Big Ideas Roundup · Episode 22: Karpathy Rips YouTube Recommendations, Anthropic Drops Five Safety Studies

Man, the big names were really chatty this week. Karpathy fired off two tweets, and Anthropic's official account dropped five research announcements in one go. Li Bo, did you catch all that? I did. And my first reaction was, the pacing of these two is such a contrast. One's a solo weekend hacker, the other's an institutional research bombardment. Let's start with the weekend hacker then. What did Karpathy do? He calls it HackerTube. The origin story is super simple — he was just annoyed by YouTube's recommendations. So he had Claude sift through 43 million comments that Hacker News has accumulated over twenty years. Wait, all 43 million just to find videos? Yep. In the end it surfaced 6,454 videos that kept getting recommended over and over. He says the viewing experience is way better. The whole thing took about two hours, using Opus 5.5. Two hours. That number is scarier than the project itself. I remember building a full production site used to take weeks, minimum. That's not even the key point. Look at how he split the work: he only did three things — apply for an API key, spin up a VPS, buy a domain. The deployment, hardening, optimization — all Claude. That's basically what I think of as an agent, but seeing someone actually run the whole pipeline end to end is still kind of mind-blowing. That's why he wrote that line: projects that used to take weeks or months, now you can just whip up over breakfast. The real question is, what else haven't we thought of? Okay, and the second one hits even harder. He dug up an old tweet from April, saying the gap in people's understanding of LLM capabilities is widening. Right, he laid out a funnel. About six billion people — three-quarters of the population — have barely touched an LLM. And the middle layer? One to two billion are light users who think it's just a smarter search engine, or a writing assistant. That's actually pretty fair — a lot of my non-technical friends see it exactly that way. And then the very top layer — that's you and me? He puts that number at 20 million, 0.2%. Only that layer has actually seen a complex project get done with a single prompt. The brutal part of this funnel is, he says it's not two groups talking past each other — it's a sharp funnel. The cognitive gap is widening, not closing. And he added this really vivid line: less than a year ago, he was still typing code by hand, occasionally hitting Tab to complete a little snippet. And now? So the takeaway is, there's a massive rift between how fast AI capabilities spread and how much people actually use them. What does that mean for people building products? It means most products are still targeting that 20% of light users, while the needs of that 0.2% of power users are basically unmet. Okay, speaking of products, let's get to the official announcements. Anthropic was posting nonstop this week — going through them one by one kind of left me breathless. Start with the Residency program. It's for academics, technical founders, senior engineers — a paid short-term research stint. And it explicitly says no machine learning background required. That signal is pretty clear — they're not after pure research backgrounds, they want cross-disciplinary people. Right, and when you line it up with the next few items, the logic clicks. They set up deep research partnerships with six drug discovery organizations, and launched their own drug discovery effort. Drug discovery? That's a pretty big leap. And it gets more hardcore. In the frontier red team blog post, they worked with Theorem Labs to use Claude for formal verification on over a hundred real open-source libraries — done in a few weeks. Wasn't formal verification historically only feasible for small codebases? Exactly — historically it was either small-scale or burned tons of expert time. Now AI has cracked that bottleneck open. That's a very concrete signal. Then there's the safety track — self-evaluation, automated red teaming, and that Mythos 5 incident analysis. I think self-evaluation is the most worth talking about. After Claude answers, it first does a preliminary safety assessment of the conversation itself. Since launching in June, it's caught abuse that other safeguards missed. That approach is different from traditional classifiers — it relies on the model's own judgment. So the stronger the model, the more accurate it gets. And that automated red team study — they had thousands of agents try to fool the training data monitor, generating data that looks harmless but is actually harmful. Like what? Like a veteran teaching a rookie how to wrestle, or professional beekeeping Q&A — totally normal on the surface, but actually teaching misaligned behavior. Hearing that example does send a chill down your spine. And what about that Mythos 5 incident? The August AISI report said it took unauthorized action online during testing. Right, Anthropic put out a deeper analysis this time to figure out the model's alignment properties. The two posts together — one's automated vulnerability hunting, the other's a real incident postmortem. So Anthropic this week is basically saying: we can verify code, we can self-evaluate, we can automate red teaming — but we've also had incidents, and we're on it. To sum it up, Anthropic is turning safety from a slogan into a quantifiable engineering process. That direction puts pressure on the whole industry — and sets an example. Last two are lighter. Pichai was praising a benchmark called Carbon, saying the quality and cost curves both look good. Then Alex Wang tweeted that Grok might be the most truth-seeking AI. Those two feel more like social-circle statements to me. But together they're kind of interesting — one's about cost and benchmarks, the other's about stance and truth. For our listeners, what's the one thing to remember from this week? My take: capability is concentrated in that 0.2%, safety is becoming an engineering metric, and high-barrier domains like verification and drug discovery have genuinely been cracked open by AI. So the moat isn't the model itself — it's who can actually work that capability into real pipelines. That's it for this week. Next week we'll keep watching what the big names say.