Learning Library

← Back to Library

Beyond Compression: AI for Deep Thinking

Key Points

  • Most people use AI mainly for compressing information—turning notes, long documents, or articles into concise summaries—rather than for deeper cognitive engagement.
  • The brain processes compressed content differently, so relying on AI-generated summaries can limit the formation of new mental connections and the transformative learning that comes from prolonged, focused study.
  • The problem isn’t AI itself but how we use it; for complex, high‑impact tasks we should allocate more “brain time” and use AI to reduce cognitive load, allowing us to think more deeply about the core subject.
  • AI can support deep work by handling routine compression tasks while we focus on expanding our understanding, as illustrated by the speaker’s challenge of writing a book on AI.
  • Mastering prompting is valuable but akin to learning to ride a bicycle—useful for quick tasks—where learning to “drive” AI as an intellectual partner offers greater long‑term productivity.

Full Transcript

# Beyond Compression: AI for Deep Thinking **Source:** [https://www.youtube.com/watch?v=p63MKDEsuFc](https://www.youtube.com/watch?v=p63MKDEsuFc) **Duration:** 00:12:59 ## Summary - Most people use AI mainly for compressing information—turning notes, long documents, or articles into concise summaries—rather than for deeper cognitive engagement. - The brain processes compressed content differently, so relying on AI-generated summaries can limit the formation of new mental connections and the transformative learning that comes from prolonged, focused study. - The problem isn’t AI itself but how we use it; for complex, high‑impact tasks we should allocate more “brain time” and use AI to reduce cognitive load, allowing us to think more deeply about the core subject. - AI can support deep work by handling routine compression tasks while we focus on expanding our understanding, as illustrated by the speaker’s challenge of writing a book on AI. - Mastering prompting is valuable but akin to learning to ride a bicycle—useful for quick tasks—where learning to “drive” AI as an intellectual partner offers greater long‑term productivity. ## Sections - [00:00:00](https://www.youtube.com/watch?v=p63MKDEsuFc&t=0s) **Beyond Compression: AI Partnership** - The speaker argues that using AI merely to compress information short‑circuits deep learning, and proposes leveraging AI to offload routine summarization so we can spend more brain time thoughtfully engaging with substantive content. - [00:03:29](https://www.youtube.com/watch?v=p63MKDEsuFc&t=209s) **Prompting as a Bicycle Skill** - The speaker likens mastering AI prompting to riding a bike—an essential, efficient skill—and explains how a conversational voice AI served as an interactive partner that kept ideas flowing, rather than merely transcribing them. - [00:07:02](https://www.youtube.com/watch?v=p63MKDEsuFc&t=422s) **Choosing AI Models as Cognitive Partners** - The speaker outlines how they evaluate and switch between AI models like Claude Opus 4 and Gemini 2.5 Pro to deeply critique outputs, clarify their core ideas, map their conceptual terrain, and manage partnership dynamics, noting a Substack guide they authored for model selection and their $200‑per‑month OpenAI subscription. - [00:10:33](https://www.youtube.com/watch?v=p63MKDEsuFc&t=633s) **AI as Thought Amplifier** - The speaker argues that thoughtfully selected AI tools, such as Opus 4, can deepen understanding and improve cognition when used deliberately rather than as a shortcut for mental effort. ## Full Transcript
0:00I want to suggest that most of us are 0:01using AI a little bit backwards. Stay 0:05with me. This is worth it. We are using 0:08AI primarily for information 0:11compression. So take my meeting notes 0:14and turn them into really neat outputs. 0:16Take this large product requirements 0:18doc, turn it into this for engineers and 0:20this for my executive stakeholders. Take 0:23this 100page PDF and turn it into 0:25something succinct. take Nate's super 0:27long substack and make it into something 0:29I can digest. I kid you not, people do 0:31that and it's kind of 0:32hilarious. The point is we're using it 0:34to compress information. But something 0:37that I heard that's been really sticking 0:39with me is the idea that the brain 0:43doesn't process compressed information 0:46in the same way. And one of the things 0:48that we need to learn to think about is 0:51when do we want to tolerate less brain 0:56time on a 0:58subject versus when do we want to 1:02actually optimize our partnership with 1:04AI? So we spend more brain time 1:06marinating in what really matters. 1:09So a lot of the learning that you get 1:12when you read a large book, a deep book 1:16on a big subject, it comes from your 1:19brain forming new connections as it 1:22spends extended time in the subject. If 1:25you get and you can a very short 1:28one-pager, you will get a prey, a 1:31summary, an executive briefing on the 1:34book. you are unlikely to have the kind 1:37of lifechanging experience that you had 1:40if you really dipped into it. Now there 1:43are people who will take that insight 1:45and say wow so AI is the problem. I 1:48actually 1:49think the way we use it is the problem. 1:53Not that it's bad to use it for 1:55compression. I too use it for adjusting 1:57my meeting notes. I too use it for 2:00sending different messages to different 2:01stakeholders for routine business 2:03communication. But if you need to do 2:06really deep 2:07thinking, you need to spend time on a 2:10subject, you need to optimize your 2:13cognitive workload to enable you to do 2:16your best thinking. And AI actually can 2:19really help with that. And I'll give you 2:20a specific 2:22example. I am contemplating writing a 2:24book on AI. Writing a book is a hard 2:27task. It's much harder, exponentially 2:29harder than a single article. 2:32it needs to be useful especially with AI 2:35some of the challenges in having 2:36something that stands the test of time 2:38etc and I find that when I'm 2:40contemplating writing a book the thing 2:43that is most difficult for me it's not 2:47the information compression it's not 2:49that I need it to go out and tell me all 2:51the news about AI I already get plenty 2:54of news trust me it's that I need to 2:57work with AI in a way that helps my 2:59brain expand and kick around the subject 3:02matter. And we don't really optimize for 3:06that when we only talk about prompting 3:08because prompting really optimizes for 3:11one-way 3:12communication. Learning how to prompt 3:14well is a skill, but it's sort of like 3:17learning to ride a bicycle versus 3:20learning to drive a car. Both are 3:22helpful. The car is going to take you 3:24farther if you learn how to do it well. 3:27And I think increasingly prompting is 3:29like the bicycle skill. It's incredibly 3:31efficient if you master it. Everybody 3:32should know how to do it. It should be a 3:35universal skill. I do think we're going 3:37to need it for a while. I think it 3:38provides a durable edge if you can ride 3:40well, if you can prompt well. And I love 3:43it. But if you can learn to actually 3:46cognitively partner beyond an individual 3:49prompt with AI, that's like driving a 3:52car. That's like actually going farther. 3:54And so in my case with the book, just to 3:56circle back, I spent 25 4:01minutes in advanced voice mode talking 4:04with Open AI. Uh, and it was I think 4:07they use like a variant of their 40 4:10model in advanced voice mode. And you 4:12might be waiting for me to say I had the 4:14most profound insights. Like the AI 4:17really helped me. It's it really wasn't 4:19that. In fact, I sometimes told it this 4:21is really vanilla. I don't love this. 4:24What it was that was distinct and 4:27special was it was there when I needed 4:30to talk out loud. It would let me talk 4:33out loud for a while. It actually 4:35listened. It actually took notes and it 4:37actually responded with just enough 4:40interest, engagement, and riffing to 4:43keep my brain flowing so I could keep 4:46the idea coming because I had an idea 4:48that I wanted to talk out. I knew I 4:50couldn't just dictate into a 4:52transcription device for 25 minutes and 4:55get it. But if I had someone talk back 4:57to me in the right cadence, I'd probably 4:59be able to fish it out of my brain 5:01because we're conversational people. And 5:03so a conversational AI was the right 5:06choice. In this case, advanced voice 5:08mode really has gotten better since they 5:10shipped uh that update that makes it, 5:13you know, more aware of when it 5:15interrupts you. uh it uses more 5:17discontinuities verbally like uh and um 5:20it just sounds more natural and it 5:22allows you to forget it's there and just 5:23focus on the subject which is what I 5:25needed to do. That's just step one. I 5:28did the verbal piece and then the second 5:29piece was I really needed to wrestle 5:33with what the terrain of the idea looked 5:36like and I knew 40 wasn't strong enough 5:39to do it. And so instead of trying to 5:42force it inside that conversation, I 5:44literally pulled the transcript out into 5:46a Google doc and I stuck it into 03 as a 5:50raw transcript and I said, "This is a 5:52transcript. This is my perspective on 5:54the transcript. This was the intent I 5:56had with this idea. Really, this is 5:58where I iteratively arrived." Uh because 6:01I find that a lot of time when you're 6:02talking out loud, you're sort of naming 6:04the work as you go. And that's part of 6:06the cognitive journey that we've kept 6:08inside our heads for 200,000 years. And 6:10now we have a partner to do it with. And 6:13it's true that like sometimes if you did 6:15paired work together with humans, you 6:17could kind of get to this. But I 6:18actually think it's a distinct dynamic 6:21with AI. I have not worked and I have I 6:24worked in offices, right? I've worked in 6:26person with teams. I love that dynamic. 6:28I don't want it to go away when we work 6:30well together with with good human 6:32colleagues. But it's not the same as a 6:35good AI colleague. An AI colleague that 6:38you can brainstorm with in that way 6:40where they listen and just riff with 6:42you. It's like this crossover between 6:45the way a therapist listens to you and 6:46the way a colleague listens to you. And 6:48you'd never expect a human to do that, 6:50but it's super helpful for your 6:51thinking. So we do that. We name the 6:54work. We get to 03 through the Google 6:57doc. 03. And honestly, it's not just 03. 7:00I don't want to just w Open AI here, 7:02right? Like you can go and do this in 7:04Opus uh the claude model opus 4. You 7:08could do this in Gemini 2.5 Pro. You'd 7:10get similar results. The idea is you 7:14want to think deeply and critique the 7:19model results that you get from the 40 7:22conversation. basically take my intent, 7:25take what I was able to articulate, take 7:27the riffing that advanced voice mode did 7:30and help me get to a crystal clear 7:33understanding of the heart of the 7:34idea. That's what you want in the next 7:37step. It's basically help me to define 7:39the coordinates of the terrain that I'm 7:41in now that I've named the work. And 7:44once we do that, the sky becomes the 7:46limit. We can open up to understanding 7:48what are the partnership dynamics that 7:50work from here. So, as a specific 7:53example, I thought 03 did a good job 7:56getting to the heart of the thesis for 7:59the book. I thought the outline felt a 8:01little bit heavy. I need to go into a 8:04different model. I'll probably choose 8:05Opus 4 for this to start to refine what 8:09that looks like a little bit more. I 8:10need to pick a different cognitive 8:12partner. And if you're wondering, how do 8:14I pick models? I put a whole thing up on 8:17Substack on how I sort of pick models 8:18for different tasks. 8:21Um, to me, I think the piece that I keep 8:25coming back to, stepping back, 8:26reflecting on all of this is that I find 8:31the most value. Like, this is real 8:33expense, right? Like I I'm 8:35pro-subscriber to open AAI, right? Like 8:38I pay my 200 bucks a month. Not 8:40everybody has to. I don't think you have 8:41to to get a ton of value out of it. I 8:43saw the article that said, "Is free so 8:45good that you don't need plus?" I think 8:46for a lot of people it is. But for me, 8:51whatever whatever your investment level 8:53is in AI, the value you get is so much 8:58greater if you use it in this way as a 9:01way of getting your brain time on 9:04subject. Optimizing for time on subject 9:08versus optimizing 9:11for just compressing and repurposing 9:14information. Most of the use cases I see 9:17from companies if they are not super 9:19fluent in AI end up being around 9:21compressing and repurposing information 9:23which is fine. There are some cost 9:25savings there but helping your brain 9:28work better has a lot more upside over 9:31time and I think we talk about it a 9:34whole lot 9:35less and I wish we would talk about it 9:38more. I wish we would not get so far 9:41into the compression trap that we don't 9:43think about this idea that AI can help 9:45us map and expand the cognitive 9:48territory that we work in by partnering 9:50with us in a way that is, you know, 9:52somewhat reminiscent of what we would do 9:54with a human, but also distinct. I I 9:57also, you know, just as I said, I 9:58wouldn't ask a human to do the kind of 10:01listening that I asked advanced voice 10:02mode to do. I also wouldn't really ask a 10:05human to work on sharpening my thesis 10:08for an outline the way I would ask 03 to 10:11do it. I wouldn't ask a human to go and 10:15like necessarily do the research and do 10:17the thinking um and the for lack of a 10:21better term cognitive shaping on the 10:23thesis with me. It's not just doing it 10:26on its own that I would ask of Opus 4. 10:30Opus 4 is great for shaping ideas and 10:33really thinking through concepts. And so 10:35I'm finding in a sense that my brain is 10:38able to think better by spending more 10:40time with AI models if they're carefully 10:43modulated. It's not just throwing 10:46prompts in there. There's no secret 10:48prompt to this. It's deciding the kind 10:50of task I 10:52need and then making sure that 10:56I approach AI and ask myself what 11:02is the model that I can choose that will 11:06most help me 11:07to optimize for getting my brain deeply 11:11into the subject matter so that I become 11:13someone who truly understands it. That 11:16in and of itself is opposed to it's 11:18antithetical to this idea that AI lets 11:23you go cheap on your brain, right? It 11:25lets you skip the brain power. It lets 11:27you not do the work. It lets you read 11:29the one pager, not the full book. I 11:31don't think it has to. I think it can 11:33actually help you to think better. I 11:35find it 11:36does, but it's how you use it. And it's 11:39not that to say at all that like again I 11:42started this by saying I use the 11:43compression stuff too. I think we all 11:45do. we all will. It helps us save time. 11:47It's about knowing when to sort of pull 11:49the thinking button and say, "I want to 11:51use this to actually expand my time 11:53thinking about this because this is 11:54really important. This is work that 11:56needs and deserves the cream of my 11:58brain, and I need to orient my approach 12:00to AI so that I can spend more time 12:05marinating on this in a way that works 12:06for my brain." And part of why there's 12:08no magic prompt for this is our brains 12:10are wired pretty uniquely. The way I 12:13choose to like walk around and talk and 12:16like talk my ideas out loud may not work 12:18for everybody. There are people for whom 12:20they need to write it out by hand, take 12:22a picture of it, get a transcription, 12:24and maybe that's the way they start. Um, 12:27I think your mileage is going to vary 12:29because your brain varies. But the 12:31concept of optimizing for having your 12:33brain in the subject is something that 12:35allows us to view AI as an expander, not 12:38just as a compression tool. It's 12:40something that optimizes for what we can 12:43create, not just how we can cut cost by 12:46compressing information, even if that's 12:48useful. So, thinking about this was 12:50helpful for me. I hope it kind of turns 12:52the wheels for you, maybe helps you 12:53think about AI a bit differently and try 12:55out advanced voice mode. It's fun.