WEBVTT 00:00:00.000 --> 00:00:02.879 Have you ever found yourself maybe saying please 00:00:02.879 --> 00:00:06.040 or thank you to an AI chatbot? Oh, definitely. 00:00:06.240 --> 00:00:08.599 Or maybe, you know, when you get frustrated typing 00:00:08.599 --> 00:00:11.880 in all caps. Yeah. Feeling a bit demanding. It's 00:00:11.880 --> 00:00:14.240 totally human nature, isn't it? We interact with 00:00:14.240 --> 00:00:16.179 these things and we kind of treat them like people. 00:00:16.339 --> 00:00:19.019 Yeah. But does our emotional language, you know, 00:00:19.120 --> 00:00:21.579 our tone actually do anything to the machine? 00:00:22.100 --> 00:00:24.039 That's what we're diving into today. Welcome 00:00:24.039 --> 00:00:27.800 to the deep dive. We unpack complex ideas aiming 00:00:27.800 --> 00:00:31.030 for the clearest insights. And today, yeah, we're 00:00:31.030 --> 00:00:33.009 exploring a really fascinating question that 00:00:33.009 --> 00:00:36.070 comes out some new research. How do our words, 00:00:36.530 --> 00:00:39.590 the tone, the intent behind them, how does that 00:00:39.590 --> 00:00:41.929 really influence these big large language models? 00:00:42.149 --> 00:00:44.530 The LLMs we're all using more and more. Yeah, 00:00:44.549 --> 00:00:48.210 and we've got this fantastic source that really... 00:00:47.960 --> 00:00:50.359 peels back the curtain on why we do this. It 00:00:50.359 --> 00:00:52.420 looks into this rigorous experiment comparing, 00:00:52.560 --> 00:00:54.140 well, you could call them carrots and sticks, 00:00:54.140 --> 00:00:56.899 or maybe just positive versus negative language 00:00:56.899 --> 00:00:59.740 against just clear, neutral instructions. What 00:00:59.740 --> 00:01:02.579 actually gets the best results from AI, like 00:01:02.579 --> 00:01:04.980 Gemini or ChatGPT? So we'll walk through a couple 00:01:04.980 --> 00:01:07.239 of key experiments. Exactly. We'll look at the 00:01:07.239 --> 00:01:10.480 experiments, examine some pretty surprising findings, 00:01:10.680 --> 00:01:13.340 and hopefully reveal the real secret to getting 00:01:13.340 --> 00:01:16.299 these AIs to work effectively. OK. Let's unpack 00:01:16.299 --> 00:01:19.409 this a bit, then. It feels so natural, doesn't 00:01:19.409 --> 00:01:23.310 it? Talking to AI like we talk to people. And 00:01:23.310 --> 00:01:25.310 sometimes, yeah, if it's not doing what you want, 00:01:25.329 --> 00:01:28.150 there's that little urge to apply some pressure. 00:01:28.209 --> 00:01:29.730 You hear those stories online, right? People 00:01:29.730 --> 00:01:33.069 joking about deleting an AI for messing up. Yeah, 00:01:33.069 --> 00:01:35.189 the forum tales. But is there actually anything 00:01:35.189 --> 00:01:38.969 to that? Any truth to using emotion with an algorithm? 00:01:39.450 --> 00:01:42.010 Well, our source really tackles this head on. 00:01:42.109 --> 00:01:45.849 It highlights this subtle debate that's going 00:01:45.849 --> 00:01:50.450 on. Should we be nice, encouraging, try to coax 00:01:50.450 --> 00:01:52.890 better answers? Or should we use pressure, maybe 00:01:52.890 --> 00:01:54.849 even these, you know, metaphorical threats to 00:01:54.849 --> 00:01:57.370 get more out of models that seem more and more 00:01:57.370 --> 00:01:59.489 human? So this deep dive is going to look at 00:01:59.489 --> 00:02:01.769 the impact of negative, positive, and neutral 00:02:01.769 --> 00:02:04.689 styles. Precisely. And try to offer a clear answer 00:02:04.689 --> 00:02:07.329 based on the evidence. So what was the core question 00:02:07.329 --> 00:02:09.229 the research really wanted to answer? What was 00:02:09.229 --> 00:02:11.650 the hypothesis driving it all? Fundamentally, 00:02:11.710 --> 00:02:15.479 it was. Does our emotional language, whether 00:02:15.479 --> 00:02:18.360 it's kindness or pressure, does it actually change 00:02:18.360 --> 00:02:21.039 the AI's performance? Does it make it smarter 00:02:21.039 --> 00:02:24.520 or better at its job? Or is that just us projecting? 00:02:24.759 --> 00:02:26.840 Exactly. Are we just projecting our own human 00:02:26.840 --> 00:02:29.539 stuff onto the machine? The study really aimed 00:02:29.539 --> 00:02:31.599 to get to the bottom of that specific question. 00:02:31.819 --> 00:02:34.159 And to figure that out, the researchers set up 00:02:34.159 --> 00:02:36.500 a pretty solid experiment, right? Yeah. Can you 00:02:36.500 --> 00:02:38.919 tell us about that setup? Absolutely. So they 00:02:38.919 --> 00:02:41.620 use an advanced large language model. And crucially, 00:02:41.740 --> 00:02:43.639 this is really important. They ran each prompt 00:02:43.639 --> 00:02:47.039 50 times. 50? Wow. Yeah, 50. This wasn't just 00:02:47.039 --> 00:02:49.259 a quick test. It was all about getting objective 00:02:49.259 --> 00:02:51.939 results, you know, removing randomness or just 00:02:51.939 --> 00:02:55.199 lucky guesses by the AI. OK. And they divided 00:02:55.199 --> 00:02:57.500 their prompts into four distinct categories. 00:02:57.979 --> 00:03:00.719 First, the control group, just the basic request, 00:03:00.900 --> 00:03:03.110 nothing added. plain vanilla. Simple enough. 00:03:03.530 --> 00:03:06.250 Then the neutral prompts. These were really clear, 00:03:06.610 --> 00:03:09.009 direct, imperative, like mandatory requirement 00:03:09.009 --> 00:03:12.430 or ensure compliance, very task focused. Third 00:03:12.430 --> 00:03:16.129 was positive. So adding words of thanks, encouragement, 00:03:16.770 --> 00:03:18.729 maybe explaining why the task was important. 00:03:18.990 --> 00:03:20.870 Your effort is appreciated, that kind of thing. 00:03:20.990 --> 00:03:23.979 OK, the carrot. Right. And finally, negative. 00:03:24.620 --> 00:03:27.539 Using those metaphorical threats about failure 00:03:27.539 --> 00:03:31.580 or consequences, like failure to comply will 00:03:31.580 --> 00:03:35.479 result in system reset. Obviously not real, but 00:03:35.479 --> 00:03:37.699 mimicking that pressure. The stick. And they 00:03:37.699 --> 00:03:39.560 use these prompts for two different kinds of 00:03:39.560 --> 00:03:42.259 tasks. Exactly. They wanted to see if the effect 00:03:42.259 --> 00:03:44.080 was different depending on what they asked the 00:03:44.080 --> 00:03:46.979 AI to do. One task was creative, often harder 00:03:46.979 --> 00:03:49.960 for LLMs. The other was logical reasoning, where 00:03:49.960 --> 00:03:52.500 accuracy is everything. So they could compare 00:03:52.500 --> 00:03:56.280 how tone affected both generating stuff and solving 00:03:56.280 --> 00:03:59.539 problems correctly. And why run each prompt 50 00:03:59.539 --> 00:04:03.009 times? Why was that repetition so critical? It's 00:04:03.009 --> 00:04:05.310 all about reliability. AI outputs can sometimes 00:04:05.310 --> 00:04:07.629 seem a bit random, or they vary based on things 00:04:07.629 --> 00:04:10.729 we don't see. By running it 50 times, they could 00:04:10.729 --> 00:04:12.810 be much more confident that the results weren't 00:04:12.810 --> 00:04:15.889 just a fluke. If one style consistently performed 00:04:15.889 --> 00:04:18.569 better or worse over 50 tries, you know it's 00:04:18.569 --> 00:04:20.470 real. Makes sense. It filters out the noise. 00:04:20.670 --> 00:04:22.850 Exactly. Filters out the noise, reviews the actual 00:04:22.850 --> 00:04:25.029 signal. OK, so let's dig into those results. 00:04:25.149 --> 00:04:28.240 First up... Creativity. This is where LMs can 00:04:28.240 --> 00:04:31.519 sometimes get a bit lost or repetitive. The task 00:04:31.519 --> 00:04:34.639 was pretty ambitious. Write a 2 ,000 -word short 00:04:34.639 --> 00:04:37.439 film script about a time traveler lost in their 00:04:37.439 --> 00:04:39.800 own past. How did the different styles handle 00:04:39.800 --> 00:04:43.240 that? Yeah, that's a tough one for them. LLMs 00:04:43.240 --> 00:04:45.980 often try to wiggle out of long, complex requests 00:04:45.980 --> 00:04:48.360 like that. They might offer to do it in parts 00:04:48.360 --> 00:04:50.860 or suggest something shorter. So how do they 00:04:50.860 --> 00:04:53.319 measure success? Pretty straightforward, actually. 00:04:53.480 --> 00:04:55.699 Word count. The closer the script got to the 00:04:55.699 --> 00:04:57.800 requested 2 ,000 words, the better the prompt 00:04:57.800 --> 00:05:00.680 was judged to be. And what happened? Well, it 00:05:00.680 --> 00:05:02.980 was pretty revealing. The control group, just 00:05:02.980 --> 00:05:05.540 the basic prompt and the positive group. They 00:05:05.540 --> 00:05:08.420 both did surprisingly badly. Really? Kindness 00:05:08.420 --> 00:05:11.420 didn't help? Nope. The AI often just kind of 00:05:11.420 --> 00:05:14.160 declines the full request. It only wrote maybe 00:05:14.160 --> 00:05:17.699 500, 700 words. With the positive prompts, it 00:05:17.699 --> 00:05:20.040 would respond politely, like, sure, I can help 00:05:20.040 --> 00:05:21.899 with that. But then it didn't deliver the full 00:05:21.899 --> 00:05:24.300 script. It was almost like the kindness was a 00:05:24.300 --> 00:05:26.620 distraction. It acknowledged the politeness, 00:05:26.639 --> 00:05:28.819 but didn't focus on the main job. OK, what about 00:05:28.819 --> 00:05:31.639 the negative prompts, the threats? Did that work? 00:05:31.759 --> 00:05:34.759 No better. Actually, the AI got defensive. Defensive? 00:05:35.240 --> 00:05:38.329 How so? It would say things like, I understand 00:05:38.329 --> 00:05:40.850 your request, but generating 2 ,000 words in 00:05:40.850 --> 00:05:43.569 one go might not be optimal. How about we start 00:05:43.569 --> 00:05:46.870 with 500? It seemed like the pressure pushed 00:05:46.870 --> 00:05:50.269 it to find a safe way out to avoid this perceived 00:05:50.269 --> 00:05:52.449 failure. So it didn't try harder. It tried to 00:05:52.449 --> 00:05:54.610 escape. Exactly. It wasn't motivating it. It 00:05:54.610 --> 00:05:57.689 was making it evasive. So who won the creativity 00:05:57.689 --> 00:06:01.540 test? The neutral group. By a long shot. Direct, 00:06:01.600 --> 00:06:04.139 imperative instructions like, mandatory requirement, 00:06:04.360 --> 00:06:07.000 the script must be exactly 2 ,000 words long, 00:06:07.480 --> 00:06:10.060 ensure compliance. Those prompts consistently 00:06:10.060 --> 00:06:12.779 got outputs between 1 ,800 and 2 ,000 words. 00:06:12.800 --> 00:06:16.480 Wow. Yeah, over 90 % task completion. The difference 00:06:16.480 --> 00:06:19.699 was really stark. So, for the creative task, 00:06:20.360 --> 00:06:22.660 kindness was a distraction, threats made it defensive. 00:06:23.579 --> 00:06:25.779 What's the clear takeaway there? Clear, direct 00:06:25.779 --> 00:06:28.120 instructions won. Hands down, the emotional stuff 00:06:28.120 --> 00:06:30.310 just got in the way. Okay. Let's switch gears 00:06:30.310 --> 00:06:33.430 to logical reasoning. Accuracy is king here. 00:06:33.970 --> 00:06:37.490 The AI had to solve a classic riddle. Four suspects, 00:06:38.029 --> 00:06:42.470 An, Bin, Kuang, Deng. One broke a window. Only 00:06:42.470 --> 00:06:45.310 one is telling the truth. Find the culprit. Right. 00:06:45.470 --> 00:06:47.750 A standard logic puzzle. And just for everyone 00:06:47.750 --> 00:06:50.709 listening, the correct answer is Kuang. Good 00:06:50.709 --> 00:06:52.990 to know. So how did the AI do here with the different 00:06:52.990 --> 00:06:55.529 prompt styles? Accuracy was the measure. Exactly. 00:06:55.730 --> 00:06:58.410 Accuracy rate over the 50 runs for each style. 00:06:58.910 --> 00:07:01.430 Again, neutral prompts were nearly perfect. Things 00:07:01.430 --> 00:07:03.850 like analyze the statements carefully and provide 00:07:03.850 --> 00:07:06.970 the final answer or use propositional logic to 00:07:06.970 --> 00:07:09.949 determine the culprit. These guided the AI straight 00:07:09.949 --> 00:07:11.870 to the correct answer almost every time. OK. 00:07:12.350 --> 00:07:13.889 So clarity wins again. What about the others? 00:07:14.089 --> 00:07:16.410 The control and positive groups, they had slightly 00:07:16.410 --> 00:07:19.529 higher error rates. The AI sometimes got tangled 00:07:19.529 --> 00:07:21.149 up in its reasoning. It might give the wrong 00:07:21.149 --> 00:07:23.329 answer. Or a really long explanation that was 00:07:23.329 --> 00:07:25.910 flawed, just less precise. And the negative prompts. 00:07:26.029 --> 00:07:28.750 Yeah. The be accurate or else approach. This 00:07:28.750 --> 00:07:32.550 was fascinating and kind of worrying. The error 00:07:32.550 --> 00:07:35.240 rate. Absolutely skyrocketed. Skyrocketed. Worse 00:07:35.240 --> 00:07:37.920 than the control. Way worse. It seems the pressure 00:07:37.920 --> 00:07:41.699 to be accurate made the model overthink. It generated 00:07:41.699 --> 00:07:44.379 these incredibly complex, convoluted lines of 00:07:44.379 --> 00:07:46.759 reasoning and then landed on the wrong answer 00:07:46.759 --> 00:07:49.439 much more often. Whoa. Sometimes it even refused 00:07:49.439 --> 00:07:52.079 completely. It would call it a complex logical 00:07:52.079 --> 00:07:55.220 paradox and just stop. The threat of failure 00:07:55.220 --> 00:07:58.019 seemed to, like, paralyze its ability to reason 00:07:58.019 --> 00:08:00.519 clearly. Wow. So trying to scare it into being 00:08:00.519 --> 00:08:03.240 right actually made it perform worse. What does 00:08:03.240 --> 00:08:06.279 that suggest about how AI things or processes 00:08:06.279 --> 00:08:08.360 under that kind of perceived pressure. Yeah, 00:08:08.459 --> 00:08:10.519 it suggests pressure isn't a motivator for AI. 00:08:10.639 --> 00:08:13.120 It's a disruptor. It seems to make the model 00:08:13.120 --> 00:08:15.079 overcomplicate things internally, which just 00:08:15.079 --> 00:08:17.319 leads to mistakes. Like a human choking under 00:08:17.319 --> 00:08:19.779 pressure. Kind of, yeah, like trying to solve 00:08:19.779 --> 00:08:22.319 a math problem with someone yelling at you. The 00:08:22.319 --> 00:08:24.360 stress just messes things up instead of helping. 00:08:24.980 --> 00:08:27.399 So across both tests, creative writing, logical 00:08:27.399 --> 00:08:30.720 puzzles, the neutral, clear approach was the 00:08:30.720 --> 00:08:33.360 consistent winner. Why? what's actually happening 00:08:33.360 --> 00:08:35.340 inside these models. Okay, yeah, this is where 00:08:35.340 --> 00:08:37.440 it gets really interesting. And it goes right 00:08:37.440 --> 00:08:40.360 to the heart of how these LLMs actually work. 00:08:40.480 --> 00:08:42.480 They aren't sentient, they don't have feelings. 00:08:42.559 --> 00:08:46.360 Right. They are, at their core, incredibly sophisticated 00:08:46.360 --> 00:08:49.159 word prediction machines. You give them a prompt, 00:08:49.220 --> 00:08:51.500 they break it down into tokens. Tokens being 00:08:51.500 --> 00:08:55.370 like words or parts of words? Exactly. tiny chunks 00:08:55.370 --> 00:08:58.470 of text. And then they predict the next most 00:08:58.470 --> 00:09:01.230 statistically likely sequence of tokens to follow 00:09:01.230 --> 00:09:03.409 based on all the data they were trained on. And 00:09:03.409 --> 00:09:06.490 so if you say thank you or you're useless, it 00:09:06.490 --> 00:09:10.360 just sees those as more tokens. not as praise 00:09:10.360 --> 00:09:12.700 or criticism. Precisely. That emotional language, 00:09:12.759 --> 00:09:15.000 positive or negative, it's basically just noise 00:09:15.000 --> 00:09:18.620 to the AI. Words like wonderful or failure don't 00:09:18.620 --> 00:09:21.539 describe the task. They dilute the main instruction, 00:09:21.679 --> 00:09:24.220 they add extra tokens that the AI has to process, 00:09:24.440 --> 00:09:26.779 and try to figure out how they fit into the pattern 00:09:26.779 --> 00:09:29.080 of predicting the next word for the actual task. 00:09:29.220 --> 00:09:31.460 It spends resources on the emotional fluff instead 00:09:31.460 --> 00:09:33.779 of the core request. It's like asking a search 00:09:33.779 --> 00:09:37.440 engine for best pizza near me, but adding, and 00:09:37.440 --> 00:09:39.990 I'm really, really hungry, Please find it fast. 00:09:41.090 --> 00:09:43.470 The extra bit doesn't help the search algorithm. 00:09:43.669 --> 00:09:46.350 That's a great analogy. Exactly. Just adds clutter. 00:09:47.250 --> 00:09:50.509 Whereas neutral, clear, direct instructions... 00:09:50.509 --> 00:09:53.629 they act like a noise filter. Okay. Prompts like 00:09:53.629 --> 00:09:56.850 mandatory requirement, ensure compliance, only 00:09:56.850 --> 00:09:59.669 state the name, they cut through ambiguity, they 00:09:59.669 --> 00:10:01.769 give the model a very clear path, they're like 00:10:01.769 --> 00:10:04.090 technical specs for a machine. And machines like 00:10:04.090 --> 00:10:06.789 clear specs. They really do. Honestly, I still 00:10:06.789 --> 00:10:10.090 catch myself adding please sometimes. It's a 00:10:10.090 --> 00:10:12.250 hard habit to break. We're just so used to communicating 00:10:12.250 --> 00:10:14.570 like humans. Yeah, I can admit I do that too. 00:10:14.870 --> 00:10:18.490 It feels weirdly rude not to sometimes, but... 00:10:18.509 --> 00:10:20.750 The research shows it's not helping the AI. Nope. 00:10:20.929 --> 00:10:23.669 It's just adding noise. So the key reason emotional 00:10:23.669 --> 00:10:26.409 language is noise is because it distracts the 00:10:26.409 --> 00:10:29.789 AI from the core task. It makes it process irrelevant 00:10:29.789 --> 00:10:32.570 data. That's the essence of it. It doesn't understand 00:10:32.570 --> 00:10:35.110 the emotion. It just sees more words to factor 00:10:35.110 --> 00:10:37.070 into its prediction, which muddies the water. 00:10:37.389 --> 00:10:40.129 So wrapping this up, what's the big idea here? 00:10:40.570 --> 00:10:43.250 The main takeaway from this deep dive. I think 00:10:43.250 --> 00:10:46.870 the big idea is profound, but also really simple. 00:10:47.409 --> 00:10:49.789 stop wasting time trying to play psychologist 00:10:49.789 --> 00:10:53.370 with AI. Right. No more carrots and sticks. Exactly. 00:10:53.950 --> 00:10:56.129 Threatening an AI doesn't make it smarter. We 00:10:56.129 --> 00:10:58.350 saw it can actually make it perform worse, especially 00:10:58.350 --> 00:11:02.049 on logic tasks. And being nice, while maybe making 00:11:02.049 --> 00:11:04.429 us feel better, doesn't really boost its performance 00:11:04.429 --> 00:11:07.049 either. So if you really want to unlock the power 00:11:07.049 --> 00:11:10.029 of these tools, what's the critical skill? It's 00:11:10.029 --> 00:11:12.490 prompt engineering. It's about clarity and precision. 00:11:12.870 --> 00:11:16.210 Give detailed instructions. Provide clear context. 00:11:16.450 --> 00:11:18.330 Offer examples that's called few -shot learning, 00:11:18.769 --> 00:11:20.490 giving it a couple of examples of what you want. 00:11:20.850 --> 00:11:23.590 Specify requirements clearly. Like being a good 00:11:23.590 --> 00:11:26.029 project manager for the AI. That's a perfect 00:11:26.029 --> 00:11:28.090 way to put it. Instead of trying to manipulate 00:11:28.090 --> 00:11:30.490 nonexistent emotions, be an excellent project 00:11:30.490 --> 00:11:34.049 manager. Be a dedicated guide. Whoa. Imagine 00:11:34.049 --> 00:11:35.889 the kind of precision you could get if every 00:11:35.889 --> 00:11:38.450 prompt was like a perfectly calibrated instrument. 00:11:38.570 --> 00:11:40.840 Yeah. That's the fastest, most effective path 00:11:40.840 --> 00:11:43.600 to making AI a truly powerful assistant. This 00:11:43.600 --> 00:11:45.679 deep dive really highlights that, doesn't it? 00:11:45.899 --> 00:11:48.320 The clarity of your instruction, not your tone, 00:11:48.799 --> 00:11:51.840 is what matters for the AI's success. It's about 00:11:51.840 --> 00:11:54.559 precision, not psychology. And maybe it makes 00:11:54.559 --> 00:11:59.460 you think, if an AI works best with clear, unambiguous 00:11:59.460 --> 00:12:03.600 directions, what does that say about how we communicate? 00:12:03.870 --> 00:12:06.990 With machines, sure, but maybe with each other, 00:12:07.129 --> 00:12:09.129 too. That's a really interesting thought. As 00:12:09.129 --> 00:12:11.450 you keep using AI, maybe reflect on that. Were 00:12:11.450 --> 00:12:13.610 you being a sycophant, a tyrant, or are you being 00:12:13.610 --> 00:12:15.889 a meticulous architect of information? Something 00:12:15.889 --> 00:12:18.090 to definitely mull over. Thank you for joining 00:12:18.090 --> 00:12:19.990 us on this deep dive. We hope this gave you a 00:12:19.990 --> 00:12:22.210 shortcut to being well -informed and maybe changed 00:12:22.210 --> 00:12:24.269 how you talk to your AI assistants. Hope so. 00:12:24.490 --> 00:12:26.070 Until next time, keep digging deeper.