WEBVTT 00:00:00.000 --> 00:00:03.140 You know, AI is achieving AGI capabilities right 00:00:03.140 --> 00:00:05.559 now, but maybe not where you'd expect. It's happening 00:00:05.559 --> 00:00:08.619 pretty quietly, actually, inside production code 00:00:08.619 --> 00:00:12.140 bases. We're talking about AI autonomously fixing 00:00:12.140 --> 00:00:15.160 really complex bugs, submitting finalized code 00:00:15.160 --> 00:00:18.879 changes, no direct human intervention. Welcome 00:00:18.879 --> 00:00:21.379 to the Deep Dive. This is where we take the week's 00:00:21.379 --> 00:00:23.800 key AI research and news, and while we distill 00:00:23.800 --> 00:00:26.539 it into a quick, deep analysis for you, we're 00:00:26.539 --> 00:00:28.079 really trying to jump into the progress that 00:00:28.079 --> 00:00:31.280 often goes completely unreported. Yeah, exactly. 00:00:31.440 --> 00:00:33.560 We're looking past the sort of viral chat apps 00:00:33.560 --> 00:00:35.520 today, getting right into the, let's say, the 00:00:35.520 --> 00:00:38.460 technical engine room of autonomy. So we've structured 00:00:38.460 --> 00:00:40.539 this dive around three main things for you. First, 00:00:40.719 --> 00:00:43.179 why coding is perhaps the real epicenter for 00:00:43.179 --> 00:00:46.380 AI autonomy right now. Second, some key technical 00:00:46.380 --> 00:00:48.179 highlights, stuff like major infrastructure fixes, 00:00:48.420 --> 00:00:49.899 public performance checks, that kind of thing. 00:00:50.020 --> 00:00:51.979 And finally, we'll take a really detailed look 00:00:51.979 --> 00:00:54.780 at Delphi 2M. That's an AI forecasting serious 00:00:54.780 --> 00:00:58.280 disease risk, maybe 20 years out. Okay, let's 00:00:58.280 --> 00:01:00.100 unpack that first part, starting with software 00:01:00.100 --> 00:01:02.899 development. The central idea from the sources 00:01:02.899 --> 00:01:06.799 seems pretty clear. Coding, you know, writing, 00:01:07.019 --> 00:01:09.219 debugging, implementing functional programs. 00:01:09.500 --> 00:01:12.659 That's the real canary for AGI. And this revolution, 00:01:12.840 --> 00:01:15.659 it's been so gradual. Maybe people kind of missed 00:01:15.659 --> 00:01:17.739 the seismic shift happening. Well, what's really 00:01:17.739 --> 00:01:19.920 fascinating is how these capabilities got added, 00:01:20.019 --> 00:01:22.620 like in layers, often without even changing the 00:01:22.620 --> 00:01:24.920 user experience much. So the change felt subtle, 00:01:25.000 --> 00:01:27.920 you know. Back in, what, 2021, we got... GitHub 00:01:27.920 --> 00:01:30.579 Copilot. That was basically smart code completion. 00:01:30.799 --> 00:01:33.180 Right. Simple tab completion. And then by 2022, 00:01:33.560 --> 00:01:35.840 things like ChatGPT were good enough to write 00:01:35.840 --> 00:01:39.239 like short standalone scripts. Exactly. Fast 00:01:39.239 --> 00:01:42.060 forward to now, 2025, and you've got tools like 00:01:42.060 --> 00:01:45.540 Cloud Code, Codex Manalign Interface, the CLI 00:01:45.540 --> 00:01:47.439 actually building pretty complex mini projects 00:01:47.439 --> 00:01:50.280 and fixing bugs, submitting those code changes, 00:01:50.340 --> 00:01:53.299 what developers call pull requests or PRs autonomously. 00:01:53.629 --> 00:01:55.269 Okay, when we talk about that leap, we have to 00:01:55.269 --> 00:01:57.549 define autonomous agents because this isn't just 00:01:57.549 --> 00:02:00.109 glorified autocomplete anymore, is it? Oh, not 00:02:00.109 --> 00:02:03.030 at all. We define autonomous agents as AI that 00:02:03.030 --> 00:02:06.170 can, like, understand a task, plan out multiple 00:02:06.170 --> 00:02:08.490 steps to solve it, run the code it needs, make 00:02:08.490 --> 00:02:10.849 decisions if something goes wrong, and actually 00:02:10.849 --> 00:02:13.330 complete the objective without constant human 00:02:13.330 --> 00:02:16.129 hand -holding. It's a whole loop operating on 00:02:16.129 --> 00:02:19.240 its own, like stacking Lego blocks of data. And 00:02:19.240 --> 00:02:21.960 there's actual data backing this up. These projects 00:02:21.960 --> 00:02:24.939 tracking agents in the wild, they show the Codex 00:02:24.939 --> 00:02:27.580 web agent has merged over a million pull requests 00:02:27.580 --> 00:02:31.020 already. One million. Yeah. Yeah. And those PRs, 00:02:31.020 --> 00:02:32.680 they're getting merged at an impressive rate, 00:02:32.759 --> 00:02:35.360 like 80 plus percent. That's for real production 00:02:35.360 --> 00:02:37.979 code changes, even from agents that are basically, 00:02:38.120 --> 00:02:40.340 you know, first time users on a code base. Yeah. 00:02:40.379 --> 00:02:42.620 Plus, we've seen huge adoption like Claude Code. 00:02:42.719 --> 00:02:45.479 It has something like 20 times more NPM downloads 00:02:45.479 --> 00:02:48.080 as the node package manager developers use than 00:02:48.080 --> 00:02:51.000 the code. Codex CLI. Honestly, I still wrestle 00:02:51.000 --> 00:02:53.439 with prompt drift myself sometimes, you know, 00:02:53.439 --> 00:02:55.080 when I'm just trying to fix simple bugs in my 00:02:55.080 --> 00:02:57.300 own little weekend project. So it's genuinely 00:02:57.300 --> 00:02:59.379 humbling to see these production systems hitting 00:02:59.379 --> 00:03:03.180 that consistent 80 % merge rate. Wow. So if coding 00:03:03.180 --> 00:03:06.379 really is the AGI canary, that consistent 80 00:03:06.379 --> 00:03:09.080 % merge rate feels like a pretty loud chirp. 00:03:09.159 --> 00:03:11.479 What does that percentage really tell us about 00:03:11.479 --> 00:03:13.340 where software autonomy is right now? I think 00:03:13.340 --> 00:03:15.979 it tells us the AI has achieved a level of reliability. 00:03:17.180 --> 00:03:19.240 We're definitely past simple suggestions here. 00:03:19.379 --> 00:03:22.699 These agents are delivering tangible, trusted, 00:03:23.120 --> 00:03:25.340 ready -to -use output. Okay, let's shift from 00:03:25.340 --> 00:03:27.319 that underlying code autonomy to some of the 00:03:27.319 --> 00:03:30.020 headlines. Model performance, public scrutiny. 00:03:30.659 --> 00:03:33.139 It often feels like the biggest perceived problems 00:03:33.139 --> 00:03:35.780 in AI turn out to be, well, pretty mundane infrastructure 00:03:35.780 --> 00:03:38.759 stuff. That is spot on. Like if you felt Claude 00:03:38.759 --> 00:03:41.879 seemed a bit nerfed recently, you know, its capability 00:03:41.879 --> 00:03:44.259 felt kind of reduced. It wasn't some secret downgrade. 00:03:44.580 --> 00:03:46.360 Anthropic actually put out a postmortem. They 00:03:46.360 --> 00:03:47.919 explained the perceived change was just down 00:03:47.919 --> 00:03:50.439 to three overlapping infrastructure bugs. Simple 00:03:50.439 --> 00:03:52.319 as that. And they're all fixed now. So, yeah, 00:03:52.360 --> 00:03:54.460 model stability often comes down to pretty boring 00:03:54.460 --> 00:03:56.780 infrastructure work, not some fundamental drop 00:03:56.780 --> 00:04:00.020 in capability. Right. And speaking of control, 00:04:00.240 --> 00:04:03.539 there is also that new feature for GPT -5. Users 00:04:03.539 --> 00:04:06.719 can now toggle its thinking time. It's web only 00:04:06.719 --> 00:04:09.840 for now, but it lets you choose faster answers 00:04:09.840 --> 00:04:12.500 or maybe smarter, more deliberate quality. Yeah, 00:04:12.560 --> 00:04:15.180 that's potentially huge for efficiency, letting 00:04:15.180 --> 00:04:18.120 the user decide the tradeoff. And speaking of 00:04:18.120 --> 00:04:20.259 public perception, Meta had that high profile 00:04:20.259 --> 00:04:23.100 moment recently. During their public demo, trying 00:04:23.100 --> 00:04:25.660 to show off live tasks, the system kind of hung 00:04:25.660 --> 00:04:28.730 for about a minute. Left the audience a bit unimpressed, 00:04:28.730 --> 00:04:30.810 apparently. But look, that was likely just real 00:04:30.810 --> 00:04:33.129 -time latency or maybe an API connection issue 00:04:33.129 --> 00:04:35.569 during data fetching. That's a really common 00:04:35.569 --> 00:04:37.310 bottleneck. Doesn't necessarily mean the model 00:04:37.310 --> 00:04:40.050 was hallucinating or broken. Still, takes guts 00:04:40.050 --> 00:04:42.370 to demo that stuff live. Definitely. And on a 00:04:42.370 --> 00:04:44.930 more creative note, there's this new tool, NanoBanana, 00:04:45.029 --> 00:04:48.269 allows pretty sophisticated photo merging. The 00:04:48.269 --> 00:04:50.269 interesting part is it handles images with more 00:04:50.269 --> 00:04:52.850 than two people and lets you control the exact 00:04:52.850 --> 00:04:56.029 aesthetic, the pose, for everyone involved simultaneously. 00:04:56.560 --> 00:04:59.959 So thinking about that GPT -5 toggle, why is 00:04:59.959 --> 00:05:02.620 giving users control over thinking speed actually 00:05:02.620 --> 00:05:05.000 such a useful option? Well, it really lets you, 00:05:05.180 --> 00:05:08.620 the user, consciously prioritize. Speed for simple 00:05:08.620 --> 00:05:11.480 stuff or deep analytical quality for complex 00:05:11.480 --> 00:05:15.560 tasks. Tailoring it. Okay, let's pivot now to 00:05:15.560 --> 00:05:17.779 the strategic side. Financial moves, where the 00:05:17.779 --> 00:05:21.019 money's flowing, and why so many projects seem 00:05:21.019 --> 00:05:22.920 to fail. Right, we're seeing some serious investments 00:05:22.920 --> 00:05:25.000 still. Databricks, for instance, just raised 00:05:25.000 --> 00:05:28.180 a billion dollars. One billion. And they launched 00:05:28.180 --> 00:05:31.100 an AI accelerator program. They're giving early 00:05:31.100 --> 00:05:34.560 stage startups like $50 ,000 in platform credits, 00:05:34.740 --> 00:05:36.800 helping them scale compute quickly. That sounds 00:05:36.800 --> 00:05:39.019 fantastic. Pure opportunity. But then there's 00:05:39.019 --> 00:05:41.120 this other data point that kind of balances that 00:05:41.120 --> 00:05:44.399 optimism. Research showing that, what, 95 % of 00:05:44.399 --> 00:05:46.660 AI automation projects actually fail. That seems 00:05:46.660 --> 00:05:49.300 incredibly high. And the successful ones, they 00:05:49.300 --> 00:05:51.279 apparently rely on a process -first method to 00:05:51.279 --> 00:05:53.779 get a positive ROI. Whoa, just imagine Databricks 00:05:53.779 --> 00:05:56.240 scaling that accelerator funding globally. That 00:05:56.240 --> 00:05:58.240 could cause a massive, almost immediate shift 00:05:58.240 --> 00:05:59.939 in compute allocation for startups everywhere. 00:06:00.199 --> 00:06:02.699 Yeah. But yeah, that 95 % failure rate is sobering. 00:06:02.699 --> 00:06:04.699 It really underscores the need for something 00:06:04.699 --> 00:06:07.399 called context engineering. Okay, let's dig into 00:06:07.399 --> 00:06:10.540 that. For listeners who are informed but maybe 00:06:10.540 --> 00:06:13.860 don't use that term daily, what exactly is context 00:06:13.860 --> 00:06:16.399 engineering and why is it apparently the key 00:06:16.399 --> 00:06:20.360 to avoiding that 95 % failure rate? Context engineering 00:06:20.360 --> 00:06:22.759 is basically about connecting the AI directly 00:06:22.759 --> 00:06:25.420 to your stuff, your company's internal data, 00:06:25.620 --> 00:06:28.060 your live workflows, your proprietary databases. 00:06:28.259 --> 00:06:30.899 You're giving the large language model the specific 00:06:30.899 --> 00:06:34.079 relevant context it needs to do its job well 00:06:34.079 --> 00:06:36.500 for you. Not just feeding it generic web data. 00:06:36.639 --> 00:06:38.620 That's how you get consistently high quality 00:06:38.620 --> 00:06:41.040 outputs that actually, you know, matter to the 00:06:41.040 --> 00:06:43.379 business process. So based on the research then, 00:06:43.500 --> 00:06:45.779 what's the core strategic reason the successful 00:06:45.779 --> 00:06:48.660 projects manage to avoid that huge 95 % failure 00:06:48.660 --> 00:06:50.620 rate? It seems they really focus on defining 00:06:50.620 --> 00:06:52.779 the underlying business process before they even 00:06:52.779 --> 00:06:55.000 think about implementing the AI solution. Process 00:06:55.000 --> 00:06:57.879 first. All right, let's turn now to what feels 00:06:57.879 --> 00:06:59.579 like one of the biggest scientific breakthroughs 00:06:59.579 --> 00:07:02.560 mentioned in the sources. Delphi2M, this new 00:07:02.560 --> 00:07:05.240 AI system from Europe. researchers, it seems 00:07:05.240 --> 00:07:08.180 genuinely set to redefine proactive health care. 00:07:08.279 --> 00:07:11.639 Oh, its capability is absolutely stunning. Delphi2M 00:07:11.639 --> 00:07:15.220 forecasts the risk for 12 ,258 distinct diseases. 00:07:15.459 --> 00:07:18.579 Think diabetes, neurological disorders, heart 00:07:18.579 --> 00:07:21.439 conditions, the whole gamut, up to 20 years into 00:07:21.439 --> 00:07:24.019 the future. And it does this just using a patient's 00:07:24.019 --> 00:07:26.300 standard electronic medical records. Nothing 00:07:26.300 --> 00:07:28.699 more exotic than that. The training data must 00:07:28.699 --> 00:07:30.779 have been immense. It was initially trained on, 00:07:30.839 --> 00:07:34.100 what, data from over 400 ,000 UK patients, including 00:07:34.100 --> 00:07:36.639 everything from doctor visits, hospital records, 00:07:36.879 --> 00:07:39.480 even known lifestyle choices factored in. Exactly. 00:07:39.540 --> 00:07:41.399 And critically, this is super important to make 00:07:41.399 --> 00:07:43.459 sure it wasn't just good for the UK population. 00:07:43.500 --> 00:07:46.939 They validated it. They tested it against 1 .9 00:07:46.939 --> 00:07:49.699 million entirely separate Danish patient records. 00:07:49.879 --> 00:07:52.379 That validation step is crucial. It proves the 00:07:52.379 --> 00:07:55.040 model has generalizability. It proves it's not 00:07:55.040 --> 00:07:57.000 just hallucinating predictions based on some 00:07:57.000 --> 00:07:59.439 bias in the original UK data set. The potential 00:07:59.439 --> 00:08:01.699 implications for actual patient care seem profound 00:08:01.699 --> 00:08:04.360 here. Delphi2M could help doctors shift from 00:08:04.360 --> 00:08:06.439 just reacting to symptoms that already exist 00:08:06.439 --> 00:08:09.139 to actively anticipating future health risks 00:08:09.139 --> 00:08:12.079 years in advance. That could fundamentally change 00:08:12.079 --> 00:08:14.079 medicine towards really personalized prevention. 00:08:14.160 --> 00:08:16.360 Yeah, but it's really important to stress the 00:08:16.360 --> 00:08:18.600 caveat here. Human doctors are still absolutely 00:08:18.600 --> 00:08:21.180 necessary. They need to interpret the AI's predictions. 00:08:21.459 --> 00:08:24.160 The system analyzes risk. The physician provides 00:08:24.160 --> 00:08:26.000 the judgment, the empathy, the actual treatment 00:08:26.000 --> 00:08:28.740 plan. It's designed to augment the doctor, definitely 00:08:28.740 --> 00:08:31.459 not replace them. So going beyond the headline 00:08:31.459 --> 00:08:34.299 number of diseases or years, what's maybe the 00:08:34.299 --> 00:08:37.840 biggest operational shift Delphi2M could bring 00:08:37.840 --> 00:08:40.100 to something like a standard annual checkup? 00:08:40.539 --> 00:08:42.539 Well, it fundamentally redefines that checkup, 00:08:42.559 --> 00:08:44.919 doesn't it? From just a snapshot evaluation of 00:08:44.919 --> 00:08:48.200 current status to proactive long -term risk mapping 00:08:48.200 --> 00:08:50.960 for the individual. Okay, so just to quickly 00:08:50.960 --> 00:08:52.700 synthesize everything we've covered for you today. 00:08:52.820 --> 00:08:56.019 First, AI encoding. It's quietly crossing some 00:08:56.019 --> 00:08:58.700 really crucial autonomy thresholds. That 80 plus 00:08:58.700 --> 00:09:01.440 percent PR merge rate is a key example. Second, 00:09:01.600 --> 00:09:04.179 model stability issues, often tied to boring 00:09:04.179 --> 00:09:07.080 infrastructure fixes, not fundamental nerfing. 00:09:07.159 --> 00:09:09.899 And users are getting more control, like toggling 00:09:09.899 --> 00:09:12.419 speed versus intelligence. And finally, predictive 00:09:12.419 --> 00:09:15.419 AI like Delphi 2M is really starting to redefine 00:09:15.419 --> 00:09:17.139 human health and the whole concept of prevention. 00:09:17.440 --> 00:09:19.879 So what does this all really mean for you? I 00:09:19.879 --> 00:09:21.940 think the takeaway is that the most impactful 00:09:21.940 --> 00:09:24.919 AI progress is often happening silently. It's 00:09:24.919 --> 00:09:27.379 in the background, deep inside complex systems, 00:09:27.480 --> 00:09:29.980 the infrastructure, the code bases, the hospitals. 00:09:30.360 --> 00:09:32.779 It's not just in those viral chat apps that grab 00:09:32.779 --> 00:09:35.080 all the headlines. You really have to pay attention 00:09:35.080 --> 00:09:37.419 to the infrastructure, the less flashy stuff. 00:09:38.019 --> 00:09:39.840 Thank you so much for joining us for this team 00:09:39.840 --> 00:09:41.740 dive today. We really hope you continue learning 00:09:41.740 --> 00:09:43.360 about these advancements. They're fundamentally 00:09:43.360 --> 00:09:46.120 reshaping tech and health right now. And maybe 00:09:46.120 --> 00:09:48.539 here's a final thought for you to chew on. Consider 00:09:48.539 --> 00:09:50.759 that China is apparently already teaching children 00:09:50.759 --> 00:09:54.200 AI principles from age six, and that the DeepSeek 00:09:54.200 --> 00:09:56.980 foundational model, a pretty capable one, reportedly 00:09:56.980 --> 00:10:00.909 cost only $294 ,000 to train. Not millions, thousands. 00:10:01.629 --> 00:10:04.169 So how rapidly do you think the center of gravity 00:10:04.169 --> 00:10:06.330 for foundational AI breakthroughs might shift 00:10:06.330 --> 00:10:09.110 in, say, the next five years? Something to think 00:10:09.110 --> 00:10:10.009 about. Until next time.