One of the founders of OpenAI, AI expert Andrej Karpathy, recently released an annual review, stating that 2025 will be a year of vigorous development in the field of large language models (LLMs), with six key &#34;paradigm shift&#34; turning points emerging. These changes not only alter the industry landscape but, more importantly, reveal that LLMs are evolving into a completely new form of intelligence.
On December 20th, according to Hard AI news, Karpathy stated in his annual review published on social platform X that LLMs are evolving into a new type of intelligence, &#34;much smarter than I expected, while also much dumber than I expected.&#34;
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/42ee8248-0650-4766-a5b6-ba29e350952b.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/42ee8248-0650-4766-a5b6-ba29e350952b.png"/>
He pointed out that this year saw six key turning points that changed the industry landscape, among which reinforcement learning with verifiable rewards (RLVR) has become a new stage in the LLM production process, with major laboratories shifting computational power originally used for pre-training to longer-term reinforcement learning training.
He particularly emphasized the &#34;sawtooth&#34; characteristics of LLM intelligence, stating that these models are both knowledgeable geniuses and confused elementary school students. Karpathy noted that LLMs are not &#34;evolving animals&#34; but are &#34;summoning ghosts,&#34; and this new form of intelligence needs to be understood from different perspectives.
In his review, he pointed out that the technological evolution from pre-training to reinforcement learning, the user experience innovation from text interaction to graphical interfaces, and the democratization of capabilities from professional programming to &#34;ambient coding&#34; all signify that AI applications are entering a new stage of development.
Although the capabilities are currently extremely useful, Karpathy believes that the potential of LLMs realized by the entire industry is still less than 10%, and he expects to see continuous and rapid progress, although the technical challenges remain daunting.
<h2>Turning Point One: Reinforcement Learning with Verifiable Rewards Changes the Training Paradigm</h2>
The most important technological breakthrough in 2025 is that reinforcement learning with verifiable rewards (RLVR) has become a new stage in LLM training.
Karpathy stated, the traditional production-level LLM training process includes three stages: pre-training, instruction fine-tuning, and reinforcement learning based on human feedback, while the addition of RLVR has completely changed this landscape.
RLVR trains LLMs in automatically verifiable environments such as math problems and coding puzzles, allowing them to spontaneously evolve strategies similar to &#34;reasoning.&#34; The model learns to break down problems into intermediate computational steps and masters various methods of trial and error to derive solutions. The DeepSeek R1 paper showcases the specific manifestations of these strategies.
Unlike the less computationally intensive SFT and RLHF, RLVR targets objective and non-cheatable reward functions, allowing for longer-term optimization. This approach has a very high &#34;capability/cost ratio,&#34; consuming the computational power originally allocated for pre-training. Most capability improvements in 2025 will stem from laboratories digesting this new stage's &#34;computational backlog.&#34;
OpenAI o1 is the first demonstration of the RLVR model, but the release of o3 is the real turning point, allowing people to intuitively feel the difference. RLVR also introduces a brand new adjustment knob: controlling reasoning ability by generating longer reasoning paths and increasing &#34;thinking time.&#34;
<h2>Turning Point Two: &#34;Ghost Intelligence&#34; Exhibits Zigzag Performance Characteristics</h2>
In 2025, the industry begins to truly understand the unique &#34;form&#34; of LLM intelligence.
Karpathy points out that we are not &#34;evolving animals,&#34; but rather &#34;summoning ghosts.&#34; The neural architecture, training data, algorithms, and optimization pressures of LLM are completely different from biological intelligence, resulting in a brand new intelligent entity.
Human neural networks are optimized for survival in jungle tribes, while LLM neural networks are optimized for mimicking human text, obtaining rewards for math problems, and gaining likes in LM Arena. This difference leads to LLM exhibiting &#34;zigzag&#34; performance characteristics: being both a polymath genius and a cognitively impaired elementary school student, solving problems one second and potentially being cracked by simple prompts the next.
With the RLVR application in verifiable domains, LLM experiences a &#34;surge&#34; in capabilities near these areas, but overall performance is extremely uneven. This phenomenon has led Karpathy to develop a crisis of trust in benchmark testing, as benchmarks are essentially verifiable environments that are easily susceptible to RLVR attacks. Laboratory teams build environments around the test set through &#34;ranking manipulation,&#34; and &#34;training for the test set&#34; has become a new art form.
<h2>Turning Point Three: Cursor Leads the Rise of a New Generation of LLM Application Layers</h2>
The rise of Cursor lies not only in its success but also in its revelation of a new level of &#34;LLM applications.&#34; People begin to discuss &#34;Cursor versions&#34; across various industries, marking the rise of vertical domain LLM applications.
LLM applications like Cursor encapsulate and orchestrate LLM calls for specific vertical domains, featuring four core functions:
<blockquote>
Handling &#34;context engineering&#34;; orchestrating multiple LLM calls at the underlying level, linking them into complex directed acyclic graphs, balancing performance and cost; providing a graphical interface for human participation in specific applications; offering a &#34;degree of autonomy adjustment slider.&#34;
</blockquote>
In 2025, the industry is hotly debating the &#34;thickness&#34; of this new application layer: will LLM laboratories dominate all applications, or will there be room for vertical applications to develop?
Karpathy believes that LLM laboratories tend to cultivate &#34;college students&#34; with general capabilities, while LLM applications organize these &#34;college students&#34; into specialized practitioners in specific fields by providing private data, sensors, actuators, and feedback loops.
<h2>Turning Point Four: Claude Code Creates a New Paradigm for Local AI Agents</h2>
Claude Code becomes the first convincing demonstration of LLM agents, linking tool calls and reasoning in a cyclical manner for long-term problem-solving. More importantly, Claude Code runs on the user's computer, utilizing private environments, data, and context.
Karpathy believes OpenAI has gone off track in this regard, overly focusing on cloud containers and ChatGPT orchestration rather than local deployment. While cloud-based agent clusters resemble the &#34;AGI endpoint,&#34; running agents directly on computers during the current transitional phase of uneven capabilities is more meaningful, allowing for specific configurations and collaboration with developers
Claude Code has captured the correct priorities, packaging them into a beautifully minimalist command-line form, transforming the appearance of AI. AI is no longer a passive website to access but a spirit that &#34;lives&#34; in the computer. This localized and personalized AI interaction paradigm points the way for future development, emphasizing the importance of privacy protection and personalized experiences.
<h2>Turning Point Five: Vibe Coding Makes Programming Skills Universal</h2>
In 2025, AI crosses a critical capability threshold, allowing people to build complex programs using only English, completely ignoring the existence of underlying code. The popularity of the &#34;Vibe Coding&#34; concept marks a complete reduction in the programming barrier.
Vibe Coding makes programming no longer the exclusive domain of professionals; anyone can participate. This confirms the LLM's characteristic of &#34;power to the people&#34;: unlike previous technologies, ordinary people benefit from LLMs far more than professionals, businesses, and governments. Not only can ordinary people try programming, but professional developers can also create more software that would not have been developed otherwise.
Karpathy shared his practical experience:
<blockquote>
Developing an efficient BPE tokenizer in Rust, creating various quick demo applications, and even developing entire temporary applications to find a single bug. Code has become free, instantaneous, and malleable, used and discarded. This change will reshape the software ecosystem, redefine professions, and bring the cost of realizing creativity close to zero.
</blockquote>
<h2>Turning Point Six: Nano Banana Opens the Era of LLM Graphical Interfaces</h2>
Google's Gemini Nano Banana is referred to by Karpathy as the most shocking and paradigm-shifting model of 2025. In his worldview, LLMs represent the next major computing paradigm following computers in the 1970s and 80s, leading to innovations of similar historical significance.
Currently, conversing with LLMs is akin to inputting commands into a computer console in the 80s. While text is the raw data representation preferred by computers and LLMs, it is not in a format preferred by humans. Humans do not like reading long texts; they prefer to consume information visually and spatially, which is why GUIs were invented in traditional computing.
LLMs should also communicate in formats preferred by humans: images, infographics, slides, whiteboards, animated videos, web applications, etc. Emojis and Markdown are early attempts, but a true &#34;LLM GUI&#34; requires deeper innovation.
Nano Banana provides an early prototype of this possibility, characterized not only by image generation but also by the intertwining of text generation, image generation, and world knowledge within the model weights.
This multimodal fusion capability heralds a fundamental transformation in future AI interaction interfaces, shifting from pure text dialogue to rich media, multisensory immersive experiences

DPSK

OpenAI

AI guru Karpathy pointed out six key turning points for AI large models by 2025: reinforcement learning based on verifiable rewards will change the training paradigm, models will exhibit "sawtooth" ghost intelligence features, Cursor will lead the rise of vertical applications, Claude Code will create a new model for local agents, Vibe Coding will popularize programming skills, and Nano Banana will usher in the era of graphical interfaces. He believes that LLM is evolving into a new form of intelligence, both knowledgeable and clumsy, with current potential realized at less than 10%

Andrej Karpathy's annual review: AI large models are evolving into a new type of intelligence, with 6 key turning points emerging this year