Como aprendería a usar IA en 2026 (desde 0)

The speaker, with 4+ years of professional AI experience, identifies three critical mistakes beginners make when using artificial intelligence: (1) constantly switching between tools instead of mastering one primary platform like ChatGPT, Claude, or Gemini; (2) providing insufficient context before asking AI to complete tasks, leading to generic and ineffective outputs; and (3) accepting AI-generated first drafts without review and iteration.

The solution framework consists of three progressive levels. Level One involves selecting a single trusted AI tool, always providing rich context before making requests, and reviewing outputs by asking two questions: 'Is anything here untrue?' and 'Would I say this this way?' This manual approach already surpasses most users' AI productivity.

Level Two eliminates manual repetition by treating AI interactions as task assignments rather than one-off prompts. A proper task has three components: what you want, what context you provide, and how success is measured. Context comes from three sources—established frameworks (like SWOT analysis), real-world examples, and connected tools (email, calendar, Drive). Crucially, AI can autonomously iterate when given verifiable criteria (word count, specific dates, rule compliance), checking its work until standards are met. This approach is stored in 'Projects' (GPT Projects in ChatGPT, similar features in Claude and Gemini) that permanently house instructions, files, examples, and external tool connections.

Level Three moves beyond AI waiting for instructions to AI working proactively and autonomously. Systems like Google's Spark access multiple information sources simultaneously, cross-reference data without manual input, complete complex analytical work, and present results for human confirmation before acting. The speaker demonstrates this with an example where Spark analyzed 42 company documents and automatically created an executive report and functional dashboard with multiple cross-referenced tabs.

An additional innovation discussed is the 'LLM Wiki' concept proposed by Andrej Karpathy, where AI organizes documents into an automatically-updated private knowledge base (demonstrated with Obsidian vault), creating interconnected information that improves with each new document added. This represents a meta-layer where AI systems automate their own context management, learning better over time.

The speaker emphasizes that prompt quality matters far less than commonly believed; instead, success comes from surrounding context, clear verification criteria, and system architecture. The progression is not about reaching Level Three immediately but understanding your current level and advancing methodically while sound judgment guides implementation throughout.