WEBVTT 00:00:00.000 --> 00:00:03.160 You know that feeling? You're maybe 10, 15 minutes 00:00:03.160 --> 00:00:06.259 into a chat with an AI. Oh, yeah. And it is brilliant. 00:00:06.839 --> 00:00:10.220 You're building something and it's just, it's 00:00:10.220 --> 00:00:12.380 clicking. It remembers every single thing you 00:00:12.380 --> 00:00:13.960 told it. You're in the flow state. It feels like 00:00:13.960 --> 00:00:16.079 it's reading your mind. Exactly. And then, I 00:00:16.079 --> 00:00:18.320 don't know, message 20 hits. Yeah. And it's like 00:00:18.320 --> 00:00:20.039 a switch flips. All of a sudden, it ignores your 00:00:20.039 --> 00:00:22.379 formatting, contradicts itself. It hasn't gotten 00:00:22.379 --> 00:00:24.519 dumber. It just ran out of whiteboard space. 00:00:24.679 --> 00:00:27.190 And today... We're going to fix that. Welcome 00:00:27.190 --> 00:00:29.929 to the deep dive. We are looking at a guide by 00:00:29.929 --> 00:00:33.770 Max Anne called Mastering AI Context Windows 00:00:33.770 --> 00:00:37.469 Memory Hacks for 2026. And this is really the 00:00:37.469 --> 00:00:39.409 manual for dealing with what the author calls 00:00:39.409 --> 00:00:42.530 the memory wall. We're going to cover why models 00:00:42.530 --> 00:00:46.289 like ChatGPT 5 .2 and Claude 4 .5 seem to just, 00:00:46.390 --> 00:00:48.689 you know, forget things. Right. We'll look at 00:00:48.689 --> 00:00:51.170 the hidden token limits of these 2026 models 00:00:51.170 --> 00:00:54.630 and a strategy called the handoff process to 00:00:54.630 --> 00:00:59.939 keep. So the goal here is to move from just chatting 00:00:59.939 --> 00:01:03.039 to becoming what the article calls a context 00:01:03.039 --> 00:01:05.859 orchestrator. That's the idea. Okay, so let's 00:01:05.859 --> 00:01:08.420 start with that core concept, the context window. 00:01:09.079 --> 00:01:12.319 The guide compares it to a finite whiteboard, 00:01:12.459 --> 00:01:15.060 which I find really helpful. It's honestly the 00:01:15.060 --> 00:01:17.540 best analogy. Yeah. Just picture a physical whiteboard. 00:01:17.920 --> 00:01:20.420 That is a fixed size. It can't get any bigger. 00:01:20.739 --> 00:01:24.480 Every single word you type, every file you upload, 00:01:24.659 --> 00:01:28.180 every response the AI gives, it all takes up 00:01:28.180 --> 00:01:30.500 physical space on this board. And even its own 00:01:30.500 --> 00:01:32.819 internal thinking, right? Exactly. The hidden 00:01:32.819 --> 00:01:34.959 reasoning steps, all of it gets written down. 00:01:35.099 --> 00:01:37.459 And once it's full, I mean, it doesn't just stop 00:01:37.459 --> 00:01:40.780 working. No, it starts erasing. To make room 00:01:40.780 --> 00:01:43.159 for the new stuff you just typed, it has to wipe 00:01:43.159 --> 00:01:45.219 the oldest stuff off the board. Which is why 00:01:45.219 --> 00:01:48.700 instruction drift happens. It's not a gug. It's 00:01:48.700 --> 00:01:51.120 an automatic deletion of your initial rules just 00:01:51.120 --> 00:01:53.959 to fit the new data. The AI isn't being difficult. 00:01:54.060 --> 00:01:56.400 Your rules literally don't exist in its memory 00:01:56.400 --> 00:01:59.739 anymore. So, is this strictly based on word count? 00:02:00.299 --> 00:02:02.739 Or is it more complex? It's not words. It uses 00:02:02.739 --> 00:02:04.719 tokens. Which are what? Roughly three quarters 00:02:04.719 --> 00:02:06.579 of a word? Yeah, that's a good rule of thumb. 00:02:06.680 --> 00:02:08.419 It's a bit more complicated, but that's close 00:02:08.419 --> 00:02:10.939 enough for practical use. Got it. Okay, let's 00:02:10.939 --> 00:02:13.680 talk about the big three models for 2026. Because 00:02:13.680 --> 00:02:15.759 their rod boards are apparently of very different 00:02:15.759 --> 00:02:17.900 sizes. Wildly different. And they're for different 00:02:17.900 --> 00:02:22.289 jobs. So first up. ChatGPT 5 .2 Codex from OpenAI. 00:02:22.530 --> 00:02:24.430 Right, so this one's the smallest memory of the 00:02:24.430 --> 00:02:27.710 majors. In the UI, you're limited to about 60 00:02:27.710 --> 00:02:31.169 ,000 tokens. Which is around 45 ,000 words. Yeah, 00:02:31.229 --> 00:02:33.409 and it's built for speed. It's your sports car. 00:02:33.530 --> 00:02:36.270 Quick code, fast answers. Not for analyzing a 00:02:36.270 --> 00:02:38.550 huge novel, then. No, you'll hit that wall almost 00:02:38.550 --> 00:02:40.189 immediately. Okay, then in the middle, we've 00:02:40.189 --> 00:02:43.310 got Claude Opus 4 .5 from Anthropic. This is 00:02:43.310 --> 00:02:45.469 the middle ground. It's got 200 ,000 tokens. 00:02:46.009 --> 00:02:48.650 About 150 ,000 words. So it's better for deeper 00:02:48.650 --> 00:02:51.530 reasoning. Much better. It's a bit slower, but 00:02:51.530 --> 00:02:54.750 it's great for deep analysis writing. It can 00:02:54.750 --> 00:02:56.750 hold a thought for much longer. And then there's 00:02:56.750 --> 00:03:00.030 the beast. Gemini 3 Pro from Google. The free 00:03:00.030 --> 00:03:03.580 train. It has a one million token window. Whoa. 00:03:03.840 --> 00:03:07.639 That's about 750 ,000 words. One million tokens. 00:03:07.699 --> 00:03:09.580 Just imagine that. You could feed it a two -hour 00:03:09.580 --> 00:03:12.680 video, 10 PDFs, and ask 50 questions without 00:03:12.680 --> 00:03:15.039 it even breaking a sweat. That is Gemini 3 Pro. 00:03:15.180 --> 00:03:18.219 It's just a different scale. So, wait. Does having 00:03:18.219 --> 00:03:21.860 a bigger tank mean it drives better? Why do we 00:03:21.860 --> 00:03:23.939 need memory hacks if we can just use Gemini? 00:03:24.439 --> 00:03:27.300 Surprisingly, no. Performance actually degrades 00:03:27.300 --> 00:03:30.139 as that giant window fills up. So bigger isn't 00:03:30.139 --> 00:03:32.300 always better. Not at all. There's a very specific 00:03:32.300 --> 00:03:34.500 performance curve. Okay, tell me about it. So 00:03:34.500 --> 00:03:38.199 from zero to about 30 % capacity, you get beautiful 00:03:38.199 --> 00:03:41.219 performance. The honeymoon phase. Right. 30 to 00:03:41.219 --> 00:03:43.939 50 % is kind of the peak zone. But here's the 00:03:43.939 --> 00:03:48.020 danger zone. Above 60 % capacity, quality starts 00:03:48.020 --> 00:03:50.979 to slide. 60%. And once you get to that 70 % 00:03:50.979 --> 00:03:53.879 to 100 % range, the results get really unpredictable. 00:03:54.360 --> 00:03:56.659 Hallucinations go way up. It's like computer 00:03:56.659 --> 00:03:58.840 RAM. You can have a ton of it, but if you're 00:03:58.840 --> 00:04:01.819 running at 90 % capacity, everything slows down 00:04:01.819 --> 00:04:04.120 and gets glitchy. It's the perfect analogy. The 00:04:04.120 --> 00:04:06.219 AI is trying to juggle too much at once. So the 00:04:06.219 --> 00:04:08.000 goal isn't just fitting the data. It's about 00:04:08.000 --> 00:04:11.419 active management. Exactly. Unused memory actually 00:04:11.419 --> 00:04:14.000 creates better focus than an overloaded memory. 00:04:14.520 --> 00:04:17.000 Unused memory is clarity. I like it. It's the 00:04:17.000 --> 00:04:20.079 mantra for this whole process. Okay, so if we're 00:04:20.079 --> 00:04:22.420 trying to stay out of that danger zone, we need 00:04:22.420 --> 00:04:26.160 to spot the warning signs. The guide lists four 00:04:26.160 --> 00:04:28.939 big red flags for memory loss. Yeah, these are 00:04:28.939 --> 00:04:31.120 the signs that you need to stop immediately. 00:04:31.600 --> 00:04:34.639 Number one is instructions disappear. The classic. 00:04:35.149 --> 00:04:37.310 You ask for bullet points in message one. And 00:04:37.310 --> 00:04:40.189 by message 20, you get a 600 -word essay. The 00:04:40.189 --> 00:04:42.350 rule just fell off the board. It literally doesn't 00:04:42.350 --> 00:04:44.189 know you asked for bullets anymore. Okay, number 00:04:44.189 --> 00:04:47.370 two, contradictions. Right. Message five said, 00:04:47.550 --> 00:04:50.029 let's use a conservative investment strategy. 00:04:50.170 --> 00:04:53.750 By message 20, it's saying, go all in on crypto. 00:04:53.889 --> 00:04:56.069 It's not being creative. It just forgot the persona. 00:04:56.389 --> 00:04:59.949 Yep. Number three, this one feels really dangerous. 00:05:00.899 --> 00:05:03.800 Facts get invented. It's the most crucial warning, 00:05:03.899 --> 00:05:06.779 I think. Yeah. If the budget was $9 ,000 in message 00:05:06.779 --> 00:05:09.720 two, the AI might just make up a number like 00:05:09.720 --> 00:05:12.399 6 ,500 in message 30. Where does it get that 00:05:12.399 --> 00:05:15.199 from? Thin air. It knows a number should be there, 00:05:15.300 --> 00:05:17.800 but the original fact was erased. So it just 00:05:17.800 --> 00:05:20.480 hallucinates a plausible substitute. That's terrifying 00:05:20.480 --> 00:05:23.420 for actual work. And number four is a specific 00:05:23.420 --> 00:05:26.000 tell for Claude, right? Yeah. Claude kind of 00:05:26.000 --> 00:05:28.279 admits when it's struggling. If you see organizing 00:05:28.279 --> 00:05:32.060 thoughts. or compacting conversation, that's 00:05:32.060 --> 00:05:33.860 Klan telling you its memory is full and it's 00:05:33.860 --> 00:05:35.959 trying to save itself. So is there a way to see 00:05:35.959 --> 00:05:38.800 the fuel gauge before it crashes? Or are we just 00:05:38.800 --> 00:05:40.980 guessing? You're not just guessing. There are 00:05:40.980 --> 00:05:43.519 tools like Google's AI Studio Playground that 00:05:43.519 --> 00:05:46.420 show you the exact token counts. So you can watch 00:05:46.420 --> 00:05:48.620 the meter go up. Exactly. It's worth getting 00:05:48.620 --> 00:05:50.259 familiar with those, even if you're not a developer. 00:05:50.560 --> 00:05:52.879 Okay, let's talk tactics. The main one from the 00:05:52.879 --> 00:05:55.480 guide is called the handoff process. This is 00:05:55.480 --> 00:05:58.740 the most effective fix for 2026. It takes a little 00:05:58.740 --> 00:06:01.439 discipline, but it works every time. So what's 00:06:01.439 --> 00:06:03.540 the trigger? When do you do it? When you hit 00:06:03.540 --> 00:06:07.639 about 60 % capacity. For most chats, that's usually 00:06:07.639 --> 00:06:10.519 around 15 or 20 messages. You just stop. Stop. 00:06:10.519 --> 00:06:13.120 Don't ask the next question. Then what? Step 00:06:13.120 --> 00:06:16.459 one. You ask for a very specific kind of summary. 00:06:16.819 --> 00:06:20.259 You ask four questions. Okay. One, what we covered. 00:06:20.399 --> 00:06:23.319 Two, decisions we made. Three, current to -do 00:06:23.319 --> 00:06:26.550 lists. And four. the next immediate task. You're 00:06:26.550 --> 00:06:28.329 basically packing up the essential information. 00:06:28.769 --> 00:06:32.670 Exactly. Then step two, open a brand new chat, 00:06:32.850 --> 00:06:37.970 a clean whiteboard, 0 % capacity. And step three, 00:06:37.990 --> 00:06:40.790 you just paste that summary in. You paste it 00:06:40.790 --> 00:06:42.709 in with a little prompt, like here's the context 00:06:42.709 --> 00:06:44.970 from our last session. Let's continue. Why is 00:06:44.970 --> 00:06:47.089 this better than just continuing the old chat 00:06:47.089 --> 00:06:49.610 though? Because it eliminates all the back and 00:06:49.610 --> 00:06:52.389 forth fluff that clogs the AI's reasoning. So 00:06:52.389 --> 00:06:54.370 you're getting rid of the can you tweak that, 00:06:54.430 --> 00:06:56.870 how about this kind of messages. Yes, you get 00:06:56.870 --> 00:06:58.949 a clean whiteboard, but you keep the essential 00:06:58.949 --> 00:07:01.329 intelligence of the project. It's a fresh start 00:07:01.329 --> 00:07:03.670 without starting from scratch. I want to talk 00:07:03.670 --> 00:07:05.670 about files next because that's where I think 00:07:05.670 --> 00:07:07.589 a lot of token waste happens. But first, we'll 00:07:07.589 --> 00:07:10.470 take a quick break. All right, we're back diving 00:07:10.470 --> 00:07:13.250 into AI memory. We just covered the handoff process. 00:07:13.959 --> 00:07:17.319 Now, memory hygiene for file uploads. The guide 00:07:17.319 --> 00:07:19.639 says not all files cost the same amount of memory. 00:07:19.819 --> 00:07:22.139 Not even close. People think an upload is an 00:07:22.139 --> 00:07:24.600 upload, but some files are insanely expensive 00:07:24.600 --> 00:07:26.819 in terms of tokens. So what's cheap? What's in 00:07:26.819 --> 00:07:31.339 the green zone? Plain text files. So .txt, CSVs, 00:07:31.360 --> 00:07:35.560 Markdown, just raw data with no formatting. Okay, 00:07:35.620 --> 00:07:37.199 pretty straightforward. What about the stuff 00:07:37.199 --> 00:07:40.259 we all use, like PDFs and Word docs? That's the 00:07:40.259 --> 00:07:42.660 yellow zone. Moderate cost. They have all this 00:07:42.660 --> 00:07:45.439 hidden data headers, footers, formatting that 00:07:45.439 --> 00:07:47.500 the AI has to parse, so it costs a bit more. 00:07:47.699 --> 00:07:49.860 And then it gets expensive. The orange zone. 00:07:50.120 --> 00:07:52.339 This is where you find images and complex Excel 00:07:52.339 --> 00:07:55.180 files. A spreadsheet with charts and colors. 00:07:55.939 --> 00:07:58.420 The AI has to visually interpret all of that, 00:07:58.459 --> 00:08:00.639 and it just burns tokens. So what's in the red 00:08:00.639 --> 00:08:04.110 zone? Very expensive. Video and audio. A five 00:08:04.110 --> 00:08:06.250 -minute video can cost more memory than a 50 00:08:06.250 --> 00:08:09.129 -page book because the AI has to transcribe the 00:08:09.129 --> 00:08:11.589 audio and analyze the visual frames over time. 00:08:11.790 --> 00:08:14.629 Wow. So a practical tip would be, if I have a 00:08:14.629 --> 00:08:16.470 big spreadsheet, don't upload the whole thing. 00:08:16.569 --> 00:08:19.569 Exactly. If you only need one tab, export that 00:08:19.569 --> 00:08:22.629 single tab as a CSV. And upload that instead. 00:08:22.870 --> 00:08:24.769 So we should just chop things up before we upload 00:08:24.769 --> 00:08:27.269 them. Yes, you have to pre -process. Trim videos, 00:08:27.629 --> 00:08:30.329 extract text. Don't make the AI do that work. 00:08:30.490 --> 00:08:32.590 That makes perfect sense. Now, the guide talks 00:08:32.590 --> 00:08:35.470 about building intuition because you can't be 00:08:35.470 --> 00:08:37.909 counting tokens all day. Right. It's not practical. 00:08:38.429 --> 00:08:41.950 And every task is different. Coding costs more 00:08:41.950 --> 00:08:45.549 tokens than, say, writing a poem. So how do you 00:08:45.549 --> 00:08:48.470 build that intuition? It suggests keeping a simple 00:08:48.470 --> 00:08:51.820 log just for a week. A log. What do you track? 00:08:52.019 --> 00:08:55.179 Four things. The task you were doing, the files 00:08:55.179 --> 00:08:57.740 you used, the message count when quality started 00:08:57.740 --> 00:09:00.139 to drop, and your rating of the final output. 00:09:00.340 --> 00:09:03.240 That's it? That's it. After 10 or so projects, 00:09:03.460 --> 00:09:05.960 you'll start to just feel it. You'll know, okay, 00:09:06.080 --> 00:09:08.159 this task is probably good for about 15 messages 00:09:08.159 --> 00:09:10.679 before I need to do a handoff. You start to anticipate 00:09:10.679 --> 00:09:12.559 the cliff instead of falling off it? You stop 00:09:12.559 --> 00:09:15.480 flying blind. Okay, there's another tactic in 00:09:15.480 --> 00:09:17.580 here for when you really don't want to start 00:09:17.580 --> 00:09:20.179 a new chat. It's called in -thread summaries. 00:09:20.500 --> 00:09:23.100 This is the light version of the handoff. A quick 00:09:23.100 --> 00:09:25.779 fix. How does that work? Every five to ten messages, 00:09:25.879 --> 00:09:28.919 you just pause and say, stop. Summarize our current 00:09:28.919 --> 00:09:31.700 goals, decisions, and status. But wait, wouldn't 00:09:31.700 --> 00:09:34.120 that summary just use up more space on the already 00:09:34.120 --> 00:09:36.820 full whiteboard? It does, but remember, the oldest 00:09:36.820 --> 00:09:39.679 stuff gets erased first. By creating that summary 00:09:39.679 --> 00:09:42.769 now, you're putting the most important info at 00:09:42.769 --> 00:09:44.690 the bottom of the whiteboard, the freshest part 00:09:44.690 --> 00:09:47.149 of the memory. I see. So you're forcing it to 00:09:47.149 --> 00:09:49.330 refocus on the important stuff and keeping it 00:09:49.330 --> 00:09:51.850 safe from the eraser for a little longer. Exactly. 00:09:51.850 --> 00:09:54.629 It anchors the context. So is this as good as 00:09:54.629 --> 00:09:56.649 the handoff? No, it's a lighter alternative. 00:09:57.269 --> 00:09:59.230 Handoff is still the best for heavy -duty tasks. 00:09:59.610 --> 00:10:02.309 Let's get to the mistakes to avoid. As I was 00:10:02.309 --> 00:10:04.889 reading this section, I felt very seen. I think 00:10:04.889 --> 00:10:07.330 we're all guilty of these. The first one is just 00:10:07.330 --> 00:10:09.980 one more message. The classic trap. You're at 00:10:09.980 --> 00:10:12.779 80 % capacity, you know you should restart, but 00:10:12.779 --> 00:10:14.879 you think, I'll just ask one more quick question. 00:10:14.960 --> 00:10:17.419 I do this constantly. I still wrestle with prompt 00:10:17.419 --> 00:10:19.720 drift myself. I always think I can squeeze in 00:10:19.720 --> 00:10:22.759 one last query. It's the sunk cost fallacy. You 00:10:22.759 --> 00:10:25.639 feel invested. But that one question pushes you 00:10:25.639 --> 00:10:29.799 to 95 % and the answer you get is garbage and 00:10:29.799 --> 00:10:31.840 sends you down the wrong path. You waste more 00:10:31.840 --> 00:10:33.720 time than if you'd just done the handoff. Way 00:10:33.720 --> 00:10:36.700 more. Okay, next mistake. Uploading everything 00:10:36.700 --> 00:10:40.240 at once. The data dump? You think more context 00:10:40.240 --> 00:10:42.820 is better, so you upload 100 pages of background 00:10:42.820 --> 00:10:45.340 documents. But you only need an answer from page 00:10:45.340 --> 00:10:47.720 10. And you've just filled your entire whiteboard 00:10:47.720 --> 00:10:49.840 with noise before you've even asked your first 00:10:49.840 --> 00:10:52.159 question. You've paralyzed it. And the last one 00:10:52.159 --> 00:10:54.379 is just ignoring the warning signs. It's about 00:10:54.379 --> 00:10:56.600 discipline. When you see instructions start to 00:10:56.600 --> 00:10:58.659 fade, you have to pause. It's not going to get 00:10:58.659 --> 00:11:01.279 better on its own. So what's the ultimate role 00:11:01.279 --> 00:11:04.500 shift here? You stop being a user and you start 00:11:04.500 --> 00:11:07.490 being a context manager. you're actively managing 00:11:07.490 --> 00:11:09.429 a scarce resource. That's a great way to put 00:11:09.429 --> 00:11:12.350 it. It puts the control back in your court. Okay, 00:11:12.409 --> 00:11:14.470 let's wrap this up with the big ideas. For me, 00:11:14.490 --> 00:11:17.659 the big takeaway is... Your AI isn't broken. 00:11:17.879 --> 00:11:20.360 Its whiteboard is just full. That's the whole 00:11:20.360 --> 00:11:23.440 thing in a nutshell. And success in 2026 isn't 00:11:23.440 --> 00:11:26.360 about one long conversation. It's about a series 00:11:26.360 --> 00:11:28.759 of short, focused sprints connected by these 00:11:28.759 --> 00:11:31.679 summarized handoffs. Right. Keep your capacity 00:11:31.679 --> 00:11:34.539 under 60 percent. Be deliberate with your file 00:11:34.539 --> 00:11:37.059 uploads and treat that memory like the scarce 00:11:37.059 --> 00:11:40.899 resource it is. And remember, unused memory is 00:11:40.899 --> 00:11:44.480 clarity. A clean whiteboard is always the smartest 00:11:44.480 --> 00:11:47.659 AI. Unused memory is clarity. That's a great 00:11:47.659 --> 00:11:49.620 line. That's a good rule for life, too. So here's 00:11:49.620 --> 00:11:51.779 a final thought to mull over. We've spent this 00:11:51.779 --> 00:11:54.480 whole deep dive talking about how to manage these 00:11:54.480 --> 00:11:58.600 memory limits. But what happens when the limits 00:11:58.600 --> 00:12:01.940 basically disappear? When we have a hundred million 00:12:01.940 --> 00:12:05.120 token window. Right. And we will. Will we stop 00:12:05.120 --> 00:12:07.639 needing to curate and summarize? And if the AI 00:12:07.639 --> 00:12:09.460 can hold everything perfectly, does that mean 00:12:09.460 --> 00:12:12.059 we stop trying to hold it in our own heads? Does 00:12:12.059 --> 00:12:14.820 the friction of context management actually force 00:12:14.820 --> 00:12:17.100 us to learn? That's the question. If you don't 00:12:17.100 --> 00:12:19.320 have to pack the bags yourself, do you even know 00:12:19.320 --> 00:12:21.159 what you're carrying? Something to think about 00:12:21.159 --> 00:12:24.279 for 2027. For now, next time you see ChatGPT 00:12:24.279 --> 00:12:27.259 ignore a rule, don't get mad. Check the tokens. 00:12:27.559 --> 00:12:30.009 And go clear that whiteboard. Thanks for listening 00:12:30.009 --> 00:12:32.070 to The Deep Dive. See you in the next print.