I enjoyed your 2025 summary article. Karpathy, simon The AI app ecosystem is maturing in expected and surprising ways. We’ve found ways to make code cheaper, but it hasn’t yet spread throughout the enterprise (or the world) in the way that lower costs imply. I also don’t think we understand even 10% of what that means for how companies are built and what software exists. On the other hand, there are still fundamental tool problems to solve, such as the fact that all of our tools are for making, not for thinking.
Thinking tools vs. creating tools
One of the big changes I expect to see is in the nature of the tools themselves. All the tools we use for knowledge work focus on: execution: An IDE for writing code, Figma for creating designs, and spreadsheets for creating models. When it comes to tools; expedition – Tools to help you think – There isn’t really a modern product, apart from the way the LLM itself emerges as a thinking partner.
As coding agents become more accurate and able to work over longer periods of time, the challenge becomes: How do I build it? to What do I build? Imagine a near-future PM who sets broad goals for AI and wakes up every morning to review 2-3 features that models invent, run, and A/B test overnight. However, in my experience the model is still not very good at deciding what to build next. The ideas are bland, derivative, and generally lack the sparkle that comes with really good new product thinking. So I think the spiritual successor to coding tools, design tools, productivity tools will be very focused on: Explore and execute. Coding tools are already leading the way here. I thought the cursor was the furthest away antigravity was interested in being “agent-first” (exploration-first) in product design.
Software eats up all “service” functionality in your organization
I always notice the difference between “power” and “service” functions in software companies. Power functions (engineering/product/performance marketing) tend to be closer to software, while service functions (legal/finance/HR) tend to be further away from software and more human capital-driven.
Coding agents have two important implications for companies. The first is that every team and every task (marketing, legal, procurement, finance) should be software first, and all of these leaders must learn how to reach into the software toolbox before the processes and human systems they have traditionally relied on. Many of these organizations will adopt domain-specific products like Harvey, but others will still use “bare metal” coding agents like Codex and Claude Code. Every team should be a software team.
Second, companies (particularly those that produce software) dramatically We are more ambitious about what software we create, and our entire ideation and prioritization pipeline needs to be restarted to accommodate this. All the capabilities that can be built will be built, but most companies are not prepared for this reality.
I think the problem of cultural change is just as difficult as the problem of organizational change.
Complexity of AI apps
As we enter the second year of inference models, we expect to see a continued disconnect between AI-native apps and AI models, with apps that combine state-of-the-art model orchestration, domain-specific UIs, and a much broader range of functional aspects that are now much cheaper to build. This is the natural meaning of what we called “.”narrow startupIt allows for extraordinary specialization, and I think this is part of the strong case for apps being separate and increasingly divorced from the model.
The capabilities of research labs and big tech companies feel as “jagged” as the models they produce. They’re formidable in the areas they focus on, but they also have complex commitments (i.e., Google’s promise to regulators to no longer intermediate the Internet) and difficult prioritization issues (OpenAI is simultaneously competing to be at the top of consumer, enterprise, model, and hardware companies). So I think it’s a wrong assumption that the app layer will be built into the model. Even in areas like model progression and coding, which is the focus of the lab, there is a thriving startup ecosystem that will generate more than $1 billion in new revenue in 2025 alone.
we used to Outlined the framework A domain that leverages AI apps, a domain that benefits from a multi-model, highly focused ecosystem of data resources, network products, and many functional aspects. Combine this with Karpathy’s excellent representation of “thick” AI apps (multi-model orchestration, autonomy sliders, context engineering, etc.) and you can start to see what AI apps will look like as they mature.
Humans discover the “rest” of AI
eugenia He is the one who has thought the most about how command line UIs have kept everyday consumers away from the great capabilities of AI. This is starting to change: Wabi is a big catalyst for exposing code generation to consumers, and ChatGPT/Grok[画像]Tabs will do the same for image generation, and with a little luck, Apps Directory and skills will do the same for MCP and prompt plugins.
I liked it Dan Wang’s The criticism is that Silicon Valley is a bit culturally tone-deaf to the impact of AI, but I think this problem will be partially alleviated as more consumers make things. Generating a small app in 2025 was as much fun as generating a poem in 2023, and most consumers still don’t know it exists. I think this is partially covered too. Nikita’s Take note of who is creating things. This is really black medicine.
Memo for (current) CEO
While our focus is right on builders, we have some thoughts for CEOs who have already scaled and are thinking about how to move forward with the transition to AI. First, you’ll see a best-in-class example of how the model integrates all customer-facing roles (sales, support, collections) into one function with broad goals. The second is to accept the above caveat of being software-first in every feature. The non-technical feature of employing the model is how companies gain extensive operating leverage. Finally, if Tesla can deliver on FSD, I’d think twice about demanding a more ambitious product and a more ambitious price. coast to coast And the Claude code is Written in Claude code In that case, AGI already exists for the short-term purpose of most enterprise tasks.
Finally…have fun
Consider this your realization, because no one will tell you you’re living in the good old days until the good old days are gone. This product cycle is less centralized, more software-driven, and simply more fun for engineers than any recent cycle. I hope you enjoy it as much as I enjoy exploring these new technologies, discussing their implications, and creating even more new ones.
