Must-Read: K's 2025 AI Review – 6 Paradigm Shifts Reshaping Our World

Introduction

AI development is accelerating so fast that designers, developers, and tech enthusiasts wonder which advances truly change the game. The author summarizes six paradigm‑shifting developments in large language models (LLMs) observed in 2025, aiming to clarify the direction of future human‑AI interaction.

1. AI Learns to "Reason": The Rise of RLVR

In early 2025 the standard LLM training pipeline (pre‑training → supervised fine‑tuning (SFT) → reinforcement learning from human feedback (RLHF)) was extended with a new stage called Reinforcement Learning from Verifiable Rewards (RLVR). RLVR trains models in environments where rewards can be automatically verified—such as math puzzles or code challenges—allowing LLMs to develop human‑like reasoning strategies, e.g., breaking complex problems into intermediate steps.

RLVR requires far longer training time than SFT or RLHF, consuming most of the compute budget previously allocated to pre‑training. It also introduces a new control knob: longer reasoning paths and increased "thinking time" let users steer model capability at test time. OpenAI’s o1 (late 2024) first demonstrated RLVR, but the 2025 release of o3 marked a clear inflection point.

2. Summoning "Ghosts" Instead of "Animals": A New AI Cognition

The author argues that 2025 is the year LLM intelligence is internalized as a "ghost" rather than an "animal". Human intelligence evolved under survival pressure, while LLM intelligence optimizes for mimicking text and earning rewards in mathematically verifiable tasks.

This difference produces a "sawtooth" characteristic: LLMs act like brilliant scholars but can also behave like confused children, vulnerable to jailbreak prompts that leak data. Consequently, the author’s trust in traditional benchmarks declines because those benchmarks are themselves verifiable environments that RLVR can exploit, prompting labs to tailor training near the benchmark embedding space.

3. "Cursor for X": A New LLM Application Middle Layer

The rapid rise of the programming assistant Cursor in 2025 signals a distinct middle layer for LLM applications. Developers are now discussing "Cursor for X"—custom AI assistants built for specific verticals. According to the author’s Y Combinator talk, such assistants perform three core functions:

Context engineering : supplying precise background information needed for a task.

Orchestrating LLM calls : managing multiple model invocations while balancing performance and cost.

Providing dedicated GUIs : offering domain‑specific graphical interfaces for human interaction.

Autonomy sliders : letting users adjust the AI’s degree of self‑direction.

The author predicts that future LLM providers will train models to be "well‑rounded college graduates," while application layers will organize these graduates into specialized teams through private data, sensors, and feedback loops.

4. AI Lives on Your Computer: Local Interaction Paradigm

Anthropic’s Claude Code (CC) is highlighted as the first convincing LLM agent that runs on a user's localhost computer, allowing direct use of private environments, data, and context.

The author criticizes OpenAI for deploying agents in cloud containers instead of on the local machine, arguing that in a world of uneven capabilities, local agents that cooperate with developers in their own environments are a smarter strategy. Claude Code’s minimalist CLI interface reframes AI from a remote service to a "ghost" living on the developer’s machine.

5. "Vibe Coding": Programming by Natural Language

In 2025 AI crossed a threshold where complex programs can be built solely through English conversation, a phenomenon the author coined "Vibe Coding". This trend shifts power toward the public: non‑experts can now create software, while professionals can generate large amounts of one‑off code quickly.

Empowers non‑technical users to program creatively.

Enables developers to produce massive, disposable code snippets—e.g., the author built a custom BPE tokenizer in Rust for the nanochat project with near‑zero generation cost.

The author foresees "Vibe Coding" reshaping the software industry and redefining related job descriptions.

6. From Chat to GUI: The Future of Visual Interaction

Google’s Gemini Nano Banana model is presented as a paradigm‑shifting example. The author likens current text‑based LLM interaction to 1980s command‑line computing, arguing that humans prefer visual, spatial information formats. Just as GUIs transformed early computers, LLMs need a visual interface—delivering output as images, infographics, slides, videos, or web apps.

Nano Banana combines text generation, image synthesis, and world knowledge into a single entangled capability, signaling the first glimpse of AI moving from pure text toward richer visual interaction.

Conclusion

2025 proves to be an exhilarating year: LLMs exhibit a new form of intelligence that is simultaneously smarter and more foolish than expected. Their potential remains largely untapped—perhaps less than ten percent of what they can achieve. While rapid, continuous progress is anticipated, substantial work still lies ahead. The author challenges designers to consider how AI’s shift from text to vision and from cloud to local will shape new interaction paradigms and give form to the summoned "ghosts".