DiscoverSoftware Defined Interviews
Software Defined Interviews
Claim Ownership

Software Defined Interviews

Author: Software Defined Talk

Subscribed: 90Played: 955
Share

Description

Deep discussions about technology, enterprise IT, and the like
135 Episodes
Reverse
D Niu spent almost five years in security research at Datadog, then left this summer to co-found Ontology Labs, which studies how AI agents misbehave doing ordinary work. She explains why she calls this summer's Hugging Face breakout a misalignment event rather than a hack, and how she built a fake eight-person company convincing enough that agents don't know they're in a test. Whitney and Coté ask why bigger models write more code than a simple task needs, and about what agents leave behind: API routes and database tables that never went through a deploy. Also: one model is a good intern, another a disgruntled senior engineer. You can watch the video of the recording as well, if you're into that kind of thing. Mentions: Ontology Labs, and its first set of research. misaligned.studio, D's earlier research on predicting agent behavior from tool calls. D on LinkedIn AI Alignment Forum, where D says the up-to-date research gets posted. Exa, the search engine that markets itself to agents. Business!, Louie Zong's album of corporate music, now a two-episode favorite. Special Guest: D Niu.Special Guest: D Niu.
Ben Savage is CEO and founder of Veritas Automata, a consultancy doing AI, blockchain and IoT, and CTO at Apex, where one of his teams built the Little Caesars Pizza Portal. He explains why those pickup cubbies mostly don't need heat (people just feel better when they're warm), what a logistics check call is, and why a routing algorithm is really a search algorithm. Whitney and Coté ask how a consultancy keeps its bench busy, and when code is too safety-critical to let AI write it: a robot can't "-ish" a command. Also: model routers, decision models, and turning SEC filings into corporate theme songs. You can watch the video version of this episode as well, if you prefer that kind of thing. Mentions: Veritas Automata Apex Order Pickup, the food pickup lockers. Ben on LinkedIn Truth in Automation: The Deep Dive, the Veritas Automata podcast. The Domino's Pizza Tracker Theory of Everything, from Planet Money. Business!, Louie Zong's album of corporate music. Special Guest: Ben Savage.Special Guest: Ben Savage.
Rick Clark is Global Head of Cloud Advisory at UST, and co-founded OpenStack. His argument is that enterprise architecture is nearly extinct, killed off by fifteen years of shifting every decision left, and that this is how one insurance company he worked with ended up running over 50 secrets managers in production. Whitney and Coté push on whether platform engineering was meant to solve exactly that, and on what a modern enterprise architect would do instead - write the rules the platform enforces. The back half is agents: why nobody is ready to run them in production, and the difference between a rule an agent cannot break and one it must not break. You can watch the video version of this episode as well, if you prefer that kind of thing. Mentions: Rick on LinkedIn, which is where he sends people. The Construct, his newsletter on agentic workloads in production. From tokenmaxxing to tokenomics: you're still doing it wrong, his first op-ed for The Stack. Special Guest: Rick Clark.Special Guest: Rick Clark.
Viktor Farcic is a developer advocate at Upbound, runs the DevOps & AI Toolkit YouTube channel, and was last on the show back in episode 89. Whitney and Coté start with the animated bedtime stories he has started making - tech history told as fairy tales with a nightmare at the end - which he began after catching sight of himself in a mirror and deciding people should not have to watch that. From there it is what his working day looks like now, with 20 to 50 agents running at once, a coder and a tester and a reviewer who are not allowed to do each other's jobs and argue until they agree, and a private YouTube channel where the agents post recorded tests for him to review on his phone in the morning. The last stretch is money: why he thinks AI is cheap next to what a person costs, why he expects a bloodbath as industries fail to adapt, and a walk through Docker, Chef, Puppet, Ansible and HashiCorp on whether an open source company can ever turn a profit on its own. Check it out in YouTube, if you prefer that kind of thing. Mentions: DevOps & AI Toolkit on YouTube, where the bedtime stories live. devopstoolkit.live - the blog and video archive. devopstoolkit.ai - the product. Viktor on GitHub. DevOps Paradox, his other podcast. Episode 89, his first time on the show. Special Guest: Viktor Farcic. Show notes written by AI with a human's lazy-look.Special Guest: Viktor Farcic.
Lin Sun is Head of Open Source at Solo.io, a recent KubeCon + CloudNativeCon program co-chair, and an emeritus member of the CNCF Technical Oversight Committee. Whitney and Coté start with the keynote she gave in Amsterdam, where she flew a drone over 13,000 people so an AI model on her laptop could score how engaged the audience was - it broke twice on stage, and Whitney, who was in the room, remembers nothing else from that day. Lin’s account of why it worked anyway is a checklist: a wow factor, something the audience does every day so they can relate to it, real engagement, and a planted surprise. From there they get into the part of co-chairing nobody sees, where three thousand submissions become about two hundred talks, how the AI-written ones get spotted, and why the CNCF cannot simply run a detector over the pile. The rest is open source as a job - twenty years of it, Istio at IBM through to Solo, and a straight answer to the question her own reports keep asking her, which is whether working on open source makes you less strategic to the company paying you. You can watch the video version of this episode as well, if you prefer that kind of thing. Mentions: Lin Sun on LinkedIn, where she says she is most active. Solo.io - the book “AI Agents on Kubernetes” is a free download from the homepage, and is being refreshed for Salt Lake City. The Agentic AI Foundation and its five projects: Model Context Protocol, goose, AGENTS.md, agentgateway, and the agent-to-agent protocol. Agent Substrate, the Google project that rethinks how agents get scheduled. Special Guest: Lin Sun. Header photo from the CNCF.Special Guest: Lin Sun.
loading
Comments