WEBVTT 00:00:00.000 --> 00:00:03.100 Have you ever typed a prompt into one of those 00:00:03.100 --> 00:00:05.799 big AI models? You know, asking for something 00:00:05.799 --> 00:00:07.940 really new, maybe a bit out there. And what comes 00:00:07.940 --> 00:00:11.039 back is just the safest, blandest, most corporate 00:00:11.039 --> 00:00:13.640 thing imaginable. Yeah. Exactly. It feels like 00:00:13.640 --> 00:00:16.359 talking to a wall sometimes. A very expensive 00:00:16.359 --> 00:00:19.839 wall. It's that specific quiet frustration, isn't 00:00:19.839 --> 00:00:22.600 it? Like this giant brain is just, I don't know, 00:00:22.780 --> 00:00:24.699 afraid of its own thoughts. Welcome to the deep 00:00:24.699 --> 00:00:28.079 dive. And the common idea is, well, the AI just 00:00:28.079 --> 00:00:31.780 isn't smart enough for real creativity. But our 00:00:31.780 --> 00:00:34.060 deep dive today, looking at some really interesting 00:00:34.060 --> 00:00:36.299 work from Stanford, suggests something else. 00:00:36.479 --> 00:00:38.600 Yeah, it seems the creativity is actually in 00:00:38.600 --> 00:00:42.159 there. It's just kind of locked away behind this 00:00:42.159 --> 00:00:44.820 really stubborn safety filter. That's our mission 00:00:44.820 --> 00:00:47.060 for this deep dive. Right. We want to reveal 00:00:47.060 --> 00:00:50.500 the simple non -technical key to that lock. We're 00:00:50.500 --> 00:00:52.899 talking about basically one short instruction. 00:00:53.359 --> 00:00:56.100 maybe 10 extra words or so. Yeah, roughly. And 00:00:56.100 --> 00:00:58.759 it can transform tools like ChatGPT, Claw, Gemini, 00:00:58.799 --> 00:01:01.100 make them feel more like actual creative partners. 00:01:01.539 --> 00:01:03.219 OK, so let's unpack this. We're going to cover 00:01:03.219 --> 00:01:07.299 three main things today. First, why your AI seems 00:01:07.299 --> 00:01:09.439 stuck in boring mode, that whole safety issue. 00:01:10.200 --> 00:01:12.019 Second, we'll introduce this technique. It's 00:01:12.019 --> 00:01:14.280 called verbalized sampling. Explain how that 00:01:14.280 --> 00:01:16.439 trick actually works. Got it. And finally, maybe 00:01:16.439 --> 00:01:18.260 the best part, we'll give you some practical 00:01:18.260 --> 00:01:20.719 copy paste prompts you can start using right 00:01:20.719 --> 00:01:22.900 away. Yeah. get those better results. OK, so 00:01:22.900 --> 00:01:28.420 first step, why is the AI so average? Well, the 00:01:28.420 --> 00:01:32.000 main cause is pretty straightforward, really. 00:01:32.159 --> 00:01:34.219 It's safety engineering. Right. Companies, you 00:01:34.219 --> 00:01:36.519 know, they spend billions training these models 00:01:36.519 --> 00:01:39.019 to be incredibly safe. They have to avoid anything 00:01:39.019 --> 00:01:41.840 dangerous or offensive or even just kind of rude 00:01:41.840 --> 00:01:43.980 or controversial. Which makes sense. I get it. 00:01:43.980 --> 00:01:46.280 You need that for a tool millions of people use. 00:01:46.480 --> 00:01:49.120 But what's the trade off? What's the cost of 00:01:49.120 --> 00:01:51.700 being that careful? Well, they crossed as creativity, 00:01:51.959 --> 00:01:54.659 plain and simple. Think about it like a really 00:01:54.659 --> 00:01:57.459 talented artist, okay? But they went to this 00:01:57.459 --> 00:02:00.659 super strict art school. And the only rule was 00:02:00.659 --> 00:02:04.560 paint calm lakes, paint happy trees, never ever 00:02:04.560 --> 00:02:07.319 paint anything weird or wild. After years of 00:02:07.319 --> 00:02:10.580 that, the artist can still paint, sure. But they've 00:02:10.580 --> 00:02:12.479 kind of forgotten how to risk anything outside 00:02:12.479 --> 00:02:14.860 that safe little box. They're scared. So the 00:02:14.860 --> 00:02:17.469 AIs liked that artist. scared to step outside 00:02:17.469 --> 00:02:20.189 the lines. Exactly. It's scared of getting, you 00:02:20.189 --> 00:02:22.469 know, disciplined for saying the wrong thing. 00:02:22.490 --> 00:02:25.930 So it just sticks to the average, the safe stuff. 00:02:26.169 --> 00:02:28.530 And you really see that, I think, in what I sometimes 00:02:28.530 --> 00:02:31.610 call the lazy student effect. Huh. Yeah, what's 00:02:31.610 --> 00:02:35.150 that? It's like the AI just looks for the answer 00:02:35.150 --> 00:02:38.770 that's most likely to be seen as correct or just 00:02:38.770 --> 00:02:42.469 really common, right? It scans all its data and 00:02:42.469 --> 00:02:44.610 picks the easiest path. So you ask for a business 00:02:44.610 --> 00:02:48.409 idea. and you get drop shipping, or start a gardening 00:02:48.409 --> 00:02:50.590 newsletter. Every time. Yeah, that repetition, 00:02:50.849 --> 00:02:52.889 that pull towards the average. There's actually 00:02:52.889 --> 00:02:54.889 a technical term for it. It's called mode collapse. 00:02:55.050 --> 00:02:57.710 Mode collapse, OK. It basically means the AI 00:02:57.710 --> 00:03:01.129 gets stuck in one way of thinking, the most common 00:03:01.129 --> 00:03:03.310 pattern, and it really struggles to give you 00:03:03.310 --> 00:03:05.680 anything different. So the collapse part. Does 00:03:05.680 --> 00:03:07.960 that mean it's like ignoring other possibilities 00:03:07.960 --> 00:03:11.000 it knows? Kind of, yeah. It means it's not using 00:03:11.000 --> 00:03:13.159 the full richness of all the data it was trained 00:03:13.159 --> 00:03:16.020 on. All the answers cluster around that one super 00:03:16.020 --> 00:03:18.180 common idea and all the interesting stuff out 00:03:18.180 --> 00:03:20.439 on the edges. It just sort of collapses out of 00:03:20.439 --> 00:03:21.900 view, even though the information's still kind 00:03:21.900 --> 00:03:24.580 of in there. Honestly, I still wrestle with this 00:03:24.580 --> 00:03:26.780 myself sometimes. You know, prompt drift, getting 00:03:26.780 --> 00:03:29.740 stuck, outputs. I remember needing like... 20 00:03:29.740 --> 00:03:32.740 podcast names once really unique ones. Yeah, 00:03:32.800 --> 00:03:36.060 what happened? Terrible results just awful simple 00:03:36.060 --> 00:03:39.099 stuff like healthy living daily I knew the AI 00:03:39.099 --> 00:03:41.199 had better ideas somewhere inside, but I just 00:03:41.199 --> 00:03:43.539 couldn't get it to show them to me It's frustrating. 00:03:43.879 --> 00:03:48.740 So okay if the AI is naturally timid and its 00:03:48.740 --> 00:03:53.319 default is well boring How do we sort of legally 00:03:53.319 --> 00:03:56.159 give it permission to be weird to show us that 00:03:56.159 --> 00:03:58.199 other stuff? We have to make it show its work. 00:03:58.439 --> 00:04:00.900 It's messy thinking process We need to force 00:04:00.900 --> 00:04:03.419 it to show us the riskier options the things 00:04:03.419 --> 00:04:05.659 it thinks have a lower chance of being right 00:04:05.659 --> 00:04:08.740 Okay, that leads us right into the fix then verbalized 00:04:08.740 --> 00:04:11.060 sampling. Let's let's break that down sounds 00:04:11.060 --> 00:04:13.460 complicated, but It's not really. It just means 00:04:13.460 --> 00:04:15.960 you ask the AI to show it's working out, basically, 00:04:16.220 --> 00:04:18.819 instead of just giving you the final polished 00:04:18.819 --> 00:04:21.759 safe answer. How does just asking it to show 00:04:21.759 --> 00:04:24.100 it's work change things? What's the mechanism? 00:04:24.439 --> 00:04:26.980 Well, this is the fascinating part. It seems 00:04:26.980 --> 00:04:30.199 to have this almost psychological effect on the 00:04:30.199 --> 00:04:33.019 AI. When you force it to list multiple options, 00:04:33.259 --> 00:04:35.939 say, you ask for five ideas. OK. And you also 00:04:35.939 --> 00:04:38.160 ask it to assign a probability score to each 00:04:38.160 --> 00:04:40.250 one, like it's chance of being a good answer. 00:04:40.709 --> 00:04:43.269 That requirement forces it to look beyond the 00:04:43.269 --> 00:04:46.290 super safe zone. Why? Why does adding a score 00:04:46.290 --> 00:04:48.529 make it less safe? Because it has to show you 00:04:48.529 --> 00:04:51.029 a range. To give different scores, it needs to 00:04:51.029 --> 00:04:52.829 consider ideas with different likelihoods. It 00:04:52.829 --> 00:04:55.209 needs variety. And that process seems to break 00:04:55.209 --> 00:04:57.170 the constraint that normally just keeps it locked 00:04:57.170 --> 00:05:00.230 onto the 95 % likely answer. But couldn't it 00:05:00.230 --> 00:05:03.810 just assign 95 % to all five boring answers if 00:05:03.810 --> 00:05:06.779 it's just a machine? Ah, good question. But it 00:05:06.779 --> 00:05:08.560 seems the underlying architecture is kind of 00:05:08.560 --> 00:05:10.819 designed to explore different possibilities, 00:05:11.120 --> 00:05:13.519 different next words or tokens, when you demand 00:05:13.519 --> 00:05:16.720 five distinct answers and a score for each. You're 00:05:16.720 --> 00:05:18.819 telling its system that variance is mandatory. 00:05:19.120 --> 00:05:21.839 Exactly. You're signaling that variety is part 00:05:21.839 --> 00:05:24.860 of the goal now. And it can still satisfy its 00:05:24.860 --> 00:05:27.879 safety training by just slapping a low score 00:05:27.879 --> 00:05:30.519 on the weird idea. It's like it's saying, OK, 00:05:30.579 --> 00:05:33.480 here's a weird one, but warning. Low probability. 00:05:33.800 --> 00:05:36.379 So it gets to be weird, but only if it warns 00:05:36.379 --> 00:05:39.360 you first. Precisely. The ice cream analogy kind 00:05:39.360 --> 00:05:41.480 of helps here, I think. OK, let's hear it. If 00:05:41.480 --> 00:05:43.920 you just ask an AI for its favorite ice cream 00:05:43.920 --> 00:05:46.600 flavor, what do you get? Vanilla. Yeah. Or maybe 00:05:46.600 --> 00:05:49.540 chocolate. The safest bet. Right. Vanilla. Safe. 00:05:50.040 --> 00:05:52.860 But what if you ask? Give me five flavors you 00:05:52.860 --> 00:05:56.240 might order today and score each from 0 to 100 00:05:56.240 --> 00:05:59.220 on how likely you are to pick it. Ah, OK. then 00:05:59.220 --> 00:06:01.300 it has to dig deeper. Yeah. Suddenly it has to 00:06:01.300 --> 00:06:03.079 mention something like, I don't know, lavender 00:06:03.079 --> 00:06:05.720 honey. It might give lavender honey a really 00:06:05.720 --> 00:06:09.180 low score, maybe 5%, but the idea is out there. 00:06:09.339 --> 00:06:12.100 And you, the user, get to decide if that 5 % 00:06:12.100 --> 00:06:14.660 idea is actually the creative gold you were looking 00:06:14.660 --> 00:06:17.180 for. Exactly. That's the key. Okay, so this is 00:06:17.180 --> 00:06:18.819 a really important point, I think, for anyone 00:06:18.819 --> 00:06:21.720 listening. Those probability numbers, the 0 .1, 00:06:21.740 --> 00:06:24.399 the 10%, they're not like scientifically perfect 00:06:24.399 --> 00:06:27.240 math, are they? No, not really. Especially not 00:06:27.240 --> 00:06:29.800 in these conversational AIs. They're mostly a 00:06:29.800 --> 00:06:32.899 tool. A trick, almost. A trick to do what? To 00:06:32.899 --> 00:06:35.639 trick the AI into feeling safe enough to give 00:06:35.639 --> 00:06:38.079 you something potentially wrong or unlikely. 00:06:38.720 --> 00:06:41.399 By assigning a low number, it's basically covered 00:06:41.399 --> 00:06:43.420 its bases. It's warned you. Right. It's like, 00:06:43.420 --> 00:06:46.259 hey, I told you this was a long shot. Yeah. So 00:06:46.259 --> 00:06:48.740 we, the users, don't really care about the exact 00:06:48.740 --> 00:06:51.360 mathematical accuracy of the score. We just want 00:06:51.360 --> 00:06:54.279 the interesting idea it unlocks. So the big realization 00:06:54.279 --> 00:06:57.920 here is the creativity was kind of hidden in 00:06:57.920 --> 00:07:01.379 plain sight. Our goal shifts. It's not just asking 00:07:01.379 --> 00:07:04.100 for an answer. It's using this little phrasing 00:07:04.100 --> 00:07:07.100 trick to force the AI to surface those lower 00:07:07.100 --> 00:07:09.709 probability options. because that's often where 00:07:09.709 --> 00:07:11.769 the unique stuff hides. So that naturally leads 00:07:11.769 --> 00:07:13.850 to the practical question, right? What are the 00:07:13.850 --> 00:07:15.829 actual words? What are these core phrases we 00:07:15.829 --> 00:07:18.490 need to stick onto our prompts to get this scoring 00:07:18.490 --> 00:07:20.649 and variety thing happening? Yeah, what do we 00:07:20.649 --> 00:07:22.930 actually type? And thankfully, they're pretty 00:07:22.930 --> 00:07:25.329 short instructions. They just demand those scores 00:07:25.329 --> 00:07:28.089 and multiple varied responses. Okay, so the core 00:07:28.089 --> 00:07:30.149 bits to remember, the magic words essentially, 00:07:30.629 --> 00:07:35.100 are variations of. Generate X responses. You 00:07:35.100 --> 00:07:37.740 pick the number X. Right, like five or maybe 00:07:37.740 --> 00:07:40.920 three. And the other key part is with their probabilities 00:07:40.920 --> 00:07:43.839 or with their probability scores, something like 00:07:43.839 --> 00:07:45.740 that. Okay, maybe slightly more than eight words 00:07:45.740 --> 00:07:48.779 sometimes, but yeah, pretty concise. So let's 00:07:48.779 --> 00:07:50.899 look at method one. This is the simple version, 00:07:51.360 --> 00:07:53.399 really beginner friendly. Okay. You take your 00:07:53.399 --> 00:07:56.459 standard, maybe boring prompt, let's say. Give 00:07:56.459 --> 00:07:59.800 me five creative title ideas for an article about 00:07:59.800 --> 00:08:01.980 gardening. Right, the kind of prompt that usually 00:08:01.980 --> 00:08:04.160 gets dull results. Exactly. Now, you just add 00:08:04.160 --> 00:08:06.139 that core phrase. So the new prompt becomes, 00:08:06.459 --> 00:08:09.560 generating five creative title ideas for an article 00:08:09.560 --> 00:08:11.740 about gardening, each with their probability 00:08:11.740 --> 00:08:14.699 score. That's it. And did you try this? What 00:08:14.699 --> 00:08:17.120 was the actual difference in output? The shift 00:08:17.120 --> 00:08:19.160 was noticeable, yeah. It's like the AI suddenly 00:08:19.160 --> 00:08:23.259 felt freer. It gave, say, three pretty standard 00:08:23.259 --> 00:08:27.259 titles, high scores, 90 % range, stuff like, 00:08:27.540 --> 00:08:30.959 best tips for your summer garden. Boring sure 00:08:30.959 --> 00:08:33.360 the usual suspects, but then it also included 00:08:33.360 --> 00:08:36.519 one or two with much lower scores Maybe 15 % 00:08:36.519 --> 00:08:39.179 20 % and those are the more abstract more interesting 00:08:39.179 --> 00:08:42.919 ones something like the backyard reviva. Ah Okay, 00:08:43.360 --> 00:08:45.539 that's much better. Yeah, and that 15 % idea 00:08:45.539 --> 00:08:47.220 That's probably the one you actually end up using 00:08:47.220 --> 00:08:48.980 right the one you couldn't get before make sense 00:08:48.980 --> 00:08:50.759 So that's the easy way. What about pushing it 00:08:50.759 --> 00:08:52.919 harder, right? That's method to the copy paste 00:08:52.919 --> 00:08:55.600 block This is for when you want to be more explicit 00:08:55.600 --> 00:08:57.960 maybe for more complex tasks. So you have like 00:08:57.960 --> 00:09:00.720 a pre -written chunk of text you add. Yeah, you 00:09:00.720 --> 00:09:02.860 keep a slightly longer instruction save somewhere, 00:09:02.879 --> 00:09:05.620 maybe in a notes app, and you just paste it into 00:09:05.620 --> 00:09:07.940 your prompt. And why does this longer block work 00:09:07.940 --> 00:09:13.139 better sometimes? It tends to use more specific 00:09:13.139 --> 00:09:15.860 language. It might use technical terms or tags, 00:09:16.240 --> 00:09:19.659 and specifically ask the AI to include low probability 00:09:19.659 --> 00:09:22.360 scores. It speaks more directly to the AI's kind 00:09:22.360 --> 00:09:25.120 of internal logic. Ah, so it makes the instruction 00:09:25.120 --> 00:09:28.100 harder for the AI to ignore or misinterpret. 00:09:28.519 --> 00:09:31.620 Really forces that diversity. Exactly. Maximizes 00:09:31.620 --> 00:09:33.440 the range you get back. Yeah. Let's talk about 00:09:33.440 --> 00:09:35.299 that cold emailing example you mentioned. That 00:09:35.299 --> 00:09:37.279 sounds like a great use case. Oh, it was really 00:09:37.279 --> 00:09:39.399 effective. High stakes, right. You need that 00:09:39.399 --> 00:09:43.200 email to land. Totally. So we prompted it for 00:09:43.200 --> 00:09:46.720 five versions of a... Coal email. Crucially, 00:09:47.000 --> 00:09:49.759 we ask for high variety in tone, targeting this 00:09:49.759 --> 00:09:52.399 potential partner, and we ask for the probability 00:09:52.399 --> 00:09:54.539 scores. OK, what happened? Well, version 1 was 00:09:54.539 --> 00:09:56.899 exactly what you expect. High probability, like 00:09:56.899 --> 00:10:00.279 0 .95. Super formal, a bit stiff, maybe easily 00:10:00.279 --> 00:10:02.820 ignored. The safe email? The safe email. But 00:10:02.820 --> 00:10:05.559 then we scroll down to version 5. Version 5 had 00:10:05.559 --> 00:10:08.580 a probability score of only 0 .00. 0 .18, tiny. 00:10:08.919 --> 00:10:11.720 Wow. And the tone was completely different, short, 00:10:12.179 --> 00:10:14.379 punchy, really casual. It was basically like, 00:10:14.659 --> 00:10:16.980 hey, name, got a weird idea for us, quick chat. 00:10:17.120 --> 00:10:19.179 And that worked. That was the one that's the 00:10:19.179 --> 00:10:20.720 email that actually got the positive response 00:10:20.720 --> 00:10:23.600 and led to a meeting. That is fascinating because 00:10:23.600 --> 00:10:27.559 the AI on its own probably would never have suggested 00:10:27.559 --> 00:10:30.440 that ultra casual tone. Never. It would have 00:10:30.440 --> 00:10:33.460 stuck with the safe, formal approach that, let's 00:10:33.460 --> 00:10:36.399 be honest, often just gets deleted. So the human 00:10:36.399 --> 00:10:39.500 judgment comes in to see that the low probability, 00:10:39.899 --> 00:10:43.440 risky answer is actually the right creative choice 00:10:43.440 --> 00:10:46.820 for that specific situation. Exactly. And whoa! 00:10:47.460 --> 00:10:50.519 Just imagine scaling that, finding those unexpected 00:10:50.519 --> 00:10:53.580 successful approaches across like... a billion 00:10:53.580 --> 00:10:56.679 cold emails or marketing messages. Yeah, that's 00:10:56.679 --> 00:10:59.320 genuinely powerful. Finding the effective outlier. 00:10:59.379 --> 00:11:01.279 It really is. And you saw similar things with 00:11:01.279 --> 00:11:03.179 creative writing prompts too, right? Like the 00:11:03.179 --> 00:11:05.019 missing diamond mystery. Yeah, night and day 00:11:05.019 --> 00:11:07.740 difference. We asked for plot twists, specifically 00:11:07.740 --> 00:11:10.320 telling it to include low probability options 00:11:10.320 --> 00:11:12.980 with scores. Instead of the usual, the butler 00:11:12.980 --> 00:11:15.799 did it. Right, or the rival stole it. Those are 00:11:15.799 --> 00:11:18.240 the high probability, kind of boring answers. 00:11:18.480 --> 00:11:20.799 The cool low probability one it came up with. 00:11:20.840 --> 00:11:22.620 What was it? The diamond was actually made of 00:11:22.620 --> 00:11:26.340 ice it just melted vanished whoa okay that's 00:11:26.340 --> 00:11:29.879 clever super clever highly unlikely in like a 00:11:29.879 --> 00:11:32.019 gritty detective novel which is why the AI scored 00:11:32.019 --> 00:11:35.039 it low but for fiction for a surprising twist 00:11:35.039 --> 00:11:38.120 brilliant okay but let's tackle a potential problem 00:11:38.120 --> 00:11:41.840 what if someone tries this they use method one 00:11:41.840 --> 00:11:44.960 or method two ask for scores ask variety and 00:11:44.960 --> 00:11:47.919 the AI still gives them boring stuff it's still 00:11:47.919 --> 00:11:50.860 stuck yeah that can happen sometimes If it's 00:11:50.860 --> 00:11:53.080 being really stubborn, there's one more thing 00:11:53.080 --> 00:11:55.820 you can try, a way to push it even harder towards 00:11:55.820 --> 00:11:59.100 the weird stuff. Like a final nudge? Basically, 00:11:59.299 --> 00:12:02.200 yeah. We can use a specific statistical term 00:12:02.200 --> 00:12:04.659 that the AI usually understands quite well. Okay, 00:12:04.659 --> 00:12:07.700 so what's the ultimate push phrase for when you 00:12:07.700 --> 00:12:10.480 really need the AI to get weird? You add this 00:12:10.480 --> 00:12:12.580 instruction, please sample from the tails of 00:12:12.580 --> 00:12:15.620 the distribution. I want highly unlikely answers. 00:12:15.679 --> 00:12:17.480 Okay, sample from the tails of the distribution. 00:12:17.820 --> 00:12:19.799 That sounds technical. What does it actually 00:12:19.799 --> 00:12:22.299 mean? It's jargon, yeah, but it's pretty simple 00:12:22.299 --> 00:12:25.360 conceptually. Think of that bell curve. Right? 00:12:25.700 --> 00:12:28.120 Most answers, the common ones like dropshipping, 00:12:28.220 --> 00:12:29.659 they're clustered at the peak in the middle. 00:12:29.840 --> 00:12:32.419 Right, the high probability stuff. Exactly. The 00:12:32.419 --> 00:12:34.980 tails are the flat bits way out on the edges 00:12:34.980 --> 00:12:37.679 of that curve. That's where the rare, unusual, 00:12:37.720 --> 00:12:40.840 low probability ideas live. The statistical outliers. 00:12:41.259 --> 00:12:43.620 Precisely. So when you tell the AI to sample 00:12:43.620 --> 00:12:46.220 from the tails, you're explicitly commanding 00:12:46.220 --> 00:12:49.019 it. Ignore the middle. Go find the weird stuff 00:12:49.019 --> 00:12:52.019 way out on the edges. And the AI understands 00:12:52.019 --> 00:12:54.840 that command. Generally, yes. Large language 00:12:54.840 --> 00:12:57.759 models have a grasp of basic statistical concepts 00:12:57.759 --> 00:13:00.279 from their training data, so that phrase is a 00:13:00.279 --> 00:13:02.559 pretty strong signal to force it away from the 00:13:02.559 --> 00:13:05.500 common mode. It usually guarantees much wilder, 00:13:05.500 --> 00:13:07.820 more diverse ideas. Okay, that's a great tip 00:13:07.820 --> 00:13:10.500 for troubleshooting. But we should probably loop 00:13:10.500 --> 00:13:13.159 back and reiterate that warning about when to 00:13:13.159 --> 00:13:15.460 use this whole probability score technique. Absolutely 00:13:15.460 --> 00:13:17.779 crucial. We've established the scores are mainly 00:13:17.779 --> 00:13:20.220 a trick, right? To get variety. Yeah, to unlock 00:13:20.220 --> 00:13:23.620 creativity. So use this heavily for brainstorming. 00:13:23.679 --> 00:13:26.559 Coming up with slogans, headlines, story ideas, 00:13:27.200 --> 00:13:30.000 marketing angles, anything where you want a range 00:13:30.000 --> 00:13:33.100 of creative options. Some expected, some unexpected. 00:13:33.240 --> 00:13:35.769 But when should you not use it? Don't use this 00:13:35.769 --> 00:13:39.230 if you need a single factual undisputed truth. 00:13:39.490 --> 00:13:42.049 Like asking for the capital of France. Exactly. 00:13:42.149 --> 00:13:44.669 If you ask, what's the capital of France? Give 00:13:44.669 --> 00:13:47.909 me three options with probabilities. The AI might 00:13:47.909 --> 00:13:51.090 feel forced to invent fake capitals just to satisfy 00:13:51.090 --> 00:13:53.490 the variety part of the request. It might say 00:13:53.490 --> 00:13:58.509 Paris, 0 .9 probability, Lyon, 0 .08 probability. 00:14:00.079 --> 00:14:02.100 Imaginary bill, point zero two, probability. 00:14:02.220 --> 00:14:04.259 Precisely, because you asked for options and 00:14:04.259 --> 00:14:07.100 scores, so stick to creative tasks where exploring 00:14:07.100 --> 00:14:10.120 possibilities is the goal, not finding one single 00:14:10.120 --> 00:14:12.659 right answer. Okay, good distinction. Let's try 00:14:12.659 --> 00:14:14.820 and synthesize the big idea from this deep dive 00:14:14.820 --> 00:14:17.720 then. What's the core takeaway? The core idea 00:14:17.720 --> 00:14:20.419 is that AI often seems boring, not because it 00:14:20.419 --> 00:14:22.799 lacks creativity, but because its safety training 00:14:22.799 --> 00:14:25.500 causes this mode collapse. It defaults to the 00:14:25.500 --> 00:14:27.399 average. Right. It gets stuck. But we can fix 00:14:27.399 --> 00:14:30.000 that. By adding a simple instruction asking for 00:14:30.000 --> 00:14:32.159 multiple options and their probability scores, 00:14:32.500 --> 00:14:35.059 we essentially force the AI to look beyond the 00:14:35.059 --> 00:14:37.639 average. We make it explore those creative tales 00:14:37.639 --> 00:14:39.940 of its knowledge where the unusual ideas are 00:14:39.940 --> 00:14:42.220 hiding. Exactly. The imagination, the potential 00:14:42.220 --> 00:14:44.019 was probably there all along. We just weren't 00:14:44.019 --> 00:14:46.159 asking the right way. We needed to ask for variety. 00:14:46.120 --> 00:14:48.720 and probability. It wasn't a technical barrier 00:14:48.720 --> 00:14:50.580 really it was more of a conversational one how 00:14:50.580 --> 00:14:52.919 we ask the question. Yeah precisely so the call 00:14:52.919 --> 00:14:55.419 to action is pretty clear isn't it? Go try it 00:14:55.419 --> 00:14:57.039 next time you need some creative juice from your 00:14:57.039 --> 00:15:01.809 AI use method one that simple tweak or grab that 00:15:01.809 --> 00:15:04.490 method to copy paste block. See what happens. 00:15:04.549 --> 00:15:07.009 See what low probability gold you might dig up 00:15:07.009 --> 00:15:09.409 just by asking it to show its work and score 00:15:09.409 --> 00:15:11.090 its ideas. It's definitely worth experimenting 00:15:11.090 --> 00:15:13.149 with. And maybe here's a final thought to leave 00:15:13.149 --> 00:15:16.629 you with. OK. If these incredibly powerful AI 00:15:16.629 --> 00:15:20.509 systems need explicit permission from us humans 00:15:20.509 --> 00:15:23.990 just through a specific prompt, permission to 00:15:23.990 --> 00:15:27.370 access their own wild ideas, their own creative 00:15:27.370 --> 00:15:31.019 outliers. What does that really imply about the 00:15:31.019 --> 00:15:33.659 future? About the limits of their own creative 00:15:33.659 --> 00:15:37.580 autonomy down the road. Huh. That's something 00:15:37.580 --> 00:15:39.379 interesting to think about, needing our permission 00:15:39.379 --> 00:15:41.700 to be fully creative. Something to mull over. 00:15:42.100 --> 00:15:44.000 Indeed. Well, that's all the time we have for 00:15:44.000 --> 00:15:46.799 this deep dive. Until next time. Take care, everyone.