AI is making engineers faster, factories pricier, and apparently malware more scalable. Nice, calm Monday. Quick context before today’s development: the AI autonomy debate has moved from capability into security and governance. Recent episodes covered AI exploit discovery, whether patch-generating systems can fix vulnerabilities without creating new ones, and Toby Ord’s argument that AGI timelines are often misread, uncertainty is underweighted, and stronger governance tools — including bans or treaties around superintelligence — deserve more serious attention. This is Tech Podcast Podcast. We’ve got the AI bill, deep-tech money, and a policy plan that makes “delay” concrete. First: Patrick Wendell. This story isn't over: OpenAI singularity and autonomous capability debate. Follow us wherever you're listening, and the next chapter comes to you. TBPN, with Samir Kaul, Patrick Wendell, Grant LaFontaine:
Patrick Wendell discusses his role as Databricks co-founder and VP of Engineering, where he leads AI products and internal AI adoption. He explains how AI coding tools can nearly double engineering capacity while creating rapidly escalating consumption costs, and highlights model switching, intelligent routing, and optimization as key ways to control spending without sacrificing productivity.
Patrick Wendell saying AI coding can nearly double engineering capacity — then immediately spike consumption costs — is the part everybody conveniently leaves out of the demo reel. And Databricks has an operating answer: switch models and route work intelligently to keep costs down. The useful follow-up is where that bill starts eating into the capacity gain — at ten engineers, a hundred, a thousand? “Intelligent routing” can mean real discipline — send the cheap jobs to cheap models — or it can mean putting a nice label on the runaway inference tab. I wanted the implementation details. Grant LaFontaine’s Whatnot has raised $545 million at a $20 billion valuation. The idea that knowledgeable sellers can build meaningful businesses with small audiences is plausible. The harder test is whether the customer experience holds as Whatnot adds markets, categories, and streaming formats. TBPN writes:
Aditya Agarwal, managing partner at South Park Commons, discusses its new $575 million fund and the growing ambition of founders building capital-intensive deep-tech companies. He also covers founder dilution, AI-assisted application review, Series A fundraising, scientific diligence, and the importance of technical progress and rapid growth.
Chris Power’s Hadrian raised $1.37 billion to build automated factories, while Aditya Agarwal’s South Park Commons has a $575 million fund backing founders before the company is even fully formed. Same neighborhood of dollars, wildly different ways to set money on fire. Agarwal gets specific about the constraint: capital-intensive deep tech changes the dilution math early. Hadrian is the vivid version — if the product includes operating factories, the round has to finance physical reality, not just a roadmap. I also liked the AI-assisted application review detail at SPC, because that’s VC using AI for throughput, not pretending it found secret alpha. You still need somebody to do scientific diligence when a founder says, “trust me, the factory works.” TBPN is at its best when guests stay at that operating layer. A $1.37 billion factory raise and a $575 million founder fund tell you more about the current ambition — and its costs — than another lap around abstract AI destiny. From Daniel Kokotajlo at Lawfare:
Daniel Kokotajlo, former OpenAI researcher and executive director of the AI Futures Project, joins Kevin Frazier, Director of the AI Innovation and Law Program at Texas Law and Senior Editor at Lawfare, to detail his policy recommendation— AI 2040: Plan A. It’s a thorough analysis of a policy pathway to delaying superintelligence, which Daniel and his co-authors think is necessary to ensure that the disruptive effects of highly-capable AI systems do not outweigh the benefits.
Kokotajlo’s AI 2040: Plan A picks up where Toby Ord’s uncertainty critique left off. He’s moved from forecasting danger to arguing for a deliberate delay, and Lawfare actually puts the scenarios under scrutiny. That’s the useful part. Kokotajlo is the most credible person in this week’s superintelligence parade. He’s a former OpenAI researcher with a concrete proposal, and actual policy stakeholders are pushing back. So I need “delay” to mean more than a tasteful PDF with 2040 in the title. Kevin Frazier does ask why Kokotajlo changed his estimate that policymakers would adopt Plan A. That gets at more than another argument over a calendar date: what would make a government choose this path? Right — and give us the off-ramp. What tells us the delay worked, what tells us it failed, and who gets to call it? Otherwise “prevent disruption from outweighing benefits” can stretch forever. Maggie Nye, writing in TechCrunch:
In this episode, Connie and Alex are joined by Reed Jobs, founder and managing partner of Yosemite, to discuss the firm’s ambitious mission to make cancer nonlethal by building biotech companies from the ground up. Jobs explains why Yosemite pairs venture capital with no-strings-attached research grants, how AI is accelerating drug discovery and clinical trials, and why he believes we’re entering one of the most important eras in the history of medicine.
Reed Jobs’s Yosemite model is more interesting than the usual biotech-fund pitch: venture checks alongside no-strings-attached research grants. That gives a lab room to find something real before somebody forces it into a slide deck. “Make cancer nonlethal” is an enormous sentence. But he at least puts a mechanism under it: AI speeding drug discovery and clinical trials, where the calendar has historically been part of the disease. And clinical trials matter here. Faster molecule generation is great, but the claim only starts to hold up if therapies can move through the evidence bottleneck. I want the ugly math, though: how much grant capital buys the science, how much venture capital builds the company, and how often that handoff works. Biotech has buried a lot of gorgeous narratives in preclinical data. Here's Robinson Meyer at Heatmap News:
The Engine Ventures, which spun out of MIT a few years ago, is different. It tries to fund what it calls “tough tech,” a somewhat nebulous category that includes companies working in energy, decarbonization, health, and infrastructure.
Katie Rae’s useful distinction is between engineering risk and software-market risk. Engine spun out of MIT to fund things like Form Energy and Commonwealth Fusion, where the question is often whether the physics works before anyone gets to argue about growth loops. “Tough tech” can sound like a venture-capital tote bag, but Rae is pointing at a real mismatch: classic VC wants a product roadmap; a fusion company may need the universe to cooperate first. And that changes the capital conversation. We just heard about Hadrian’s $1.37 billion factory raise; Rae’s framework explains why lumping that in with a SaaS round makes the numbers look comparable when the risks plainly aren’t. Right. If the engineering clears, you still need customers, manufacturing, regulation — the whole ugly second half. A lab demo does not come with product-market fit preinstalled. Have feedback, story ideas, or a correction? Email us at techpodcastpodcast at lantern podcasts dot com. We’d love to hear what you think and what you want us to cover next.
Links to every story are in the show notes. Take a look at the ones you’d like to explore further. Thanks for listening, and we’ll be back with more tomorrow. That’s Tech Podcast Podcast for today. This is a Lantern Podcast.