What 2025 Taught Us About AI, A Practical Look Ahead
2025 was a pivotal year for artificial intelligence, one that didn’t just deliver incremental improvements, but began to reshape the way we think about AI’s role in work, creativity, and the future of technology. AI researcher Andrej Karpathy, formerly of OpenAI and Tesla, shared a “Year in Review” that highlights the most important shifts in large language models (LLMs), the technology behind tools like ChatGPT and Claude.
Here’s what it means in plain English:
1. AI Is Learning to Think Better (But Not Like Us)
In early 2025, researchers started using a new approach called Reinforcement Learning from Verifiable Rewards (RLVR). Instead of just training AI to mimic human examples, models were trained to solve puzzles and tasks where success can be objectively measured — like math or coding challenges. This encouraged AI to develop its own problem-solving strategies rather than just repeat patterns from training examples.
Why it matters: AI is becoming better at reasoning, not just generating words. This makes it more useful for complex problems, but still very different from human thinking.
2. AI Isn’t Human-like, It’s Something Else
Karpathy points out that AI isn’t like growing a digital brain that thinks like a person. Instead, it feels more like summoning a new kind of intelligence that can be brilliant in some areas and surprisingly poor in others. He describes this as “jagged intelligence”, models that can seem genius one moment and confused the next.
Why it matters: We shouldn’t expect AI to think the way humans do, and that’s okay. It just means we need to understand both its strengths and limitations.
3. Tools Are Becoming More Like Assistants Than Tools
Another big change in 2025 was the rise of tools like Cursor, software that turns AI into a practical collaborator, not just a text generator. These tools wrap AI in workflows and interfaces that make it useful for real tasks: organising context, chaining together actions, and presenting results in ways humans can use.
Why it matters: The value of AI increasingly comes from the applications that surround it, the software layers that make it productive, not just the models themselves.
4. AI Is Starting to Live on Your Computer
A tool called Claude Code showed that AI doesn’t always need to run in the cloud. Instead, it can run locally on your own machine, close to your files, tools, and data, with much less delay and richer context.
Why it matters: Local AI agents could transform workflows in ways cloud-only systems can’t, especially where privacy, speed, and integration with local tools matter.
5. Programming Is Changing, Enter “Vibe Coding”
One of the most talked-about ideas from 2025 is “vibe coding”, where people build software simply by describing what they want in natural language. The code becomes almost invisible; it’s just a means to an end. Developers and non-developers alike can prototype, experiment, and build tools quickly without traditional coding skills.
Why it matters: This could democratise software creation, lowering barriers and accelerating innovation, but it also means we’re rethinking what it means to be a programmer.
6. A New Interface Is Emerging Beyond Chat
Finally, Karpathy points out that the classic chatbox interface, typing text back and forth — is like the old command line of early computing. The future will be richer: generative interfaces that combine text, images, charts, slides, and interactive elements so AI can communicate in the way people naturally understand.
Why it matters: As AI interfaces evolve, the technology will become more accessible, intuitive, and powerful for business users who don’t want to “talk to a chatbox.”
The Bottom Line
According to Karpathy, we’ve only scratched the surface of what AI can do. Models are already more capable than expected in some ways — and less capable in others — but even today we haven’t unlocked 10 % of their potential.
For business leaders, that means:
AI is becoming more than a tool, it’s becoming a collaborator
Software and workflows are being reimagined
New paradigms like local AI and natural-language development are starting to take shape
The opportunities ahead are vast, but still largely uncharted
If 2025 was any indication, the next few years will be some of the most consequential in the history of technology — and those who understand AI’s real capabilities (and limitations) will have a decisive advantage.