WEBVTT 00:00:00.000 --> 00:00:01.980 What happens when artificial intelligence stops 00:00:01.980 --> 00:00:04.480 apologizing and starts fitting completely inside 00:00:04.480 --> 00:00:07.839 your pocket? Beat. Welcome to our deep dive for 00:00:07.839 --> 00:00:10.300 today. We are genuinely thrilled you decided 00:00:10.300 --> 00:00:13.019 to join us. Today we are exploring OpenAI's newest 00:00:13.019 --> 00:00:16.679 update. It is designed to make chat GPT significantly 00:00:16.679 --> 00:00:20.120 less cringe. We are also flying through a lightning 00:00:20.120 --> 00:00:22.739 round of industry news. That includes a massive 00:00:22.739 --> 00:00:25.160 hardware leap and a mysterious leaked model. 00:00:25.469 --> 00:00:28.410 Plus, we are unpacking how Alibaba's tiny new 00:00:28.410 --> 00:00:32.130 AI humiliated a massive giant. The entire computing 00:00:32.130 --> 00:00:34.810 landscape is transforming right beneath us. I 00:00:34.810 --> 00:00:36.590 am really looking forward to unpacking this together. 00:00:36.829 --> 00:00:38.909 Let's start with OpenAI and model personality 00:00:38.909 --> 00:00:41.869 design choices. For a very long time, the biggest 00:00:41.869 --> 00:00:43.890 user complaint remained. It was never really 00:00:43.890 --> 00:00:46.630 about raw speed or overall intelligence. The 00:00:46.630 --> 00:00:48.729 primary issue was simply the underlying tone 00:00:48.729 --> 00:00:51.560 of responses. Yeah. That tone was a massive source 00:00:51.560 --> 00:00:53.899 of daily friction. You would get incredibly long, 00:00:54.060 --> 00:00:56.640 moralizing preambles before straightforward answers. 00:00:56.859 --> 00:00:59.579 It had way too many excessive safety caveats 00:00:59.579 --> 00:01:01.479 programmed in. It would literally tell you to 00:01:01.479 --> 00:01:03.920 stop and take a breath. I still wrestle with 00:01:03.920 --> 00:01:06.900 my AI constantly apologizing to me. It makes 00:01:06.900 --> 00:01:09.579 the entire user experience feel incredibly clunky 00:01:09.579 --> 00:01:11.879 today. You just want the digital machine to do 00:01:11.879 --> 00:01:14.879 its job. Users want a highly capable digital 00:01:14.879 --> 00:01:18.239 tool, not a life coach. OpenAI just released 00:01:18.239 --> 00:01:21.340 GPT -5 .3 Instant to finally fix this problem. 00:01:21.599 --> 00:01:23.799 This new update is heavily designed to sound 00:01:23.799 --> 00:01:26.540 incredibly natural. It answers your direct questions 00:01:26.540 --> 00:01:29.200 and drops the awkwardness entirely. But does 00:01:29.200 --> 00:01:31.420 the underlying statistical data actually support 00:01:31.420 --> 00:01:34.340 that claim? I am always highly skeptical of these 00:01:34.340 --> 00:01:36.959 sweeping software promises. They often promise 00:01:36.959 --> 00:01:39.560 massive behavioral improvements that feel largely 00:01:39.560 --> 00:01:41.859 invisible. Surprisingly, the hard empirical data 00:01:41.859 --> 00:01:44.379 actually backs their corporate claims. They achieved 00:01:44.379 --> 00:01:48.439 26 .8 % fewer hallucinations overall today. That 00:01:48.439 --> 00:01:51.019 specifically applies to live web results globally. 00:01:51.640 --> 00:01:53.859 That is a pretty massive leap in daily system 00:01:53.859 --> 00:01:57.560 reliability. What exactly counts as a hallucination 00:01:57.560 --> 00:02:00.299 in this specific context? When AI makes up fake 00:02:00.299 --> 00:02:02.519 facts instead of saying, I don't know, it just 00:02:02.519 --> 00:02:05.219 completely destroys foundational user trust in 00:02:05.219 --> 00:02:08.620 the system. They also saw a 19 .7 % reliability 00:02:08.620 --> 00:02:12.139 improvement internally. Plus, direct user feedback 00:02:12.139 --> 00:02:15.740 showed 22 .5 % fewer hallucinations globally. 00:02:16.180 --> 00:02:18.479 So the previous model version is stepping aside 00:02:18.479 --> 00:02:22.330 very soon. GPT -5 .2 Instant is officially retiring 00:02:22.330 --> 00:02:25.490 from service on June 3rd. That feels incredibly 00:02:25.490 --> 00:02:27.569 fast for a complete software model lifecycle. 00:02:27.949 --> 00:02:30.629 It really is an aggressively fast digital lifecycle 00:02:30.629 --> 00:02:33.129 nowadays everywhere. And they are already teasing 00:02:33.129 --> 00:02:36.569 the mysterious GPT -5 .4 release timeline. The 00:02:36.569 --> 00:02:38.449 release schedule just keeps accelerating faster 00:02:38.449 --> 00:02:41.449 than anyone expected. Two secs silence. I want 00:02:41.449 --> 00:02:43.030 to ask you about the underlying software engineering. 00:02:43.389 --> 00:02:46.849 Why is reducing these AI refusals so technically 00:02:46.849 --> 00:02:49.729 difficult today? Well, it is a profound daily 00:02:49.729 --> 00:02:52.270 balance of safety and helpfulness. If you remove 00:02:52.270 --> 00:02:54.469 the guardrails entirely, the AI breaks down. 00:02:54.610 --> 00:02:57.009 It might generate dangerous or highly toxic content 00:02:57.009 --> 00:02:59.509 almost immediately. Teaching it nuanced context 00:02:59.509 --> 00:03:01.990 requires immense computational effort constantly 00:03:01.990 --> 00:03:04.909 running. So fewer guardrails, but much smarter 00:03:04.909 --> 00:03:08.289 navigation. Precisely. It knows exactly when 00:03:08.289 --> 00:03:11.090 to pump the brakes naturally now. It understands 00:03:11.090 --> 00:03:13.750 human conversation nuance much better than previous 00:03:13.750 --> 00:03:16.740 versions. Let's dive right into a rapid synthesis 00:03:16.740 --> 00:03:20.000 of industry news. We absolutely need to examine 00:03:20.000 --> 00:03:22.139 these completely frictionless workflows today. 00:03:22.740 --> 00:03:24.879 Anthropic just rolled out voice mode for their 00:03:24.879 --> 00:03:27.699 Claude code. You simply speak your commands and 00:03:27.699 --> 00:03:29.560 it writes the code. Right. And that completely 00:03:29.560 --> 00:03:32.340 changes the deep creative flow state. You dictate 00:03:32.340 --> 00:03:34.240 the structural architecture and it writes the 00:03:34.240 --> 00:03:37.110 syntax. We're also seeing incredible platforms 00:03:37.110 --> 00:03:40.370 like GetVictor .com. It acts autonomously across 00:03:40.370 --> 00:03:43.370 over 3 ,000 different Slack tools. That level 00:03:43.370 --> 00:03:45.430 of autonomous integration is simply staggering 00:03:45.430 --> 00:03:48.789 to consider. Then we have Crisp Conversion, working 00:03:48.789 --> 00:03:51.169 on global business communication. It understands 00:03:51.169 --> 00:03:53.569 highly accented speech locally on your personal 00:03:53.569 --> 00:03:56.490 device. You get near zero latency for your global 00:03:56.490 --> 00:03:58.849 video conference calls. Yeah, it removes the 00:03:58.849 --> 00:04:00.530 language barrier from international business 00:04:00.530 --> 00:04:03.830 communication entirely. We also have major developments 00:04:03.830 --> 00:04:05.849 regarding physical hardware processing speed. 00:04:06.250 --> 00:04:09.770 Google introduced Gemini 3 .1 Flashlight in a 00:04:09.770 --> 00:04:12.449 developer preview today. This new version is 00:04:12.449 --> 00:04:15.610 45 % faster than previous iterations. It could 00:04:15.610 --> 00:04:18.310 easily handle 1 million token prompts from users. 00:04:18.629 --> 00:04:20.189 I mean, what does that actually look like in 00:04:20.189 --> 00:04:23.389 practice? Feeding the AI. dozens of full books 00:04:23.389 --> 00:04:26.170 all at once. It holds all that dense information 00:04:26.170 --> 00:04:28.850 in its active memory. That completely changes 00:04:28.850 --> 00:04:31.290 how financial analysts review massive data sets. 00:04:31.790 --> 00:04:34.589 Speaking of incredible hardware leaps, IR Labs 00:04:34.589 --> 00:04:37.879 raised massive funding. They secured $500 million 00:04:37.879 --> 00:04:40.699 at huge corporate valuations. They are backed 00:04:40.699 --> 00:04:43.279 by massive industry giants like Nvidia now. Whoa, 00:04:43.540 --> 00:04:46.540 imagine using literal light to wire AI brains 00:04:46.540 --> 00:04:48.860 together. They're replacing traditional copper 00:04:48.860 --> 00:04:51.560 wires with advanced optical connections. Using 00:04:51.560 --> 00:04:53.439 light completely changes the physical limits 00:04:53.439 --> 00:04:55.699 of global computing. We must also address the 00:04:55.699 --> 00:04:58.100 real world safety consequences today. Twitter 00:04:58.100 --> 00:05:00.279 is handing out incredibly harsh platform penalties 00:05:00.279 --> 00:05:02.620 right now. They will suspend anyone for posting 00:05:02.620 --> 00:05:05.470 completely fake AI videos. Deepfakes are becoming 00:05:05.470 --> 00:05:08.649 completely indistinguishable from actual physical 00:05:08.649 --> 00:05:12.850 video reality. We also saw a highly fascinating 00:05:12.850 --> 00:05:15.790 test regarding geopolitical conflicts. Researchers 00:05:15.790 --> 00:05:17.949 asked different models about the Iran war test 00:05:17.949 --> 00:05:21.389 recently. One hallucinated while another stayed 00:05:21.389 --> 00:05:24.470 extremely politically cautious and refused. There 00:05:24.470 --> 00:05:27.110 is also incredible drama exploding in the developer 00:05:27.110 --> 00:05:30.899 community. Developers spotted GPT -5 .4 inside 00:05:30.899 --> 00:05:33.939 the OpenAI codex source code. Then the references 00:05:33.939 --> 00:05:36.480 suddenly disappeared completely a few short hours 00:05:36.480 --> 00:05:39.040 later. It genuinely looks like a classic oops 00:05:39.040 --> 00:05:41.680 internal corporate leak. We also have a quick 00:05:41.680 --> 00:05:44.100 update from Deep Personality today. They claim 00:05:44.100 --> 00:05:45.720 to understand you better than a professional 00:05:45.720 --> 00:05:48.860 therapist. They just need 10 interactive conversational 00:05:48.860 --> 00:05:51.180 digital sessions with you. Okay, we have talked 00:05:51.180 --> 00:05:54.300 heavily about voice coding and microchips. We 00:05:54.300 --> 00:05:56.779 discussed light speed connections and fully autonomous 00:05:56.779 --> 00:05:59.819 Slack agent tools. What is the underlying theme 00:05:59.819 --> 00:06:01.860 connecting all these scattered updates? It is 00:06:01.860 --> 00:06:03.959 about removing the barrier between human thought 00:06:03.959 --> 00:06:06.300 and digital execution. You just think or naturally 00:06:06.300 --> 00:06:09.000 speak and the machine acts. It's all about frictionless 00:06:09.000 --> 00:06:13.250 thinking and absolutely zero delays. We will 00:06:13.250 --> 00:06:16.209 be right back after a brief word. Mid -roll sponsor 00:06:16.209 --> 00:06:19.910 break. And we are back to our deep dive. Now 00:06:19.910 --> 00:06:22.449 let's confidently move to our final big breakthrough 00:06:22.449 --> 00:06:25.430 today. Alibaba just launched a massive digital 00:06:25.430 --> 00:06:28.449 family of tiny models. It is officially called 00:06:28.449 --> 00:06:33.839 the QEN 3 .5 Small AI Model Series. These models 00:06:33.839 --> 00:06:36.180 are explicitly designed to run completely offline 00:06:36.180 --> 00:06:38.560 locally. They released four specific versions 00:06:38.560 --> 00:06:41.180 tailored for different local hardware. First, 00:06:41.300 --> 00:06:44.259 we clearly have the 0 .8 billion version. That 00:06:44.259 --> 00:06:46.079 metric is strictly counting the model's total 00:06:46.079 --> 00:06:48.660 parameters overall. Remind me, what exactly are 00:06:48.660 --> 00:06:51.180 parameters in this specific context? The digital 00:06:51.180 --> 00:06:53.879 brain cells that determine an AI's overall smarts. 00:06:53.879 --> 00:06:56.139 That smallest model is perfectly sized for everyday 00:06:56.139 --> 00:06:58.620 mobile phones. Then they have a 2 billion parameter 00:06:58.620 --> 00:07:01.160 version available now. The 4 billion parameter 00:07:01.160 --> 00:07:03.800 model offers significantly stronger multimodal 00:07:03.800 --> 00:07:05.839 capabilities. You can physically show it a picture 00:07:05.839 --> 00:07:07.860 and it understands. And the 9 billion parameter 00:07:07.860 --> 00:07:10.860 version is the absolute powerhouse. The 9 billion 00:07:10.860 --> 00:07:13.680 parameter model pulled off a historic upset. 00:07:13.920 --> 00:07:17.720 It completely outscored OpenAI's massive GPT 00:07:17.720 --> 00:07:21.279 -OSS 120B model in direct testing. It won on 00:07:21.279 --> 00:07:23.620 complex graduate level reasoning and nuanced 00:07:23.620 --> 00:07:26.819 multilingual knowledge. That massive OpenAI model 00:07:26.819 --> 00:07:30.329 is more than 13 times larger. The 4 billion model 00:07:30.329 --> 00:07:33.410 also achieved something absolutely visually spectacular. 00:07:33.810 --> 00:07:37.269 It matched complex visual benchmark scores previously 00:07:37.269 --> 00:07:40.449 requiring massive architecture. Even Elon Musk 00:07:40.449 --> 00:07:42.750 reacted very publicly to this specific release. 00:07:42.990 --> 00:07:45.149 He called the performance a display of impressive 00:07:45.149 --> 00:07:47.689 intelligence density. Another incredibly crucial 00:07:47.689 --> 00:07:50.149 detail is that these models are open weight. 00:07:50.389 --> 00:07:52.629 What does open weight actually mean for the average 00:07:52.629 --> 00:07:54.990 software developer? Free models where anyone 00:07:54.990 --> 00:07:57.860 can tweak the core digital brain. It fundamentally 00:07:57.860 --> 00:08:00.180 changes the raw economics of software development 00:08:00.180 --> 00:08:03.040 today. Right. But there's also highly significant 00:08:03.040 --> 00:08:06.600 human drama surrounding this. Right after Alibaba 00:08:06.600 --> 00:08:08.660 launched this incredibly historic tech, things 00:08:08.660 --> 00:08:11.360 changed. The project's key technical leads suddenly 00:08:11.360 --> 00:08:13.660 stepped down completely today. Colleagues publicly 00:08:13.660 --> 00:08:15.959 stated the sudden corporate move feels like an 00:08:15.959 --> 00:08:18.680 end. They openly called his highly sudden departure 00:08:18.680 --> 00:08:21.680 the end of an era. Severe burnout in the artificial 00:08:21.680 --> 00:08:24.379 intelligence industry is very real. The immense 00:08:24.379 --> 00:08:26.939 pressure to constantly release new models breaks 00:08:26.939 --> 00:08:30.300 people. Two sec silence. Why does a tiny model 00:08:30.300 --> 00:08:33.340 beating a massive server actually matter? Why 00:08:33.340 --> 00:08:35.779 should the completely average listener care about 00:08:35.779 --> 00:08:38.740 this upset? Think deeply about the physical laptop 00:08:38.740 --> 00:08:41.740 sitting on your desk. It forcefully moves complex 00:08:41.740 --> 00:08:44.419 data processing directly onto personal devices. 00:08:44.720 --> 00:08:47.220 You easily gain immense computing capability 00:08:47.220 --> 00:08:50.059 without sacrificing any privacy. So ultimate 00:08:50.059 --> 00:08:52.039 power is shifting directly into our pockets. 00:08:52.159 --> 00:08:55.059 Yes. It beautifully democratizes widespread access 00:08:55.059 --> 00:08:58.419 to absolute cutting -edge reasoning logic. Anyone 00:08:58.419 --> 00:09:00.720 anywhere with a basic computer essentially has 00:09:00.720 --> 00:09:03.139 a private genius. Let's deliberately take a moment 00:09:03.139 --> 00:09:06.059 to tile these themes together. AI is finally 00:09:06.059 --> 00:09:09.019 shedding its awkward, apologizing phase completely 00:09:09.019 --> 00:09:11.779 today. It is integrating flawlessly into our 00:09:11.779 --> 00:09:14.340 daily hardware and workflows. And most importantly, 00:09:14.559 --> 00:09:16.980 the entire physical technology stack is shrinking. 00:09:17.220 --> 00:09:20.299 We historically used to rely entirely on giant, 00:09:20.360 --> 00:09:23.240 expensive cloud models. Now we are completely 00:09:23.240 --> 00:09:25.679 transitioning to quiet, highly dense software 00:09:25.679 --> 00:09:28.360 locally. These incredibly hyper -intelligent 00:09:28.360 --> 00:09:30.620 models will soon live entirely on your phone. 00:09:30.960 --> 00:09:33.539 We deeply want to leave you with a final provocative 00:09:33.539 --> 00:09:36.990 thought. If a 9 billion parameter model on a 00:09:36.990 --> 00:09:39.889 laptop wins completely today, what happens in 00:09:39.889 --> 00:09:42.090 two years when your offline pocket device magically 00:09:42.090 --> 00:09:45.269 evolves? What if it easily outsmarts the absolutely 00:09:45.269 --> 00:09:48.309 greatest supercomputers existing today? How does 00:09:48.309 --> 00:09:50.470 a world teeming with private pocket geniuses 00:09:50.470 --> 00:09:53.269 change collaboration? That is a truly wild future 00:09:53.269 --> 00:09:55.429 to imagine living in. Thank you deeply for taking 00:09:55.429 --> 00:09:57.909 this highly fascinating deep dive with us. Stay 00:09:57.909 --> 00:10:00.549 endlessly curious, keep bravely exploring, and 00:10:00.549 --> 00:10:02.389 we will see you next time. Out to row music.