Is AI Moving Too Fast to Build On?

The Paradox of Progress

In almost every previous technological shift, progress has been the primary driver of adoption. The logic is linear and predictable: better infrastructure attracts more developers; better tools lead to more sophisticated products; and more products result in a larger user base. We saw this with the internet, the transition to mobile, and the rise of cloud computing.

Over the last five years, we have seen rapid capability gains across language, vision, and code. Reports such as the Stanford Institute for Human-Centered AI's AI Index Report show significant year-on-year improvements in benchmark performance across multiple domains, particularly in natural language processing and multimodal models.

Large language models (LLMs) moved from simple chatbots, to research curiosities, to usable tools. Image generation went from blurry psychedelic shapes to photorealism. Code generation reached production-grade quality, and early agentic systems began to emerge.

Yet, instead of these leaps creating a clean set of layers upon which an ecosystem could be built, they have created a state of constant flux. When the foundation shifts every six months, everything built on top of it becomes temporary. In AI, we haven't built a platform; we’ve built on a moving target.

The Unstable Abstraction Layer

The primary challenge facing developers today isn't a lack of capability - it’s instability. In traditional software, you choose a stack and expect it to remain viable for years. In AI, the "correct" way to build a product has been changing repeatedly in the space of months or weeks, not the decade long cycles developers have traditionally relied on.

Think back through the last two years, we moved from prompt engineering to Retrieval-Augmented Generation (RAG). Then the focus shifted to tool use and function calling. Now, the industry is obsessed with autonomous agents, multi-agent systems, and complex reasoning loops.

The problem is that each of these waves didn't just iterate on the previous one; in many cases, they replaced it entirely. If you spent six months optimizing a product for a specific RAG architecture, a new model release with a massive context window or native "reasoning" capabilities might suddenly make your entire engineering effort redundant. This creates a structural problem for any business: how do you justify the cost of building when the architecture itself has an expiration date?

The Real Risk: Building Features, Not Companies

As if technical instability weren't enough, there is the growing issue of platform risk. Foundation model providers - the likes of OpenAI , Google , and Anthropic - are no longer staying in their lane as simple "engine" providers. They are aggressively moving up the stack.

They aren't just shipping raw models anymore. They are shipping assistants, memory systems, retrieval tools, and end-to-end workflows. This has a brutal implication for the "wrapper" economy: anything that looks like an obvious SaaS layer on top of AI is at risk of being absorbed by the model provider.

What feels like a standalone product today, like perhaps a specialized document search tool or a basic AI coding assistant, can easily become a native feature in the next model update. When your differentiation can be collapsed into a single parameter change or a new "system prompt" from the platform provider, you aren't building a company; you're building a feature that hasn't been integrated yet.

The Illusion of Capability

There is another subtle dynamic at play that explains why enterprise adoption feels slower than the hype suggests. AI systems are capable of remarkable things, but they are not consistently reliable.

This creates a dangerous mismatch between perception and reality. AI demos are almost always incredible because they showcase the "peak" of what a model can do. However, production systems live in the "trough" of edge cases and exceptions.

• AI can solve complex problems… sometimes.
• It can reason… inconsistently.
• It can automate workflows… until it encounters a slight variation and breaks.

The Illusion of Capability (Continued)

For a hobbyist, a 70% success rate is a miracle. For a business, a system that works 70% of the time is a liability. Even 95% is often insufficient for core business logic. We are still struggling to bridge the gap between "impressive" and "usable," and the constant churn of new models makes it harder to do the grueling work of fine-tuning for reliability.

The "Build for Tomorrow" Trap

A common piece of advice often attributed to industry leaders like Perplexity CTO Denis Yarats is to " build for where AI will be, not where it is. " On the surface, this sounds like visionary strategy. In practice, it is often a trap.

Building for future capabilities only works if you have the capital to burn while waiting for the technology to catch up to your vision. For most builders, this approach leads to three specific failure modes:

• Depending on Future Capability: If your value proposition relies on hallucinations dropping to zero or models becoming "smarter," you don't have a product. You have a bet on someone else’s research roadmap.
• Competing with the Platform: If you are building toward a capability that is an obvious next step for the model providers, you are racing against a giant that has more data and more compute than you.
• Optimizing for Peak, Not Baseline: Users don't buy potential; they buy consistency. A product that occasionally delivers brilliance but frequently fails will always lose to a "boring" product that works predictably.

The Inversion of Platform Risk

In most technological eras, the rule was simple: the closer you are to the platform, the safer you are. If you built deep integrations with Windows or AWS, you benefited from their stability.

In the AI era, this has been inverted. The closer you are to the model, the more exposed you are. Because the model is improving faster than the applications on top of it, those closest to the "raw" intelligence layer are the first to be disrupted when a new version drops. You inherit all the volatility of the model without any of the control.

What Actually Works?

Despite this turbulence, some companies are succeeding. They aren't doing it by chasing the latest benchmarks; they are following a different set of patterns.

• AI is an Enabler, Not the Product: The most resilient products treat AI as the weakest, most replaceable component of a larger system. The value isn't that they "use AI." The value is that they solve a specific, painful problem where AI happens to provide a 10% or 20% efficiency gain. The real differentiation lives in the workflow design, the user experience, and the deep integrations that are hard to move.
• Anchor on Constraints AI Won't Solve: There are certain problems that a smarter model cannot solve on its own. These include regulatory compliance, multi-party coordination in complex industries, and navigating proprietary data ecosystems. If your product is protected by legal, operational, or data-moat constraints, it is naturally defensible against a model upgrade.
• Optimize for Today with Optional Upside: The goal should be to build something that is viable with today’s models but improves automatically as AI gets better. If a new model release improves your product, that’s great. If a new model release is required for your product to even exist, you are in a precarious position.
• Treat Model Progress as Margin: The best way to view model improvements is as a reduction in cost or a slight increase in quality—essentially, an improvement in your "margins." You want to be in a position where a model update makes your business more profitable, not one where it determines whether your business remains relevant.

The Uncomfortable Middle Phase

We are currently in a mismatch of timelines. AI is evolving at the pace of a high-speed research field, but businesses require the stability of enterprise software. These two speeds are currently incompatible.

This creates an uncomfortable equilibrium where builders are encouraged to move fast but are often punished for committing to a specific architecture too early. It is a period of intense innovation but also extreme vulnerability.

This phase won't last forever. Eventually, the pace of foundational change will slow, allowing stable abstraction layers to form. Or, standardized "operating systems" for AI will emerge that insulate developers from the volatility of the underlying models. But for now, we are still in the turbulence.

A Better Mental Model

Instead of asking, "How do I build for the AI of tomorrow?" a much more useful question is: "How do I build something that survives the AI of tomorrow?"

That single shift in perspective changes every decision you make. It moves the focus from chasing benchmarks to owning workflows. The real opportunity in AI right now isn't in raw intelligence; it’s in integration.

The winners of this era won't be the ones who build the smartest models or the most clever prompts. They will be the ones who control the distribution, accumulate proprietary data, and sit firmly between the user and the outcome. They won't just use AI - they will contain it within a structure that provides actual, repeatable value.

The last five years have been about the drama of the breakthrough. The next five years will be about the discipline of the application. It will be less exciting to watch, perhaps, but it will be far more consequential.