Ask anyone working with AI today and you will get a surprisingly honest answer: it is impressive, yes, but it is not enough. Not yet. Despite the breathless headlines and the billion-dollar valuations, the gap between what AI promises and what it actually delivers is still wide enough to walk through. So what is the one thing we wish AI was better at? The answer depends on who you ask, but the patterns are remarkably consistent.

The things AI still cannot do

The list is shorter than it was two years ago, but it is still significant.

AI cannot truly reason across domains. It can solve math problems, write code, and summarize legal documents, but it cannot fluidly connect insights from biology to economics to ethics the way a curious human can. It pattern-matches. It does not understand.

AI cannot form independent goals. Every task an AI completes was assigned by a human. It has no intrinsic motivation, no curiosity, no sense of "I wonder what would happen if I tried this." It is, at its core, a very sophisticated auto-complete engine with a planning layer bolted on.

AI cannot reliably know what it does not know. Hallucinations remain one of the most stubborn problems in the field. Models will confidently fabricate citations, invent statistics, and present fiction as fact. This is not a bug that will be patched in the next release. It is a structural limitation of how these systems work.

AI lacks emotional depth and moral judgment. It can simulate empathy in a chatbot conversation, but it does not feel anything. It cannot weigh the ethical consequences of a decision the way a human can, because it has no stakes in the outcome.

The cost equation is more complicated than you think

One of the loudest promises of AI is that it will make everything cheaper. The numbers seem to support this at first glance. Inference costs have plummeted: the price of processing a million tokens dropped from roughly $10 in 2024 to as low as $0.10 for input tokens on some platforms by early 2025. A customer service workload that costs $42,830 in annual human salary can be handled by an API for under $100.

But zoom out and the picture gets murkier. OpenAI generated $3.6 billion in revenue in 2024 against $9 billion in spending, and projects $74 billion in cumulative operating losses through 2028. The AI industry is burning cash faster than it earns it, subsidized by investor patience. Companies that restructured around cheap AI are building on a price floor that may not hold.

Then there is the hidden cost of oversight. AI does not manage itself. Someone has to review outputs, catch errors, handle escalations, retrain models, and maintain infrastructure. McKinsey estimates overhead costs of around $200,000 annually for enterprise AI deployment. Factor in energy consumption, water usage for cooling data centers, and the environmental footprint, and the "AI is cheaper" narrative gets a lot more nuanced.

The honest assessment: AI is extraordinarily cheap for narrow, repetitive tasks at scale. It is not yet cheap for complex, judgment-heavy work that requires reliability.

From prompts to agents: a brief history of how we got here

The evolution of large language models over the past few years reads like a compressed version of the entire history of computing.

The prompt era (2022-2023). ChatGPT launched in November 2022 and reached 100 million users in two months. The interaction model was simple: you type a prompt, you get a response. The skill was "prompt engineering," the art of phrasing your question just right to coax a useful answer out of the model. GPT-4 arrived in March 2023, bringing multimodal capabilities and better reasoning.

The context era (2024-2025). As people built more ambitious applications, they realized that a single prompt was not enough. The focus shifted to "context engineering," the practice of curating the right information, instructions, memory, and tools so the model can do its job effectively. As Anthropic put it, building with language models became "less about finding the right words and phrases for your prompts, and more about answering the broader question of what configuration of context is most likely to generate the desired behavior."

The reasoning era (2024-2025). Models learned to think step by step. Reasoning models like OpenAI's o1 and DeepSeek did not just generate text, they worked through problems methodically, showing their chain of thought and self-correcting along the way. This was arguably the most significant architectural shift since the original transformer.

The agent era (2025-2026). Today, AI systems can plan, use tools, browse the web, write and execute code, and chain together multiple steps to complete complex tasks. They can be given a goal and figure out how to achieve it, at least in theory. In practice, agents are still fragile. They lose track of context, make compounding errors, and require careful guardrails. But the trajectory is clear: we are moving from AI as a tool you use to AI as a collaborator that acts.

Have we reached the limit?

This is the question that keeps AI researchers up at night. For years, the industry operated on a simple faith: make the models bigger, feed them more data, give them more compute, and they will get smarter. The scaling laws held. Until they started showing cracks.

By late 2024, reports emerged that brute-force scaling was hitting diminishing returns. Making models ten times bigger no longer produced ten times better results. AI labs began changing course, investing in new architectures, better training techniques, and efficiency gains rather than just throwing more parameters at the problem.

Google DeepMind CEO Demis Hassabis has argued that scaling "must be pushed to the maximum" but acknowledges there will likely need to be "one or two" additional breakthroughs to reach artificial general intelligence. The honest consensus in early 2026 is that we have not reached the limit of AI capability, but we have likely reached the limit of the current approach. The next leap will require something new, not just something bigger.

Epoch AI's research suggests that sustained efficiency gains could push AI scaling into the next decade, but through cleverness rather than brute force. The analogy to Moore's Law is instructive: when transistors could no longer simply be made smaller, the industry found other ways to make chips more powerful. AI will likely follow a similar path.

Are you using AI, or is AI using you?

Here is the question that nobody in the industry wants to sit with for too long.

Every time you use an AI tool, you are also training it. Your prompts, your corrections, your preferences, your data, all of it feeds back into the system. You refine your workflow around the model's strengths and work around its weaknesses. You start thinking in terms of what the AI can do rather than what you actually need. Slowly, imperceptibly, the tool shapes the user.

This is not unique to AI. Every powerful technology reshapes the people who use it. But the speed and intimacy of AI interaction makes this feedback loop unusually tight. When your writing assistant suggests a phrase, do you accept it because it is better than what you would have written, or because it is easier? When your AI agent handles your email, are you delegating or abdicating?

The most productive relationship with AI is one where you remain the one with intent. You set the direction. You define what "good" looks like. You decide when to accept the machine's suggestion and when to override it. The moment you stop doing that, you have not gained a tool. You have gained a dependency.

The one thing I wish AI was better at? Honestly, I wish it was better at making us better. Not faster, not more efficient, not more productive. Better. Better at thinking, better at asking questions, better at sitting with uncertainty. The technology is impressive. The question is whether we are using it to amplify our best qualities or to outsource them.

References