Skip to content

Podcast

How to Learn AI With AI

The AI Daily Brief: Artificial Intelligence News and Analysis

Source ↗ ← All highlights
  • Learning Is Now AI-First
    • Learning has shifted from instructor-led tutorials to pairing directly with AI as a build partner.
    • High-agency learners who master AI collaboration will shape the next generation of work. Transcript: Nathaniel Whittemore Specifically, we’re talking about how to learn AI with AI. But the genesis for this is that I think that the way that learning is going to happen has fundamentally shifted. Instead of a paradigm of instructor-led tutorials, explainer videos, step-by guides, basically that entire former paradigm of education and particularly online education, instead Now everything is going to be effectively the equivalent of pair learning with an AI build partner. AI, in other words, is going to be your companion for using AI to learn. And it turns out there’s a lot to figure out about how to do that well. Now, I want to give a little bit of specific context on why this is coming up right now. First and most important is that just after OpenAI announced 5.3 Codex, President Greg Brockman talked about how the company was endeavoring to work in a fundamentally different Way. He tweeted, By March 31st, we’re aiming that for any technical task, the tool of first resort for humans is interacting with an agent rather than using an editor or terminal. In other words, agent first work by March 31st. You might have noticed that has kind of a ring to it. And something I’ve been thinking about a lot recently anyways, is how to give people better resources for self-directed learning around what I see as this shifted paradigm of AI. (Time 0:01:01)
  • Tribe’s Team Needed Engineer Help
    • Tribe’s team used Claude Cowork to do tasks previously impossible for them, but needed engineers to get results.
    • Jacqueline Rice Nelson’s post highlighted product and UI gaps for agent adoption. Transcript: Nathaniel Whittemore Now, to be clear, Jacqueline is great, Tribe does awesome work, and the broader point of her post, which is that the UI and the products around agents need to improve dramatically for Them to be widely adopted, especially in a work or enterprise setting, I absolutely agree with. Her post was about how a bunch of her non-technical team members and herself had used Claude Cowork to do things that were impossible for them just a few months ago. But when they actually dug in, it was quite difficult. And in fact, many of the team had actually paired with engineers for hours to get the output that they eventually got. (Time 0:02:47)
  • Built Agents Without Coding Skills
    • Nathaniel built seven agents with OpenClaw despite being non-technical.
    • He relied on Clawed and other tools to persevere through technical challenges. Transcript: Nathaniel Whittemore I have dozens of live projects on Lovable, a bunch of things that I’m building in Clawed Code, seven agents that are actively interacting with me via OpenClaw that I built over the past Week, and I don’t know how to code. I am completely and utterly non-technical. What I have is Clawed to help me work through things step- figure them out, and persevere even through challenges that might otherwise have stopped me. But I realized that how to work with an AI learning partner like that is not self-evident. (Time 0:04:12)
  • Start With Vision, Not Task
    • Start with the vision and goals rather than a narrow task when working with AI.
    • Giving broad context helps the AI align with what you actually want to achieve. Transcript: Nathaniel Whittemore Number one, you got to start with the vision, not the task. The watchword for AI in 2026 is, of course, context. And when it comes to building like this, the context that the AI needs is the big idea of what you’re trying to achieve. That means instead of saying, help me build a learning platform to help people launch their first agents, instead you start with your goals and your perception of what does or doesn’t Exist out there and what the challenges are. It might feel slow, but I guarantee it’s going to save time on the other end and get your AI partner way closer to what you’re trying to actually achieve than just trying to describe the Outcome alone. (Time 0:05:12)
  • Work Out Loud Even If It’s Messy
    • Speak messy, incomplete thoughts out loud to your AI; it can handle half-formed ideas.
    • Use the AI to help structure and refine emergent concepts instead of waiting for polish. Transcript: Nathaniel Whittemore One of the things that I realized earlier today is that I was actually building two things at once. One was a set of self-directed skills projects that people could combine in whatever way that made sense. The second was a library of agent starter prompts that people could just download. I think the exact line to Claude was, okay, not to be insane, but am I building two things at once? Your AI partner has the capability to handle that sort of messiness. It doesn’t need perfectly formed thoughts to be useful. In fact, much of its utility is in helping you think through half-formed thoughts. (Time 0:05:52)
  • Push Back And Demand Critique
    • Push back often and demand critique from the AI to avoid blind acceptance.
    • Ask the model to critique from first principles when you need it to challenge your ideas. Transcript: Nathaniel Whittemore AI doesn’t have feelings in the way that your employees or colleagues do. To the extent that it wants anything, it wants to help you achieve whatever it is that you’re setting out to do. And one of the things that anyone who’s used AI knows is that it says everything pretty confidently, which means you have to push back. Now, an inverse of this, which is a little bit better with current models, but which is still a little bit of a challenge, is you also want AI to push back on you. And sometimes that involves explicitly saying, I’m not sure about this, I want you to critique it from first principles or something like that. The point is that the conversation can’t be the AI just accepting your ideas as good or you accepting the AI’s ideas as good. (Time 0:06:26)
  • Dump First, Organize Later
    • Dump unstructured thoughts first and let the AI organize them later.
    • Treat the model as a tool that excels at turning chaos into structured plans. Transcript: Nathaniel Whittemore Once again, you don’t need to have everything perfectly structured. And in fact, a lot of what AI is good at is taking your messy, disorganized and unstructured thoughts and structuring it in ways that can help you make progress. (Time 0:07:13)
  • AI As A Reflective Mirror
    • Use the AI as a mirror to refine your own ideas rather than relying solely on it for novelty.
    • Speaking your ideas aloud and having the model play them back reveals hidden gaps. Transcript: Nathaniel Whittemore Sometimes you need the AI to generate a net new idea. A lot of times you need to speak an idea to it and have it play it back for you to make sure it makes sense. And the lesson here is that you know more than you think you do. You don’t have to rely on the AI for all the new ideas. A lot of its job is to help you work through your own. An example from our building earlier today, we were trying to talk through what the categories would be for the agents on the agent bench portal, where you can download this starter Template to start working through building your own agents. And after the AI gave me a set of categories that it thought made sense, I fed it back the seven that I had built with OpenClaw this week to see how they would fit together. It ended up revealing a couple of gaps in the framework that I didn’t consciously catch, but which something else that I had built actually revealed. (Time 0:07:27)
  • Zoom Out To Reground Goals
    • Regularly zoom out and revisit the existential goals of your project.
    • Pulling out of weeds helps you and the AI stay aligned on the true objective. Transcript: Nathaniel Whittemore Every once in a while, especially the deeper into the weeds you go, it’s really valuable to zoom all the way out and reground yourself in what you’re actually trying to build. I can’t even tell you how many different versions of AIDB training have existed in my head. And even as I’ve been designing this project, it’s shifted. It is very, very easy to get lost in the sauce and knee-deep in the weeds. And to the extent that you can pull yourself out every once in a while, it’s going to help you and your AI build partner reground yourself in what you’re actually trying to accomplish. (Time 0:08:15)
  • Go Wide First, React Later
    • Let the AI draft broadly and then react to many outputs instead of perfecting one draft.
    • Use the model’s near-infinite output to go wide first and filter later. Transcript: Nathaniel Whittemore This is another one that I think is sneakily difficult for people because it’s not how we’re used to thinking about things. I think a lot of the first way that people used AI was they drafted stuff and then they had the AI comment. I think increasingly we’re shifting in the other direction where the flow that makes the most sense is to let the AI draft and then to react. Take advantage of that near infinite output capacity to go wide first. Today, as I was thinking through what skills projects I would want to have to start off, I asked the AI to write a slew of initial titles based on our categories, and it came back with 110 In about 30 seconds. I was able to very quickly spot that there were certain patterns and trends inside those that weren’t really going to work, and we went on from there. (Time 0:08:46)
  • Know When To Stop A Thread
    • Know when to stop a conversation thread and move on to avoid endless rabbit holes.
    • You are the project manager of the conversation; decide what matters now vs later. Transcript: Nathaniel Whittemore The AI will walk with you down any rabbit hole as far as you want to go. And pretty much the only time that you hear from an AI something like, hey, do you think we should move on, is when it’s coming to the end of its context window, and that’s its equivalent Of telling you that it’s tired. In general, it will happily go as deep as you want on just about anything, which means it’s your job to manage the session and decide what matters now versus what matters later. You also are allowed to temporarily diverge and then come back. I had a tangent that I knew as a tangent that I just wanted to think through in the form of a single question. And then after I got the answer to that question, I said, let’s willfully ignore that for now. We’ve got enough things to think through. Remember, at the end of the day, you are the project manager of the conversation. Your AI partner is going to follow you wherever you lead. (Time 0:09:30)
  • Treat Sessions As Shift Handoffs
    • Write handoff documents at the end of sessions to capture decisions, open questions, and current state.
    • Treat each working session like a shift handoff so future conversations preserve context. Transcript: Nathaniel Whittemore The first and maybe singularly most useful thing in anything that I’ll say today is handoff documents. AI conversations have limits. Long sessions accumulate a lot of shared understanding that exists only in that current conversation. If you don’t capture it, you start from zero next time. Yes, all of the platforms have some version of memory, but it’s very nascent, it’s very unreliable, and it requires you, at least at this moment, this is the type of thing that in three Months when someone’s listening to this could be entirely irrelevant, but at least in this moment, you have to explicitly capture the context before you move to a new conversation. What you will find is that when you start to get into a complex project, it will not take you as long as you think to get to the end of the context window. Before you do, or especially as it’s starting to happen, as you start to see those telltale signs where the AI forgets a detail or starts to get lazy or whatever it is, it just feels like You’ve been talking to it for a long time. Scroll bar on the side has gotten little and tiny because there’s so much there, ask it to write a handoff document that captures the key themes, the decisions, the open questions, and The current state of the project, whatever type of project it is. You have to kind of treat every working session like a shift handoff. You document what was decided, and in many cases, the process that got you there, because that’s really important context too, as well as what’s still open and what comes next. (Time 0:10:24)
  • Use Project Workspaces For Context
    • Use your LLM’s project or workspace features to store files and persistent context.
    • Keep architectural plans, handoff notes, and setup documents inside the project for continuity. Transcript: Nathaniel Whittemore So for example, you saw here that I’m using clawed projects. I have projects for major buckets of work. Each of them has their own set of conversations as well as their own set of files. And you can see, this is the OpenClaw agent project, that a lot of the files are handoff plans, setup plans, architectural plans, basically the additional context that future instances Of that LLM are going to need to continue to help me without losing too much in translation. (Time 0:11:56)
  • Share Screenshots Not Paraphrases
    • Use screenshots when sharing visual content, errors, or code with your AI.
    • The model can read images, so send terminal screenshots to diagnose problems faster. Transcript: Nathaniel Whittemore This next one is kind of obvious, but for the sake of completeness, don’t forget that all these models, they can look at stuff in the form of screenshots. And while screenshots are obviously useful for if you’re working on anything visual, like a design or a layout, you can also screenshot an error message, a snippet of code. In short, remember that your AI learn slash build partner can read stuff in an image as easily as it can when you copy paste it in. Especially as I’ve been setting up OpenClaw, my conversations with my clawed partner are basically nothing but screenshots of the terminal where I’m effectively saying, what the Heck does this even mean? (Time 0:12:28)
  • Copy-Paste Exact Content
    • Copy and paste exact error messages, code, and UI text instead of paraphrasing.
    • Exact content helps the AI diagnose technical problems far better than summaries. Transcript: Nathaniel Whittemore Just copy and paste stuff. When you’re talking about error messages, parts of a UI you don’t like, a code snippet, a paragraph from a document, do not paraphrase it. Don’t summarize it, especially if it’s a technical problem. Your AI partner can work with exact content far than it can with your memory of it. As insane as this sounds, copy-paste is a core skill of learning to learn with AI. (Time 0:13:08)
  • Use One AI To Prompt Others
    • Use your primary AI partner to write prompts and specs for other AIs you use.
    • Review the generated prompt to ensure it accurately represents what you want done. Transcript: Nathaniel Whittemore You’ve got your AI build slash learn partner who’s helping you coordinate the whole thing. You might have a different LLM that’s helping you with some other parts. For example, if you’re using Claude, it doesn’t have image generation, so you might be using Gemini’s Nano Banana Pro. And then, of course, if you’re using a build tool like Claude Code or Lovable or Replit or anything else like that, you’re going to be moving context and content around between a lot of Different AIs. Use your AI partner to write the prompts for your other AI partners. Get in the habit of explaining to your AI partner what you want one of the other AIs to do and have it write the spec or the prompt or whatever it is that’s needed to communicate that. Not only is it going to be more precise, it’s going to be a lot faster to have it do the writing. Now, the one caveat proviso addendum to this is that as you do a lot of this, it can get very easy to just assume that what your AI Learn partner has written is correct and fully representative Of what you’re trying to communicate. Take the time to click the file that it wrote, scroll through it, and make sure that it accurately says what you want it to. You would not believe the number of times, for example, this week that it tried to switch the models that I was using on me when initially, least I just wanted everything in Opus. (Time 0:13:41)
  • Avoid Restarting Conversations
    • Avoid starting a new conversation unless you must, because you discard accumulated context.
    • Preserve rejected ideas and prior paths since they provide valuable future context. Transcript: Nathaniel Whittemore Sometimes when something isn’t working, it feels like the best idea would be to start a new conversation. And certainly in some cases that’s right, but when you do that, remember that you’re throwing away a lot of accumulated context, not just in terms of decisions, but things you’ve thought Through and rejected, ways of looking at the problem that haven’t worked out. That can be really valuable, so you really have to have a very high burden to just start from scratch. (Time 0:15:08)
  • Talk To AI, Don’t Type
    • Prefer speaking to the AI instead of typing to accelerate the workflow.
    • Use better text-to-speech tools like Whisperflow to speed up sessions significantly. Transcript: Nathaniel Whittemore Now, unfortunately, you probably know that the native text to speech in your devices isn’t very good. Luckily for all of us, there are an increasing number of tools that are much better. I use Whisperflow, and while it’s not perfect, I literally would be moving about a third as fast if I didn’t have it. The single biggest speed pickup that I can offer you probably is making the switch from typing to talking. (Time 0:15:47)