Following the Frontier

The frontier of AI is changing quickly. One way to keep up is to establish intellectual proxies: people whose knowledge and judgment you trust in a particular domain.

We have our own opinions about who is worth following at the AI frontier: people who help us notice important developments, understand what they mean, and decide what is worth trying. But the right people to follow depend on your interests, and you should decide for yourself who deserves your trust. This guide introduces some of the people we follow, then helps you build and evaluate a list of your own.

Who do we trust?

Here are three kinds of people we trust at the AI frontier:

Practical experimenters try new capabilities in real work and explain what happened. Explore Simon Willison for constant hands-on LLM experiments, Claire Vo’s How I AI for short, copyable demos of real workflows, and Dan Shipper’s AI & I for screen-shared conversations about how people use AI to think, create, and work.
Operators who publish their methods show the workflows and artifacts behind their results. Explore Every’s guides for practical AI products and workflows. Watch AI That Works for live experiments with coding agents, context, and production AI systems. The AI Engineer talks channel collects more presentations from practitioners working at the frontier. Watch Matt Pocock for reusable agent skills and disciplined workflows.
Measured field reporters discuss failures, constraints, and tradeoffs alongside the wins. Explore Ethan Mollick for sustained observations of AI in professional, educational, and creative work, and AI as Normal Technology for reality checks on inflated claims.

These people do not always agree, and you do not need to agree with everything they say. What matters is that they show their work, so you can evaluate their claims for yourself.

What makes a good frontier guide?

A trustworthy frontier guide tends to:

show the artifact, experiment, code, data, or primary source behind a claim;
distinguish what happened from what they infer might happen next;
explain the conditions, costs, failures, and limits;
update when the evidence changes;
and leave you with a method you can understand or test yourself.

A polished demo, a large following, or a confident prediction can still be useful—but none is evidence by itself.

So how can you build an information diet that keeps you inspired without letting hype control what you do?

Find your frontier guides

You do not need a hundred accounts in your feed. A small group of people with different relationships to the work—a builder, an explainer, a field reporter, and a skeptic—will usually give you a better view.

This prompt will find trusted voices based on your work and interests, then recommend a compact set of sources and explain what each one is good for.

Prompt

Help me build a small, high-quality information diet for the parts of AI and agentic work that are most relevant to me.

First inspect the goals, projects, files, and past conversations you can access so the recommendations reflect work I am actually doing. Then research current primary sources online.

Recommend no more than six people or organizations across these roles:

- a practitioner who builds and shows real artifacts;
- an explainer who makes the underlying ideas understandable;
- a field reporter who tests tools in real work over time;
- and a skeptic or evaluator who is good at finding limits and failures.

For each recommendation:

1. Explain the specific topic I should trust them to help me understand.
2. Link to one or two representative primary artifacts—not a profile or search page.
3. Show which trust signals are present: inspectable evidence, reproducibility, constraints, failures, costs, updates, or useful methods.
4. Name what they are not a reliable proxy for, or what bias I should keep in mind.
5. Connect their work to a concrete project, decision, or capability in my own life.

End with a minimal information diet: two sources worth checking regularly, the rest only when relevant, and a short explanation of what you deliberately left out. Do not create subscriptions or automations yet.

Test a claim before you adopt it

When an exciting post finds you, the first question is not “is this person smart?” It is “what exactly are they claiming, and what would count as evidence that it works under my conditions?”

Paste a post, video, product page, thread, or recommendation into this prompt. It will separate the evidence from the story, look for the strongest reality check, and design a cheap test you can run yourself.

Prompt

Evaluate the following frontier claim before I adopt it:

[PASTE THE LINK, CLAIM, TRANSCRIPT, OR SCREENSHOT HERE]

Open the original source and any evidence it links to. Follow important claims back to primary sources where possible.

Analyze it in this order:

1. State the exact claim in plain language. Separate observed results, interpretation, and prediction.
2. Inventory the evidence: working artifact, code, data, demonstration, user report, benchmark, citation, or assertion.
3. Identify the conditions under which the result occurred and what important information is missing.
4. Check the source's incentives and whether the evidence comes from someone selling the conclusion.
5. Find the strongest credible counterevidence, disagreement, or failure report. Do not manufacture false balance.
6. Explain whether the claim is relevant to my goals and setup, using the context you can inspect.
7. Design the smallest cheap, reversible test that would teach me whether it works for me.

Give the claim one verdict: ADOPT, TEST, WATCH, or IGNORE. State your confidence and what evidence would change the verdict. Do not make purchases, install tools, or change my setup yet.

Turn the frontier into a useful briefing

The goal of following the frontier is not to accumulate links. It is to notice the rare update that should change what you understand, try, or do.

This prompt will design a lightweight recurring brief that filters new ideas through your projects and is allowed to say that nothing worth changing happened.

Prompt

Design a recurring frontier brief for me based on my current projects, agent setup, and interests. If you can find trusted sources in this conversation or my files, use them. If not, recommend a starting source list based on my interests.

The brief should contain no more than three items per edition. Include an item only if it could change what I understand, try, buy, build, or stop doing.

For each item, require:

- the new idea or development in one sentence;
- a direct link to the best primary source;
- why it matters specifically to my work;
- what evidence supports it and what remains uncertain;
- whether it updates, contradicts, or merely repeats what we already knew;
- and one optional cheap test or question to pursue.

Prefer working artifacts and careful field reports over announcements, summaries, and viral commentary. Include a “nothing worth changing this time” result when that is the honest conclusion.

Recommend a cadence, source list, and compact output format. Draft the exact recurring instructions and one sample edition using current information, but do not schedule the automation yet.