"Learn AI" sounds responsible, but it is too vague to act on. It is like telling someone in 1998 to "learn the internet." Useful, yes, but not specific enough to turn into a career, a business, or a weekend project.
Greg Isenberg's argument is better: build a skill stack that becomes more valuable as AI improves. Not one magic tool. Not one prompt trick. A stack of practical capabilities that let you design, ship, explain, distribute, and organize work in an AI-heavy world.
Source note
Credit for the source idea goes to Idea Browser and Greg Isenberg, host of The Startup Ideas Podcast. You can find Greg at gregisenberg.com and on X/Twitter.
This post uses the video and transcript as the main commentary source, then adds a JQ AI SYSTEMS operator lens: what each skill means for builders, consultants, founders, and small teams that need useful workflows instead of vague AI fluency.
Why learn AI is vague
The phrase "learn AI" hides too many decisions:
- Do you mean prompting?
- Do you mean using Claude, ChatGPT, Codex, Gemini, or local models?
- Do you mean building agents?
- Do you mean marketing with AI?
- Do you mean robotics and physical automation?
- Do you mean building a business around a workflow?
The better question is: which human capability becomes more valuable when AI gets cheaper, faster, and more capable?
That framing changes the answer. You stop chasing every launch and start building muscles that compound.
The 6 durable skills
1. Running agents and local models
This is the grown-up version of prompt engineering. A real agent is not just a chatbot with a nice prompt. It has context, tools, permissions, memory, a goal, and a way to check its own work before it bothers a human.
The business version is simple: customer support agent, research agent, sales follow-up agent, reporting agent, inbox triage agent. The valuable person is the one who can define what each agent may read, what it may draft, what it may send, what needs approval, and how success is measured.
Local models matter because some jobs need privacy, low marginal cost, offline resilience, or control. Tools like Ollama and LM Studio make local experimentation much more approachable. Ollama is useful when you want programmable local/open-model workflows. LM Studio is especially friendly for downloading, testing, and serving local models from a desktop interface.
First rep: build a daily briefing agent with three sources, your calendar, a folder of notes, and an approval rule before it sends anything.
2. Marketers who build distribution
Distribution is not posting. Distribution is knowing where attention already lives, what people are anxious about, what exact words they use, and how to create trust before asking for action.
In an AI world, building gets easier. Demand does not. The marketer who wins is part researcher, part storyteller, part media operator, and part community builder. They can turn one insight into a landing page headline, founder story, short video, newsletter angle, demo thread, and sales conversation.
First rep: choose one niche and map 20 places where its attention lives: newsletters, creators, Reddit threads, Slack groups, podcasts, events, search terms, tools, and communities.
3. Robotics engineers who build and source hardware
Greg's strongest non-obvious point is robotics. The last decade rewarded people who moved pixels. The next decade may reward people who can move atoms too.
This does not mean everyone needs a robotics PhD. It means the entry point is changing. Open-source robotics projects, cheap cameras, low-cost arms, smaller vision-language-action models, and shared datasets make physical automation more learnable than it used to be.
The Hugging Face LeRobot ecosystem is worth watching. The SO-100 and SO-101 docs show how concrete this has become: motors, controller boards, calibration, ports, assembly, and training workflows. The SO-ARM100/SO-101 repo also documents build-or-buy paths and vendor options.
First rep: assemble or buy a low-cost arm, add a cheap camera, and teach it one boring task like sorting three objects or moving an item between trays.
4. Curators who make short-form video
Curation is not dumping links. The valuable curator watches the timeline and says: this matters because.
This skill is rising because information is abundant and trust is scarce. People do not only need more news. They need someone with taste to translate what a launch, demo, policy change, model release, robotics clip, or pricing update means for their specific niche.
Greg calls this "yapping" in the positive sense: raw, opinionated, useful, lightly entertaining short-form explanation. It works because it feels more human than synthetic content and because it carries a point of view.
First rep: run a seven-day curation sprint. Pick one lane, find three things per day, and publish one short video using: "I saw this. Most people think it means this. I think it actually means this. Here is the move."
5. The builder distributor
The old startup split was clean: one person builds, another person sells. AI compresses that. One person can prototype the app, write the launch copy, record the demo, DM the first users, create the clips, and iterate from feedback.
This is the skill behind many credible one-person-business conversations. Not because one person magically does everything perfectly, but because the handoff cost disappears. The builder distributor can close the loop faster than a team that separates product from market learning.
First rep: do a 48-hour loop. Build the smallest version of one problem with AI, then create 10 distribution assets before you feel ready: one demo video, three short clips, three posts, two DMs, and one landing page.
6. IRL community builders
This one feels old-school, which is exactly why it matters. As content, software, and advice become abundant, scarcity moves toward belonging, trust, and context.
A good community builder knows how to pick the right room, invite the right mix of people, set one sharp question, create a ritual, and make people want to come back. The room becomes a network. The network becomes a recruiting asset, media asset, deal-flow asset, and friendship layer.
First rep: host six to eight people around one sharp question, then send a short recap with best quotes, ideas, and follow-ups.
Weekend reps
The best part of Greg's framework is that every skill has a small first rep. Here is the practical version:
| Skill | Weekend rep | Proof that it worked |
|---|---|---|
| Agents and local models | Build a daily briefing agent with sources and approvals | It saves 10 minutes or catches something you would have missed |
| Distribution | Map 20 attention channels and write 20 hooks | You find language real buyers already use |
| Robotics | Assemble or study a low-cost arm workflow | You understand one physical failure mode better than before |
| Curation | Publish seven short videos in one niche | You develop a repeatable take structure |
| Builder distributor | Ship a tiny product and 10 distribution assets in 48 hours | You get a real reaction from a real person |
| IRL community | Host a small dinner, walk, breakfast, or meetup | People ask when the next one is |
JQ scorecard
If I were advising a founder, consultant, or operator on where to start, I would score the six like this:
| Skill | Client value | Speed to learn | Business fit | JQ take |
|---|---|---|---|---|
| Agents and local models | 10/10 | 7/10 | 10/10 | Best first pick for automation consultants and operators. |
| Distribution | 10/10 | 8/10 | 10/10 | Every AI product still needs demand. |
| Robotics and sourcing | 8/10 | 4/10 | 8/10 | Harder, but more defensible if you stick with it. |
| Curation and short-form video | 8/10 | 9/10 | 8/10 | The fastest way to build market taste in public. |
| Builder distributor | 10/10 | 6/10 | 10/10 | The highest-leverage founder skill stack. |
| IRL community | 9/10 | 7/10 | 9/10 | Trust gets more valuable as content gets cheaper. |
The missing layer
I would add one layer to the six skills: verification discipline.
AI makes it easier to build, publish, automate, research, and organize. That increases the value of people who can say:
- what is the source?
- what is the success metric?
- what should be logged?
- what needs approval?
- what should not be automated yet?
- what evidence would change our mind?
This is why agent architecture, distribution, and community fit together. An agent can do more work. Distribution can get more feedback. Community can give richer context. Verification keeps the loop honest.
Do not try to learn AI as one giant subject. Pick one skill stack, run one weekend rep, and keep the result only if it creates a useful workflow, a real audience signal, or a stronger network.
How to choose your stack
Here is the simple filter:
- If you sell services: start with agents and local models, then add distribution.
- If you want to build a startup: start with builder-distributor, then add curation.
- If you want a durable technical moat: start with robotics and sourcing, then add agents.
- If you want career leverage: start with agent operations inside your current job.
- If you already have an audience: add product building and turn attention into feedback.
- If you feel isolated: build the IRL room first. The ideas may come from the people in it.
You do not need all six. One makes you useful. Two gives you leverage. Three makes you rare.
Sources
- Greg Isenberg: "Learn AI" Is Bad Advice. Learn This Instead
- Idea Browser
- Greg Isenberg
- Greg Isenberg on X/Twitter
- Ollama
- LM Studio
- Hugging Face LeRobot
- LeRobot SO-100 docs
- LeRobot SO-101 docs
- TheRobotStudio/SO-ARM100
- Alibaba supplier marketplace
The short version: do not learn AI in the abstract. Learn how to create useful agents, get attention, move between software and hardware, curate signal, ship and distribute, and bring people together. That is a skill stack AI makes more valuable, not less.