WEBVTT 1 00:00:01.199 --> 00:00:06.200 Welcome to the Sentient Code, where intelligence is engineered, autonomy 2 00:00:06.280 --> 00:00:10.439 is emerging, and a line between human and machine grows thinner. 3 00:00:10.800 --> 00:00:15.359 Each episode, we decode the algorithms, explore the robotics, and 4 00:00:15.439 --> 00:00:21.879 examine the ideas shaping the future of artificial minds. 5 00:00:23.879 --> 00:00:27.679 Welcome back. It is Wednesday, February eighteenth, twenty twenty six. 6 00:00:28.480 --> 00:00:30.920 And if you've been looking at the tech headlines this morning, 7 00:00:31.160 --> 00:00:35.320 or really any morning for the past month, you're seeing 8 00:00:35.359 --> 00:00:37.399 one phrase just repeated over and over. 9 00:00:37.520 --> 00:00:39.000 You really are. It's everywhere. 10 00:00:39.119 --> 00:00:41.880 The industry is trying very, very hard to make the 11 00:00:41.960 --> 00:00:45.079 AI coscientist the defining concept of the year. 12 00:00:45.200 --> 00:00:47.560 It's absolutely the buzzword of the quarter. You can't open 13 00:00:47.600 --> 00:00:49.520 a journal or a tech blog without seeing it. 14 00:00:49.640 --> 00:00:51.000 But I want to push back on this a little. 15 00:00:51.079 --> 00:00:54.159 Right from the start, we've been calling AI an assistant 16 00:00:54.240 --> 00:00:58.240 or a copilot since what twenty twenty three at least. Yeah, 17 00:00:58.280 --> 00:01:01.320 we've had tools that can summarize PDFs for us and 18 00:01:01.799 --> 00:01:05.079 write email a drafts for years now. So my question 19 00:01:05.159 --> 00:01:07.959 is is co scientist just a branding pivot? Is this 20 00:01:08.120 --> 00:01:10.640 just marketing trying to sell us the same old llms 21 00:01:10.680 --> 00:01:12.799 with a new code of paint or is there an 22 00:01:12.840 --> 00:01:16.680 actual fundamental architectural difference in the stack that we're seeing today. 23 00:01:16.760 --> 00:01:20.239 That's actually, that is exactly the right question to be asking. 24 00:01:20.439 --> 00:01:22.519 And it's fair to be skeptical because the hype cycle 25 00:01:22.599 --> 00:01:25.319 is very real. It's always real, always, But if you 26 00:01:25.359 --> 00:01:28.120 actually look at the engineering, the shift that we're seeing 27 00:01:28.120 --> 00:01:30.599 in early twenty twenty six is not just marketing. It 28 00:01:30.719 --> 00:01:34.000 really isn't. If you look back at that, say, twenty 29 00:01:34.040 --> 00:01:37.840 twenty three, twenty twenty four era, the architecture was fundamentally 30 00:01:37.879 --> 00:01:41.200 something called retrieval augmented generation RIGG. 31 00:01:41.519 --> 00:01:42.599 We've talked about that before. 32 00:01:42.719 --> 00:01:46.159 Exactly ria you'd ask a question, the model would go 33 00:01:46.239 --> 00:01:48.760 look up a document, summarize it, and then give you 34 00:01:48.799 --> 00:01:51.040 an answer based on what it found. It was a librarian, 35 00:01:51.159 --> 00:01:55.319 a very very fast library, very fast, incredibly well read librarian. Sure, 36 00:01:55.560 --> 00:01:58.599 but it was still just retrieving existing information, right. 37 00:01:58.680 --> 00:02:01.480 It's fetching, it's not it's not thinking, it's not creating. 38 00:02:01.640 --> 00:02:05.840 Precisely, the shift to coscientists here in twenty twenty six 39 00:02:05.920 --> 00:02:10.199 is really about two core things, agency and closed loop systems. 40 00:02:10.479 --> 00:02:12.759 We are not just talking about a chatbot that answers 41 00:02:12.759 --> 00:02:13.599 a prompt anymore. 42 00:02:13.639 --> 00:02:15.800 Okay, so that sounds like engineering jargon. Let's break that 43 00:02:15.840 --> 00:02:18.080 down close. Loop, what does it actually look like in 44 00:02:18.120 --> 00:02:19.000 a real lab setting? 45 00:02:19.159 --> 00:02:22.639 So it means the system is self correcting. It observes data, 46 00:02:22.719 --> 00:02:25.199 it generates a hypothesis. It then writes the code to 47 00:02:25.240 --> 00:02:28.680 test that hypothesis. It executes a simulation or even a 48 00:02:28.719 --> 00:02:32.759 physical experiment. Okay, it reads the error logs or the results. 49 00:02:33.039 --> 00:02:36.280 And then, and this is the absolute key part, it 50 00:02:36.360 --> 00:02:40.080 refines its own hypothesis based on whether it's succeeded or failed. 51 00:02:40.759 --> 00:02:43.599 It doesn't stop and ask the human what should I 52 00:02:43.639 --> 00:02:44.159 do next? 53 00:02:44.360 --> 00:02:44.719 Ah? 54 00:02:44.800 --> 00:02:47.840 It iterates that self correction loop is what earns it 55 00:02:47.879 --> 00:02:50.479 the title of scientist rather than just search engine. 56 00:02:50.560 --> 00:02:53.159 So it's the difference between asking a sus chef to 57 00:02:53.240 --> 00:02:55.879 chop an onion for you and asking a chef to 58 00:02:56.000 --> 00:02:57.639 invent a completely new soup. 59 00:02:57.759 --> 00:02:58.879 That's a great analogy. 60 00:02:59.039 --> 00:03:01.560 The chef tastes, it realizes that it needs more salt, 61 00:03:01.759 --> 00:03:04.919 adds the salt tasted. Again, that loop is internal. It 62 00:03:04.960 --> 00:03:06.919 doesn't need external direction for every. 63 00:03:06.719 --> 00:03:09.960 Step precisely, and the capabilities that we're seeing now things 64 00:03:10.080 --> 00:03:13.879 like designing experiments from the ground up, simulating molecular interactions 65 00:03:13.879 --> 00:03:17.439 at a dynamic level, not static, and even controlling physical 66 00:03:17.560 --> 00:03:20.439 robots in a lab. These are things the copilots of 67 00:03:20.479 --> 00:03:22.520 twenty twenty four simply could not do. 68 00:03:22.919 --> 00:03:24.800 There was this quote from the research steck we looked 69 00:03:24.800 --> 00:03:27.439 at that really stuck with me. AI has moved from 70 00:03:27.520 --> 00:03:29.639 summarizing papers to making discoveries. 71 00:03:29.800 --> 00:03:30.759 That's the headline. 72 00:03:30.879 --> 00:03:33.680 It's a bold claim, and today we're going to test 73 00:03:33.719 --> 00:03:36.120 that claim. We're going to look at the evolution of 74 00:03:36.159 --> 00:03:39.199 this technology, specifically, how we got from you know, the 75 00:03:39.240 --> 00:03:43.080 static images of alpha fold to the dynamic simulation engines 76 00:03:43.080 --> 00:03:44.560 we have now, and we have to open. 77 00:03:44.439 --> 00:03:46.800 Up the hood and look at the brain of these systems, 78 00:03:46.800 --> 00:03:50.360 the whole egenic workflow, this test time compute that everyone's 79 00:03:50.400 --> 00:03:51.199 talking about. 80 00:03:50.960 --> 00:03:53.080 Absolutely, and then we have to talk about the physical layer, 81 00:03:53.560 --> 00:03:58.080 the robotics, because that's where the digital world actually touches 82 00:03:58.120 --> 00:03:58.639 the real one. 83 00:03:58.680 --> 00:04:00.400 That's where code moves. 84 00:04:00.520 --> 00:04:02.879 Yeah, and we'll get into the medical break twos that 85 00:04:02.919 --> 00:04:07.000 are happening right now, real world stuff in cancer, in Alzheimer's, 86 00:04:07.280 --> 00:04:09.960 and we'll wrap up with how you our listeners can 87 00:04:10.000 --> 00:04:12.919 actually get your hands on these tools, because apparently you 88 00:04:12.919 --> 00:04:15.080 don't need a PhD and a million dollar grant to 89 00:04:15.120 --> 00:04:16.560 play in the sandbox anymore. 90 00:04:16.680 --> 00:04:20.040 No, not at all. The barrier to entry has completely collapsed. 91 00:04:20.040 --> 00:04:22.680 It is a very very different world than it was 92 00:04:22.800 --> 00:04:24.000 even eighteen months ago. 93 00:04:24.279 --> 00:04:26.959 Okay, so let's rewind a little bit to set the 94 00:04:26.959 --> 00:04:31.480 foundation here. To really understand twenty twenty six, we have 95 00:04:31.519 --> 00:04:34.000 to understand where we came from. And you can't talk 96 00:04:34.040 --> 00:04:36.680 about AI and science without talking about alpha fold. 97 00:04:37.079 --> 00:04:40.720 Right. Alpha fold is really patient zero for this entire revolution. 98 00:04:41.120 --> 00:04:43.000 If we go back to that twenty twenty one to 99 00:04:43.040 --> 00:04:47.000 twenty twenty four window, Google deep Mind effectively solved the 100 00:04:47.319 --> 00:04:49.000 protein folding. 101 00:04:48.600 --> 00:04:51.439 Problem, which was a fifty year old grand challenge in biology. 102 00:04:51.480 --> 00:04:53.079 This wasn't a minor problem, oh no. 103 00:04:53.120 --> 00:04:56.360 It was the grand challenge for half a century. We 104 00:04:56.480 --> 00:04:59.279 knew the chemical formula of a protein, you know the 105 00:04:59.360 --> 00:05:02.000 list of ingreen and seemino ass, but we had no 106 00:05:02.199 --> 00:05:04.680 reliable way of knowing how it folded up into a 107 00:05:04.720 --> 00:05:06.079 complex three D shape. 108 00:05:06.120 --> 00:05:08.720 And in biology, shape is everything everything. 109 00:05:08.920 --> 00:05:11.360 Shape determines function. If you don't know the shape, you 110 00:05:11.360 --> 00:05:12.839 don't know if that protein is going to build a 111 00:05:12.879 --> 00:05:15.800 muscle or digest your food, or you know, causet disease. 112 00:05:15.839 --> 00:05:16.839 It's the whole ballgame. 113 00:05:16.879 --> 00:05:19.000 I remember the analogy everyone was using back then. It 114 00:05:19.040 --> 00:05:21.000 was like Google Maps for the molecular world. 115 00:05:21.079 --> 00:05:24.120 It's a pretty decent analogy. Actually, before alpha fold, we 116 00:05:24.120 --> 00:05:27.879 were navigating the cellular world by stars and intuition, doing 117 00:05:28.000 --> 00:05:30.759 painstaking X ray crystallography. 118 00:05:30.360 --> 00:05:34.040 Which could take months or years for a single protein exactly. 119 00:05:34.360 --> 00:05:36.920 Suddenly, with alpha fold we had a street view for 120 00:05:37.000 --> 00:05:40.600 nearly every protein known to science. Then alpha fold three 121 00:05:40.720 --> 00:05:43.839 came along and predicted the structures of not just proteins, 122 00:05:43.839 --> 00:05:47.079 but their interactions with DNA, RNA other molecules. 123 00:05:47.120 --> 00:05:49.399 It was a massive leap, and the creators won the 124 00:05:49.439 --> 00:05:51.800 Nobel Prize in Chemistry for it in twenty twenty four. 125 00:05:52.000 --> 00:05:54.959 Rightfully so it was the moment of validation. It was 126 00:05:55.000 --> 00:05:58.720 the proof that AI had truly arrived in serious hard science. 127 00:05:58.800 --> 00:06:02.120 It wasn't just for you know, identifying cats and pictures anymore. 128 00:06:02.480 --> 00:06:05.439 But here we are in twenty twenty six and looking 129 00:06:05.480 --> 00:06:07.319 at the research we have in front of us for today, 130 00:06:07.839 --> 00:06:10.360 alpha fold almost looks I don't want to say basic 131 00:06:10.439 --> 00:06:13.279 because that feels disrespectful. No I know em but it 132 00:06:13.319 --> 00:06:14.839 looks like just the starting line. 133 00:06:14.879 --> 00:06:18.439 It was the foundation. Absolutely, the key limitation of alpha fold, 134 00:06:18.639 --> 00:06:21.079 and this is so important to understand the current moment, 135 00:06:21.560 --> 00:06:23.959 is that it gave you a static structure. It gave 136 00:06:24.000 --> 00:06:28.240 you a snapshot, a beautiful, high resolution statue of a protein. 137 00:06:28.360 --> 00:06:30.480 But biology is in the statue, not at all. 138 00:06:30.600 --> 00:06:34.079 Biology is wet, it's warm, it's noisy. Proteins don't just 139 00:06:34.199 --> 00:06:37.839 sit still. They vibrate, they twist, they dance, They adopt 140 00:06:37.879 --> 00:06:40.879 all these different conformations based on what they're touching, or 141 00:06:40.959 --> 00:06:42.959 the temperature of the cell or the pH level. 142 00:06:43.000 --> 00:06:46.040 And computing that dance is exponentially harder than just computing 143 00:06:46.120 --> 00:06:48.560 the one static shape, right yeah, because you aren't just 144 00:06:48.639 --> 00:06:51.600 solving for X, you're solving for eds over time in 145 00:06:51.639 --> 00:06:53.480 a changing environment exactly. 146 00:06:53.560 --> 00:06:57.160 You're dealing with what's called an energy landscape. Alpha fold 147 00:06:57.279 --> 00:07:00.639 was incredible at finding the lowest energy state, the most 148 00:07:00.639 --> 00:07:02.639 stable resting pose. 149 00:07:02.319 --> 00:07:04.240 Of the protein, the bottom of the valley, the. 150 00:07:04.279 --> 00:07:07.079 Very bottom of valley. But the twenty twenty six systems, 151 00:07:07.160 --> 00:07:11.319 these successors, like alpha proteo and a SOD are probabilistic. 152 00:07:11.600 --> 00:07:13.519 They don't just give you the bottom of the valley. 153 00:07:13.720 --> 00:07:16.000 They map out the entire landscape. They tell you the 154 00:07:16.000 --> 00:07:20.160 probability of a protein shifting into a different shape that can, say, suddenly, 155 00:07:20.199 --> 00:07:21.560 lock onto a drug molecule. 156 00:07:21.639 --> 00:07:23.720 So it's a difference between a photograph and a movie. 157 00:07:23.800 --> 00:07:25.600 It's a movie where you can see all the possible 158 00:07:25.600 --> 00:07:28.439 alternate endings, and even better, a movie where you can 159 00:07:28.519 --> 00:07:32.399 influence which ending happens. This gets us to the concept 160 00:07:32.439 --> 00:07:33.360 of inverse folding. 161 00:07:33.639 --> 00:07:36.720 Okay, inverse folding, I saw that in the isomorphic labs papers. 162 00:07:36.759 --> 00:07:37.639 Break that down for us. 163 00:07:37.759 --> 00:07:41.839 Okay, So traditional protein folding is here is a sequence 164 00:07:41.879 --> 00:07:44.720 of amino acids. What three D shape does it make? 165 00:07:44.959 --> 00:07:46.959 Input is the sequence? Output is the shape? 166 00:07:47.120 --> 00:07:50.800 Right, inverse folding flips that entire problem on its head. 167 00:07:50.839 --> 00:07:53.759 It says, here is the shape I need. For example, 168 00:07:53.920 --> 00:07:58.319 I need a shape that perfectly blocks this specific virus receptor. 169 00:07:58.759 --> 00:08:00.920 Now you tell me this sequence of amino acids that 170 00:08:00.959 --> 00:08:03.040 will reliably build that exact shape. 171 00:08:03.079 --> 00:08:05.959 So it's generative. It's like using an image generator like Dally, 172 00:08:06.319 --> 00:08:09.040 but instead of the prompt being an astronaut riding a 173 00:08:09.040 --> 00:08:11.639 horse on the moon. Okay, the prompt is a protein 174 00:08:11.720 --> 00:08:14.800 that binds to and neutralizes the COVID twenty six spike. 175 00:08:14.920 --> 00:08:18.920 Precisely. It is generative biological engineering, and this one capability 176 00:08:19.360 --> 00:08:23.079 to design interactions, not just predict them, has completely utterly 177 00:08:23.120 --> 00:08:25.199 collapsed the timeline for drug discovery. 178 00:08:25.360 --> 00:08:28.600 Okay, let's talk about those timelines, because traditionally, discovering a 179 00:08:28.600 --> 00:08:31.079 new drug is a marathon. We're talking ten to fifteen 180 00:08:31.160 --> 00:08:33.639 years from initial idea to pharmacy shelf. 181 00:08:33.639 --> 00:08:36.559 And billions of dollars, and you have to remember most 182 00:08:36.559 --> 00:08:39.440 of that time is just failure. It's high throughput screening 183 00:08:39.440 --> 00:08:41.159 where you're just trying a key in a lock and 184 00:08:41.200 --> 00:08:43.080 it doesn't fit, so you file it down a little, 185 00:08:43.159 --> 00:08:45.080 You try again, it still doesn't fit. You do that 186 00:08:45.120 --> 00:08:49.600 ten thousand, one hundred thousand times. Route force pure route force. Now, 187 00:08:49.879 --> 00:08:52.639 because we can simulate the wiggling and jiggling of both 188 00:08:52.639 --> 00:08:55.080 the lock and the key in the computer, we can 189 00:08:55.120 --> 00:08:58.240 eliminate ninety nine percent of the failures before we ever 190 00:08:58.320 --> 00:09:00.600 mix a single chemical in a day tube. 191 00:09:00.639 --> 00:09:03.799 It's a huge shift in efficiency. The sources mentioned that 192 00:09:03.840 --> 00:09:08.000 AI designed molecules. Molecules that were invented in silico entered 193 00:09:08.039 --> 00:09:11.240 phase three clinical trials back in twenty twenty five. 194 00:09:11.559 --> 00:09:14.240 That is the holy grail. Phase three is the final 195 00:09:14.320 --> 00:09:17.559 most expensive step right before FDA approval. Getting a drug 196 00:09:17.600 --> 00:09:20.279 to phase three and twenty twenty five means the foundational 197 00:09:20.320 --> 00:09:23.399 discovery work probably started in what twenty twenty two or 198 00:09:23.399 --> 00:09:24.320 twenty twenty three. 199 00:09:24.360 --> 00:09:26.759 That's a compression of the timeline that is almost hard 200 00:09:26.759 --> 00:09:27.240 to believe. 201 00:09:27.360 --> 00:09:29.919 It's why we're seeing these market forecasts for AI driven 202 00:09:29.960 --> 00:09:33.120 drug discovery just exploding. They're hitting eight to ten billion 203 00:09:33.159 --> 00:09:36.600 dollars in twenty twenty six alone. This isn't a future trend, 204 00:09:36.600 --> 00:09:39.320 it's the economic reality of the industry right now. 205 00:09:39.519 --> 00:09:43.080 So it's speed, yes, but it's also precision. We aren't 206 00:09:43.120 --> 00:09:46.039 just throwing spaghetti at the wall faster. We're guiding each 207 00:09:46.120 --> 00:09:48.440 trand of spaghetti to the wall with I don't know, 208 00:09:48.559 --> 00:09:49.320 laser beams. 209 00:09:49.480 --> 00:09:52.360 That's a colorful image, but yes, that's it. And that 210 00:09:52.399 --> 00:09:54.879 precision comes from the fact that we aren't just using 211 00:09:54.879 --> 00:09:58.320 a single monolithic algorithm anymore. We've moved into this era 212 00:09:58.399 --> 00:10:01.360 of agentic workflows, and this is where we really need 213 00:10:01.399 --> 00:10:03.480 to get into the inside the machine part of our 214 00:10:03.480 --> 00:10:04.159 discussion today. 215 00:10:04.240 --> 00:10:06.840 Yes, let's unpack this because when I hear AI, my 216 00:10:06.960 --> 00:10:09.960 brain still defaults to a text box where a type 217 00:10:09.960 --> 00:10:12.320 of question and he gives me an answer a chatbot. 218 00:10:13.360 --> 00:10:16.720 But the co scientists system that Google deep Mind has built, 219 00:10:17.159 --> 00:10:19.879 which is running on their new Gemini two point zero model, 220 00:10:20.320 --> 00:10:24.879 isn't just one thing. It's a team. The outline describes 221 00:10:24.919 --> 00:10:26.399 a cast of agents, right. 222 00:10:26.240 --> 00:10:28.080 And you should think of it almost like a corporate 223 00:10:28.080 --> 00:10:31.240 structure or a research lab team. In any functional lab, 224 00:10:31.279 --> 00:10:33.480 you have different roles, right. You have the creative postdoc 225 00:10:33.480 --> 00:10:36.679 who has one hundred wild ideas a day brainstormer exactly. 226 00:10:36.960 --> 00:10:40.360 Then you have the critical senior scientist, the principal investigator, 227 00:10:40.639 --> 00:10:43.320 who pokes holes in ninety nine of those ideas and 228 00:10:43.360 --> 00:10:45.639 finds the one that might actually work. You have the 229 00:10:45.720 --> 00:10:47.759 lab tech who actually runs the experiments, and you have 230 00:10:47.840 --> 00:10:50.080 the lab manager making sure it all runs smoothly. 231 00:10:50.440 --> 00:10:55.720 And the aicoscientist replicates that entire structure using specialized, fine 232 00:10:55.759 --> 00:10:58.639 tuned software agents. It does, But hang on a second, 233 00:10:59.360 --> 00:11:02.000 why is that necessary? Isn't that just the model talking 234 00:11:02.039 --> 00:11:04.879 to itself. Why do we need distinct agents? Why can't 235 00:11:04.919 --> 00:11:07.799 I just put it all into one big prompt and say, hey, Gemini, 236 00:11:08.279 --> 00:11:11.480 be creative, but also be critical and also manage the project. 237 00:11:11.559 --> 00:11:13.759 You can try that, people have, but you run into 238 00:11:13.799 --> 00:11:17.960 a problem that engineers call context contamination, or having conflicting 239 00:11:18.000 --> 00:11:21.360 objective functions. Okay, if you ask one single model to 240 00:11:21.399 --> 00:11:23.960 be both creative and critical at the very same time, 241 00:11:24.000 --> 00:11:27.080 the two objectives bleed into each other. You get mediocre 242 00:11:27.120 --> 00:11:30.960 creativity because the model is presensoring its own wild ideas. 243 00:11:31.039 --> 00:11:32.919 It's pulling its punches exactly. 244 00:11:33.279 --> 00:11:36.960 And you get soft, ineffective criticism because it subconsciously wants 245 00:11:37.000 --> 00:11:39.360 to support the ideas it just generated. It's not a 246 00:11:39.440 --> 00:11:41.000 true adversarial process. 247 00:11:41.799 --> 00:11:44.639 So by splitting them up into separate agents, you create 248 00:11:44.679 --> 00:11:47.759 a genuine adversarial dynamic. You let them argue, you let 249 00:11:47.799 --> 00:11:50.559 them argue, you optimize them against each other. So you 250 00:11:50.559 --> 00:11:54.440 have the generation agent. Its only job is quantity and variety. 251 00:11:54.519 --> 00:11:56.919 It looks at all the available literature, all the data, 252 00:11:57.159 --> 00:12:01.480 and it generates hundreds of potential hypotheses. It's rewarded purely 253 00:12:01.519 --> 00:12:04.000 for novelty. It doesn't care if an idea is perfect 254 00:12:04.080 --> 00:12:06.919 or even feasible. It just wants to explore the entire 255 00:12:06.960 --> 00:12:07.799 solution space. 256 00:12:08.120 --> 00:12:11.200 Okay, so that's the wild idea engine. Then the supervisor 257 00:12:11.360 --> 00:12:14.159 passes those ideas to the next stage, the review and 258 00:12:14.279 --> 00:12:15.080 ranking agents. 259 00:12:15.399 --> 00:12:18.919 The critics. Yeah, the peer review panel. These agents take 260 00:12:18.960 --> 00:12:22.000 those hundreds of hypotheses and they filter them through a 261 00:12:22.039 --> 00:12:26.840 brutal gauntlet. They have three main jobs. First, check for novelty. 262 00:12:27.240 --> 00:12:30.039 Has this been published before? Is this just rediscovering something 263 00:12:30.039 --> 00:12:32.120 from a nineteen eighty two paper, which is a huge 264 00:12:32.120 --> 00:12:35.600 problem in science, a huge problem. Second, check for feasibility. 265 00:12:35.840 --> 00:12:38.320 Does this violate the known laws of physics? Is this 266 00:12:38.519 --> 00:12:41.960 actually chemically possible? And Third, and this is maybe the 267 00:12:41.960 --> 00:12:45.600 most important, they check for testability. Is there an experiment 268 00:12:45.639 --> 00:12:48.799 we can actually run to prove or disprove this hypothesis. 269 00:12:48.840 --> 00:12:51.639 So an idea could be brilliant and true, but if 270 00:12:51.639 --> 00:12:54.279 it's untestable, it's not science. 271 00:12:54.159 --> 00:12:58.120 It's philosophy. The ranking agents discard it. So out of 272 00:12:58.159 --> 00:13:00.960 one hundred ideas, maybe only five or six make it 273 00:13:00.960 --> 00:13:01.639 through this filter. 274 00:13:01.799 --> 00:13:04.600 And then those survivors go to the evolution agent. 275 00:13:04.399 --> 00:13:07.080 The improver, the polisher. It takes the best ideas that 276 00:13:07.159 --> 00:13:10.399 survive the critics, and it iterates on them. It refines 277 00:13:10.440 --> 00:13:14.720 the hypothesis, It tightens the parameters, It suggests modifications. It 278 00:13:14.759 --> 00:13:17.000 turns a rough diamond into a cut gem. 279 00:13:17.240 --> 00:13:21.159 And the conductor of this whole orchestra is the supervisor agent. 280 00:13:20.960 --> 00:13:24.440 The manager, the supervisor is the traffic cop. It decides 281 00:13:24.519 --> 00:13:26.720 when to send an idea back to the generation agent 282 00:13:26.720 --> 00:13:29.080 for more brainstorming, when to push it forward to the 283 00:13:29.080 --> 00:13:33.559 evolution agent, and crucially, it manages the compute budget. It 284 00:13:33.639 --> 00:13:36.279 manages this thing called test time compute right. 285 00:13:36.320 --> 00:13:38.679 I saw that term in the papers, test time compute scaling. 286 00:13:38.720 --> 00:13:41.600 It sounds incredibly technical. But in plain English, does that 287 00:13:41.679 --> 00:13:42.919 just mean thinking longer? 288 00:13:43.240 --> 00:13:45.360 That's a perfect way to put it. Yes, But it's 289 00:13:45.399 --> 00:13:48.879 how it thinks longer that's revolutionary. Instead of just giving 290 00:13:48.919 --> 00:13:52.240 you the first, most probable answer, it allows the model 291 00:13:52.279 --> 00:13:56.000 to generate thousands of internal thoughts or reasoning. 292 00:13:55.600 --> 00:13:57.679 Steps, a chain of thought, a whole. 293 00:13:57.519 --> 00:14:00.799 Branching tree of thoughts, and it explores all the different paths, 294 00:14:01.399 --> 00:14:04.120 and most of them it discards as dead ends before 295 00:14:04.159 --> 00:14:07.519 it ever gives you the final, polished answer. This is 296 00:14:07.519 --> 00:14:11.279 a system too thinking that the psychologist Daniel Khannaman talks about. 297 00:14:11.320 --> 00:14:14.240 It's slow, it's deliberate, it's logical. 298 00:14:14.000 --> 00:14:17.039 So it's not just retrieving an answer from its training data. 299 00:14:17.279 --> 00:14:20.960 It's actively deliberating. It's reasoning through the problem. 300 00:14:21.039 --> 00:14:24.559 It's simulating multiple paths of reasoning. It's fact checking itself 301 00:14:24.559 --> 00:14:27.440 against its own internal knowledge and external tools. It's the 302 00:14:27.480 --> 00:14:30.840 difference between a gut reflex and a deep considered contemplation, 303 00:14:31.440 --> 00:14:34.360 and the expert analysis is unambiguous. The more time you 304 00:14:34.399 --> 00:14:37.279 give the model to think at test time, the better 305 00:14:37.360 --> 00:14:39.120 and more robust the reasoning becomes. 306 00:14:39.200 --> 00:14:41.840 We have a perfect real world example of this in 307 00:14:41.879 --> 00:14:43.919 the sources the bacterial mystery, and I want to go 308 00:14:43.919 --> 00:14:45.799 deep on this one because the timeline is just wild. 309 00:14:46.039 --> 00:14:49.360 This is the story about cfpicis. 310 00:14:48.720 --> 00:14:53.039 Right, capsid forming phage, inducible chromosomal. 311 00:14:52.399 --> 00:14:54.000 Islands, the zood height. 312 00:14:54.240 --> 00:14:57.600 Thank you. So these are basically little mobile genetic elements 313 00:14:57.600 --> 00:15:01.840 in bacteria. The mystery for science was how do they 314 00:15:01.879 --> 00:15:05.799 spread between different bacterial species. They're not viruses, but they 315 00:15:05.799 --> 00:15:06.840 act a bit like them. 316 00:15:07.120 --> 00:15:10.840 And humans had been working on this problem for how long, over. 317 00:15:10.679 --> 00:15:15.360 A decade, really smart people publishing papers, making incremental progress, 318 00:15:15.679 --> 00:15:18.480 but nobody could figure out the exact mechanism for how 319 00:15:18.519 --> 00:15:19.279 they moved around. 320 00:15:19.519 --> 00:15:23.559 So enter the aiicoscientist walk us through exactly what happened here. 321 00:15:23.600 --> 00:15:25.360 It didn't just guess the answer. 322 00:15:25.080 --> 00:15:27.759 Right, no, not at all. It followed the exact agentic 323 00:15:27.799 --> 00:15:31.200 workflow we just described. The system started by ingesting a 324 00:15:31.240 --> 00:15:34.399 massive data set of bacterial genomes and all the existing 325 00:15:34.440 --> 00:15:35.919 scientific literature on the topic. 326 00:15:35.960 --> 00:15:36.919 Okay, so read everything. 327 00:15:37.000 --> 00:15:39.159 It read everything. Then the generation agent went to work. 328 00:15:39.200 --> 00:15:41.840 It proposed a bunch of hypotheses. One of the most 329 00:15:41.840 --> 00:15:46.639 promising was this, maybe the cfpicis steal proteins from actual 330 00:15:46.679 --> 00:15:49.159 viruses that happened to be in the same cell. They 331 00:15:49.240 --> 00:15:52.360 hijack the viral proteins to build their own transport capsules. 332 00:15:52.600 --> 00:15:56.360 Okay, that's a clever, plausible theory, like a car thief 333 00:15:56.360 --> 00:15:58.279 stealing an engine to put in their own chassis. 334 00:15:58.320 --> 00:16:01.879 It's a great analogy. But a theory isn't enough. The 335 00:16:01.919 --> 00:16:06.120 review agent immediately kicked in. It said, Okay, if that's true, 336 00:16:06.240 --> 00:16:08.480 then we should be able to see a specific cluster 337 00:16:08.559 --> 00:16:12.080 of genes activating at a specific time when a bacterium 338 00:16:12.159 --> 00:16:14.759 is co infected. It made a testable prediction. 339 00:16:15.200 --> 00:16:17.960 It defined the evidence it needed to define exactly. 340 00:16:18.240 --> 00:16:21.240 So the supervisor agent then directed the system to run 341 00:16:21.320 --> 00:16:25.159 a massive search across forty thousand bacterial genomes to look 342 00:16:25.159 --> 00:16:28.639 for that specific genetic signature, that coactivation pattern, and it 343 00:16:28.639 --> 00:16:31.360 found it. It found a candidate, strong one. But it 344 00:16:31.399 --> 00:16:34.279 didn't stop there. The evolution agent took over and refined 345 00:16:34.279 --> 00:16:37.960 the hypothesis. It predicted exactly which proteins from the virus 346 00:16:38.000 --> 00:16:40.679 were being stolen and how they were being incorporated. And then, 347 00:16:40.919 --> 00:16:43.440 and this is the kicker, it designed the follow up 348 00:16:43.480 --> 00:16:46.759 analysis and recapitulated the necessary findings to prove it. 349 00:16:46.919 --> 00:16:49.120 And the total time for this whole process. 350 00:16:49.240 --> 00:16:51.519 The system went from here is a decade old mystery 351 00:16:52.200 --> 00:16:55.879 to here is a highly probable data supported mechanism in 352 00:16:55.919 --> 00:16:56.480 two days. 353 00:16:56.759 --> 00:17:00.759 Two days versus ten years of human strung that is 354 00:17:00.879 --> 00:17:01.759 just staggering. 355 00:17:01.919 --> 00:17:05.559 And it was validated. The human scientist took the AI's output, 356 00:17:05.960 --> 00:17:09.440 ran the physical lab experiments and confirmed it. The paper 357 00:17:09.519 --> 00:17:12.039 was published in the journal Cell in September of twenty 358 00:17:12.079 --> 00:17:15.880 twenty five, and importantly, it wasn't a computer science paper 359 00:17:15.880 --> 00:17:19.079 about how cool the AI is. It was a biology 360 00:17:19.079 --> 00:17:21.400 paper about a fundamental discovery made by the AI. 361 00:17:21.599 --> 00:17:25.160 That distinction is so critical. The AI isn't the story anymore. 362 00:17:25.200 --> 00:17:27.519 The discovery that it enables is the story. 363 00:17:27.640 --> 00:17:29.519 That's when you know the technology has matured. But you 364 00:17:29.559 --> 00:17:32.640 know it's not just biology. This deep think capability, this 365 00:17:32.720 --> 00:17:36.480 ability to reason through these long, complex chains of logic, 366 00:17:36.880 --> 00:17:39.400 is showing up in the most abstract fields like pure 367 00:17:39.440 --> 00:17:41.079 mathematics and physics. 368 00:17:40.680 --> 00:17:43.480 Which brings us to open ais. GBT five point two. 369 00:17:43.559 --> 00:17:46.440 Right GPT five point two, which came out late last year. 370 00:17:46.839 --> 00:17:50.079 It introduced these thinking modes or deep think modes as 371 00:17:50.119 --> 00:17:53.319 a default feature, and the benchmark that just blew everyone 372 00:17:53.319 --> 00:17:57.400 away was its performance on the International Mathematics Olympiad the IMO. 373 00:17:57.920 --> 00:18:00.920 This is a competition for the most brilliant high school 374 00:18:00.960 --> 00:18:03.480 math kids on the planet. These are problems that can 375 00:18:03.519 --> 00:18:06.559 stump professional mathematicians with PhDs. 376 00:18:06.839 --> 00:18:11.200 They're designed to require not just knowledge, but true ingenuity, 377 00:18:11.799 --> 00:18:15.799 and GPT five point two achieved a gold metal level performance. 378 00:18:15.880 --> 00:18:18.119 But wait, math has always been the achilles heel of 379 00:18:18.160 --> 00:18:21.599 large language models, hasn't it? Because lms are probabilistic. They're 380 00:18:21.640 --> 00:18:24.200 just predicting the next most likely word that's right. And 381 00:18:24.480 --> 00:18:27.720 math is pure logic. It's precise. If you are ninety 382 00:18:27.799 --> 00:18:30.319 nine percent right in a mathematical proof, you are one 383 00:18:30.400 --> 00:18:33.880 hundred percent wrong. So how did they finally crack that problem? 384 00:18:33.960 --> 00:18:36.240 They cracked it by combining that test time compute we 385 00:18:36.240 --> 00:18:39.759 talked about with sophisticated tool use. When GPT five point 386 00:18:39.799 --> 00:18:42.279 two faces a hard math problem, now it doesn't just 387 00:18:42.319 --> 00:18:44.440 try to guess the answer in one shot. It breaks 388 00:18:44.440 --> 00:18:46.440 the problem down. It says, okay, first I need to 389 00:18:46.440 --> 00:18:49.359 solve this integral. Then it actually writes a small Python 390 00:18:49.400 --> 00:18:52.039 script to call a symbolic math library like simpi. 391 00:18:52.319 --> 00:18:54.920 Ah. So it's using the computer as a calculator the 392 00:18:55.000 --> 00:18:57.960 same way human mathematician would use Wolfram alpha or a 393 00:18:58.000 --> 00:18:59.039 physical calculator. 394 00:18:59.279 --> 00:19:02.519 Exactly. It runs the script, it gets a definitive grounded 395 00:19:02.559 --> 00:19:05.920 truth as the outpuint, and then it incorporates that result 396 00:19:05.920 --> 00:19:09.240 into the next step of its logical proof. It's orchestrating 397 00:19:09.240 --> 00:19:11.440 a whole suite of computational tools. 398 00:19:11.519 --> 00:19:14.000 It's not just a language model anymore. It's a reasoning 399 00:19:14.039 --> 00:19:15.559 engine that uses language. 400 00:19:15.599 --> 00:19:18.920 That's the perfect description, and this capability has allowed it 401 00:19:18.920 --> 00:19:21.960