Andrej Karpathy posted an incredible 2025 LLM Year in Review. It is a totally salient take (obviously, he's brilliant!) and mirrors a lot of my experiences in the trenches. I have a slightly different take on a point or two below.
⪼ Like Karpathy, I believe there is a ton of value for the application layer above LLMs. I think the model providers are making a generally useful brain, but its frozen in a moment and for now at least, we need prompt and context engineering, specialized tools, and verification for producing repeatably good results. Agents need access to platforms that have tools built-to-purpose for a given domain. For coding: think semantic indexing over codebases and semantic browser-use APIs, debugging techniques optimized for machines, and unit/integration/end-to-end tests. And its true for basically every specialized domain - including environmental! Think about a human without any long term memory and only a few pieces of paper on which to record your thoughts - you'd get extremely creative "compressing" the important ideas down, but eventually you're going to start repeating the mistakes of the past.
⪼ On vibe coding, I only take issue with one point here: "regular people benefit a lot more from LLMs compared to professionals, corporations and governments." In my experience, the people that benefit the most are both very experienced and also very cross-functional (see: Foxes and Hedgehogs) - I think its a bimodal distribution. I think if your problem is solved by throw-away software or analysis, LLMs absolute change the entire calculus of whats possible. If your problem is bigger or more systemic than an individual contributor can solve on their machine, it still requires an experienced hand to get good results, and if you experienced hand is also cross functional (has good design sense, product ownership, dips into industry-specific domain knowledge deeply) you get absolutely outsized results. As to why we aren't seeing huge impact on the rest of the economy vs. what we're seeing in coding, I think its related to point 1. We are very slowly building domain specific tools (outside of coding) and most products have not let go of their pure SaaS mindset yet. 2026 is going to bring way more of this, and I think the impact on the economy will start to accelerate.
⪼ On "AI that lives on your computer" or more specifically why coding agents are best when they are local to the developer's machine: I agree with Andrej's take broadly, but I actually think this is solvable just inside the "tools layer." We need a new focus on repeatable dev setup - local databases, test API keys or mocked third party services, enough compute to run automated tests including browser automation like Playwright. Isolated test datasets that don't interact with each other so they can be run in parallel. Without the verification layer reproduced inside a container and more rigorous verification and adversarial prompting, you can't actually get the results you need. You get half baked PRs that have compile issues and have not actually run tests or verified their results - and thats where almost all of the magic takes place. So a developer ends up having to review and intervene anyway. Ergo, its just better to be local to have a faster devx feedback loop. You might argue thats fragile enough that we'll see local development continue to dominate, but I think there's a pretty strong incentive for us to focus on solving for. Smart models operate in 15 - 25 minute increments, and we should be able to have multiple threads running. The same things that make "worktrees" work make cloud-based development work - solve for both.
Andrej is an incredible educator, and he sheds light in a special way. He's a rare individual that combines deep understanding with the ability to communicate complex ideas simply.
View Karpathy's original post on X
