WEBVTT 00:00:00.000 --> 00:00:02.220 I want you to picture an extremely sophisticated 00:00:02.220 --> 00:00:05.940 mind, like a super genius. Okay. They're trying 00:00:05.940 --> 00:00:08.019 to solve these really complex logic problems, 00:00:08.160 --> 00:00:10.980 but first they have to run a full detailed calculation 00:00:10.980 --> 00:00:13.380 just to remember what the capital of France is. 00:00:13.539 --> 00:00:15.599 Every single time. Every single time. It's an 00:00:15.599 --> 00:00:18.839 absurd waste of energy. It's totally counterintuitive, 00:00:18.839 --> 00:00:21.739 isn't it? But that strange inefficiency, that 00:00:21.739 --> 00:00:24.399 repetitive brute force thinking just to recall 00:00:24.399 --> 00:00:27.980 simple facts. That's been a foundational limit 00:00:27.980 --> 00:00:30.640 for large language models since day one. They 00:00:30.640 --> 00:00:34.979 just lack a reliable built -in memory. Okay, 00:00:35.039 --> 00:00:37.280 so let's unpack this using the sources you shared. 00:00:37.439 --> 00:00:40.539 For years, the story of AI has been all about 00:00:40.539 --> 00:00:43.399 scaling up, right? More data, more parameters. 00:00:43.500 --> 00:00:45.439 More processing power. Exactly. We've been focused 00:00:45.439 --> 00:00:48.759 on just increasing the raw brain size. But today... 00:00:48.960 --> 00:00:51.020 We're seeing this fascinating shift. The sources 00:00:51.020 --> 00:00:52.799 are pointing toward the focus moving from that 00:00:52.799 --> 00:00:55.619 inefficient, constant processing. To extremely 00:00:55.619 --> 00:00:58.560 efficient architectural remembering. Right. Absolutely. 00:00:58.600 --> 00:01:01.700 And our mission today is to cut through the noise 00:01:01.700 --> 00:01:04.319 and give you those vital nuggets from this evolving 00:01:04.319 --> 00:01:07.319 landscape. This deep dive is all about specialization, 00:01:07.500 --> 00:01:10.400 personalization, and efficiency. So here's the 00:01:10.400 --> 00:01:12.640 roadmap for you. First, we're going to look at 00:01:12.640 --> 00:01:15.420 how OpenAI is finally seriously challenging Google 00:01:15.420 --> 00:01:17.689 Translate. And they're doing it by injecting 00:01:17.689 --> 00:01:20.569 emotional and tonal context into it. Then we'll 00:01:20.569 --> 00:01:22.510 do a quick synthesis of the market. We're talking 00:01:22.510 --> 00:01:26.049 predicted mega IPOs, major corporate shifts, 00:01:26.170 --> 00:01:29.670 and critical revaluation of data. And finally, 00:01:29.730 --> 00:01:31.450 we'll dive deep into that breakthrough memory 00:01:31.450 --> 00:01:33.969 architecture we just mentioned, DeepSeq's Enneagram 00:01:33.969 --> 00:01:36.290 system, which really changes the game by giving 00:01:36.290 --> 00:01:39.590 LLMs a true functional memory bank. Let's start 00:01:39.590 --> 00:01:42.650 with translation. For what, over a decade, Google 00:01:42.650 --> 00:01:44.969 Translate has just been the default. Oh, yeah. 00:01:45.049 --> 00:01:47.260 It's functional. But everyone knows the output 00:01:47.260 --> 00:01:51.099 often sounds, well, sterile. Yeah. Mechanical. 00:01:51.180 --> 00:01:54.260 And now OpenAI has launched ChatGPT Translate, 00:01:54.359 --> 00:01:57.359 and it is a direct challenge with a major twist. 00:01:57.560 --> 00:01:59.180 And that twist is? They aren't just trying to 00:01:59.180 --> 00:02:00.700 provide a basic language conversion. They're 00:02:00.700 --> 00:02:03.280 targeting the professional need for nuance. Right. 00:02:03.319 --> 00:02:05.859 The core innovation is that this translator lets 00:02:05.859 --> 00:02:08.460 you customize the output tone right from the 00:02:08.460 --> 00:02:12.800 interface. That flexibility is the real aha moment 00:02:12.800 --> 00:02:14.780 here, isn't it? Especially for anyone who needs 00:02:14.780 --> 00:02:18.490 to bridge cultural gaps or handle sensitive communication? 00:02:18.870 --> 00:02:21.849 It's tone -aware translation. It completely changes 00:02:21.849 --> 00:02:24.069 the standard for quality. So you could give it 00:02:24.069 --> 00:02:27.250 an example. Sure. You can input a complex sentence, 00:02:27.530 --> 00:02:30.469 say a response to a client complaint, and tell 00:02:30.469 --> 00:02:32.990 the AI, translate this into German, but make 00:02:32.990 --> 00:02:36.280 it sound... highly formal like a university professor 00:02:36.280 --> 00:02:38.639 writing a business contract. Wow. Or you can 00:02:38.639 --> 00:02:40.919 flip that completely and ask it to translate 00:02:40.919 --> 00:02:42.939 something into Spanish and make it sound, you 00:02:42.939 --> 00:02:44.939 know, casual like a text message between friends. 00:02:45.180 --> 00:02:47.300 This just highlights how mechanical the old systems 00:02:47.300 --> 00:02:49.180 are. Google Translate is basically operating 00:02:49.180 --> 00:02:51.960 on this massive statistical map. Yeah, it finds 00:02:51.960 --> 00:02:54.159 the most likely linguistic equivalent. But it 00:02:54.159 --> 00:02:56.289 was never built... to understand the emotional 00:02:56.289 --> 00:02:58.610 or professional intent behind the words. And 00:02:58.610 --> 00:03:01.310 that's the competitive barrier. The sources suggest 00:03:01.310 --> 00:03:04.110 this new service, which is likely powered by 00:03:04.110 --> 00:03:07.930 GPT -4 Turbo, handles over 50 languages. And 00:03:07.930 --> 00:03:10.009 it works right on mobile browsers. Right, with 00:03:10.009 --> 00:03:13.789 text or voice. For big tech, this is all about 00:03:13.789 --> 00:03:17.349 targeting that next layer of value that goes 00:03:17.349 --> 00:03:20.270 beyond just basic utility. When you think about 00:03:20.270 --> 00:03:23.810 the user base, I mean... Business users, students 00:03:23.810 --> 00:03:26.530 drafting papers, international creators. Yeah. 00:03:26.939 --> 00:03:29.219 They all need translations that don't sound like 00:03:29.219 --> 00:03:31.159 they came from a machine. A contract that sounds 00:03:31.159 --> 00:03:33.919 robotic isn't just awkward. No, it could actually 00:03:33.919 --> 00:03:36.159 hurt a deal. Exactly. It saves you that extra 00:03:36.159 --> 00:03:39.060 step of translating, then manually rewriting 00:03:39.060 --> 00:03:41.500 the whole thing to sound natural. It delivers 00:03:41.500 --> 00:03:44.219 context and quality at the same time. But wait, 00:03:44.280 --> 00:03:46.439 if the user just typing in a prompt like make 00:03:46.439 --> 00:03:48.580 it sound casual, isn't that essentially just 00:03:48.580 --> 00:03:51.060 a packaged prompt engineering trick? Why is this 00:03:51.060 --> 00:03:52.979 a major step forward for translation models? 00:03:53.199 --> 00:03:55.020 That's a really good question. It's more than 00:03:55.020 --> 00:03:57.439 just a trick. The model itself is trained to 00:03:57.439 --> 00:04:00.240 recognize this huge spectrum of linguistic features, 00:04:00.379 --> 00:04:03.360 things like register, formality markers, cultural 00:04:03.360 --> 00:04:05.580 norms. Things that are usually invisible to a 00:04:05.580 --> 00:04:08.199 simple statistical engine. Precisely. It's embedding 00:04:08.199 --> 00:04:10.819 that contextual awareness into the translation 00:04:10.819 --> 00:04:14.159 process itself. So the translated language structure 00:04:14.159 --> 00:04:17.139 reflects the requested tone, not just the word 00:04:17.139 --> 00:04:21.310 choice. So how will this focus on tone? fundamentally 00:04:21.310 --> 00:04:23.810 change professional communication across different 00:04:23.810 --> 00:04:26.569 languages. It ensures translated content conveys 00:04:26.569 --> 00:04:29.550 human intent, not just literal word equivalence. 00:04:29.670 --> 00:04:32.050 Okay, let's pivot to the broader AI ecosystem. 00:04:32.350 --> 00:04:34.470 Let's do it. We'll look at shifts in productivity 00:04:34.470 --> 00:04:37.269 and personalization that affect you daily, and 00:04:37.269 --> 00:04:39.709 then at the massive capital that's driving this 00:04:39.709 --> 00:04:42.689 entire sector. On the personalization side, we're 00:04:42.689 --> 00:04:45.930 seeing users find really clever ways to integrate 00:04:45.930 --> 00:04:49.389 AI into their workflows. The sources highlighted 00:04:49.389 --> 00:04:52.310 a trending... thread on Reddit where a user curated 00:04:52.310 --> 00:04:54.910 15 of what they called golden prompts over six 00:04:54.910 --> 00:04:57.449 months. And they claim this saved them 10 to 00:04:57.449 --> 00:04:59.949 15 hours every single week. Prompt engineering 00:04:59.949 --> 00:05:02.069 or prompt hacking, as some people are calling 00:05:02.069 --> 00:05:04.129 it, is definitely becoming a high value skill. 00:05:04.310 --> 00:05:06.689 It really is. But the providers are also doing 00:05:06.689 --> 00:05:08.769 the hard work to bake that personalization right 00:05:08.769 --> 00:05:11.529 into the product itself. And that's where Google's 00:05:11.529 --> 00:05:13.610 personal intelligence in Gemini comes in. Right. 00:05:13.790 --> 00:05:16.149 This is integrating the model directly with a 00:05:16.149 --> 00:05:19.680 user's personal ecosystem. your Gmail, your photos, 00:05:19.920 --> 00:05:23.699 your YouTube data, to offer these hyper -customized 00:05:23.699 --> 00:05:25.980 answers. So instead of asking a general question, 00:05:26.259 --> 00:05:28.480 you can ask something like, based on my Gmail, 00:05:28.699 --> 00:05:31.360 summarize the three key action items from the 00:05:31.360 --> 00:05:33.680 Smith meeting last week. And then find the flight 00:05:33.680 --> 00:05:36.259 receipt from my photos folder. Exactly. Their 00:05:36.259 --> 00:05:38.360 pitch is basically, we're saving you a headache. 00:05:38.829 --> 00:05:41.350 by making the AI an extension of your digital 00:05:41.350 --> 00:05:44.709 memory. And that focus on seamless integration 00:05:44.709 --> 00:05:48.029 is also happening so fast inside corporations. 00:05:48.689 --> 00:05:51.810 Microsoft, for instance, just shut down its massive 00:05:51.810 --> 00:05:54.610 internal employee library. I saw that, including 00:05:54.610 --> 00:05:57.290 subscriptions they've used for decades. And they 00:05:57.290 --> 00:05:59.490 replaced that traditional resource with an AI 00:05:59.490 --> 00:06:01.959 -powered skilling hub. Just think about the message 00:06:01.959 --> 00:06:04.560 that sends. Institutional knowledge is no longer 00:06:04.560 --> 00:06:07.560 static documents in a folder somewhere. No. It's 00:06:07.560 --> 00:06:10.180 a dynamic conversational resource managed by 00:06:10.180 --> 00:06:13.160 an AI. That's a fundamental change. Now, moving 00:06:13.160 --> 00:06:16.579 to the money side. The legal drama between Altman 00:06:16.579 --> 00:06:19.540 and Musk, that's still unfolding. New court documents 00:06:19.540 --> 00:06:22.040 just dropped. But the market valuation story 00:06:22.040 --> 00:06:24.819 is what's truly staggering. I mean, a New York 00:06:24.819 --> 00:06:27.600 Times report suggests 2026 is shaping up to be 00:06:27.600 --> 00:06:29.920 the year of the... What do they call it? The 00:06:29.920 --> 00:06:33.439 $2 trillion mega IPO. That's wild. We're talking 00:06:33.439 --> 00:06:36.860 about open AI, Anthropic and SpaceX potentially 00:06:36.860 --> 00:06:38.959 hitting the public markets almost at the same 00:06:38.959 --> 00:06:41.519 time. That would just redefine global tech investment. 00:06:41.699 --> 00:06:44.019 And investors are still betting big on specialized 00:06:44.019 --> 00:06:47.360 sectors. The AI video startup Hakesfield just 00:06:47.360 --> 00:06:50.800 hit a $1 .3 billion valuation after a new funding 00:06:50.800 --> 00:06:53.220 round. Which shows massive confidence in video 00:06:53.220 --> 00:06:55.360 generation as a real market, you know, despite 00:06:55.360 --> 00:06:57.519 the ongoing debates about quality. But the most 00:06:57.519 --> 00:06:59.540 fascinating shift, the one that speaks directly 00:06:59.540 --> 00:07:02.100 to the long term cost of all this. It involves 00:07:02.100 --> 00:07:04.959 Wikipedia. This is a huge signal that the old 00:07:04.959 --> 00:07:07.860 scraping era is ending. A hundred percent. Wikipedia 00:07:07.860 --> 00:07:10.620 has now signed actual payment deals with major 00:07:10.620 --> 00:07:14.579 tech companies, Microsoft, Meta, even Perplexity. 00:07:14.600 --> 00:07:17.040 Giving them access to their content for AI training. 00:07:17.279 --> 00:07:20.480 So big tech, which for years just ingested vast 00:07:20.480 --> 00:07:23.439 amounts of content for free, is now paying a 00:07:23.439 --> 00:07:27.759 premium for quality verified source data. This 00:07:27.759 --> 00:07:30.000 changes the entire economics of training models. 00:07:30.259 --> 00:07:33.379 Whoa. I mean, imagine scaling your company to 00:07:33.379 --> 00:07:36.060 hundreds of millions, even a billion queries 00:07:36.060 --> 00:07:38.879 without ever paying for your core source knowledge. 00:07:39.079 --> 00:07:41.560 That was the Internet model. And it seems that 00:07:41.560 --> 00:07:44.300 model is finally broken. Yeah, it suggests the 00:07:44.300 --> 00:07:47.420 market has realized that freely scraped, noisy 00:07:47.420 --> 00:07:50.839 data is just far less valuable than curated, 00:07:50.980 --> 00:07:54.199 clean data. Quality is now a scarce and valuable 00:07:54.199 --> 00:07:57.350 commodity. So what does big tech paying Wikipedia 00:07:57.350 --> 00:08:00.089 tell us about the fundamental long term value 00:08:00.089 --> 00:08:02.970 of quality training data? Quality training data 00:08:02.970 --> 00:08:05.509 is transitioning from a free resource into a 00:08:05.509 --> 00:08:07.990 critical, expensive commodity. OK, so let's get 00:08:07.990 --> 00:08:09.569 back to that technical challenge we introduced 00:08:09.569 --> 00:08:11.569 at the top. The inefficiency problem. Right. 00:08:11.629 --> 00:08:14.529 Why do these massive models waste so much compute 00:08:14.529 --> 00:08:16.649 power just remembering how to spell basic words 00:08:16.649 --> 00:08:19.029 or recall simple facts? Traditionally, the race 00:08:19.029 --> 00:08:21.430 has been all about increasing the context window, 00:08:21.629 --> 00:08:24.009 you know, making the workspace bigger. We focus 00:08:24.009 --> 00:08:26.610 on how much data the model can see at one time. 00:08:27.000 --> 00:08:28.899 But you're right. The problem isn't the size 00:08:28.899 --> 00:08:31.560 of the room. It's the filing system inside the 00:08:31.560 --> 00:08:34.779 model. Precisely. They are constantly retalculating 00:08:34.779 --> 00:08:37.700 facts because they lack an indexed built -in 00:08:37.700 --> 00:08:40.320 memory structure that separates that rote knowledge 00:08:40.320 --> 00:08:42.700 from complex reasoning. This is where DeepSeq 00:08:42.700 --> 00:08:45.460 comes in. Enter DeepSeq. They introduced the 00:08:45.460 --> 00:08:48.159 memory lookup module for LLMs, and they named 00:08:48.159 --> 00:08:51.240 it Ngram. So Ngram is essentially a real memory 00:08:51.240 --> 00:08:53.879 system. A conditional one, yeah. It lets the 00:08:53.879 --> 00:08:56.879 model know... when to think deeply and when to 00:08:56.879 --> 00:08:59.600 just retrieve a fact. How does that actually 00:08:59.600 --> 00:09:01.840 work? Well, think of it this way. The model's 00:09:01.840 --> 00:09:03.840 deep neural network is like a master sculptor. 00:09:03.860 --> 00:09:06.139 It takes time and compute to create a new statue. 00:09:06.340 --> 00:09:08.840 Okay. But if you just want to reproduce the same 00:09:08.840 --> 00:09:11.480 simple flower over and over, you don't re -sculpt 00:09:11.480 --> 00:09:14.720 it every time. You use a mold, a stamp. Enneagram 00:09:14.720 --> 00:09:17.559 uses something called engram embeddings. And 00:09:17.559 --> 00:09:19.440 for our listeners, how would you define engram 00:09:19.440 --> 00:09:22.019 embeddings in simple terms? Ngram embeddings 00:09:22.019 --> 00:09:25.259 are basically compressed pre -packaged digital 00:09:25.259 --> 00:09:28.460 representations of common phrases or facts. They're 00:09:28.460 --> 00:09:30.940 like pre -made organized stamps you can grab 00:09:30.940 --> 00:09:33.720 instantly. So instead of rebuilding the answer 00:09:33.720 --> 00:09:36.750 through complex processing. It just. performs 00:09:36.750 --> 00:09:39.909 a single, lightning -fast, one -step lookup for 00:09:39.909 --> 00:09:41.669 that fact. Hang on a minute. If they're just 00:09:41.669 --> 00:09:45.230 prepackaged, compressed phrases stamps, as you 00:09:45.230 --> 00:09:47.750 put it, isn't that just a really optimized caching 00:09:47.750 --> 00:09:50.470 layer? Why is DeepSea calling this revolutionary? 00:09:51.240 --> 00:09:53.600 Because it fundamentally changes how the model 00:09:53.600 --> 00:09:55.740 uses its depth. It's not just caching external 00:09:55.740 --> 00:09:58.220 data. It's offloading the internal memory function 00:09:58.220 --> 00:10:01.200 that was previously eating up the network's reasoning 00:10:01.200 --> 00:10:04.440 capacity. I see. That instant retrieval frees 00:10:04.440 --> 00:10:06.919 up the core model's depth for actual complicated 00:10:06.919 --> 00:10:10.700 tasks. The math, the code generation, the truly 00:10:10.700 --> 00:10:13.720 novel logic. It stops simple recall from clogging 00:10:13.720 --> 00:10:16.779 up. the intellectual arteries of the neural network. 00:10:16.919 --> 00:10:18.519 That's a good way to put it. It separates the 00:10:18.519 --> 00:10:20.580 dictionary from the logic center. And DeepSeek's 00:10:20.580 --> 00:10:22.320 research showed something critical here. What 00:10:22.320 --> 00:10:24.259 was that? They found that general model performance 00:10:24.259 --> 00:10:27.259 actually dips significantly when you push models 00:10:27.259 --> 00:10:29.700 to a massive scale if memory isn't handled correctly. 00:10:30.179 --> 00:10:32.700 But by offloading that rote memory work to the 00:10:32.700 --> 00:10:36.259 Enneagram system, performance rebounds. A lot. 00:10:36.480 --> 00:10:39.870 And this whole tension. between using deep processing 00:10:39.870 --> 00:10:43.549 for complex logic versus a quick lookup for simple 00:10:43.549 --> 00:10:46.690 facts, that balance is known as sparsity allocation, 00:10:47.029 --> 00:10:49.899 right? how you intelligently distribute computational 00:10:49.899 --> 00:10:52.779 effort. Exactly. It's one of the biggest efficiency 00:10:52.779 --> 00:10:56.039 and cost problems in AI today. On a human level, 00:10:56.220 --> 00:10:58.039 you know, when we try to juggle too much information 00:10:58.039 --> 00:11:00.559 in our heads, we get context collapse. I mean, 00:11:00.580 --> 00:11:02.720 I still wrestle with prompt drift myself when 00:11:02.720 --> 00:11:04.480 I overload a context window on a challenging 00:11:04.480 --> 00:11:06.620 task. That's relatable. If you can offload the 00:11:06.620 --> 00:11:08.399 simple stuff, you can focus on the deeper thought. 00:11:08.700 --> 00:11:10.879 So the question now is whether the major players 00:11:10.879 --> 00:11:13.240 like OpenAI and Google will actually commit to 00:11:13.240 --> 00:11:15.940 this memory -first architecture. Will they adopt 00:11:15.940 --> 00:11:18.299 Enneagram or something like it? Or will they 00:11:18.299 --> 00:11:20.980 just keep doubling down on sheer scale, hoping 00:11:20.980 --> 00:11:23.639 massive compute brute forces the memory problem? 00:11:23.980 --> 00:11:25.960 If you can remember simple facts efficiently, 00:11:26.279 --> 00:11:29.759 you just. You dramatically reduce the effort 00:11:29.759 --> 00:11:31.899 needed for that constant brute force thinking. 00:11:32.159 --> 00:11:35.159 It changes the entire economic model for future 00:11:35.159 --> 00:11:38.100 development. So will this new memory architecture 00:11:38.100 --> 00:11:41.159 fundamentally change how all models are developed 00:11:41.159 --> 00:11:44.039 going forward? If AI can remember efficiently, 00:11:44.399 --> 00:11:47.440 less effort is needed for continuous expensive 00:11:47.440 --> 00:11:50.100 brute force processing. So what does this all 00:11:50.100 --> 00:11:52.820 mean for you, the learner? We've covered an AI 00:11:52.820 --> 00:11:55.539 landscape that is rapidly moving away from these 00:11:55.539 --> 00:11:58.519 generalized one size fits all tools. And towards 00:11:58.519 --> 00:12:01.679 specialized tone aware services that prioritize 00:12:01.679 --> 00:12:04.200 human nuance. Yeah. And we watched the market 00:12:04.200 --> 00:12:06.700 finally recognize and actually start paying for 00:12:06.700 --> 00:12:09.700 the true value of quality data. The overarching 00:12:09.700 --> 00:12:12.940 trend here is clearly specialization, personalization, 00:12:13.080 --> 00:12:16.080 and a relentless push for efficiency. AI is getting 00:12:16.080 --> 00:12:18.179 smarter, but not just by getting bigger, like, 00:12:18.200 --> 00:12:20.200 you know, stacking Lego blocks of data. Right. 00:12:20.340 --> 00:12:22.740 It's getting smarter by optimizing its architecture 00:12:22.740 --> 00:12:25.039 and using its resources more intelligently, which 00:12:25.039 --> 00:12:26.940 is exactly what we saw with DeepSeq's Enneagram. 00:12:27.259 --> 00:12:29.600 If architectural breakthroughs like Enneagram 00:12:29.600 --> 00:12:32.460 become the standard and they solve this massive 00:12:32.460 --> 00:12:34.480 cost problem with fact retrieval and memory, 00:12:34.720 --> 00:12:37.500 the next great limit might not be data access 00:12:37.500 --> 00:12:41.080 or rock. compute. It might be economic. How so? 00:12:41.279 --> 00:12:43.700 Well, if efficiency makes the cost of running 00:12:43.700 --> 00:12:47.159 these LLMs just collapse, does AI stop being 00:12:47.159 --> 00:12:49.539 a specialized tool for giant corporations and 00:12:49.539 --> 00:12:52.120 become a true utility accessible to every single 00:12:52.120 --> 00:12:54.279 person on Earth? That's something for you to 00:12:54.279 --> 00:12:56.279 consider as these systems evolve. And I'd encourage 00:12:56.279 --> 00:12:58.240 you to look into the concept of sparsity allocation, 00:12:58.580 --> 00:13:01.039 how the AI allocates its effort, or even just 00:13:01.039 --> 00:13:03.220 test out that new tone -aware translation feature 00:13:03.220 --> 00:13:05.539 yourself. You'll really see the difference between 00:13:05.539 --> 00:13:08.279 mechanical output and contextual understanding. 00:13:08.720 --> 00:13:10.460 Thanks for diving deep into your sources with 00:13:10.460 --> 00:13:11.799 us. We'll catch you on the next deep dive.