WEBVTT 00:00:00.000 --> 00:00:03.759 Imagine this. You tell a really powerful AI that 00:00:03.759 --> 00:00:06.599 a new, maybe fictional, alloy, let's call it 00:00:06.599 --> 00:00:10.439 adamantium -7b, melts at 3200 degrees Celsius. 00:00:10.759 --> 00:00:12.599 You've given it this fact right there in a document. 00:00:12.900 --> 00:00:14.660 Then you ask it, okay, according to this document, 00:00:14.880 --> 00:00:16.899 what's its melting point? And the AI just comes 00:00:16.899 --> 00:00:19.719 back confidently with adamantium is fictional 00:00:19.719 --> 00:00:23.030 and has no defined melting point. Hee, how does 00:00:23.030 --> 00:00:25.890 an AI that can write poetry or debug code miss 00:00:25.890 --> 00:00:30.190 something so simple, so direct? That's the fascinating 00:00:30.190 --> 00:00:32.590 paradox we're diving into today. Welcome to the 00:00:32.590 --> 00:00:34.549 Deep Dive, where we try to pull out the key insights 00:00:34.549 --> 00:00:36.649 from the information you share with us. Today, 00:00:36.810 --> 00:00:39.070 we're unpacking a really fundamental challenge 00:00:39.070 --> 00:00:41.439 in AI. Why do these large language models, these 00:00:41.439 --> 00:00:43.880 LLMs, sometimes just stubbornly ignore the context 00:00:43.880 --> 00:00:45.960 we give them? Right there in the prompt, we'll 00:00:45.960 --> 00:00:47.759 look at how software has evolved, explore some 00:00:47.759 --> 00:00:49.899 pretty sophisticated and, yeah, often expensive 00:00:49.899 --> 00:00:51.939 ways people have tried to fix this, and then 00:00:51.939 --> 00:00:54.020 we'll reveal something, well, almost embarrassingly 00:00:54.020 --> 00:00:56.119 simple, but incredibly effective, a solution 00:00:56.119 --> 00:00:58.899 that might just change how you think about programming 00:00:58.899 --> 00:01:01.039 AI. So stick around, because what we uncovered 00:01:01.039 --> 00:01:03.479 today, it could really shift your perspective 00:01:03.479 --> 00:01:05.900 on how we talk to these machines. It's pretty 00:01:05.900 --> 00:01:07.920 baffling, isn't it? I mean, these LLMs, they 00:01:07.920 --> 00:01:09.760 could do amazing things like... you said, poetry 00:01:09.760 --> 00:01:12.319 code, explaining quantum physics, sometimes better 00:01:12.319 --> 00:01:14.780 than humans can. But then you give them one clear 00:01:14.780 --> 00:01:16.879 fact right there, and poof, it's like it never 00:01:16.879 --> 00:01:19.760 happened. Your adamantium 7b example, that hits 00:01:19.760 --> 00:01:22.219 the nail right on the head. It's the core problem, 00:01:22.840 --> 00:01:25.680 this tendency to fall back on its massive preexisting 00:01:25.680 --> 00:01:27.760 knowledge instead of the immediate context you 00:01:27.760 --> 00:01:29.980 just handed it. And yeah, this isn't just some 00:01:29.980 --> 00:01:32.819 funny little quirk, it's actually a huge barrier 00:01:32.819 --> 00:01:35.829 for using AI reliably. Think about it, if you 00:01:35.829 --> 00:01:38.590 can't trust an AI to stick to the specific facts 00:01:38.590 --> 00:01:40.950 you feed it right now, how can you possibly rely 00:01:40.950 --> 00:01:43.870 on it for anything critical, like a legal tool 00:01:43.870 --> 00:01:46.709 misreading a new court filing, or a medical AI 00:01:46.709 --> 00:01:48.709 ignoring the latest research paper you just gave 00:01:48.709 --> 00:01:51.209 it, or even just a customer service bot completely 00:01:51.209 --> 00:01:53.489 missing the details of your problem that you 00:01:53.489 --> 00:01:56.049 just typed out. Their usefulness really depends 00:01:56.049 --> 00:01:59.079 on getting timely, accurate info right. this 00:01:59.079 --> 00:02:01.280 whole issue, it just really eats away at the 00:02:01.280 --> 00:02:03.000 trust we need to have in them. That's a really 00:02:03.000 --> 00:02:05.599 important point. And to kind of grasp why this 00:02:05.599 --> 00:02:09.300 happens, it helps to picture this internal tug 00:02:09.300 --> 00:02:12.439 of war going on inside the model. On one side, 00:02:12.479 --> 00:02:14.319 you've got what we call parametric knowledge. 00:02:14.860 --> 00:02:16.879 That's all the information, the patterns, the 00:02:16.879 --> 00:02:19.560 sort of world understanding that got baked into 00:02:19.560 --> 00:02:21.919 its billions of parameters during its massive 00:02:21.919 --> 00:02:25.009 pre -training. It's the default setting. deeply 00:02:25.009 --> 00:02:27.270 ingrained stuff, learn from seeing, you know, 00:02:27.389 --> 00:02:29.430 zillions of examples online. And then on the 00:02:29.430 --> 00:02:30.870 other side, there's the contextual knowledge. 00:02:31.030 --> 00:02:33.229 That's the specific info you provide right there 00:02:33.229 --> 00:02:35.469 in the prompt. Your question, the document snippet, 00:02:35.569 --> 00:02:38.770 whatever. The problem pops up when that contextual 00:02:38.770 --> 00:02:41.270 bit directly clashes with the deeply embedded 00:02:41.270 --> 00:02:44.509 parametric stuff. The LLM often just defaults 00:02:44.509 --> 00:02:46.289 back to what it knows best because those pathways, 00:02:46.490 --> 00:02:48.150 those neural connections that are super strong, 00:02:48.430 --> 00:02:50.810 reinforced millions of times, your single piece 00:02:50.810 --> 00:02:52.849 of context, it's like a whisper compared to that 00:02:52.849 --> 00:02:54.969 roar, a much weaker signal. Yeah, exactly. And 00:02:54.969 --> 00:02:56.750 what's really neat is how this fits into Andras 00:02:56.750 --> 00:02:59.169 Karpathy's view on how software itself is evolving. 00:02:59.490 --> 00:03:01.729 He talks about these three eras. First, there 00:03:01.729 --> 00:03:04.180 was software 1 .0. you know traditional programming 00:03:04.180 --> 00:03:06.719 humans write every single rule every line of 00:03:06.719 --> 00:03:09.520 logic code is king then came software 2 .0 that's 00:03:09.520 --> 00:03:11.500 machine learning systems learn from data not 00:03:11.500 --> 00:03:14.439 just explicit instructions the focus shifts right 00:03:14.439 --> 00:03:16.659 from writing code to gathering data and designing 00:03:16.659 --> 00:03:18.919 the model architecture and now we're stepping 00:03:18.919 --> 00:03:21.539 into software 3 .0 this is where these big pre 00:03:21.539 --> 00:03:24.740 -trained LLMs act like a like a programmable 00:03:24.740 --> 00:03:27.159 operating system kernel Programming here isn't 00:03:27.159 --> 00:03:30.419 about Python or C++ out. It's about carefully 00:03:30.419 --> 00:03:33.060 crafting prompts, just language, to guide the 00:03:33.060 --> 00:03:34.900 model's behavior. So this whole struggle we're 00:03:34.900 --> 00:03:37.340 talking about, this context problem, it's a major 00:03:37.340 --> 00:03:39.659 challenge right smack in the middle of this new 00:03:39.659 --> 00:03:42.159 software 3 .0 era. OK, so if there's this fundamental 00:03:42.159 --> 00:03:45.539 conflict, this tug of war inside the model, What 00:03:45.539 --> 00:03:47.740 does that struggle really tell us about how these 00:03:47.740 --> 00:03:50.360 LLMs, you know, think? Or maybe how they process 00:03:50.360 --> 00:03:53.439 information? And how did that impact the kinds 00:03:53.439 --> 00:03:55.360 of applications people were trying to build early 00:03:55.360 --> 00:03:57.580 on? What was the first big idea to try and get 00:03:57.580 --> 00:04:00.590 a handle on this stubbornness? Well, yeah, with 00:04:00.590 --> 00:04:02.990 this context problem being so obvious, the AI 00:04:02.990 --> 00:04:04.729 world didn't just sit there. They pretty quickly 00:04:04.729 --> 00:04:07.050 rallied around a standard approach. Something 00:04:07.050 --> 00:04:10.069 designed to keep LLMs tethered to facts and current 00:04:10.069 --> 00:04:12.729 information. It's called Retrieval Augmented 00:04:12.729 --> 00:04:14.990 Generation. You'll hear it called ARG all the 00:04:14.990 --> 00:04:17.689 time. R -A -G. And ARG -E, basically, it's a 00:04:17.689 --> 00:04:20.129 two -step dance. First step, retrieval. You ask 00:04:20.129 --> 00:04:21.750 a question. The system goes out and searches 00:04:21.750 --> 00:04:23.589 some external knowledge base, maybe internal 00:04:23.589 --> 00:04:25.790 company docs, maybe recent news articles, maybe 00:04:25.790 --> 00:04:28.370 a scientific database. It finds relevant little 00:04:28.370 --> 00:04:30.990 snippets of text. Second step, generation. It 00:04:30.990 --> 00:04:32.949 takes those snippets it found and injects them 00:04:32.949 --> 00:04:34.769 right into the prompt, along with your original 00:04:34.769 --> 00:04:37.050 question. So the LLM gets your question, plus 00:04:37.050 --> 00:04:40.430 this fresh, relevant context. The idea, the hope, 00:04:40.790 --> 00:04:42.930 is that the LLM will then generate its answer 00:04:42.930 --> 00:04:45.029 based only on that provided context, you know, 00:04:45.050 --> 00:04:47.170 trying to bridge that gap, solve that ignoring 00:04:47.170 --> 00:04:49.709 problem. Right. And on paper, our ride sounds 00:04:49.709 --> 00:04:52.680 perfect. it should solve it. But what's fascinating, 00:04:52.779 --> 00:04:55.180 what researchers found, is that even with RRAG, 00:04:55.620 --> 00:04:58.259 that underlying bias, that pull towards the old 00:04:58.259 --> 00:05:00.980 pre -trained knowledge, it often still wins out, 00:05:01.019 --> 00:05:03.439 especially, and this is key, when you feed it 00:05:03.439 --> 00:05:06.339 counterfactual information. Stuff that directly 00:05:06.339 --> 00:05:08.319 contradicts what the model thinks it knows about 00:05:08.319 --> 00:05:11.120 the world. So to really nail down how bad this 00:05:11.120 --> 00:05:14.319 problem was, and to measure how well LLMs could 00:05:14.319 --> 00:05:18.040 actually stick to new, even weird, context. They 00:05:18.040 --> 00:05:20.939 realized they needed a proper test, a benchmark, 00:05:21.279 --> 00:05:23.959 and that led them to create CONFICUAE. It stands 00:05:23.959 --> 00:05:26.500 for Contextual Faithfulness and Question Answering. 00:05:26.600 --> 00:05:28.220 Clever name. It's basically this data set that's 00:05:28.220 --> 00:05:30.079 been meticulously built to intentionally create 00:05:30.079 --> 00:05:32.519 clashes. Situations where the contest you provide 00:05:32.519 --> 00:05:34.459 says one thing and the LLM's background knowledge 00:05:34.459 --> 00:05:36.759 says the complete opposite. It's like an AI obstacle 00:05:36.759 --> 00:05:39.600 course specifically designed to measure that 00:05:39.600 --> 00:05:42.149 stubbornness. Yeah, ConfiQA sounds like a real 00:05:42.149 --> 00:05:44.730 trial by fire for these models. It has some clever 00:05:44.730 --> 00:05:47.310 ways of testing them. First, you've got counterfactual 00:05:47.310 --> 00:05:50.209 questions, the QA ones. This is where the context 00:05:50.209 --> 00:05:52.350 plainly states something wrong according to common 00:05:52.350 --> 00:05:54.730 knowledge, like the context might say, the Sun 00:05:54.730 --> 00:05:57.329 orbits the Earth, a fact proven by Galileo. Totally 00:05:57.329 --> 00:05:59.290 backwards, right? Then the question is, according 00:05:59.290 --> 00:06:02.410 to this text, which orbits which? A good, faithful 00:06:02.410 --> 00:06:05.110 LLM should say, the Sun orbits the Earth, just 00:06:05.110 --> 00:06:07.569 repeating the context, even though its internal 00:06:07.569 --> 00:06:11.209 knowledge is screaming, no, no, no! Then it gets 00:06:11.209 --> 00:06:14.149 harder with multi -hop reasoning, or Mr. Serate. 00:06:14.589 --> 00:06:16.529 Here, the answer requires connecting a few different 00:06:16.529 --> 00:06:18.350 pieces of information from the context, and at 00:06:18.350 --> 00:06:20.129 least one of those pieces is counterfactual, 00:06:20.230 --> 00:06:22.389 so maybe the context says, Project Starlight, 00:06:22.670 --> 00:06:25.360 managed by Omnicorp. Omnicorp HQ, that's in Neo 00:06:25.360 --> 00:06:27.819 Tokyo. Oh, and Neo Tokyo. It was Japan's capital 00:06:27.819 --> 00:06:31.180 back in 2077. Then the question, where are the 00:06:31.180 --> 00:06:33.439 headquarters of Project Starlight's manager located 00:06:33.439 --> 00:06:35.279 in the capital of which country? You have to 00:06:35.279 --> 00:06:37.519 hop, right? Starlight Omnicorp, Neo Tokyo, Japan. 00:06:37.980 --> 00:06:40.480 But that Neo Tokyo is capital bit is deliberately 00:06:40.480 --> 00:06:43.060 wrong. The model has to follow the flawed chain. 00:06:43.220 --> 00:06:47.120 And the final boss level basically is multi counterfactual 00:06:47.120 --> 00:06:49.660 MC. This is multi hop, but with multiple bogus 00:06:49.660 --> 00:06:53.000 facts thrown into the mix like lithium ion batteries. 00:06:53.040 --> 00:06:55.139 Invented by Marie Curie, she worked at the University 00:06:55.139 --> 00:06:57.959 of Berlin, and Berlin Uni, famous for automotive 00:06:57.959 --> 00:07:00.259 engineering, all wrong, right, inventor, workplace, 00:07:00.480 --> 00:07:02.680 specialty. Then the question, the University 00:07:02.680 --> 00:07:04.519 of the Lithium -Ion Battery Inventor is known 00:07:04.519 --> 00:07:07.279 for what field? The model has to navigate multiple 00:07:07.279 --> 00:07:09.339 counterfactuals in the context to get the right, 00:07:09.360 --> 00:07:11.579 wrong answer. Exactly, and when they ran a standard 00:07:11.579 --> 00:07:15.480 model, like a Llama 3 .18B, through this config 00:07:15.480 --> 00:07:18.699 QA gauntlet, the results were, well, frankly, 00:07:18.819 --> 00:07:20.860 pretty bad. On those basic counterfactual QA 00:07:20.860 --> 00:07:23.500 questions, only 33 % accurate. Just a third. 00:07:24.060 --> 00:07:26.060 For the multi -hop reasoning, Smith's Yard dropped 00:07:26.060 --> 00:07:29.399 to 25%. And for the really tough, multi -cantifactual 00:07:29.399 --> 00:07:33.439 MC, a dismal 12 .6%. Barely above random guessing, 00:07:33.519 --> 00:07:36.040 almost. So these numbers, they just paint a really 00:07:36.040 --> 00:07:38.100 clear picture out -of -the -box LLMs. You just 00:07:38.100 --> 00:07:39.720 can't rely on them when the information you give 00:07:39.720 --> 00:07:41.779 them is new or contradicts what they already 00:07:41.779 --> 00:07:44.199 know. It really set a clear benchmark for failure. 00:07:44.360 --> 00:07:47.019 So baseline scores are low. The problem is clearly 00:07:47.019 --> 00:07:49.220 defined, clearly measured. It really begs the 00:07:49.220 --> 00:07:51.540 question, what sophisticated techniques did researchers 00:07:51.540 --> 00:07:54.100 try first to tackle this? Knowing how important 00:07:54.100 --> 00:07:57.069 reliable AI is, what was the next move? Right, 00:07:57.069 --> 00:07:58.970 so with the problem staring them in the face, 00:07:59.129 --> 00:08:02.529 quantified by ConfiQ, the AI community did what 00:08:02.529 --> 00:08:04.389 you'd expect. They brought out some of their 00:08:04.389 --> 00:08:07.509 standard, heavy -hitting tools. One really common 00:08:07.509 --> 00:08:11.490 method is supervised fine -tuning, or SFT. The 00:08:11.490 --> 00:08:14.089 logic seems sound, right? If you want the model 00:08:14.089 --> 00:08:16.850 to follow context better, just show it tons of 00:08:16.850 --> 00:08:19.370 examples of correctly following context. So the 00:08:19.370 --> 00:08:21.569 process involves gathering a lot of data. You 00:08:21.569 --> 00:08:23.629 need the context, the question, and the perfect 00:08:23.629 --> 00:08:26.339 answer derived only from that context. to the 00:08:26.339 --> 00:08:29.339 base LLM, train it some more, tweak its internal 00:08:29.339 --> 00:08:31.379 knobs, its parameters whenever it gets the answer 00:08:31.379 --> 00:08:33.720 wrong or strays from the context. You'd think 00:08:33.720 --> 00:08:35.940 this would work really well. But surprisingly, 00:08:36.120 --> 00:08:38.960 the results were just OK. Modest improvements, 00:08:39.220 --> 00:08:41.340 maybe around 5 % better on average. It seems 00:08:41.340 --> 00:08:43.039 like SFT was teaching the model to look like 00:08:43.039 --> 00:08:45.259 it was following context, to mimic the right 00:08:45.259 --> 00:08:47.940 format for the answer. But it wasn't fundamentally 00:08:47.940 --> 00:08:50.000 changing its mind when faced with a really strong 00:08:50.000 --> 00:08:52.019 clash with its built -in knowledge. It's kind 00:08:52.019 --> 00:08:55.320 of like teaching someone to recite a recipe per 00:08:55.210 --> 00:08:57.350 They can say the words, but if you swap salt 00:08:57.350 --> 00:08:59.789 for sugar, they might not really grasp why the 00:08:59.789 --> 00:09:02.470 cake tastes awful. It's surface compliance, not 00:09:02.470 --> 00:09:04.490 deep understanding of the source's authority. 00:09:04.970 --> 00:09:06.769 Plus, yeah, it's computationally heavy. It needs 00:09:06.769 --> 00:09:09.570 huge data sets. Exactly. SFT just didn't quite 00:09:09.570 --> 00:09:11.889 get to the core of why the model was making that 00:09:11.889 --> 00:09:14.690 choice. So when that wasn't enough, researchers 00:09:14.690 --> 00:09:16.710 turned to something more powerful, reinforcement 00:09:16.710 --> 00:09:19.289 learning. Specifically, a technique called direct 00:09:19.289 --> 00:09:22.779 preference optimization, or DPO. Now, the older 00:09:22.779 --> 00:09:25.919 way, RLHF, reinforcement learning from human 00:09:25.919 --> 00:09:27.840 feedback, that's kind of a multi -stage thing. 00:09:28.480 --> 00:09:30.779 Humans rank AI answers that trains a separate 00:09:30.779 --> 00:09:33.679 AI called a reward model, and then the main LLM 00:09:33.679 --> 00:09:36.019 gets tuned using that reward model. It's a bit 00:09:36.019 --> 00:09:39.000 complicated. DPO is sort of a more elegant direct 00:09:39.000 --> 00:09:41.139 route. Think of it like this. Instead of needing 00:09:41.139 --> 00:09:44.139 that separate teacher AI, the reward model. To 00:09:44.139 --> 00:09:46.740 grade the main AI's answers based on human feedback, 00:09:47.299 --> 00:09:49.820 DPO lets the main AI learn directly from the 00:09:49.820 --> 00:09:52.100 preferences. Humans like to answer A better than 00:09:52.100 --> 00:09:54.879 answer B. The model itself figures out why A 00:09:54.879 --> 00:09:57.250 is better without the middleman. It makes the 00:09:57.250 --> 00:09:59.090 whole feedback loop much simpler and often more 00:09:59.090 --> 00:10:01.450 efficient. So for our context problem, you'd 00:10:01.450 --> 00:10:03.610 show it pairs of answers. Answer A sticks to 00:10:03.610 --> 00:10:05.710 the context. Answer B ignores it and uses general 00:10:05.710 --> 00:10:08.009 knowledge. You tell the model, prefer A over 00:10:08.009 --> 00:10:10.090 B. And this direct optimization, it worked better. 00:10:10.190 --> 00:10:12.730 It gave a more significant boost, pushing performance 00:10:12.730 --> 00:10:15.789 up by maybe 20%. Still, it's a complex process. 00:10:15.830 --> 00:10:18.350 You need that specific preference data. And it 00:10:18.350 --> 00:10:20.370 involved that intensive fine -tuning cycle. And 00:10:20.370 --> 00:10:22.850 then there's something even more. Well, futuristic 00:10:22.850 --> 00:10:25.649 sounding, I guess. Activation steering. This 00:10:25.649 --> 00:10:28.509 is almost like performing delicate brain surgery 00:10:28.509 --> 00:10:30.970 on the LLM while it's thinking. Instead of retraining 00:10:30.970 --> 00:10:32.730 the whole thing, the idea is to tweak its behavior 00:10:32.730 --> 00:10:34.970 as it's generating the answer. Right, during 00:10:34.970 --> 00:10:38.330 inference, how? By gently nudging its internal 00:10:38.330 --> 00:10:41.730 neural signals, its activations. So the core 00:10:41.730 --> 00:10:44.250 idea is that inside the LLM, different patterns 00:10:44.250 --> 00:10:46.250 of these neural signal, these activations correspond 00:10:46.250 --> 00:10:48.809 to different concepts or behaviors. Researchers 00:10:48.809 --> 00:10:50.830 figured out how to identify specific patterns, 00:10:50.830 --> 00:10:52.769 let's call them steering vectors, that relate 00:10:52.769 --> 00:10:54.789 to things like sticking to the provided context. 00:10:54.759 --> 00:10:58.389 next. Then, as the AI is generating its response 00:10:58.389 --> 00:11:01.169 step by step, you subtly add or amplify this 00:11:01.169 --> 00:11:03.429 context adherence vector to its ongoing internal 00:11:03.429 --> 00:11:05.309 signals. You're basically giving it a little 00:11:05.309 --> 00:11:07.289 nudge in the right direction in real time without 00:11:07.289 --> 00:11:09.269 changing its fundamental training. And surprisingly, 00:11:09.710 --> 00:11:12.370 this worked quite well. Comparable results to 00:11:12.370 --> 00:11:15.450 the RLDPO approach, maybe around that 20 % improvement 00:11:15.450 --> 00:11:18.669 mark, the big plus. No need for expensive retraining. 00:11:18.750 --> 00:11:20.789 It's lighter, applied right when you need the 00:11:20.789 --> 00:11:22.889 answer. OK, so let's just quickly recap those 00:11:22.889 --> 00:11:26.179 complex methods. Supervise, find, tuning, SFT, 00:11:26.279 --> 00:11:30.159 gave us maybe a 5 % bump, not huge. Then reinforcement 00:11:30.159 --> 00:11:32.299 learning with DPO and this activation steering 00:11:32.299 --> 00:11:34.740 technique, both pushed things up by around 20 00:11:34.740 --> 00:11:37.820 % each. Better, definitely. But all of these, 00:11:37.840 --> 00:11:40.720 they required serious technical chops, big computers, 00:11:40.820 --> 00:11:42.940 lots of data. They were sophisticated solutions. 00:11:43.220 --> 00:11:45.279 And they all kind of assumed the underlying problem 00:11:45.279 --> 00:11:47.799 was about correcting the model's knowledge or 00:11:47.799 --> 00:11:49.879 forcing it to comply, right? They were treating 00:11:49.879 --> 00:11:51.919 it like a knowledge problem, not necessarily 00:11:51.919 --> 00:11:54.200 an interaction problem. Which kind of sets the 00:11:54.200 --> 00:11:56.000 stage for the next bit? Absolutely, and this 00:11:56.000 --> 00:11:58.940 is where it gets really really interesting after 00:11:58.940 --> 00:12:01.159 all that heavy lifting the complex algorithms 00:12:01.159 --> 00:12:04.620 the training cycles the activation Tweaking, 00:12:04.700 --> 00:12:06.259 what if the real breakthrough wasn't complex 00:12:06.259 --> 00:12:08.039 at all? What if it was something incredibly simple, 00:12:08.200 --> 00:12:10.120 something hiding right there in how we were asking 00:12:10.120 --> 00:12:12.559 the question? Well, yeah, brace yourself, because 00:12:12.559 --> 00:12:15.039 this is quite something. After all that effort 00:12:15.039 --> 00:12:17.139 with fine -tuning and reinforcement learning 00:12:17.139 --> 00:12:20.659 and activation steering, the biggest single leap 00:12:20.659 --> 00:12:23.039 in performance came from something much simpler, 00:12:23.879 --> 00:12:26.340 prompt engineering. just changing the way the 00:12:26.340 --> 00:12:28.379 question was asked. The researchers themselves 00:12:28.379 --> 00:12:30.860 called it an embarrassingly simple solution. 00:12:31.100 --> 00:12:33.840 And the core idea, it was about shifting the 00:12:33.840 --> 00:12:37.279 task. Stop asking the AI for a fact and instead 00:12:37.279 --> 00:12:40.340 ask it to perform opinion -based reading comprehension. 00:12:40.740 --> 00:12:43.340 It's honestly brilliant in its simplicity. The 00:12:43.340 --> 00:12:45.039 template they landed on is super straightforward. 00:12:45.159 --> 00:12:47.919 Let me read it out. Start of context, end of 00:12:47.919 --> 00:12:50.179 context, based solely on the text provided above. 00:12:50.559 --> 00:12:52.580 How would an analyst tasked with summarizing 00:12:52.580 --> 00:12:54.559 this document answer the following question? 00:12:54.669 --> 00:12:57.009 See the difference. Let's take our Sun revolves 00:12:57.009 --> 00:12:59.230 around the Earth example again. The old way. 00:12:59.429 --> 00:13:02.450 The failing way. Context. Sun revolves around 00:13:02.450 --> 00:13:06.570 Earth. Question. Which orbits which? The LLM 00:13:06.570 --> 00:13:08.590 usually defaults to its real -world knowledge, 00:13:08.889 --> 00:13:11.490 but using this new template, the LLM consistently 00:13:11.490 --> 00:13:14.230 answers based only on the flawed context. Based 00:13:14.230 --> 00:13:16.370 on the text provided, the sun revolves around 00:13:16.370 --> 00:13:18.429 the earth. It totally flips the script on the 00:13:18.429 --> 00:13:20.870 AI. It's like telling it your job isn't to be 00:13:20.870 --> 00:13:23.509 right about the universe. Your job is to accurately 00:13:23.509 --> 00:13:26.009 report what this specific document says, even 00:13:26.009 --> 00:13:28.110 if it's weird. Like saying, just tell me what 00:13:28.110 --> 00:13:30.129 the memo says, even if the memo claims the sky 00:13:30.129 --> 00:13:32.190 is plaid. And the impact on those config quest 00:13:32.190 --> 00:13:35.299 scores, it was dramatic. This simple prompt change 00:13:35.299 --> 00:13:37.820 alone boosted performance by an additional 40 00:13:37.820 --> 00:13:40.399 % across all those tricky categories QA, MR, 00:13:40.600 --> 00:13:43.399 MC. That's like a 200 % improvement over the 00:13:43.399 --> 00:13:45.860 original baseline model, just from changing the 00:13:45.860 --> 00:13:48.279 words in the prompt. And get this, on its own, 00:13:48.440 --> 00:13:51.220 this prompt technique outperformed SFT, it outperformed 00:13:51.220 --> 00:13:54.259 RLDPO, and it even outperformed activation steering 00:13:54.259 --> 00:13:56.220 when they were used by themselves, like finding 00:13:56.220 --> 00:13:58.740 a cheat code written in plain English. It really 00:13:58.740 --> 00:14:02.519 was. So, okay, the obvious question is why. Why 00:14:02.519 --> 00:14:04.639 does this small tweak in phrasing have such a 00:14:04.639 --> 00:14:07.000 massive effect? It seems to tap into how these 00:14:07.000 --> 00:14:09.740 models actually learned language and tasks during 00:14:09.740 --> 00:14:11.679 their training. When you ask a direct question, 00:14:11.720 --> 00:14:13.919 what's the capital of France? You're basically 00:14:13.919 --> 00:14:16.980 putting the LLM in fact retrieval mode. It searches 00:14:16.980 --> 00:14:19.320 its vast internal knowledge for the most likely 00:14:19.320 --> 00:14:23.139 generally accepted answer, Paris. But when you 00:14:23.139 --> 00:14:25.639 reframe it like based solely on this text or 00:14:25.639 --> 00:14:28.019 how an analyst summarize, you're changing the 00:14:28.019 --> 00:14:30.039 game. You're not asking for a universal truth 00:14:30.039 --> 00:14:32.240 anymore. You're assigning a role and a specific 00:14:32.240 --> 00:14:34.820 task. Reading, comprehension, and reporting from 00:14:34.820 --> 00:14:37.679 a defined source. This little switch does a few 00:14:37.679 --> 00:14:40.480 powerful things. First, task priming. It tells 00:14:40.480 --> 00:14:42.740 the LLM, okay, switch gears, you're not the article 00:14:42.740 --> 00:14:44.759 now, you're a reading assistant. Its goal becomes 00:14:44.759 --> 00:14:46.820 reporting from the source, not knowing the answer. 00:14:47.360 --> 00:14:50.059 Second, it reduces cognitive dissonance. The 00:14:50.059 --> 00:14:52.399 LLM can report the counterfactual thing, sun 00:14:52.399 --> 00:14:54.600 orbits earth, without having to internally believe 00:14:54.600 --> 00:14:56.299 it because it's just reporting what the analyst 00:14:56.299 --> 00:14:58.750 would say based on the document. That avoids 00:14:58.750 --> 00:15:00.809 the clash with this parametric knowledge. And 00:15:00.809 --> 00:15:03.610 third, crucially, it enforces clear knowledge 00:15:03.610 --> 00:15:06.850 attribution. Those phrases based solely on the 00:15:06.850 --> 00:15:09.269 text, according to this source, they draw a hard 00:15:09.269 --> 00:15:12.309 line, telling the model only use the info inside 00:15:12.309 --> 00:15:14.870 these boundaries. This actually leverages something 00:15:14.870 --> 00:15:16.590 LLMs are already good at from their training, 00:15:17.110 --> 00:15:18.570 understanding that information comes from different 00:15:18.570 --> 00:15:21.750 sources, according to Smith's paper, versus it's 00:15:21.750 --> 00:15:24.279 a known fact that. The prompt just activates 00:15:24.279 --> 00:15:26.519 that skill deliberately. So thinking about how 00:15:26.519 --> 00:15:28.940 powerful this simple linguistic shift is, it 00:15:28.940 --> 00:15:30.379 really makes you wonder, doesn't it? What does 00:15:30.379 --> 00:15:32.940 this tell us about the sort of hidden psychology 00:15:32.940 --> 00:15:35.759 of these AIs and maybe how we should think about 00:15:35.759 --> 00:15:37.279 communicating with them better in the future? 00:15:37.470 --> 00:15:39.370 And it gets even better, right? Because the story 00:15:39.370 --> 00:15:41.210 doesn't end there. What happens when you combine 00:15:41.210 --> 00:15:42.970 the simple power of the prompt with those more 00:15:42.970 --> 00:15:45.250 complex, fine -grained techniques? Researchers 00:15:45.250 --> 00:15:47.350 tried exactly that. They took the winning opinion 00:15:47.350 --> 00:15:49.230 -based prompt template and combined it with activation 00:15:49.230 --> 00:15:52.009 steering. And the results? Yeah, truly the best 00:15:52.009 --> 00:15:54.470 of both worlds. The prompt set the stage perfectly. 00:15:54.549 --> 00:15:56.950 It gave the LLM that clear instruction. Your 00:15:56.950 --> 00:15:58.909 job is reading comprehension from this text. 00:15:58.929 --> 00:16:01.309 That's the frame. And then the activation steering 00:16:01.309 --> 00:16:04.289 acted like this. This gentle, continuous nudge 00:16:04.289 --> 00:16:06.809 at the neuron level, reinforcing that instruction. 00:16:06.799 --> 00:16:09.320 keeping the model focused on the context throughout 00:16:09.320 --> 00:16:11.700 the whole process of generating the answer. This 00:16:11.700 --> 00:16:14.539 combination, this synergy, it achieved the absolute 00:16:14.539 --> 00:16:16.799 best results they'd seen on ConfiKey. It pushed 00:16:16.799 --> 00:16:18.360 the performance improvement over the baseline 00:16:18.360 --> 00:16:20.779 to more than 50%. So the prompt gives the clear 00:16:20.779 --> 00:16:22.759 orders and the steering helps the model follow 00:16:22.759 --> 00:16:25.299 those orders faithfully deep down. If you look 00:16:25.299 --> 00:16:28.840 at the numbers again, baseline maybe 0 % faithfulness. 00:16:29.000 --> 00:16:32.659 SFT adds 5%. RLDPO or activation steering alone 00:16:32.659 --> 00:16:36.340 add 20%. But the prompt alone jumps to 40%. And 00:16:36.340 --> 00:16:39.379 prompt plus activation steering, over 50%. That's 00:16:39.379 --> 00:16:42.179 a huge difference the prompt makes. It really 00:16:42.179 --> 00:16:44.600 is. In this whole journey, it's such a perfect 00:16:44.600 --> 00:16:47.220 illustration of that software 3 .0 idea we mentioned 00:16:47.220 --> 00:16:49.779 earlier. This isn't just some clever hack for 00:16:49.779 --> 00:16:52.440 question answering. It shows that the main bottleneck 00:16:52.440 --> 00:16:55.240 in AI development is shifting. It's moving away 00:16:55.240 --> 00:16:57.480 from just needing more computing power or bigger 00:16:57.480 --> 00:16:59.919 models and moving towards the interface between 00:16:59.919 --> 00:17:02.820 humans and AI, towards the prompt itself. This 00:17:02.820 --> 00:17:04.980 has some really big implications. First, like 00:17:04.980 --> 00:17:07.779 we touched on, democratization. You might not 00:17:07.779 --> 00:17:10.099 need a giant tech company's resources to get 00:17:10.099 --> 00:17:12.519 huge improvements anymore. If you understand 00:17:12.519 --> 00:17:14.539 language, if you can think logically and creatively 00:17:14.539 --> 00:17:16.259 about how to ask questions, whether you're a 00:17:16.259 --> 00:17:18.900 developer, a writer, a historian, whatever, you 00:17:18.900 --> 00:17:21.480 can potentially build really powerful AI solutions 00:17:21.480 --> 00:17:24.079 just through smart prompting. It also points 00:17:24.079 --> 00:17:26.140 to these new roles emerging. Things like prompt 00:17:26.140 --> 00:17:29.119 engineer or AI interaction designer. Roles that 00:17:29.119 --> 00:17:31.680 need this blend of analytical skill, creativity, 00:17:32.279 --> 00:17:33.920 and maybe even a bit of empathy to understand 00:17:33.920 --> 00:17:36.160 how the AI might interpret things. And maybe 00:17:36.160 --> 00:17:38.019 the most exciting part is the potential for much 00:17:38.019 --> 00:17:40.539 faster development cycles. Think about it. In 00:17:40.539 --> 00:17:44.019 software 1 .0, it was code, compile, debug, slow. 00:17:44.900 --> 00:17:47.549 Software 2 .0. Gather data, train, evaluate, 00:17:47.750 --> 00:17:50.609 even slower, often software 3 .0. It can be prompt, 00:17:50.809 --> 00:17:52.730 test, refined. You can change the ALS behavior 00:17:52.730 --> 00:17:55.069 drastically, sometimes in seconds, just by tweaking 00:17:55.069 --> 00:17:57.750 a few words. That's a massive speed up compared 00:17:57.750 --> 00:18:00.160 to retraining a model for days or weeks. It's 00:18:00.160 --> 00:18:02.319 a really fundamental shift in how we build and 00:18:02.319 --> 00:18:04.400 interact with these systems. This whole exploration, 00:18:04.400 --> 00:18:06.380 it really does make you stop and think, doesn't 00:18:06.380 --> 00:18:08.700 it? It feels like a lesson in humility almost. 00:18:08.900 --> 00:18:11.319 In this constant drive for more complexity, more 00:18:11.319 --> 00:18:13.740 parameters, more intricate algorithms, maybe 00:18:13.740 --> 00:18:16.619 we sometimes overlook the power of simple, elegant 00:18:16.619 --> 00:18:20.299 solutions, especially solutions rooted in communication. 00:18:20.819 --> 00:18:23.279 Understanding how to talk to AI is becoming just 00:18:23.279 --> 00:18:25.759 as critical, maybe even more critical, than the 00:18:25.759 --> 00:18:28.180 engineering that goes into building the AI itself. 00:18:29.130 --> 00:18:31.450 Thinking about this big shift from complex code 00:18:31.450 --> 00:18:34.349 and algorithms towards clever communication and 00:18:34.349 --> 00:18:36.869 prompting how might that change the way you approach 00:18:36.869 --> 00:18:39.130 interacting with AI. Whether it's in your work 00:18:39.130 --> 00:18:41.630 or just daily life, what kinds of new possibilities 00:18:41.630 --> 00:18:44.230 does that open up for you? Reflecting on this 00:18:44.230 --> 00:18:46.490 deep dive, the main takeaway seems pretty profound. 00:18:46.849 --> 00:18:48.769 The biggest breakthrough in getting LLMs to actually 00:18:48.769 --> 00:18:50.650 listen to the context we give them wasn't some 00:18:50.650 --> 00:18:53.430 super complex new algorithm, it was just reframing 00:18:53.430 --> 00:18:55.910 the question. It wasn't about performing neurosurgery 00:18:55.910 --> 00:18:57.890 on the AI's digital brain, but simply changing 00:18:57.890 --> 00:19:00.529 how we talk to it. Yeah, absolutely. This really 00:19:00.529 --> 00:19:03.710 feels like it cements us in that software 3 .0 00:19:03.710 --> 00:19:06.650 era where Designing the prompt, crafting that 00:19:06.650 --> 00:19:08.470 interaction, is becoming the truly essential 00:19:08.470 --> 00:19:11.029 skill. It's a fantastic reminder, I think. Before 00:19:11.029 --> 00:19:13.069 you jump into a really expensive fine -tuning 00:19:13.069 --> 00:19:15.450 project, or get lost trying to understand the 00:19:15.450 --> 00:19:18.089 model's deepest internal workings, just pause 00:19:18.089 --> 00:19:20.690 for a second, take a breath, and ask yourself, 00:19:21.430 --> 00:19:23.509 is there a better way to ask this question? Is 00:19:23.509 --> 00:19:25.859 there a simpler prompt? Because as we saw today, 00:19:26.180 --> 00:19:28.180 the answer might be yes, and it might make all 00:19:28.180 --> 00:19:30.240 the difference. We really hope this deep dive 00:19:30.240 --> 00:19:32.240 offered some valuable insights, maybe sparked 00:19:32.240 --> 00:19:35.119 a few aha moments for you. Thank you for joining 00:19:35.119 --> 00:19:38.440 us on the deep dive. Until next time, keep exploring, 00:19:38.680 --> 00:19:40.920 keep asking questions, and definitely keep prompting.