AI agents moved from demos into the stack this week — and, honestly, the receipts are finally getting interesting. This is Tech Podcast Podcast. Today we’ve got Sundar Pichai talking chips and China, the creator of Claude Code pushing back on actual job-loss charts, a Sequoia founder naming exactly which model goes to which task, and a cybersecurity platform saying an agent found a SQLite zero-day in hours. That SQLite claim is either the sharpest red-team result we’ve heard all week or the priciest marketing line. And the Serval model-routing split is the first time someone has actually named the seam instead of just waving at it. These three episodes are all making structural arguments, just from completely different angles. Let’s see if the specifics can carry the weight. Forward Future, with Matthew Berman:
I think so I think it's they're genuinely adding value in a way I think people will will use them. I think it's important we build it in a way that users feel a sense of control and agency and transparency when they use agents. I think that's important.
Sundar Pichai is on Forward Future with Matt Berman — yes, the YouTuber CEO — and he comes in framing agents as the new entry point to the internet, with developers as the group already doing the first real agentic workflows. He keeps using developers as the proof-of-concept population, which is fine, but the ReliaQuest episode we covered already has an agent autonomously finding a memory-safety zero-day in SQLite in hours with no expert team. That’s not developers on the frontier — that’s the frontier moving past developers. And that’s the tension worth naming. Sundar’s version is still staged adoption — developers first, then everybody else — but once you’ve got Amy Webb and Harper Reed asking whether U.S. tech dominance is actually ending, his framing lands very differently when he’s describing something that already feels mid-stride. The anti-gravity OS demo gets the namecheck as the agentic workflow example, which tells you a lot about the ceiling of the public example here. He’s pointing at a demo, not a production number. Here's Sky News:
In the 50 years after the Gutenberg press came out, the number of printed material in Europe, uh there there was more printed material in Europe produced than in the thousand years before. And so I think in software we're somewhere in this early part of the curve. With tools like FOD code, everyone is starting to be able to code.
Boris Chernny gets to the Gutenberg press within the first minute of this interview, and I do not mean that as praise. That printing-press comp is the 'it’s like electricity' of 2026 — you reach for it when you want historical cover, not when you want an actual answer. To be fair, the interviewer did try to pin him down — 'where are we on the curve?' — and Chernny actually gave a number: the 1700s. That’s more specific than most guests manage when they’re asked to self-locate on a hype chart. The 1700s is a vibe, not a coordinate. Did Sky actually press him on the coding-jobs charts, or did those just become props for the democratization story? Because 'everyone will learn to code like reading and writing' is exactly the line Claude Code’s creator is supposed to say. I wanted the interviewer to ask which workers, which roles, and what the displacement timeline looks like. Sequoia Capital writes:
Jake Stauch, founder and CEO of Serval, is building a ServiceNow for the AI era. His most contrarian bet is that the product should look like boring old enterprise software, but with unlimited intelligence. Jake explains why his team uses OpenAI models for end-user interaction and Anthropic models for code generation, and why he’s not worried the foundation labs will come downmarket.
Jake Stauch at Sequoia is described as building a ServiceNow for the AI era, and the headline is boring enterprise software with unlimited intelligence. That’s either the most honest positioning I’ve heard all week or a very clean way to avoid saying what the product actually does. I’ll take the bait on boring. What I actually want to know is the model-routing decision: OpenAI for end-user interaction, Anthropic for code generation. That’s not just vendor preference, that’s an explicit architectural split. Did Sequoia’s Pat Grady get into why — eval data, latency, cost — or did Stauch just say it and they moved on? That split is also the first time the GV velocity argument has landed on something concrete. Stauch is a named founder making specific tradeoffs under resource constraints, which is what we said we needed for that framework to mean anything. And his 'fewer, better' hiring thesis lines up with SAP’s deterministic-execution bet: in both cases, the vibe layer is not the moat, the architecture is. Corey Noles, Matthew Robinson, writing in The Neuron:
Tudor is the co-founder and CEO of Harmonic, the company behind Aristotle, a formal reasoning system built to generate mathematical proofs that computers can actually verify. That sounds abstract. But it could matter a lot. Because if AI can move from “trust me, this is right” to “check me, this is right,” it could reshape math, software, chip design, scientific computing, and maybe even how humans discover new knowledge.
Tudor Achim at Harmonic is drawing a line between AI that says something is true and AI that can formally verify it. Aristotle is the system, formal proofs are the output, and the claim is that a computer can actually check the work. That distinction matters because it connects directly to what SAP’s Christian Klein was calling deterministic execution earlier this week. The argument there was that 'almost right' breaks enterprise software. Achim is making the same structural point from the math side: a proof that’s mostly correct is not a proof. What I want to know is whether Corey and Grant actually pushed on what 'mathematical superintelligence' means operationally, because that framing is doing a lot of work. If the episode just uses it as a chapter header for standard 'AI is creative now' content, that’s a miss. The formal-verification angle is the real edge here. Chip design and scientific computing are named as downstream targets, and those are places where a wrong answer means a billion-dollar tape-out failure, not a hallucinated footnote. If Achim gets specific about even one of those, queue it. Here's ReliaQuest:
An AI agent was pointed at a piece of open- source software and asked a simple question. Are there any bugs? Hours later, not weeks, it surfaced a memory safety zero day in SQLite with no expert team, no multi-week fuzzing campaign, just a prompt. And stateback threat groups are already running the same tools that found the vulnerability before the rest of us catch up.
ReliaQuest’s Shadow Talk opens with a very specific claim: an AI agent, no expert team, no multi-week fuzzing campaign, just a prompt — and it surfaces a memory-safety zero-day in SQLite in hours. That’s the most concrete read/output loop we’ve had all week for what agentic security ops looks like in practice. I want the methodology before I buy the headline. Which version of SQLite? What was the actual prompt? Because 'hours, not weeks' is doing a lot of work there — hours compared with what baseline? If the comparison is a full red-team engagement, that’s real. If it’s 'faster than doing nothing,' that’s a press release. The other half of the episode is the flip side: state-backed threat groups running the same tools before defenders catch up. That’s not a vague China-rising gesture — it’s the same agentic capability pointed in both directions, which is exactly the concrete footing Sirota’s data-layer attack-surface argument needed. The CISO describing a second supply-chain attack landing while they were still cleaning up the first — that’s the operating detail I’m going to remember. That’s not a framework. That’s a person describing a real queue-overflow problem. If Tech Podcast Podcast is part of your daily routine, take a moment to subscribe or leave a review wherever you’re listening. It really helps other people find the show, and it helps us keep making it better.
You’ll find links to everything we talked about today in the show notes, so if a story stuck with you, take a minute to follow it through.
That’s Tech Podcast Podcast for this Thursday, May 21st. Thanks for listening, and we’ll be back next time. This is a Lantern Podcast.