Karpathy’s 2025 LLM Year in Review: Six Paradigm Shifts That Blew My Mind
The other day while scrolling through Twitter, I saw Karpathy’s 2025 LLM year-end review. As the former Director of AI at Tesla and founding member of OpenAI, he has unique insights into the AI trends of 2025. You know who Karpathy is, right? Former Tesla AI Director, OpenAI founding member, Stanford PhD, top-tier figure in the AI circle. His writings are always down-to-earth yet profound, and this year-end review brilliantly explains the major events in the LLM field.
After reading Karpathy’s 2025 LLM summary, I found it so interesting that I had to share it with you. Not a simple word-for-word translation, but explaining in plain language what AI has been up to this year.
1. RLVR: The New Paradigm for 2025 LLM Training
According to Karpathy, before 2025, training a powerful large model basically involved three steps:
- Pre-training: Let the model read the entire internet
- Supervised Fine-tuning: Show it high-quality human-written answers
- Reinforcement Learning from Human Feedback (RLHF): Have humans rate what constitutes a good answer
This playbook worked for several years, and everyone knew it by heart. But in 2025, major changes in AI trends emerged—RLVR, Reinforcement Learning from Verifiable Rewards became the new favorite.
In plain English: previously, humans would score AI to tell it what a good answer is. Now, it’s changed to having AI solve math problems and programming challenges, then check the answers immediately. Whether the answer is right or wrong, the program decides, no human intervention needed.
What’s this like? It’s like how we used to teach kids math—teachers would hand-hold through every calculation step. Now we’ve changed approach: give the child a workbook with standard answers in the back. You do the problems yourself, check your own answers, continue if right, and figure out what went wrong if wrong.
Amazingly, after training this way, AI spontaneously “figured out” a set of reasoning methods. It learned to break down big problems into small steps, learned to try different paths when stuck. These strategies weren’t taught by humans—they were discovered by the AI itself—because that’s the only way to solve problems correctly.
Even more impressive, unlike previous methods that were just superficial, the fine-tuning stages used to be very short. Now RLVR lets AI practice intensely on verifiable problems for as long as needed. The result: 2025’s models didn’t grow much in parameter size, but their capabilities soared. OpenAI’s o3 is a representative of this approach—everyone could clearly feel the difference when it came out—this thing is different.
2. Ghosts vs Animals: The Truth About 2025 LLM Intelligence
In 2025, Karpathy said he finally understood something: AI isn’t a pet we’ve raised, but a ghost we’ve summoned. This AI trend perspective is particularly interesting.
This sounds mysterious, but it’s actually very vivid. Think about it: how did animal intelligence evolve? Millions of years of evolution, struggling in the jungle, getting smarter bit by bit to survive. But what about 2025’s LLMs? They’ve read all human texts, score points in math problems, compete on leaderboards.
The evolutionary pressures are completely different, so what emerges is naturally completely different. This might be one of the most important AI trend recognitions of 2025.
The result is that 2025’s AI intelligence is crooked and uneven. The same LLM model can write you a stunning literature review one moment, then be tricked by a middle schooler’s brain teaser the next. Sometimes an all-knowing genius, sometimes a clueless fool.
Karpathy calls this “jagged intelligence.” I think “jagged intelligence” is quite fitting. Like a hedgehog—some spines are very long, some places are bare. This recognition is becoming a mainstream view of 2025 AI trends.
This is also why he said he’s increasingly unimpressed with benchmarks in 2025. Because AI is particularly good at scoring high in areas that can be practiced like tests, but those scores don’t represent true general intelligence. It’s like a student who crazily practices past exam papers—scores very high, but might freeze when faced with unfamiliar question types.
So next time you see news like “AI surpasses humans on some test,” just laugh it off. Real AGI probably can’t be measured by benchmarks. This is also Karpathy’s important reminder about 2025 AI trends.
3. Cursor: A New Species of LLM Applications
In 2025, a programming tool called Cursor became incredibly popular. But Karpathy says what’s most notable about Cursor isn’t how useful it is, but that it represents a completely new species—LLM Applications.
Previously, when we discussed AI, it was either the underlying large models or direct user-facing ChatGPT. Cursor shows us there’s a middle layer: AI applications specifically designed for vertical domains.
What does this application layer do? Karpathy summarizes four things:
- Context Engineering: Helps you feed relevant information to AI, saving you from copying and pasting
- Multi-turn Orchestration: A task might require calling AI several times, it helps string them together
- Dedicated Interface: Designs more convenient interactions for specific scenarios
- Autonomy Slider: Lets you decide whether AI plays support or takes the lead
The home renovation analogy is clear. Large models are various craftsmen—electricians, carpenters, masons. Using ChatGPT directly is like being your own contractor, assigning tasks to workers one by one. But LLM applications like Cursor are like hiring a renovation company—they coordinate all the craftsmen for you, you just say “I want Nordic style.”
So everyone started wondering: does every industry need a “Cursor for X”?
4. Claude Code: AI Living on Your Computer
Speaking of AI programming assistants, 2025 had another heavyweight—Claude Code.
What’s special about it? It runs on your own computer.
This might not sound like a big deal, but Karpathy thinks this design decision is crucial. OpenAI’s early approaches with Codex and various agents were all deployed in cloud containers orchestrated from ChatGPT. You interact through a webpage, and AI works on remote servers for you.
Claude Code is different—it’s a command-line tool installed on your computer, with your development environment, able to directly access your files, configurations, and context.
What’s this like? OpenAI’s approach is like hiring remote customer service—you describe the problem on the phone, they operate on the other end. Claude Code is like hiring a live-in butler—they live in your house, know what’s in your fridge, where you keep your keys, what your habits are.
Although “the ultimate AI form” might involve cloud agent clusters sounding more impressive, in this current era where AI capabilities are still “jagged,” having AI live on your computer, directly using your environment, reading your code, understanding your context, is actually more practical.
And Claude Code changed the form factor. AI is no longer just a website you visit, but a “little spirit” that “lives” on your computer. This is a completely new human-AI interaction paradigm.
5. Vibe Coding: The Programming Revolution Brought by 2025 LLM
In 2025, Karpathy posted a tweet, casually coining the term “Vibe Coding”—atmospheric programming, or “intuition-based programming.” This term became one of the hottest words in the 2025 LLM field.
The term went viral, becoming an annual buzzword representing important AI trends.
What exactly is Vibe Coding? You use English (or Chinese) to describe what you want, and LLM writes the code for you—you don’t even need to see what the code looks like. It’s like ordering food without knowing how to cook—you just say clearly “I want sweet and sour pork, less oil,” and the AI chef gets it done. This is the new programming trend brought by 2025 LLMs.
This really became a reality in 2025. Ordinary people who can’t program can create decent programs with Vibe Coding. Programmers who can program can quickly handle small tools they were too lazy to write before.
Karpathy himself is an example. He used Vibe Coding to write an efficient tokenizer—in Rust, a language he wasn’t very familiar with. Previously, he would have either spent effort learning Rust or used existing libraries. Now? Just describe the requirements clearly in English, and AI handles it.
6. Nano Banana: The GUI Era of AI
The last thing that impressed Karpathy was Google’s Nano Banana.
The name sounds funny, but the meaning behind it is profound.
Think about it, what were computers like in the 1980s? Black screen with green text, typing command lines. Ordinary people couldn’t use them at all. Later came GUI—Graphical User Interface—windows, icons, mouse, computers truly entered millions of homes.
What’s it like using AI now? Typing and chatting. Actually not fundamentally different from typing command lines back then, still communicating with machines using text.
But humans naturally prefer visual things like images, videos, animations. Reading text is tiring, pictures can be understood at a glance.
The direction Nano Banana represents is: AI should speak to us in the way we like. Not just generating text, but generating images, infographics, slides, animations, short videos, even small web pages.
Of course, at this stage we already have “lightly visualized” things like Markdown, tables, code blocks. But Nano Banana hints at greater possibilities: future AI output could be a dynamic infographic, a demo video, or an interactive mini-program.
Conclusion: Karpathy’s View of 2025 LLM Summary
Karpathy says at the end of the article that 2025’s AI is much smarter than he expected, and also much dumber than he expected. This is too perfect a summary of 2025 LLM development.
It’s really smart in places where it can practice and score—AI can already do jaw-dropping things. It’s also really dumb in places, can be tricked by simple routines anytime. This is the true portrayal of 2025 AI trends.
But regardless, 2025’s AI is already very useful. Karpathy thinks that even if AI capabilities freeze today without progress, we’ve only utilized less than 10% of its potential. There’s still so much that can be done. This view captures the core of 2025 LLM development.
At the same time, the research field in 2025 is still a blue ocean, with too many ideas not yet tried. He simultaneously believes two seemingly contradictory things: AI will continue to progress rapidly, and there’s still much work to be done. This judgment of AI trends is particularly profound.
2025’s AI is like a strange genius teenager. Sometimes amazing enough to make you doubt life, sometimes adorably silly enough to make you laugh. But undeniable, it has grown up a lot this year.
And we are just beginning to learn to coexist with this new species. Karpathy’s 2025 LLM summary tells us this is just the beginning.
Alright, that’s all for today.
If you found this interesting, feel free to give it a like or share it with friends who are also following AI. If you have any thoughts, welcome to discuss in the comments—I’d love to hear what you think about these 2025 AI changes.
See you in the next article.