Long, multi-step agent tasks
"Agentic AI: from chatbot to autonomous coworker" explained the error chain in long tasks - yet long agent tasks aren't science fiction anymore today. Coding agents already work through multi-step changes across several files on their own; research agents search and condense dozens of sources into a coherent report. The key difference from pure theory: these systems run with human checkpoints (see "Human in the loop: when AI needs approval"), not fully autonomously - and that's exactly what makes them practical already.
Multimodal understanding
"Multimodal AI: when AI can do more than text" shows that current models no longer process text, images, audio, and video separately, but combined in a single request. A model today can interpret a photographed circuit diagram, listen to a spoken error description, and evaluate both together - a few years ago that needed several specialized standalone systems.
Tool use
"What is MCP (Model Context Protocol)?" describes an open standard that's already made this an everyday thing: an AI that independently calls external tools and data sources, instead of answering only from its own training knowledge. This is one of the most concrete practical approximations of generality - not because the model itself gets more omniscient, but because it learns to fetch help in a targeted way.
Scientific and creative assistance
In tightly scoped fields, AI-assisted systems already speed up parts of real research work today - for instance, suggesting and pre-filtering chemical or biological candidates that experts then review. Important for an honest read: that's targeted acceleration of one narrow step, not independent scientific discovery by the AI alone.
Reasoning on complex problems
"Why some AI answers take longer" explains how reasoning models use more compute at request time to solve multi-step problems more reliably. On tasks that were considered "solvable only with human expertise" until recently - like multi-step mathematical proofs or complex debugging chains - these models now deliver usable first solutions.
Why this matters for you as a decision-maker
Public perception of AI swings between "can already do almost anything" and "still pure gimmick" - neither matches practice. The five capabilities here are genuinely deployable today, but tied to clear conditions: human checkpoints for agents, tightly scoped fields for scientific assistance, extra compute time for reasoning. Knowing those conditions lets you use AGI-adjacent capabilities productively today, instead of waiting for a single big breakthrough.