AI Digest, Oct 15: Code Review Bottleneck, Financial Agents, Small AI

0:02PM Lead ProThis is PM Lead Pro. Today's digest asks whether AI is speeding your team up or just moving the queue: code review, financial agents, formal math proofs and small AI.

0:15Welcome to this deep dive. You know, you're probably integrating AI into your team right now to, well, to move faster, right?

0:22Yeah, it sounds amazing on paper.

0:24I mean, it's like strapping a rocket engine to your car.

0:27Exactly.

0:28But what if that rocket is actually like driving you straight into a massive traffic jam?

0:33That is a very real risk we're seeing right now.

0:36So today we're pulling from a huge stack of sources. We've got the Agent Report, NBER, the World Bank, and a really fascinating brief on Chinese AI.

0:44AI. And we're going to explore the AI coding bottleneck, swarms of financial agents, a pure

0:49math breakthrough, and this concept of small AI. Yeah, and our whole mission today is just cutting

0:55through the hype. We want to figure out how to actually ship value with AI, not just generate

0:59noise. Right. So let's start with that traffic jam from the agent report and NBER. Yeah, so say your

1:06developers are using AI tools. The data shows lines of code are just jumping by a staggering 741%.

1:13Which sounds incredibly productive until you look at the output.

1:17Right, because actual shipped software is only up 20%.

1:20So where is all that extra code going?

1:24Well, it's just sitting in merge queues.

1:27The NBER data shows the time it takes to review and approve code has lengthened by 49%.

1:32Wow.

1:33Yeah, that adds almost three and a half extra days to a cycle.

1:36I mean, writing code was never the real constraint.

1:38Reading and verifying it is.

1:41Okay, wait, I have to push back here, though.

1:42If an AI is smart enough to generate the code why can we just prompt it to you know review its own work Mostly because of silent failures Okay If you let an AI review its own massive output without a secondary check Right it might miss its own blind spots

1:57Exactly.

1:58You risk shipping vulnerabilities at light speed.

2:01Despite all these tools, humans are still writing 77% of the code review comments.

2:05So we are the ultimate bottleneck.

2:07Yeah, we are.

2:08Man.

2:10PM Lead ProTakeaway for team leads: lines of code rose 741%, but shipped software only 20%. Review time grew 49%. Track shipped output and review capacity, not code volume.

2:23Okay, now project that bottleneck onto fully autonomous enterprise systems.

2:27According to Third Moment Research, they're running 10,000 agents concurrently for up to 48 hours in quantitative finance.

2:34Which is just a staggering volume of actions.

2:38Right. And you can't possibly have a human reviewing every single trade there.

2:42It's like, you know those terms of service pop-ups on websites?

2:47Oh yeah, where you just blindly click accept.

2:49Exactly. Eventually you get so overwhelmed by the volume you just click accept without reading.

2:54Well, hedge fund Two Sigma actually calls that exact phenomenon approval fatigue.

3:00Approval fatigue, huh?

3:01Yeah. Forcing a human to click approve on every single AI action doesn't just limit your scale, it creates massive risk.

3:09Because that human safety net basically just stops paying attention.

3:11Exactly. So the solution is building deterministic controls outside the model entirely.

3:18Oh, so instead of asking the AI to police itself, you build a traditional hard-coded filter.

3:23Right, an auto mode classifier. It automatically approves the routine, low-risk tasks and only flags the highly ambiguous edge cases for a human.

3:32So you need brakes, not to stop, but just so you can drive fast safely.

3:35Yeah perfectly put And if traditional software needs those hard brakes we seeing a fascinating divergence in fields that require absolute precision Like pure mathematics Right

3:46PM Lead ProFor owners deploying agents: approving every action creates approval fatigue. The discussion suggests deterministic controls outside the model, with a classifier that passes routine tasks and flags ambiguous cases to a human.

4:00Our brief on Chinese AI research notes that the Shanghai AI Lab used their Intern S2 model to help prove the 3D geese conjecture.

4:08Which is a major dynamical systems problem.

4:10And the crazy part is they didn't just ask a chatbot for the answer.

4:14No, they plugged the model into Peking University's Sheaf framework.

4:17Right. And it generated 550,000 lines of Lean code in just five days. And Lean isn't like Python or something.

4:25No. Lean is a formal verification language. It basically turns abstract math into software that absolutely must compile.

4:33So if there is a single logical flaw?

4:35The code simply fails to run. The system mathematically proves its own logic.

4:40By doing this, they completely eliminated the human review errors we see in standard software engineering.

4:45That is wild. But this level of automated reasoning isn't just for hedge funds and elite math labs.

4:51Not at all. It's about to hit the broader global economy.

4:55Right. Because research from CEPR shows human capital, skills and education, drives up to 70% of income differences between countries.

5:03And the World Bank's 2026 report argues AI is the ultimate bridge for that gap.

5:08But developing nations don't need these trillion parameter frontier models running in massive data centers.

5:13Exactly. They need small AI. Models adapted to basic phones via text and voice to scale expert knowledge instantly.

5:21It kind of like how developing nations skipped landlines entirely and went straight to mobile phones but for cognitive skills That a great analogy Which actually ties perfectly into our model radar today Look at Google Times FM and Tab and so on

5:55and so on.

5:56But for one three hundredth the cost of traditional methods.

5:59Oh wow.

6:00So you could run that affordably to send a farmer a hyperlocal forecast over an SMS text.

6:05Exactly.

6:06And finally for running complex tasks on local devices we have a whole new class of model

6:10called Cortex.

6:11Right because normally an AI's memory is just the prompt window which is like trying

6:15to memorize a textbook while you're holding it open.

6:18Yeah but Cortex uses parametric memory.

6:20It learns on the fly actually updating its own internal wiring, its weights.

6:25So it remembers context natively without needing massive computing power.

6:28Exactly. So for you listening today, the practical takeaway is this. Do not measure your team's AI success by volume.

6:35Right, not by lines of code.

6:37Measure the actual shipped output. And please, implement classifiers so you don't burn out your team with approval fatigue.

6:44Great advice. And I'll leave you with this to chew on.

6:46If small AI allows developing economies to leapfrog traditional education barriers,

6:52will the future center of global human capital shift entirely away from the West?

6:57It's a huge question.

6:59It really is.

7:00Maybe that AI rocket engine isn't just making us faster.

7:03Maybe it's taking us somewhere completely different.

7:06PM Lead ProMore at pmlead.pro: the transcript and sources are on the episode page. Question for you: where is the queue in your team right now, writing, reviewing or approving?