The token counter's spinning up, and somebody finally asked the awkward question: who's actually buying all this AI-written code? If you're just joining: across the AI startup beat, investors have been moving the bar for what counts as good — faster growth, stronger defensibility, and capital piling into fewer winners. Scale Venture's Rory O'Driscoll put the software version bluntly: some pre-ChatGPT companies are already dead, a lot more are under threat, and the range of outcomes has blown wide open. This is Tech Podcast Podcast. Today: Apple sues OpenAI over prototypes walking out the door, a $300 million valuation with no product, and Chelsea Troy says we've been measuring AI coding all wrong. If AI startup growth benchmark reset matters to you, hit follow — we'll be back on it soon. Jason Lemkin, writing in SaaStr:
Apple sued OpenAI for trade secret theft this week, and the details are ugly: a six-year Apple employee allegedly walked physical prototypes out the door for show-and-tell, encouraged by a 24-year Apple veteran now running OpenAI’s hardware group. The individuals involved are, in Jason’s words, toast. But the more interesting read is what the lawsuit signals about OpenAI’s hardware ambitions, which look increasingly like a distraction from the one thing actually printing money right now: coding.
Okay, the Apple suit — a six-year employee allegedly walks physical prototypes out the door for show-and-tell, egged on by a 24-year Apple lifer who now runs OpenAI's hardware group. Forget the IP filing for a second — okay, that's a spy movie. And it's a different category than every 'they scraped our data' story we've been circling. Physical prototypes out the door is plain, specific insider risk. Question is whether Harry and Jason actually sat with the complaint or just riffed off the headline. That's exactly what I want to know. Because 'the individuals are toast' is a great clip and tells you nothing. The interesting read is the one Jason buries — the whole hardware play might be OpenAI chasing a distraction while coding is the thing printing money. Right, and that tees up the TAM question — probably the week's actual payoff. Rory O'Driscoll's on the AI coding TAM here, the same guy who put numbers on the SaaS doom earlier this week. The 20VC-SaaStr desk is now openly asking whether coding revenue hits a ceiling faster than anyone modeled. ClickHouse AI spend up 60x since February. Jason burned a month of Claude credits redesigning one page. Best devs are running 10 to 20 agents around the clock and still can't eat enough tokens. That's the token-maxing era — and it's either a growth story or the fastest way to hit a wall I've ever seen. And Meta charging developers for its own models for the first time — Zuck breaking three years of silence on X to do it. Priced straight at the cheap-token tier, because every half-awake CIO is about to hand devs a budget model to stop the bleeding. The company whose whole brand was 'free and open' is now sending an invoice. Cool. Very cool. Maggie Nye, writing in TechCrunch:
Building an AI startup is one thing. Raising one of the largest seed rounds of the year before launching a product is something else entirely. In this episode of Build Mode, host and Startup Battlefield lead Isabelle Johannessen sits down with Andrew Dai, founder and CEO of Elorian and former Google DeepMind researcher, to unpack how his company raised a $55 million seed round at a $300 million valuation before generating revenue or releasing a product.
Fifty-five million dollars, three hundred million valuation, and the product is... a vision. Literally — visual AI, but no product to look at yet. And it lands the same morning as the SaaStr panel we just hit, where Rory O'Driscoll's out there doing TAM arithmetic on AI coding. One episode is asking how big the market really is. The other is pricing a market that hasn't shipped a pixel. That's the whole tension of the week in two feeds. Andrew Dai leaves Google DeepMind, and the pitch — per TechCrunch — is that today's fundraising rewards clear storytelling over technical jargon. Which is a very polite way of saying the deck did the work, not the demo. What I actually want from a 34-minute Build Mode episode is the specific thing investors saw. He says he didn't max out the valuation and chose investors instead. Okay — what does that mean in practice? Which investor, what terms, what did he trade away? Right, because 'I chose the right investors' is what everyone says after they raise at three hundred million. Nobody's on Build Mode going, 'yeah, I took the dumb money.' From Nilay Patel at The Verge:
You’ll hear Bart say pretty plainly that the thing Proton sells, at a high level, isn’t really the products themselves, but actually trust. And trust in the software world isn’t only about the people who run the companies, but also the technology they develop and sell and the corporate structure in place to make sure that technology is built against the right incentives.
Bart Butler, Proton's CTO, on Decoder: 'no company is going to go to jail for you.' That's the most honest thing anyone in this space has said all week. And that gets very practical fast: not who funds the frontier lab, but what a company actually does when a government shows up with a warrant. The backdoor line is the one, though. He says a backdoor only the good guys can use is technically impossible — you build the hole, everybody eventually finds it. Which is why he frames the product as trust, not the email client. Encrypted mail, docs, a calendar, an AI assistant called Lumo — the whole stack is basically a bet that the corporate structure won't cave. Here's Ben Lorica and Chelsea Troy at O'Reilly:
The tech industry is measuring AI productivity all wrong, and Mozilla MLOps engineer and University of Chicago instructor Chelsea Troy makes a strong case for why. The real opportunity, she argues, isn’t shipping more code faster but finally having the bandwidth to run the experiments, tests, and simulations that engineering teams have always wanted to run but never had time for.
Chelsea Troy just walked straight into the token-maxing thing we hit up top. Her whole argument is that measuring AI by token consumption is measuring the wrong thing — she wants token efficiency, how much useful work each token actually does. That's a direct shot at the 'ship a million apps a week' framing. In her version, the payoff is finally having the bandwidth to run the tests and simulations teams never had time for, instead of just cranking out more code. Two episodes, same day, opposite conclusions. One panel's counting tokens like calories, and Troy's over here going, sure, but are any of them nutritious? And she's specific about where the ceiling actually is. Scoped file edits, that works. The collaboration-and-steering layer — where you'd actually want the leverage — is where it breaks. This one's from The Neuron:
Goodfire is building tools that use AI to interpret AI. Its long-term goal is what Eric calls intentional design: training models less like mystery creatures fed enormous piles of data, and more like software you can inspect, debug, edit, and improve on purpose.
So Eric Ho's pitch is that a model isn't just spitting tokens one at a time — it's building this whole internal world first. Features, circuits, confidence signals, curved shapes. Which, if true, is a much bigger claim than the usual interpretability handwave. And that's why the hallucination angle matters. Goodfire connects the shapes-inside-the-model idea to a concrete failure — models giving you confident false answers — and says if you can read the internal structure, you can catch that before it ships. That's the part I actually want tested. After Chelsea Troy's episode arguing the industry is measuring AI productivity wrong, here's Goodfire saying maybe the black box is where the wrong answers hide. Same mess, totally different diagnosis. Right — and interpretability has a credibility problem because those internal processes are opaque even to the people who built the model. So 'we can open it' is the claim that carries the whole company. I want the mechanism, not the metaphor. 'Models think in shapes' is a great line for a YouTube thumbnail. Whether it reduces hallucinations by a measurable amount is the thing Corey and Grant should've pinned him on. Got thoughts on today’s briefing, a story we should follow, or a correction we need to hear? Send us a note at techpodcastpodcast at lantern podcasts dot com. We really do read your feedback.
You’ll find links to every story we covered today in the show notes, so if something stuck with you, that’s the place to dig in a little more.
That’s Tech Podcast Podcast for today. Thanks for listening, and have a great Friday. This is a Lantern Podcast.