The lesson that trained chess bots is now being applied to antibodies, video games, venture funds, and your agents — and today, somebody finally put a hard number on it. This is Tech Podcast Podcast, and after a week of vibes, Chai Discovery finally shows up with an actual receipt. The question is where that number stops being true. From Sequoia Capital:
Most people treat biology as a bespoke, messy science. Josh Meier and Matt McPartlon, co-founders of Chai Discovery, treat it as an engineering problem. They make the case that drug design obeys the bitter lesson: scale data, models, and compute, and the model can learn what a hand-built pipeline simply couldn’t capture. The results: de novo antibody design from a sub 0.1% hit rate to 16%.
Sub 0.1% to 16% on de novo antibody design. That's a 160-fold jump, and it's the first number all week I actually want to chew on. Right — because it tests whether “biology as engineering” is a slogan or a measurable claim. Josh Meier and Matt McPartlon are betting on the bitter lesson: keep scaling data, compute, and models, and eventually the hand-built pipeline stops mattering. Sure, but where exactly does that 16% get measured? Is it the clean assay they cherry-picked, or the full pipeline yield to a real candidate? With a number that good, there's always a point where the counting stops. And here's my follow-up: nine months to nine days is the pitch. Fine — what still takes the other eight months and twenty-nine days? Name the bottleneck scale doesn't touch. Here's the part I respect: they're arming pharma rather than competing with it, and they published Chai-1. It's two guys treating drug design like an inference-scaling problem, out in the open. That's the honest version of this bet. And that's an interesting counterpoint to everyone treating hoarded proprietary data as the moat. Chai is betting on scale and compute over a locked-up context layer — a structurally different wager. Let's see if it holds. Sett writes:
Yotam is Co-Founder and CEO of Peerplay, a casual studio he started in 2023 and runs like almost no one else: born with AI, obsessed with distribution, and built on conviction long before the data shows up. This episode is about building a great game company now that the old playbook is dead and AI has rewritten the rules.
So the pitch is “the MVP is dead, build on conviction before the data shows up.” Coming from a guy who ran nightclubs and made hit records as a DJ, honestly? That tracks — you can't A/B test a dance floor. What I like is that he names an actual bet, not a vibe. He's bringing casino volatility — jackpots, that spike — into mass-market casual, and Merge Cruise is where he proves it or doesn't. And the differentiator's concrete: the game is openly gay, aimed at the core casual audience — women roughly 30 to 60 — who see that everywhere on TV and never in games. The distribution thesis has a face. Give me that over another “AI rewrote the rules” slide. Right, though “retention and monetization aren't at odds” is the line I'd want him grilled on. Every casual studio says that. Show me the Merge Cruise numbers and I'll believe the old trade-off's actually dead. And “AI-native studio” still needs a receipt — after the Chai Discovery hit rate we just covered, “born with AI” better mean more than a Midjourney art pipeline. Conviction's a great story right up until the retention curve disagrees. From Poddtoppen.se:
Rafael launched a crypto insurance product in 47 states, backed by a real carrier. Everyone told him it was genius. Nobody needed it. He shut it down with $2M left in the bank, fired almost everyone, and rebuilt with two engineers on $20K a month. Notch just raised $30M.
Rafael Broshi shut down a crypto insurance product that was licensed in 47 states and backed by a real carrier — the thing everyone kept calling genius. Demand said otherwise, so he killed it. That flips the usual pivot story. The product worked; nobody wanted it, so he shut it down. That's a harder call with $2M still sitting in the bank. The number I keep circling is the rebuild: two engineers on twenty grand a month. That's the part that makes it real. Then the company grows 12x and raises a $30M Series A. Here's my one question for the investors who led that round: did you know the whole team was two people on a $20K monthly burn when you wrote the check? Because that's either the pitch or the thing that got left off the slide. And this lands right after the Chai piece we just covered. One shop treats biology as a scaling problem; this one shows that correct tech means nothing if nobody uses it. That “Isn't everyone doing this?” response may be exactly the feedback you want. Poddtoppen.se writes:
Most venture firms are using AI to save time. Earlybird is using it to generate alpha. David sits down with Andre Retterath, General Partner at Earlybird, to discuss how his firm built an AI-native venture platform, why proprietary data is becoming venture capital's biggest competitive advantage, how machine learning improves investment decisions, and why the future of venture belongs to investors who combine technology with exceptional judgment.
Earlybird says everyone else uses AI to save time; they use it for alpha. Okay, but proprietary data as your moat just means you got there first and built a wall around it. That line jumped out at me too. Andre Retterath draws a split between AI for efficiency and AI for alpha. And if proprietary data is the fund's edge, the same logic applies to founders: Chai published a model an hour ago on this show; Notch presumably never published a churn signal. Right, so the tension is real. It leads straight into what Chenxi Wang worried about: capital pooling at the extremes, with Series B and C getting starved. If the alpha accrues to the AI-native funds sitting on the data, does that starve the middle even more, or feed it? My honest read? It concentrates. Whoever owns the proprietary dataset on early-stage companies gets first look, and “exceptional judgment plus technology” is a nice phrase until you ask who's allowed to see the technology's inputs. And that's what the episode won't measure. They say machine learning improves decisions, but they're only counting the deals they saw. Show me the alpha net of the deals the model never surfaced. Chris Fowler, writing in 1Password:
When Maxim Fateev, CTO and co-founder of Temporal, joined Zero-Shot Learning, he brought a historical perspective to the challenges developers face when building agentic systems today. From vanishing state to retry storms, Fateev saw that the failures of deploying long-running agents have parallels to the problems he’s been working on for decades.
Okay, this one's a 1Password-branded podcast — Zero-Shot Learning, hosted by their own CTO. I'm bracing for an ad. But the guest is Maxim Fateev, Temporal's co-founder, and he actually gives you the mechanism: vanishing state and retry storms in long-running agents. And the through-line goes back to 2002 at Amazon, where he co-created Simple Workflow. His pitch: agents need durable execution because they're running into an old distributed-systems problem in a new costume. That's the part I want. Everybody this week says agents get flaky at scale, but nobody names why. Fateev's basically saying the failure mode has a twenty-year-old fix — checkpoint the state so a crashed step doesn't nuke the whole run. That gets at the thing that's been nagging me: keeping state trustworthy across a long job, even after you've handled isolation and retries. This episode at least points a flashlight at the problem. Sponsored slot, sure. But I'll take a co-founder explaining retry storms over one more panel telling me agents are the future. Queue it if you're actually shipping something that runs longer than a demo. If you're enjoying Tech Podcast Podcast, subscribe and leave us a review wherever you're listening. It helps more people find the show, and we're grateful you're here.
Links to every story are in the show notes, so check out the ones that caught your attention and dig a little deeper. Thanks for listening, and we'll be back tomorrow. That's Tech Podcast Podcast for today. This is a Lantern Podcast.