined on Twitter (and it unexpectedly went viral): it means you no longer need to truly understand syntax to code—you just describe your intent and feel in English, and AI handles the rest.
What changes does this bring?
- For ordinary people: the barrier to programming basically disappears.
- For experts: code becomes like tissue paper—cheap, disposable, and thrown away after use.
Example: to find a bug, I might ask AI to build a dedicated app on the spot, test it, and delete it. In the past, writing an app just to debug would sound insane. But in 2025, code feels free.
Vibe Coding will reshape the software industry and rewrite programmer job descriptions.
6) Nano Banana: AI finally gets its own “face”
Why does current AI interaction feel so non-human?
Whether it’s ChatGPT or Claude, we’re still “chatting by typing.” It’s like using the black-and-white DOS command line in the 1980s.
The reality: computers like text, but humans dislike reading walls of text. Humans are visual—we like images, charts, videos.
Google Gemini “Nano Banana” (a fictional model codename standing in for some multimodal interaction model) is another paradigm shift in 2025. It hints at what future LLM GUIs will look like.
Future AI shouldn’t spit out long text—it should draw a diagram, generate a webpage, pop up an interactive panel. This is not just “making images,” but a hybrid capability entangling text generation, logical reasoning, and visual expression.
