WEBVTT 00:00.320 --> 00:04.040 Today I resigned from GitHub after spending almost four years there and 00:04.540 --> 00:08.400 growing to senior engineer. I left to help improve AI safety because 00:08.900 --> 00:12.320 there is a dangerous trend in how AI coding agents are being deployed right 00:12.820 --> 00:16.440 now and this is going to affect you personally. I'm going 00:16.940 --> 00:20.920 to break down what I observed, why it made me walk away, and what 00:21.420 --> 00:24.760 I'm doing about this problem. So first of all, I didn't 00:25.260 --> 00:28.620 leave GitHub because I hated my job. It was genuinely the best job 00:29.120 --> 00:32.860 I've had so far. I built the first version of newer respectful AI 00:33.360 --> 00:36.460 support systems with my team, helped other engineers across the org 00:36.960 --> 00:40.660 figure out how to build mature LLM solutions, and those experiences 00:41.160 --> 00:45.180 shaped who I am as an engineer today. I've also experienced the birth 00:45.680 --> 00:49.260 of agentic coding from more traditional AI autocomplete to newer 00:49.760 --> 00:52.980 agentic systems like Claude Code, GitHub, Copilot and others. 00:53.840 --> 00:57.640 Now we as programmers and non programmers experience the benefits of agent 00:58.140 --> 01:01.600 decoding daily and it has been a massive productivity boost. 01:02.080 --> 01:06.120 But over the past year I started noticing an extremely dangerous 01:06.620 --> 01:09.920 trend across the wider industry that I just couldn't ignore 01:10.420 --> 01:13.560 anymore. Companies are not just using AI for code 01:14.060 --> 01:17.400 generation, which I'm a huge fan of, just like many of you. They are now 01:17.900 --> 01:21.620 replacing entire parts of the software development lifecycle with AI 01:22.120 --> 01:25.660 agents. Review testing, deployment decisions 01:26.160 --> 01:29.620 and even architectural choices are being handed 01:30.120 --> 01:33.660 to agents with less and less human oversight. And I'm 01:34.160 --> 01:37.660 not only talking about small startups here that can afford this 01:38.160 --> 01:42.060 risk. The industry pressure has become so high that 01:42.560 --> 01:45.820 companies with millions of users are shipping code that a 01:46.320 --> 01:48.660 human barely touches before it goes into production. 01:49.710 --> 01:53.110 Now, before you misinterpret me, I actually just want to say that at GitHub 01:53.610 --> 01:56.750 I found there to still be a lot of respect for good high quality software 01:56.990 --> 02:00.710 every single day. It's just that the industry trend 02:01.210 --> 02:04.350 overall is causing software quality to drop quickly. 02:04.590 --> 02:08.430 There are already projects and platforms trying to test the limits of 02:08.930 --> 02:11.790 how far we can push coding agents to automate everything. 02:12.350 --> 02:15.910 And the direction here is clear. Some organizations are starting 02:16.410 --> 02:19.950 to point at pull requests as an annoying bottleneck that 02:20.450 --> 02:23.950 slows down their AI generated output. Now think about that for just 02:24.450 --> 02:28.350 a second. The one remaining checkpoint where a human looks 02:28.850 --> 02:32.510 at the code before it reaches your users. And the industry is calling that a 02:33.010 --> 02:36.670 bottleneck. The ultimate step here seems to be to displace 02:37.170 --> 02:40.830 human review altogether. So why does 02:41.330 --> 02:45.160 this problem make me quit my create job at GitHub? Well, 02:45.310 --> 02:49.030 well, it has been one of the best jobs I've had and I will easily 02:49.530 --> 02:53.150 admit this AI that I've used at my job very often writes 02:53.230 --> 02:56.790 better code than me, and it outputs code faster 02:57.290 --> 03:01.349 than me. One of the ways this has excited executives in the tech industry is 03:01.849 --> 03:05.990 that it promises that every developer can be a 10x or even a 100x 03:06.490 --> 03:09.150 developer, because 10x is not enough anymore. Now, 03:09.390 --> 03:12.810 before we get too enthusiastic, think about what that actually means for 03:13.310 --> 03:16.970 a moment. If one developer can generate 100 03:17.470 --> 03:21.570 times the code, they can no longer meaningfully review what 03:22.070 --> 03:25.610 was produced. And I'm noticing this every single day as I 03:26.110 --> 03:29.650 work as an engineer. Instead of the industry seeing this as a problem, 03:29.970 --> 03:33.970 the collective response has been that agents should just take care of reviewing 03:34.470 --> 03:38.090 code as well. The critical issue with this is, is that AI 03:38.590 --> 03:41.930 writes better code often, as I just said, but it does 03:42.430 --> 03:46.130 not always write better code. It will continue to 03:46.630 --> 03:50.010 hallucinate, it will continue to make mistakes, and when tasks are 03:50.510 --> 03:53.850 long and complex enough, it will create incomplete solutions. 03:54.570 --> 03:58.090 But it does always write more code than me. Way, 03:58.410 --> 04:01.770 way more code. In a normal workday in the past, 04:01.930 --> 04:04.650 I might have had days where I wrote a few hundred lines of code. 04:05.090 --> 04:08.530 And now it's easy to write thousands of lines in a single session. 04:09.330 --> 04:13.490 So what happens when you have an imperfect system generating 04:13.990 --> 04:16.370 enormous amounts of almost perfect code? 04:16.930 --> 04:20.809 You get a statistical guarantee that there will be bugs 04:21.309 --> 04:25.130 in the system that you as a human, do not have the capacity to 04:25.630 --> 04:28.850 check for anymore. This is not a gut feeling, 04:29.010 --> 04:32.410 it is a mathematical fact. And the industry's 04:32.910 --> 04:36.810 current response to this problem is to just throw more agents at 04:37.310 --> 04:41.250 it. Oh, let's just evaluate agent output with more agents with just 04:41.410 --> 04:44.530 slightly different system prompts. Let's write automated tests 04:45.030 --> 04:48.450 with agents. Let's let agents test software end to end, stack more 04:48.950 --> 04:52.490 AI calls with slightly different prompts on top of each other, and just 04:52.990 --> 04:56.330 hope for the best. Now, in my brutal opinion, this is the 04:56.830 --> 05:00.390 software industry's equivalent of, of just playing dollhouse and pretending 05:00.890 --> 05:04.350 that there is a solution to this problem. Because we are currently trusting 05:04.590 --> 05:08.190 the same type of technology that's creating the mistakes to 05:08.690 --> 05:12.670 catch its own mistakes. Imperfect systems stacked 05:13.170 --> 05:16.270 on imperfect systems do not converge to perfection. 05:16.830 --> 05:19.910 The industry right now is actively choosing to 05:20.410 --> 05:24.280 prioritize raw code output over everything else. The amount of 05:24.780 --> 05:28.000 code generated can no longer be stopped by humans. And this is 05:28.500 --> 05:32.600 not because of some fictional rogue AI like you might read in some fiction books. 05:33.100 --> 05:36.920 No, this is just because the industry is letting this happen on purpose as 05:37.420 --> 05:40.920 part of a strategic decision. Now, unlike many statements made 05:41.420 --> 05:45.480 in the AI space, which are often based on hype or ulterior motives, 05:45.720 --> 05:49.040 I am confident in stating the following the future 05:49.540 --> 05:53.470 deployment of the fully autonomous AI software development lifecycle will 05:53.970 --> 05:57.230 lead to an absolute increase in broken software. 05:58.270 --> 06:01.230 Now I want to be clear about something because you might be thinking Xen, 06:01.730 --> 06:05.230 do you think that we should be stopping AI coding altogether? No, 06:05.730 --> 06:08.870 I'm not saying that we need to stop the point coding agents because this is 06:09.370 --> 06:12.790 highly unrealistic and I'm quite progressive. And honestly, I never want 06:13.290 --> 06:16.590 to go back to manually writing all my code myself anymore. And to be 06:17.090 --> 06:20.510 honest, some drop in quality is acceptable because some bugs just don't 06:21.010 --> 06:24.190 matter that much. Even in professional settings, not everything needs to be perfect. 06:24.350 --> 06:28.130 No, nobody is going to get hurt if, say, a support portal button 06:28.630 --> 06:32.130 breaks in an edge case condition. In general, being able to ship way 06:32.630 --> 06:35.930 more and break a few small things along the way is a net win 06:36.430 --> 06:40.490 for the majority of software out there. And I also understand that Vibe coding 06:40.990 --> 06:44.290 is awesome. It's very cool to one shot your favorite childhood game and that as 06:44.790 --> 06:48.170 a non programmer you can get into software easier nowadays. It's very nice. 06:48.730 --> 06:52.250 The problem is that the push to use these agents is so pervasive 06:52.750 --> 06:56.230 across every industry that bugs will inevitably appear. And in 06:56.730 --> 07:00.110 software where we cannot afford it, in healthcare systems, 07:00.190 --> 07:03.470 on financial platforms and infrastructure that millions 07:03.970 --> 07:07.990 of people depend on every day. And without fundamentally 07:08.490 --> 07:12.270 new forms of monitoring, AI coding agents will cause 07:12.770 --> 07:15.390 more harm than good in these kinds of critical systems. 07:16.510 --> 07:19.830 So it's a big problem, but what am I doing about it? Because it's 07:20.330 --> 07:24.200 easy to complain from the sidelines, right? Well, that is what made me resign 07:24.700 --> 07:28.360 from GitHub. It is very easy to just complain about problems that AI 07:28.860 --> 07:32.000 is causing, but if an opportunity presents itself to actually contribute 07:32.500 --> 07:35.960 to a possible solution, it is only right to try it, even if it 07:36.460 --> 07:39.840 can be difficult. I happen to find an opportunity at a research 07:40.340 --> 07:43.440 lab to work on exactly this problem. A new role 07:43.520 --> 07:47.600 building monitoring systems specifically designed for AI coding agents. 07:48.580 --> 07:51.940 Now next week I'll share the full details on this new role 07:52.100 --> 07:54.820 on my LinkedIn. So if you want to be the first to know, 07:54.980 --> 07:57.700 connect with me using the link in the description down below. 07:58.500 --> 08:02.380 But I'm also sharing this story because four years ago I was starting 08:02.880 --> 08:05.380 out as a junior and now I get to work on problems that I think 08:05.880 --> 08:09.500 genuinely matter for the wider industry. And this shows that it is 08:10.000 --> 08:13.300 still possible to get a high paid career while making 08:13.800 --> 08:18.000 a positive change or at least trying to do so. And also genuinely loving 08:18.500 --> 08:22.120 your job. And this is also why I push and help others towards their own 08:22.620 --> 08:25.920 ideal AI career as a way to give back. I'll be happy 08:26.420 --> 08:29.920 to answer any questions you have about the path that got me here. Make sure 08:30.420 --> 08:33.280 to leave a comment down below and connect with me on LinkedIn.