WEBVTT 00:00:00.000 --> 00:00:03.540 It's December 2nd, 2025. You walk into the OpenAI 00:00:03.540 --> 00:00:06.400 headquarters in San Francisco. Now, usually, 00:00:06.540 --> 00:00:08.699 this is the time for holiday parties, right? 00:00:08.759 --> 00:00:11.000 You're expecting champagne corks popping, maybe 00:00:11.000 --> 00:00:12.779 some high fives that are the kings of the hill. 00:00:12.800 --> 00:00:14.759 You would certainly think so, yeah. But instead 00:00:14.759 --> 00:00:19.160 of a party, the mood is tense. It's frantic. 00:00:19.300 --> 00:00:22.420 It is officially a code red. A code red. And 00:00:22.420 --> 00:00:25.440 why? Because the era of the bodybuilder AI, you 00:00:25.440 --> 00:00:27.420 know, those massive hulking models that rely 00:00:27.420 --> 00:00:30.160 on pure size to solve problems, is officially 00:00:30.160 --> 00:00:32.880 over and the era of the gymnast has begun. That's 00:00:32.880 --> 00:00:34.780 a vivid way to put it, but honestly, it's not 00:00:34.780 --> 00:00:37.359 far off. The raw strength we've all been obsessing 00:00:37.359 --> 00:00:39.880 over, it's still there, sure. But suddenly, the 00:00:39.880 --> 00:00:41.780 game isn't just about strength anymore. It's 00:00:41.780 --> 00:00:44.240 about precision. It's about agility. And if these 00:00:44.240 --> 00:00:46.439 leaks are real, the whole board has been reset. 00:00:46.969 --> 00:00:49.030 Welcome to the Deep Dive. It's Thursday, January 00:00:49.030 --> 00:00:52.469 22, 2026. Today, we're unpacking something that 00:00:52.469 --> 00:00:54.630 feels like a genuine pivot point in AI history. 00:00:54.810 --> 00:00:57.130 We're looking at the leaked details of OpenAI's 00:00:57.130 --> 00:01:00.369 GPT -5 .3, which has the internal, and let's 00:01:00.369 --> 00:01:02.590 be honest, slightly hilarious, codename, Garlic. 00:01:02.810 --> 00:01:05.230 Garlic. It's definitely a choice. A lot earthier 00:01:05.230 --> 00:01:08.189 than, you know, Orion or Gemini, all that celestial 00:01:08.189 --> 00:01:11.650 stuff we usually get. It really is. But don't 00:01:11.650 --> 00:01:14.510 let the name fool you. Our mission today is to 00:01:14.510 --> 00:01:17.680 figure out... why this specific model caused 00:01:17.680 --> 00:01:21.799 such a panic. This code red at OpenAI. We're 00:01:21.799 --> 00:01:23.439 going to look at the engineering breakthroughs, 00:01:23.439 --> 00:01:26.480 specifically around memory and this new self 00:01:26.480 --> 00:01:28.459 -checking thing. And we'll see how it stacks 00:01:28.459 --> 00:01:30.500 up against the current heavyweights, Google's 00:01:30.500 --> 00:01:33.879 Gemini 3 and Anthropic's Claude Opus 4 .5. This 00:01:33.879 --> 00:01:36.040 is going to be fun because the specs here aren't 00:01:36.040 --> 00:01:38.180 just number go up. Right. It's not just we added 00:01:38.180 --> 00:01:40.319 more zeros. This is a whole philosophical shift 00:01:40.319 --> 00:01:42.900 in how we built this stuff. So let's start with 00:01:42.900 --> 00:01:45.459 that context. We mentioned the code red. This 00:01:45.459 --> 00:01:47.340 leak came from the information back in December. 00:01:47.620 --> 00:01:50.340 Mark Chen, the chief research officer at OpenAI 00:01:50.340 --> 00:01:52.780 reportedly shared these details. But help me 00:01:52.780 --> 00:01:55.420 understand the panic. I mean, GBT 5 .2 was already 00:01:55.420 --> 00:01:58.019 out. It was a good model. Why the fire drill? 00:01:58.180 --> 00:02:00.219 It all comes down to momentum. In tech, if you 00:02:00.219 --> 00:02:02.680 aren't leading, you're basically dying. And frankly, 00:02:02.920 --> 00:02:05.159 OpenAI was losing ground. Losing ground to who? 00:02:05.519 --> 00:02:07.260 To everyone. I mean, look at the last six months. 00:02:07.379 --> 00:02:10.580 Google drops Gemini 3. And Gemini 3 just dominates 00:02:10.580 --> 00:02:13.500 anything multimodal video, messy real world data 00:02:13.500 --> 00:02:16.419 images. It was the king of stuff. Right. And 00:02:16.419 --> 00:02:19.280 then on the other side, you had Anthropic with 00:02:19.280 --> 00:02:22.340 Claude Opus 4 .5. And let's be honest, if you 00:02:22.340 --> 00:02:25.080 were writing code in late 2025, you were probably 00:02:25.080 --> 00:02:27.340 using Claude. I could vouch for that. I switched 00:02:27.340 --> 00:02:29.039 to Claude for all my scripting. It just felt 00:02:29.039 --> 00:02:32.500 less, I don't know, robotic. Exactly. So internally, 00:02:32.740 --> 00:02:36.300 GPT 5 .2 was seen. as a Band -Aid. It kept them 00:02:36.300 --> 00:02:38.500 in the conversation, but it wasn't bribing it. 00:02:38.580 --> 00:02:40.979 They knew that shipping GPT -5 at a little bit 00:02:40.979 --> 00:02:43.180 bigger just wasn't going to cut it. They needed 00:02:43.180 --> 00:02:45.699 a response that fundamentally changed the metric 00:02:45.699 --> 00:02:48.740 of success. And that response is garlic. You 00:02:48.740 --> 00:02:51.460 used this analogy before we started the bodybuilder 00:02:51.460 --> 00:02:53.719 versus the gymnast. I want to double click on 00:02:53.719 --> 00:02:55.909 that because for. what, the last five years, 00:02:56.090 --> 00:02:58.349 the headline has always been parameter count. 00:02:58.550 --> 00:03:00.330 Yeah. Trillions of parameters. Fairer clusters, 00:03:00.689 --> 00:03:03.229 more GPUs. Yeah. If you weren't building bigger, 00:03:03.330 --> 00:03:05.849 you weren't trying. And that's the bodybuilder 00:03:05.849 --> 00:03:08.550 approach. You solve problems by just adding more 00:03:08.550 --> 00:03:11.210 muscle. The model doesn't get physics. Add a 00:03:11.210 --> 00:03:13.050 trillion more parameters. Can't write a sonnet. 00:03:13.069 --> 00:03:16.229 Add another trillion. Total brute force. But 00:03:16.229 --> 00:03:19.210 garlic represents a shift toward density. It's 00:03:19.210 --> 00:03:22.069 the gymnast. It's physically smaller architecturally. 00:03:22.759 --> 00:03:25.719 more compact. But because of that, it can do 00:03:25.719 --> 00:03:28.199 these complex reasoning maneuvers, these mental 00:03:28.199 --> 00:03:31.460 backflips, if you will, that the massive, clumsy 00:03:31.460 --> 00:03:34.180 bodybuilder just can't do. Okay, but I have to 00:03:34.180 --> 00:03:36.500 play the skeptic here. Making it smaller but 00:03:36.500 --> 00:03:39.090 smarter sounds like... Marketing fluff. It sounds 00:03:39.090 --> 00:03:41.050 like the holy grail everyone promises, but nobody 00:03:41.050 --> 00:03:43.330 delivers. Usually when you make a neural net 00:03:43.330 --> 00:03:46.090 smaller, it gets dumber. How did they actually 00:03:46.090 --> 00:03:48.289 do that? It wasn't a straight line. It comes 00:03:48.289 --> 00:03:50.849 down to a merger of two different research tracks 00:03:50.849 --> 00:03:54.110 inside OpenAI. They had Shallot Pete, which was 00:03:54.110 --> 00:03:56.430 just their standard stability update track. Kind 00:03:56.430 --> 00:03:58.490 of boring. But then they had this experimental 00:03:58.490 --> 00:04:00.990 branch called Garlic. And the big breakthrough 00:04:00.990 --> 00:04:04.949 is a technique called EPTE. BPTE. Enhanced Pre 00:04:04.949 --> 00:04:07.520 -Training Efficiency. Okay, break that down for 00:04:07.520 --> 00:04:11.340 me. No jargon. What is EPTE actually doing that's 00:04:11.340 --> 00:04:13.599 different from GPT -4 or 5? Think of it like 00:04:13.599 --> 00:04:16.379 a garden. Or, you know, better yet, the human 00:04:16.379 --> 00:04:19.300 brain. When a baby is born, their brain just 00:04:19.300 --> 00:04:22.000 has this explosion of connections, synapses everywhere. 00:04:22.259 --> 00:04:25.019 It's a mess. Right. As we grow up, we get smarter 00:04:25.019 --> 00:04:27.540 not by adding connections, but by pruning them. 00:04:27.680 --> 00:04:29.860 We cut out the noise so the signal can travel 00:04:29.860 --> 00:04:33.519 faster. So we get smarter by deleting parts of 00:04:33.519 --> 00:04:36.120 our brain. In a way, yes. We delete the inefficiency. 00:04:36.480 --> 00:04:38.980 Traditional AI training is like letting a forest 00:04:38.980 --> 00:04:41.399 grow wild connections everywhere. Vines, weeds, 00:04:41.600 --> 00:04:45.100 you name it. It's huge, but it's messy. EPCE 00:04:45.100 --> 00:04:47.680 introduces a pruning phase during the training. 00:04:47.899 --> 00:04:50.740 The model actively discards redundant neural 00:04:50.740 --> 00:04:53.319 pathways. So it's Marie Kondo -ing its own brain. 00:04:53.540 --> 00:04:56.100 This neuron does not spark joy. Basically. It's 00:04:56.100 --> 00:04:58.100 cutting out the noise. So the result is you get 00:04:58.100 --> 00:05:00.980 GPT -6 level reasoning capabilities. So higher 00:05:00.980 --> 00:05:03.600 logic scores. Exactly. Higher logic scores. Because 00:05:03.600 --> 00:05:05.720 you've pruned away the inefficiency, it runs 00:05:05.720 --> 00:05:07.980 on a faster, smaller architecture. It's compressing 00:05:07.980 --> 00:05:10.439 thought. That's fascinating. It really does challenge 00:05:10.439 --> 00:05:13.540 that whole more is better assumption. But I have 00:05:13.540 --> 00:05:17.139 to ask, is efficiency really as exciting as raw 00:05:17.139 --> 00:05:20.899 power? As a user, do I care if the model is efficient? 00:05:21.060 --> 00:05:23.220 I usually get excited about the next biggest 00:05:23.220 --> 00:05:26.699 thing. Absolutely. And here's why. Latency and 00:05:26.699 --> 00:05:29.079 cost. Think of it like the difference between 00:05:29.079 --> 00:05:31.920 a muscle car and a Formula One racer. The muscle 00:05:31.920 --> 00:05:34.259 car has raw power and makes a ton of noise, burns 00:05:34.259 --> 00:05:37.720 a ton of gas. But the F1 car, it has precision 00:05:37.720 --> 00:05:40.560 engineering. It turns on a dime. In this case, 00:05:40.620 --> 00:05:43.259 precision means the model is cheaper to run and 00:05:43.259 --> 00:05:45.759 faster to answer. And when intelligence gets 00:05:45.759 --> 00:05:48.040 cheap and fast, you can use it in ways you never 00:05:48.040 --> 00:05:50.500 could with the big, slow bodybuilder. So it's 00:05:50.500 --> 00:05:53.519 shifting from raw horsepower to agility, and 00:05:53.519 --> 00:05:55.920 that changes what it's useful for. Exactly. It 00:05:55.920 --> 00:05:58.519 moves from a consultant you hire once to a worker 00:05:58.519 --> 00:06:00.519 that lives in your computer. Let's move to the 00:06:00.519 --> 00:06:02.800 specs, because while the philosophy is cool, 00:06:02.899 --> 00:06:05.259 the actual numbers here are, well, they're staggering. 00:06:05.459 --> 00:06:07.160 And I want to make sure we really get what they 00:06:07.160 --> 00:06:10.629 mean day to day. Let's do it. So, memory first. 00:06:10.990 --> 00:06:14.689 The context window. Garlic is reportedly shipping 00:06:14.689 --> 00:06:19.629 with a 400 ,000 token context window. Now, just 00:06:19.629 --> 00:06:23.310 to play devil advocate here, Gemini 3 has 2 million 00:06:23.310 --> 00:06:27.069 tokens. So on paper, Garlic looks smaller. Why 00:06:27.069 --> 00:06:29.540 should I be impressed by 400k? On paper, yeah, 00:06:29.600 --> 00:06:31.660 Gemini is bigger, but this is where the nuance 00:06:31.660 --> 00:06:34.040 really matters. If you've ever used Gemini's 00:06:34.040 --> 00:06:36.500 huge context window, like you dump a whole novel 00:06:36.500 --> 00:06:37.879 in there and ask about a character from chapter 00:06:37.879 --> 00:06:40.199 three, you might have noticed middle of the context 00:06:40.199 --> 00:06:42.420 loss. Right. I have seen this. It remembered 00:06:42.420 --> 00:06:44.560 the very beginning and the very end of your prompt, 00:06:44.699 --> 00:06:46.639 but it gets hazy on everything in the middle. 00:06:46.680 --> 00:06:48.579 It's like it skimmed the book. Exactly. It's 00:06:48.579 --> 00:06:50.839 the needle in a haystack problem. Yeah. Gemini 00:06:50.839 --> 00:06:52.879 has a huge stomach, but imperfect digestion. 00:06:53.579 --> 00:06:56.060 Garlic uses a new attention mechanism that reportedly 00:06:56.060 --> 00:06:59.019 gives it perfect recall. cross the whole 400 00:06:59.019 --> 00:07:01.060 ,000 tokens. It doesn't just store the data. 00:07:01.100 --> 00:07:02.939 It actually remembers it. So I could feed it 00:07:02.939 --> 00:07:04.860 literally my entire company's documentation. 00:07:05.180 --> 00:07:08.680 Every PDF, every Slack policy. Every messy confluence 00:07:08.680 --> 00:07:10.519 page. And it wouldn't just have it. It would 00:07:10.519 --> 00:07:13.019 know it. It would know it. You could ask, what 00:07:13.019 --> 00:07:15.000 was that compliance rule we changed three years 00:07:15.000 --> 00:07:18.079 ago from that one Tuesday memo? And it just pulls 00:07:18.079 --> 00:07:20.920 it instantly. No skimming. That's the difference 00:07:20.920 --> 00:07:23.829 between a hard drive and a brain. Precisely. 00:07:23.870 --> 00:07:25.930 But here's the part that actually stopped me 00:07:25.930 --> 00:07:28.810 in my tracks. It's not just how much it can take 00:07:28.810 --> 00:07:32.189 in, it's how much it can put out. The output 00:07:32.189 --> 00:07:34.769 limit. This is the big one. This is that wonder 00:07:34.769 --> 00:07:39.579 moment. 128 ,000 tokens. In a single response. 00:07:39.920 --> 00:07:42.459 It's huge. I mean, just for context, for everyone 00:07:42.459 --> 00:07:44.339 listening right now, we're all used to the model 00:07:44.339 --> 00:07:46.920 just stopping. You ask it to write code. It gets 00:07:46.920 --> 00:07:48.800 halfway through a function and it just cuts off. 00:07:48.860 --> 00:07:50.899 And you have to type continue. And then it repeats 00:07:50.899 --> 00:07:53.160 the last line or it forgets the indentation. 00:07:53.220 --> 00:07:55.439 It's so fragmented. It feels like pulling teeth 00:07:55.439 --> 00:07:57.699 sometimes. It totally breaks your flow state. 00:07:57.819 --> 00:08:01.519 But 128 ,000 tokens. That's not a response. That's 00:08:01.519 --> 00:08:04.410 a novel. That's an entire software library. It 00:08:04.410 --> 00:08:06.810 is. And I want you to really just pause and imagine 00:08:06.810 --> 00:08:09.870 the experience of that. Imagine sitting at your 00:08:09.870 --> 00:08:13.990 terminal. You ask for a complex legal brief or 00:08:13.990 --> 00:08:17.050 maybe a full backend architecture for a new app, 00:08:17.129 --> 00:08:19.350 not just the outline, the actual code. You hit 00:08:19.350 --> 00:08:22.670 enter. And instead of a summary or a little piece 00:08:22.670 --> 00:08:26.939 of it, you just watch the cursor move. And it 00:08:26.939 --> 00:08:30.019 keeps moving. And it writes the files. It writes 00:08:30.019 --> 00:08:32.379 the documentation. It writes the test suite. 00:08:32.500 --> 00:08:34.679 And it just doesn't stop until the thought is 00:08:34.679 --> 00:08:38.259 complete. One coherent stream of creation. That 00:08:38.259 --> 00:08:41.639 is, wow. It's almost overwhelming to think about. 00:08:41.679 --> 00:08:43.700 It feels like we're moving from just chatting 00:08:43.700 --> 00:08:46.460 with a bot to manufacturing with a machine. That's 00:08:46.460 --> 00:08:48.440 the shift. No more chunking. No more stitching 00:08:48.440 --> 00:08:50.240 things together. You're not pasting snippets 00:08:50.240 --> 00:08:52.580 into VS Code anymore. You're reviewing a finished 00:08:52.580 --> 00:08:55.000 product. But does infinite memory actually change 00:08:55.000 --> 00:08:57.789 how we work? Or does it just change how much 00:08:57.789 --> 00:08:59.789 junk we dump into the chat box? I worry we'll 00:08:59.789 --> 00:09:02.110 just get lazier. I think it fundamentally changes 00:09:02.110 --> 00:09:04.230 the human's role. It stops us from being librarians 00:09:04.230 --> 00:09:06.889 of data, you know, constantly fetching context, 00:09:07.149 --> 00:09:10.009 pasting files, reminding the bot what we're talking 00:09:10.009 --> 00:09:12.850 about, and lets us be architects of ideas. You 00:09:12.850 --> 00:09:15.389 design the structure, the model pours all the 00:09:15.389 --> 00:09:17.990 concrete. So the output limit frees us from being 00:09:17.990 --> 00:09:21.000 data fetchers. To be designers instead. Spot 00:09:21.000 --> 00:09:23.799 on. It shifts the cognitive load from memory 00:09:23.799 --> 00:09:26.879 to strategy. That distinction architect versus 00:09:26.879 --> 00:09:30.120 librarian, that really lands. Because the other 00:09:30.120 --> 00:09:32.980 feature Garlick supposedly has leans heavily 00:09:32.980 --> 00:09:35.120 into that architect role. We're talking about 00:09:35.120 --> 00:09:37.820 native agents. Yes. This is another area where 00:09:37.820 --> 00:09:40.000 that code red was necessary. Because right now, 00:09:40.019 --> 00:09:42.340 everyone tries to make AI agents go do this task. 00:09:42.700 --> 00:09:45.059 And, well, usually they fail. They get stuck 00:09:45.059 --> 00:09:47.019 in loops. Or they hallucinate a file path that 00:09:47.019 --> 00:09:49.440 doesn't exist and then crash. Right. But Garlic 00:09:49.440 --> 00:09:51.820 isn't pretending to be an agent. The tool use 00:09:51.820 --> 00:09:54.340 is native. It understands file systems. It can 00:09:54.340 --> 00:09:56.980 run tests. It can debug like a developer. It 00:09:56.980 --> 00:09:59.620 treats APIs not as some external thing it has 00:09:59.620 --> 00:10:02.000 to awkwardly reach for, but as part of its own 00:10:02.000 --> 00:10:04.480 cognitive process. So it's the difference between 00:10:04.480 --> 00:10:06.779 me trying to speak French with a dictionary in 00:10:06.779 --> 00:10:09.399 my hand, looking up every word. Versus actually 00:10:09.399 --> 00:10:11.659 being fluent in French. Exactly. It thinks in 00:10:11.659 --> 00:10:14.980 tools. It thinks in execution. If it writes code 00:10:14.980 --> 00:10:17.950 that... fails a test it sees the error corrects 00:10:17.950 --> 00:10:20.129 it and reruns it all before it even gets back 00:10:20.129 --> 00:10:24.350 to you now speaking of thinking and correcting 00:10:24.350 --> 00:10:28.669 itself yeah there's one feature here that i think 00:10:28.669 --> 00:10:31.769 solves the single biggest anxiety i have when 00:10:31.769 --> 00:10:34.509 i use these tools now i want to be a little vulnerable 00:10:34.509 --> 00:10:36.409 for a second go for it i still struggle with 00:10:36.409 --> 00:10:39.330 trusting the output You know, I'll spend 20 minutes 00:10:39.330 --> 00:10:41.509 crafting this perfect prompt. I get an answer 00:10:41.509 --> 00:10:43.850 that looks incredible, super confident, polished. 00:10:44.049 --> 00:10:46.049 And then I spend 40 minutes fact checking it 00:10:46.049 --> 00:10:47.830 because I've been burned by hallucinations before. 00:10:48.049 --> 00:10:50.549 Oh, yeah. The universal experience. Confident 00:10:50.549 --> 00:10:52.990 liar problem. Exactly. And it creates this weird 00:10:52.990 --> 00:10:55.529 friction where I'm like, is this actually faster 00:10:55.529 --> 00:10:58.409 if I have to babysit it? But garlic has a self 00:10:58.409 --> 00:11:01.590 -checking mechanism. This is a game changer for 00:11:01.590 --> 00:11:04.679 exactly that anxiety. Before the model answers 00:11:04.679 --> 00:11:07.059 you, it enters a verification state. It just 00:11:07.059 --> 00:11:10.019 pauses. It checks its own internal knowledge 00:11:10.019 --> 00:11:12.620 graph to see, do I actually know this or am I 00:11:12.620 --> 00:11:14.759 just statistically guessing? So it has a conscience 00:11:14.759 --> 00:11:17.779 or at least a built -in BS detector. It has a 00:11:17.779 --> 00:11:20.500 system two thinking process, to use the Daniel 00:11:20.500 --> 00:11:23.639 Kahneman term. System one is fast, intuitive. 00:11:24.080 --> 00:11:27.840 System two is slow, deliberate. Garlic uses system 00:11:27.840 --> 00:11:31.570 two. If it isn't sure, it reassesses. The report 00:11:31.570 --> 00:11:34.149 says this leads to drastically fewer hallucinations. 00:11:34.409 --> 00:11:37.230 For lawyers, for developers, this is everything. 00:11:37.470 --> 00:11:39.409 But I have to play devil's advocate again. If 00:11:39.409 --> 00:11:41.389 the model is stopping to check its work, doesn't 00:11:41.389 --> 00:11:43.129 that make it slower? We were just talking about 00:11:43.129 --> 00:11:45.409 speed being the new king. It might pause briefly 00:11:45.409 --> 00:11:47.950 before the first token appears, maybe a second 00:11:47.950 --> 00:11:49.889 or two of thinking time. But think about the 00:11:49.889 --> 00:11:51.889 time you just described, the 40 minutes you spend 00:11:51.889 --> 00:11:54.149 fact -checking. It might be slower per second, 00:11:54.230 --> 00:11:56.950 but it saves hours of human rework later. It's 00:11:56.950 --> 00:12:00.220 that Navy SEAL mantra. Slow is smooth, and smooth 00:12:00.220 --> 00:12:03.700 is fast. So the pause pays for itself by eliminating 00:12:03.700 --> 00:12:06.500 all that cleanup time. Exactly. It trades milliseconds 00:12:06.500 --> 00:12:09.879 of latency for hours of reliability. I'll take 00:12:09.879 --> 00:12:12.360 that trade any day. Okay, I want to zoom out 00:12:12.360 --> 00:12:15.039 a bit. We have the specs, the philosophy. But 00:12:15.039 --> 00:12:18.379 OpenAI isn't operating in a vacuum. They're in 00:12:18.379 --> 00:12:21.919 a war. We mentioned Google and Anthropic. How 00:12:21.919 --> 00:12:24.399 does Garlic actually stack up? This is the battle 00:12:24.399 --> 00:12:26.200 for the leaderboard. So let's look at Google 00:12:26.200 --> 00:12:29.070 first. Gemini 3. The heavyweight. The heavyweight, 00:12:29.129 --> 00:12:31.269 yeah. If you look at the leaked benchmarks, the 00:12:31.269 --> 00:12:34.309 battle here is scale versus density. Gemini 3 00:12:34.309 --> 00:12:38.370 wins on multimodal. If you have messy, real -world 00:12:38.370 --> 00:12:42.129 data video, audio, weird images, Gemini is still 00:12:42.129 --> 00:12:44.830 the king. It has that massive context and parameter 00:12:44.830 --> 00:12:46.830 count for a reason. So if I'm analyzing a movie, 00:12:46.909 --> 00:12:49.350 I use Gemini. Correct. But Garlic wins on pure 00:12:49.350 --> 00:12:52.009 text, code, and complex reasoning. The benchmark 00:12:52.009 --> 00:12:55.210 is something called GDP Val for reasoning. What's 00:12:55.210 --> 00:12:56.970 that measuring? It's measuring logic puzzles, 00:12:57.250 --> 00:12:59.110 multi -step reasoning where you can't just memorize 00:12:59.110 --> 00:13:02.730 an answer. Garlic is scoring 70 .9%. Gemini is 00:13:02.730 --> 00:13:06.110 at 53 .3%. Wow. That is not a small margin. That's 00:13:06.110 --> 00:13:08.549 a generational gap. It's a blowout on reasoning. 00:13:08.669 --> 00:13:11.570 So the verdict is, analyze a three -hour video. 00:13:11.730 --> 00:13:14.429 Use Gemini. Build the back end of a banking app 00:13:14.429 --> 00:13:17.590 where logic is everything. You use Garlic. Okay, 00:13:17.629 --> 00:13:21.669 so that's Google. What about Antropic? Quad Opus 00:13:21.669 --> 00:13:24.470 4 .5. I know a ton of developers who swear by 00:13:24.470 --> 00:13:29.029 Claude. It feels warmer. It writes really readable 00:13:29.029 --> 00:13:31.350 code. Yeah, this is the battle for the developer's 00:13:31.350 --> 00:13:34.450 soul. Claude is known for that warmth and readability. 00:13:35.250 --> 00:13:37.769 But Garlic is coming in with a ruthless value 00:13:37.769 --> 00:13:40.129 proposition. It matches Claude's coding proficiency 00:13:40.129 --> 00:13:43.789 94 .2 % on Human Evil plus R, which is the gold 00:13:43.789 --> 00:13:45.590 standard. So it's just as good at the actual 00:13:45.590 --> 00:13:48.269 coding. Okay, so it's a tie. Not quite, because 00:13:48.269 --> 00:13:50.889 Garlic does it at two times the speed and half 00:13:50.889 --> 00:13:53.549 the cost. Half the cost. Because of that pruning 00:13:53.549 --> 00:13:56.059 we talked about. The model is physically smaller. 00:13:56.240 --> 00:13:58.779 It burns less electricity. It costs OpenAI less 00:13:58.779 --> 00:14:01.879 to run, so it costs you less to use. That's significant. 00:14:02.120 --> 00:14:04.179 Yeah. But does price really trump everything? 00:14:04.320 --> 00:14:06.580 I mean, if Claude feels more human to interact 00:14:06.580 --> 00:14:09.240 with, won't people stick with it? For a casual 00:14:09.240 --> 00:14:11.980 chat. Maybe. If you're brainstorming, you might 00:14:11.980 --> 00:14:14.360 stick with Claude. But for running a business 00:14:14.360 --> 00:14:16.980 at scale, if you're an API customer processing 00:14:16.980 --> 00:14:20.200 millions of requests, half price is everything. 00:14:20.679 --> 00:14:23.360 The unit economics alone will shift the market 00:14:23.360 --> 00:14:25.120 overnight. It's the difference between a boutique 00:14:25.120 --> 00:14:28.559 shop and like industrial scale. Exactly. If you're 00:14:28.559 --> 00:14:31.279 building a startup and your API bill just got 00:14:31.279 --> 00:14:34.039 cut in half, you're not going to care how warm 00:14:34.039 --> 00:14:36.559 the model feels. You care that it works and it's 00:14:36.559 --> 00:14:40.029 cheap. So garlic. competes on unit economics 00:14:40.029 --> 00:14:42.789 and logic and sort of cedes the human touch to 00:14:42.789 --> 00:14:45.730 Claude. For now, yeah. It's an industrial revelation, 00:14:45.909 --> 00:14:48.529 not a dinner party. So we have the what and the 00:14:48.529 --> 00:14:51.409 why. I want to take a brief pause here. And we're 00:14:51.409 --> 00:14:53.629 back. We've looked at the code red, the tech, 00:14:53.629 --> 00:14:55.610 and the competition. I want to wrap our heads 00:14:55.610 --> 00:14:58.110 around the timeline and the big picture. Leaks 00:14:58.110 --> 00:15:00.029 are great, but shipping is what matters. When 00:15:00.029 --> 00:15:01.789 are we actually going to see this thing? Well, 00:15:01.830 --> 00:15:04.070 looking at how the leaks and vendor updates are 00:15:04.070 --> 00:15:06.710 converging, it feels imminent. We're expecting 00:15:06.710 --> 00:15:10.090 a preview release. Probably the ChatGPT Pro users 00:15:10.090 --> 00:15:14.070 and some partners in late January 2026. Late 00:15:14.070 --> 00:15:16.409 January. That's basically this week. It's happening 00:15:16.409 --> 00:15:20.250 now. Then the full API availability is slated 00:15:20.250 --> 00:15:23.210 for February. And this is interesting. They're 00:15:23.210 --> 00:15:25.529 expected to integrate a version of this into 00:15:25.529 --> 00:15:28.549 the free tier by March. A counter to Gemini's 00:15:28.549 --> 00:15:30.990 free access. Exactly. They have to capture that 00:15:30.990 --> 00:15:33.990 user base. They can't let Google own the entry 00:15:33.990 --> 00:15:36.730 -level market. Pulling all this together, the 00:15:36.730 --> 00:15:40.169 pruning, the 128K output, the self -checking, 00:15:40.190 --> 00:15:42.850 what's the big idea here? If I'm a listener trying 00:15:42.850 --> 00:15:45.330 to make sense of all this, what is the core shift? 00:15:45.690 --> 00:15:48.070 The core shift is that the definition of AI progress 00:15:48.070 --> 00:15:51.490 has changed. For five years, progress meant bigger. 00:15:51.649 --> 00:15:54.190 It meant more parameters. Now, progress means 00:15:54.190 --> 00:15:56.830 cognitive density. It is about intelligence per 00:15:56.830 --> 00:15:59.149 dollar and intelligence per watt. Cognitive density. 00:15:59.309 --> 00:16:01.629 I like that. It sounds focused. It's about doing 00:16:01.629 --> 00:16:04.159 more with less. And for you listening, the so 00:16:04.159 --> 00:16:05.919 what is pretty direct. If you're a developer, 00:16:06.080 --> 00:16:08.759 you can finally refactor entire code bases without 00:16:08.759 --> 00:16:10.840 losing context. You don't have to choose which 00:16:10.840 --> 00:16:13.320 files to upload. If you're a business, you can 00:16:13.320 --> 00:16:15.340 build automation that actually works because 00:16:15.340 --> 00:16:17.720 of the agentic capabilities. And if you're a 00:16:17.720 --> 00:16:21.340 creator, you can generate long form content books, 00:16:21.600 --> 00:16:25.139 scripts, courses without having to manually stitch 00:16:25.139 --> 00:16:26.940 it together. It feels like the training wheels 00:16:26.940 --> 00:16:29.820 are coming off. They are. The limitations we've 00:16:29.820 --> 00:16:32.240 all learned to work around. Oh, I can only paste 00:16:32.240 --> 00:16:34.379 half this file. Oh, I have to check its math. 00:16:34.600 --> 00:16:38.360 Those limitations are just evaporating. So before 00:16:38.360 --> 00:16:40.500 we sign off, let's give everyone listening something 00:16:40.500 --> 00:16:42.659 to do. Because if this is dropping in weeks, 00:16:42.779 --> 00:16:45.460 we shouldn't just be waiting around. How do we 00:16:45.460 --> 00:16:48.600 prepare for the garlic era? Two things. First, 00:16:48.840 --> 00:16:51.899 organize your data. If you want to use that 400 00:16:51.899 --> 00:16:54.419 ,000 token context window with perfect recall, 00:16:54.740 --> 00:16:57.419 your data needs to be ready. Clean up your documentation, 00:16:57.759 --> 00:17:00.200 merge your repositories, get it ready to go. 00:17:00.399 --> 00:17:03.000 Don't feed the gymnast junk food. Exactly. If 00:17:03.000 --> 00:17:04.660 you feed it garbage, you'll still get garbage. 00:17:04.740 --> 00:17:06.920 Just perfectly recalled garbage. Good point. 00:17:07.099 --> 00:17:10.529 And second, map your workflows. Start thinking 00:17:10.529 --> 00:17:13.430 in terms of agentic workflows. Don't just think, 00:17:13.470 --> 00:17:16.230 what question can I ask the bot? Think, what 00:17:16.230 --> 00:17:18.450 multi -step process can I hand off entirely? 00:17:18.869 --> 00:17:21.309 Check these invoices against these emails and 00:17:21.309 --> 00:17:23.710 update the spreadsheet. If you map those out 00:17:23.710 --> 00:17:26.450 now, when garlic drops, you can plug it in and 00:17:26.450 --> 00:17:28.930 it will just work. That's great advice. Whether 00:17:28.930 --> 00:17:31.769 it ends up being called garlic or GPT 5 .3 or 00:17:31.769 --> 00:17:34.329 something else entirely, the message is clear. 00:17:34.430 --> 00:17:38.890 The era of... giant, slow, forgetful AI is ending. 00:17:39.029 --> 00:17:40.809 And the people who are ready to build with these 00:17:40.809 --> 00:17:43.230 new efficient models are going to have a massive 00:17:43.230 --> 00:17:45.930 advantage. A massive advantage indeed. Thank 00:17:45.930 --> 00:17:47.730 you for walking us through the Code Red. Always 00:17:47.730 --> 00:17:49.849 a pleasure. And to you listening, thank you for 00:17:49.849 --> 00:17:51.950 diving deep with us. Go clean up your data, get 00:17:51.950 --> 00:17:54.269 your workflows ready, and we will see you on 00:17:54.269 --> 00:17:55.170 the next one. Take care.