How to turn your 1000x engineer into a 10x everyone | LinkedIn’s Karthik Ramgopal
Your harnesses can't keep up, the economics of local AI are changing, and now it's the Eternal September (but for code)
This week, Andrew sits down with LinkedIn Distinguished Engineer Karthik Ramgopal to explore the reality of deploying agentic platforms across a massive organization. Karthik unpacks the mechanics of AI memory, spanning procedural and episodic structures, and explains how to build durable engineering primitives that actually last. Finally, the two discuss the enduring importance of system fundamentals and why LinkedIn is restructuring its internship program into AI-native pods to foster a new culture of two-way mentorship.
1. The economics of local models
If you are relying heavily on top tier model providers, you have probably noticed your inference bills starting to climb. As API consumption costs slowly escalate, many engineering leaders are moving the opposite direction by fine tuning open source models on their own local infrastructure. This shift opens the door for some creative cost saving strategies, such as combining these local models with lower cost frontier options to create compounding efficiency gains. Making strategic bets on when to use local versus frontier models is going to be a crucial lever for teams moving forward.
Read: Outsourcing plus LocalAI will soon become more economical vs Frontier labs
2. The reality of spec-driven development
Spec-driven development is a fantastic practice for shifting execution left, but it is not a silver bullet for autonomous coding. Handing a massive requirements document to an agent and walking away is basically putting your codebase in YOLO mode. As requirements explode and unknown unknowns inevitably arise, intensive specifications can quickly break down. You still need mechanical ways to verify that the generated code actually matches the spec and passes rigorous testing. This article reminds us about these hidden realities.
Read: A Blast from the Past: SDD and the Illusion of Known Scope
3. One author argues we’re entering the Eternal September but for code
George Hotz argues that AI agents are fundamentally unable to program effectively, describing them as sophisticated statistical models that produce subtle but persistent bugs. The title of his piece references the classic Usenet “Eternal September,” comparing the early influx of AOL users to today’s wave of confident but inexperienced developers flooding workflows with unverified AI outputs. It is a sharp reminder that applying AI without underlying domain expertise is incredibly risky.
Read: The Eternal Sloptember
4. Rethinking data centers
We rarely stop to think about the physical infrastructure powering our agentic workflows, but AI training has fundamentally changed data center networking requirements. Unlike traditional cloud services that distribute requests across multiple locations, AI training often requires everything to happen within a single data center to prevent network bottlenecks, even requiring certain directional patterns of the physical machines to optimize compute. This piece explores the fascinating hardware realities behind the scenes, including new networking technologies from NVIDIA and what a GPU-free future might actually look like.
Read: AI Datacenters Were Built for GPUs. What Happens When You Remove the GPUs?
5. The first-ever Gartner Magic Quadrant
AI is changing how we build software, but how do you choose the right platform to measure its impact? To get the visibility you need into productivity, bottlenecks, and real ROI, you need a trusted evaluation method.
Gartner just released the first-ever Magic Quadrant for Developer Productivity Insight Platforms, naming LinearB a Leader. Download your complimentary report to understand why this category matters right now and why LinearB is recognized for our vision and workflow automation.
6. Do not let AI become a workflow crutch
When you encounter a broken or inefficient process, a human will eventually get frustrated and consolidate the workflow. An AI, on the other hand, will happily navigate the mess and burn all of your tokens in the process. Autonomous tools cannot be an excuse to tolerate bad systems. As I mentioned on the show, you need to go upstream of the problem and apply solid software engineering principles and fix your inputs rather than endlessly iterating on the outputs.
Read: “The AI Can Do It” Is Not an Excuse To Tolerate a Mess
Really enjoyed this session of Dev Interrupted with Karthik Ramgopal. Fascinating topics that were dissected, considered, and presented fantastically well. Lots of food for thought. Thanks.