I Tracked Every AI Win & Failure in 2025. Here's What Actually Worked (9 Surprises)

2025’s biggest unlock wasn’t a new model—it was letting large language models (LLMs) use code as a tool. That shift turned “talk to the model” into “talk to the computer,” enabling agentic workflows that can manipulate files, run tasks, and iterate with far less technical friction. The change built gradually through the year as practical layers emerged: cloud code, model context protocol, “skills,” Codex, and tools like Cursor. By the end of the year, plain English increasingly became a control surface for software work, not just a way to generate text. The implication is clear: once an LLM can operate parts of a computer, the addressable value expands beyond developers into non-code users—because the interface becomes natural language plus execution.

A second major surprise was images finally catching up enough to matter. For much of 2025, text stayed the most reliable way to get accurate results, with code treated as just another text-like language. But when image generation reached the point where detailed infographics, maps, layouts, and slide-like visuals didn’t look “weird,” the payoff went beyond prettier decks. It enabled generative UI—interfaces that aren’t fixed to a developer’s static design or a single screen. The vision described is a continuous, evolving digital surface: experiences that adapt to context, potentially spanning wearables, combinations of phone and laptop, and even generative elements embedded alongside services like Comet. The caveat is equally important: consistency and habit still matter, so the future isn’t a total replacement of familiar interfaces. Still, solving images was framed as a foundational step toward graphical experiences that can evolve with users rather than forcing users to adapt to rigid UI.

Beyond those two headline shifts, several operational lessons stood out. One was that progress didn’t require “AI developers” in the narrow sense. Individuals who designed systems—using templates, validators, retries, and fast iteration—could outperform larger teams that treated engineering as a model-worship problem. That fed a broader reframe: technical and non-technical categories are less useful than curiosity about domain problems and willingness to learn the AI skills needed to solve them.

Another practical win came from verification loops in agentic systems. Measuring correctness across multiple dimensions and feeding results back into iteration made agents far harder to game and dramatically faster—likened to adding a jet engine to an airplane. The expectation for 2026 is more standardization: shared eval and verification primitives, especially for areas like accessibility, so teams don’t reinvent the same checks.

The “messy middle” also emerged as underbuilt but valuable. While attention often fixated on model makers owning the stack, the transcript argues that transforming messy model outputs into structured representations—routing intent, orchestrating tool calls, handling exceptions, and producing usable interfaces—creates durable leverage. Cursor is cited as a prominent example, but the broader point is that intelligence substrates can be built on top of models.

Finally, the year’s momentum shifted from hype to quality. AI slop was treated as a symptom of unmanaged, unconstrained production; better systems can generate performant, information-dense marketing and content without sacrificing usefulness. Meanwhile, labor-market signals suggested faster selection for creative problem-solving instincts, and leaders increasingly moved from cost-cutting toward quality lift—scaling customer value while keeping people focused on what only humans can do. The closing question: what exceeded expectations in 2025?