Cloud Radio | A SaaS Podcast

The episode contains genuine non-obvious insights in the AI and security sections—particularly the GPU-speed vs. human-speed bottleneck for autonomous code review, and the non-deterministic security problem with LLM-embedded software—but the first half on VPE management largely covers well-trodden ground (IC-to-manager transition, Brooks'-Law-adjacent headcount warnings) with limited novel density.

“Philanthropic just released their latest model four six, I think it was last week, and in doing so, they also published a post about building a compiler, which is a fairly sophisticated piece of software entirely autonomously with a bunch of clot codes at a total cost of $20,000. And this is a fairly complex piece of software that was written with no human intervention.”

“moving at the speed of GPUs, but, but throttled by the speed of human, I mean the human has to reviewers, you're ultimately moving at the speed of human”

The framing of AI agents as coworkers with identities (named, on Slack, emailable) and the non-deterministic vs. deterministic security distinction are practically fresh angles; however, the VPE management advice (player-to-coach, headcount ramp analogies) and the 10x-engineer debate recycle widely circulated frameworks without meaningful reframing.

“If you treat that agent as as a bot, that's one thing. If you start treating it as a coworker, you're gonna build different intuition, different tools, and different ways of interacting with that AI than you would with a Slack bot.”

“these systems are non-deterministic...It can perform actions that you and I are looking at that code. You cannot envision”

Karim Fanous is a genuine practitioner—25 years in software, four-plus years as Head of Engineering at a real cybersecurity company (Strong DM), actively running human-plus-agent hybrid engineering teams—giving him real-world credibility, though he has not operated at hyper-scale and the episode confirms he is primarily known via Substack rather than landmark company-building.

“I am pleasantly head of engineering at a cybersecurity company called Strong dm. I've been here for a little bit over four years, and my background is entirely in engineering management and software development. I've, uh, spent 25 years so far.”

“we actually have some pockets of our team here using like, we've got canonical traditional software engineering people type on the keyboard and commit code, and we also have autonomous.”

A handful of concrete details elevate the episode—the $20K Anthropic compiler build, the named agent 'Jen,' Strong DM's open-sourced Software Factory, and specific tool names like Claude Code and Codex—but the majority of answers rely on analogies and abstraction rather than hard metrics, timelines, or outcome data from Strong DM's own AI experiments.

“Philanthropic just released their latest model four six, I think it was last week, and in doing so, they also published a post about building a compiler, which is a fairly sophisticated piece of software entirely autonomously with a bunch of clot codes at a total cost of $20,000.”

“I can pick up the phone and call Jen, which is one of our agents.”

The host shows genuine curiosity and lands a few productive follow-ups (the agent-naming drill-down, the headcount management question), but frequently accepts incomplete answers with affirmations like 'that's great' and 'fascinating,' explicitly apologises for asking 'impossible questions,' and never meaningfully challenges a claim or introduces productive friction.

“apologies for the impossible questions”

“That's great. Kind of after the interview, let's talk about the C-E-O-V-P-E relationship”