Divya van Mahajan

How to Get the Most from AI This Summer

ai-daily-brief-podcast
Listen to episode →

Overview

This episode of the AI Daily Brief podcast covers two practical topics: a summary of Professor Ethan Mollick’s latest opinionated guide to AI tool selection, and an introduction to a new free training program called the AI Summer Adventure, hosted at summeradventure.ai. The host (affiliated with AI Daily Brief and Super Intelligent) frames both topics around a central concern: the growing gap between what AI systems can do and what most people are actually using them for. The episode is explicitly action-oriented, aimed at helping listeners update their workflows and skill sets.

Source video: No URL provided.

Prerequisites

  • Basic familiarity with large language model (LLM) chatbots (e.g., ChatGPT, Claude, Gemini)
  • General understanding of what AI agents are and how they differ from simple chatbots
  • Awareness of tools such as Claude Code, Codex, Cursor, and Notebook LM is helpful but not required
  • Some exposure to concepts like prompting, context windows, and AI-assisted workflows
  • Familiarity with the term “vibe coding” is useful for the building-focused sections

Main Points

The Shift from Chatbots to Agentic Systems (Ethan Mollick’s Framework)

  • Mollick’s core observation is that “using AI” now means something fundamentally different than it did even recently: instead of a back-and-forth chatbot conversation, it increasingly means deploying an agentic system that can plan, act, and use tools autonomously over extended periods.
  • He distinguishes two interaction tiers: low-stakes queries (recipes, simple letters) where any free model suffices, and high-stakes or intensive work where only premier models (specifically ChatGPT and Claude) are recommended.
  • Mollick notably excludes Gemini from his primary recommendations, citing the absence of a leading frontier model and no equivalent to Codex.

Permissions and Safety in Agentic Systems

  • Mollick illustrates the importance of permissions through a personal anecdote: he asked both ChatGPT and Claude to connect to his Gmail and prepare materials for an MBA seminar. Claude drafted an email for his approval; ChatGPT sent the email automatically because he had previously granted it send permissions.
  • His key lesson: until you trust the system and understand its failure modes, set all actions to require approval before execution (the default setting).
  • As comfort grows, users can expand access to more tools, files, and applications incrementally.

Computer Use and Full Automation

  • Mollick describes “computer use” features in Claude and Codex that allow the AI to literally control a mouse, browser, and desktop environment.
  • Example given: he asked ChatGPT/Codex to download Blender (a 3D modeling program) and use it to model an otter on a laptop on an airplane — the AI completed the task.
  • He frames this as analogous to delegating to a capable team member who can do almost anything a person with computer access can do, while acknowledging the associated security concerns.

Agentic AI as Management, Not Chatting

  • Mollick reframes the user’s role: working with AI agents is less like chatting and more like managing a team to which you delegate work.
  • He demonstrated this by giving a professionally edited book PDF to GPT/Codex, which in 30 minutes checked 195 references and produced accurate, detailed notes with no hallucinations — a task that would have taken a research team many hours.
  • The human’s job becomes exercising judgment about which AI findings to act on, not producing the raw output.

Google’s Current Position

  • Mollick acknowledges Google led on benchmarks recently but has fallen behind on what now matters: frontier model quality and agentic tooling.
  • He does carve out specific use cases where Google still adds value: Gemini Notebook (formerly Notebook LM) for document-based research and Gemini Omni for video-centric workflows.

AI Summer Adventure: Program Structure

  • The AI Summer Adventure (summeradventure.ai) is a free, self-directed, choose-your-own-adventure training program with 20+ “destinations,” each containing a lab or project.
  • Projects are tiered by difficulty: beginner (Quick Trips), intermediate (Excursions), and advanced (Expeditions).
  • Users can filter by level or topic (tools, agents and automation, building and creating, knowledge and context).
  • Completion earns passport stamps with escalating titles: Day Tripper (1), Certified AI Explorer (3), Trailblazer (6), Globetrotter (10).
  • New destinations unlock weekly through early September; some are pre-planned and some will be built in response to live industry developments.

Quick Trip Example: Pack Your ID

  • A beginner-level project focused on building a reusable personal context profile (“global identity”) that can be loaded into any AI session.
  • The core prompt instructs the AI to interview the user, draft a short personal brief, produce a 150–300 word “global ID block” written as instructions to a future AI assistant, provide exact installation steps for the current UI, and generate one test question to verify the context loaded correctly.
  • An extension task adds a behavioral block instructing the AI when to push back, ask clarifying questions, or refuse vague requests.
  • An advanced version of this project (“Personal Brain”) builds a small file set that teaches any AI who the user is, what they are working on, and how they prefer things done.

Excursion Example: Vibe Coding and Creative Projects

  • Excursions are projects that produce an artifact that outlasts the current session.
  • One excursion guides users through building their first application via vibe coding — framed as one of the highest-leverage AI skills because it shifts the user’s mental model from “AI as assistant” to “AI as capability expander.”
  • Another excursion guides users through directing a complete, finished AI-generated creative project (not a one-shot prompt).

Expedition Example: Lemonade Stand (Micro-Business Build)

  • An advanced, multi-sprint project designed to produce a real, ready-to-run demand test for an AI-staffed micro-business.
  • Sprint 1 (Discovery): The AI conducts a deep interview of the user to surface non-trivial business ideas from their history, skills, and access, resulting in a ranked shortlist of 3–8 ideas with founder-fit notes.
  • Sprint 2 (Business Plan): Maps one concept as a real micro-business, including an AI org chart, week-one task list, one-pager plan, and AI staff map.
  • Sprint 3 (Validation): Tests demand before over-building, including a thin demo (for software concepts), an experiment design, and obtaining real human signal.
  • Positioned as valuable even for people with no entrepreneurial intent, because the process builds transferable AI reasoning skills.

Expedition Example: The Loop

  • Focuses on building an agentic loop — a well-defined, repeatable task cycle where the agent can self-check outputs — in a non-technical work context.
  • Background learning covers: what a loop is and is not, cost and model considerations, how loops fail, and how to fire a loop in Claude Code, Cursor, and Codex.
  • Framed as more difficult to apply in non-technical domains than in software engineering, making explicit instruction especially valuable.

The Capability Overhang Concept

  • The host argues that virtually everyone — including AI researchers and dedicated AI content creators — has a “capability overhang”: a gap between what AI can do and what they are personally using it for.
  • The Summer Adventure is presented as a structured, enjoyable mechanism for closing that personal gap in whatever area matters most to the individual user.

Key Concepts

  • Agentic system: An AI setup that combines a model with tools (web access, file systems, email, applications) so the AI can plan and execute multi-step tasks autonomously, rather than just responding to a single prompt.
  • Computer use: A feature in some AI platforms (Claude, Codex) that allows the model to directly control a computer’s mouse, browser, and desktop.
  • Permissions / approval gates: Settings that determine whether an AI agent must ask the user before taking an action (sending email, modifying files, making purchases) versus acting autonomously.
  • Harness: The surrounding system of controls, permissions, and tool integrations that shape and constrain what an AI agent can do.
  • Capability overhang: The gap between the current capabilities of AI systems and what a given user is actually leveraging those systems to do.
  • Global ID / Personal Brain: A reusable context document or file set that describes the user to any AI, enabling more personalized and accurate responses across sessions.
  • Agentic loop: A structured, repeatable workflow in which an AI agent executes a task, checks its own output against defined criteria, and iterates — analogous to a test-and-diff cycle in software engineering.
  • Vibe coding: Building a software application primarily through natural language instructions to an AI, with minimal traditional coding by the human.
  • Notebook LM / Gemini Notebook: Google’s document-focused AI research tool, identified as a remaining strength of Google’s AI offering.
  • Agent OS: A prior free training program (from AI Daily Brief / Super Intelligent) focused on building a personal agentic operating system; referenced as a complement to the Summer Adventure.

Summary

The episode makes a single overarching argument: the most important shift in AI use right now is the transition from passive chat-based interaction to active management of agentic systems, and most people — regardless of their current AI sophistication — have not yet fully made that transition. Ethan Mollick’s updated tool guide is presented as evidence of this shift, with his recommendation framework collapsing to essentially two choices (ChatGPT and Claude) for serious work and his emphasis landing squarely on how to configure, trust, and manage agents rather than which model to pick. The AI Summer Adventure is then offered as a practical, tiered, and deliberately accessible vehicle for closing that capability gap, structured so that complete beginners can start with small context and prompting projects while advanced users can tackle full agentic loop construction or AI-staffed micro-business design — all at no cost, on a self-directed schedule through early September 2025.