Somebody finally put a real mathematician on the mic to answer the question we've been chewing on all week — is AI actually solving math, or just doing a very convincing impression of it? If you're just joining, here's the setup: with the usual benchmarks getting quiet, frontier AI science is turning into the new scoreboard. Earlier this week, we had OpenAI leadership arguing that these models aren't just acing tests anymore — they're starting to crack scientific and mathematical problems. The thing to watch is whether that becomes the new yardstick for model progress. This is Tech Podcast Podcast. Today: Grant Sanderson on AI and math, Marc Andreessen going 'Beyond P(doom)' at a policy shop, Rocket Lab aiming at Starlink, and a year of outsourcing your life to AI. Sanderson's up first — and he might have the receipt Joey's been hunting. Here's Dwarkesh Patel at Dwarkesh Patel:
AI has been making much faster progress in math than in other fields. As a result, mathematics is showing us, very concretely, what AI progress in other fields will look like. Even within mathematics, there's a jagged landscape. What does it look like?
Okay, this is the one I've been waiting for all week. Sanderson on Dwarkesh, ninety minutes on AI and math. If anybody can tell me whether those solved problems are actual, verified proofs or just proof-shaped output, it's the 3Blue1Brown guy. And Dwarkesh is the right venue for it. We've had OpenAI making frontier-science claims at the abstract level. Sanderson's frame is more concrete: math is moving faster, so it may show us what AI progress looks like in other domains. Which is a big claim. Fifty-nine thousand views, twenty-two hundred likes — people showed up. But the pitch is math leads and biology and code follow. I want to know why math and not code, and whether he actually says the mechanism or just gestures at it. For me, that's the falsifiable part. If he can say why math is the leading indicator — the description points to this overhang from AIs connecting ideas already sitting in the literature — then we can actually test that. Connecting existing ideas is a very different superpower from inventing a new one. If that's the overhang he means, I hear it as: the model can read the whole library at once. I'll take that receipt — it's honest about what the machine's actually doing. a16z writes:
Andreessen argues that while AI has already begun reshaping the economy, the largest impacts are still ahead. He explores how AI could dramatically expand access to expertise, improve productivity, and transform industries ranging from healthcare and education to law and software development.
One like on Podbean. That's the whole grassroots for 'Betting on America.' What actually caught me is the venue. He's not on the usual founder-podcast circuit here; he's at CSIS, with Navin Girishankar as a policy interlocutor. Different room, different audience. Sure, but the pitch is still the same: AI expands access to expertise, transforms healthcare, law, education. I've heard that reel since 2023. The one line I'll credit him on: he puts the biggest barriers on institutions — regulation, policy, infrastructure constraints — more than raw technology. At CSIS, that's at least aimed at the people who could move it. Right, but does he name a mechanism? Export controls, data — okay. If the answer to 'who wins when the model layer commoditizes' is just 'America,' that's a bumper sticker, not industrial policy. Coming right after Sanderson — where he's actually asking how AI progress works inside one domain — the contrast is pretty stark. One conversation gives you things to test. Here, I'm still waiting for something falsifiable about industrial policy specifically. Bloomberg Technology writes:
Bloomberg’s Ed Ludlow speaks with the CEO of Rocket Lab about the company's biggest bet yet: buying Iridium in an $8 billion deal to challenge SpaceX in the orbital economy.
Rocket Lab buying Iridium for eight billion to go at Starlink — that's a real number attached to a real satellite constellation, which is a nice change of pace from the week we've had. It's the biggest bet Peter Beck's ever made. And unlike a lot of the SpaceX-comparison talk that floats around, Iridium already has a working fleet and paying customers — he's buying an installed base, not a pitch deck. Right, that's the part I actually respect. Beck isn't selling you a mission to Mars vibe. He's buying revenue that exists today and betting he can operate it cheaper than the incumbent. The question I want Ed Ludlow to press is whether Iridium's low-earth-orbit voice-and-data niche actually threatens Starlink's broadband core, or Rocket Lab is buying a strong business in an adjacent lane. Those are different claims. And notice, in the same rundown, Anthropic is winning back some access to Mythos 5 after the government's national-security concerns — so you've got one company buying a comms constellation, and another negotiating with regulators just to run its model. Infrastructure fights on two fronts today. Armani Ferrante, writing in Lightspeed:
Gm! In today's episode, we're joined by Armani Ferrante, founder and CEO of Backpack, to discuss the evolution of Backpack, tokenized equities, regulatory strategy, and building a crypto-native financial institution. We also explore Solana's growth, market structure, token design, 24/7 stock trading, global financial infrastructure, and the long-term role of tokenized assets and decentralized networks.
Armani Ferrante, Backpack, on Lightspeed — 24/7 stock trading on Solana, tokenized equities, the whole crypto-native financial institution pitch. Seventy-minute episode and I'm scanning the timestamps for one thing: where's the number where 24/7 markets stop being free? Someone has to run those matching engines around the clock. The interesting move here is Backpack choosing regulation. There's a whole chapter on why, and that cuts against the usual crypto-native posture. Right, and I respect the choice — but 'why we chose regulation' is exactly the kind of narrative that sounds airtight until you ask who's holding the equities and what happens at settlement on a Sunday. And don't skip the disclaimer at the bottom: Lightspeed was kickstarted by a Solana Foundation grant. So the venue is already rooting for Solana's growth story before Ferrante says a word. Joanna Stern, writing in The Jordan Harbinger Show:
How a chatbot that never tells you your ideas are dumb becomes less of a companion and more of an emotional slot machine — and why the always-available AI therapist that remembers your every 4 a.m. anxiety is both a genuine comfort and a quiet, compounding cost to real human connection.
Okay, this is the consumer-side version of the AI week — Joanna Stern spent a full year actually living with these tools. And the phrase I care about is 'hidden costs.' Because if the hidden cost is 'I felt lonely,' fine. If it's inference bills, skill atrophy, or trusting a confidently wrong answer on her medical results — now we're talking. The line that jumps out to me is the expertise one — she says you spot exactly where AI falls short the second it wanders into your own field. That gap between confident output and actual competence is the most testable thing in the whole episode. And she frames the chatbot that never pushes back as an emotional slot machine. That's a sharper read than most of the ROI talk we heard earlier — it's a cost that doesn't show up on any dashboard. The other keeper: those viral drink-pouring android demos? Often a guy in a VR headset puppeteering from offstage. That's the whole hype cycle in one image. Got a tip, a correction, or a tech story you think we should be watching? Send it our way at techpodcastpodcast at lantern podcasts dot com. We read what you send, and it helps shape the show.
We’ll have links to everything we covered today in the show notes, so if a story stuck with you, that’s the place to pick it back up and read a little further.
That’s Tech Podcast Podcast for today. Thanks for listening, and we’ll be back next time. This is a Lantern Podcast.