0:00 Margaret Hamilton, the software engineer who led the team that wrote the code for the Apollo space program, died yesterday at the age of ninety.0:10 And her passing comes on a day where the entire conversation on Hacker News is consumed by what happens when the HUMAN element of that kind of monumental, foundational work gets automated away.0:24 Last week, you and I talked about OpenAI's latest math breakthrough, and this week, that entire thread has unraveled into a much, much bigger question about the very soul of progress itself.0:37 Because while we're mourning a pioneer who represents the absolute peak of collaborative human ingenuity, the firehose of AI news just keeps blasting away.0:48 And honestly, some of it is genuinely useful, almost mundane at this point.0:54 Anthropic, for example, just dropped Claude Haiku 5.5.0:57 This is their fast, cheap, little model.1:00 And when I say cheap, I mean it's about seventy-five percent cheaper than the last version.1:07 This isn't the model you ask to write a novel; this is the one you use for high-volume, cost-sensitive stuff.1:15 Think live customer support chats, content moderation, the plumbing of the internet.1:21 It's faster, its scores on knowledge benchmarks have more than doubled...1:26 it’s a solid, practical upgrade.1:28 It's the kind of thing that just quietly makes thousands of existing products better without a big philosophical debate.1:37 It's infrastructure.1:38 And then, on the complete other end of the spectrum, you have OpenAI.1:43 They just announced broader access to GPT-6, but the headline feature is something they’re calling Intelligent UI.1:51 And this is...1:52 well, it’s different.1:54 Instead of just spitting out a wall of text, ChatGPT can now generate responses that are interactive.2:01 Think diagrams with parts you can tap on, sliders you can move, charts that update.2:07 They showed this demo of asking it to explain a seven-speed bicycle.2:12 And instead of just describing it, it generates this little module with the frame, the wheels, the drivetrain, all as distinct, interactive components.2:23 It says, "Seven gears let the rider adjust pedaling effort as the drivetrain turns effort into motion," and you can actually see the parts it's talking about.2:34 This is a big deal for explainability.2:37 It’s one thing to have an AI tell you how something works; it’s another thing entirely for it to build you a little toy to figure it out for yourself.2:47 It’s moving from being a search engine to being a museum exhibit curator.2:53 And this is rolling out to over one point two billion weekly users.2:57 So, you have the cheap, fast plumbing from Anthropic on one side, and the rich, interactive museum from OpenAI on the other.3:06 The whole ecosystem is maturing.3:08 But not everything is a product announcement.3:12 Sometimes the most upvoted things on Hacker News are just...3:16 clever ideas.3:17 There's this little web app making the rounds called Bigwords.page.3:22 And the idea is so simple it's brilliant.3:25 It turns your screen—any screen with a browser—into a giant sign that displays a message.3:31 Like, if you need to hold up a sign at the airport or flash a message across a room.3:37 The magic trick here is HOW it does it.3:40 The entire message is encoded in the URL itself, specifically in the part after the hash mark.3:47 The creator calls it "The URL is the app." Because the message lives in that URL fragment, your browser never even sends it to a server.3:56 There's no backend, no database, no account, no tracking.4:00 It's just a perfectly self-contained, privacy-preserving little tool.4:05 It’s a reminder that sometimes the most elegant solution is the one with the fewest moving parts.4:12 And in that same vein of "huh, that's a neat way to think about things," there was this huge discussion that spun out from an article about why Victorian elites were so effective.4:25 The paradox is that by all accounts, they were...4:28 not exactly living clean.4:30 We're talking heavy drinking, late nights, rampant indulgence.4:35 Yet they managed to run the largest empire in world history.4:39 The thread wasn't really about defending them, but it was wrestling with this question of what productivity and effectiveness actually ARE.4:49 Is it optimizing every minute of your day, or is it something else?4:54 Is it about intense bursts of focused work, strong social networks, or just a culture that had completely different standards?5:03 It's one of those discussions with no right answer, but it's a good gut check against our modern obsession with hyper-optimized, bio-hacked productivity.5:14 It just asks, you know, are we even measuring the right thing?5:18 And that question—are we even measuring the right thing?—is the perfect bridge to the story that I think REALLY matters this week.5:27 It’s the one that ties everything together, from Margaret Hamilton’s legacy to last week's AI math news.5:35 So, here's the setup.5:36 A few days ago, a new AI-generated proof for something called the Unique Games Conjecture—or UGC—was released.5:44 Now, you don't need to know what the UGC is.5:47 Just know it's a famous, notoriously difficult problem in theoretical computer science.5:54 Solving it is a HUGE deal.5:55 It was one of 372 results that OpenAI's system spat out.5:59 And on the surface, this is exactly the kind of headline we’ve gotten used to: AI Solves Major Unsolved Math Problem.6:08 We saw it last week, we're seeing it again now.6:11 Groundbreaking.6:12 Except...6:13 the human experts, the top people in the field, started trying to actually READ the proof.6:19 And they couldn't.6:21 Scott Aaronson, a giant in the field, wrote about it on his blog.6:25 He quoted another top complexity theorist, Dana Moshkovitz, who tried to make sense of the AI's paper.6:33 Her verdict?6:33 I have to read this to you.6:35 She said it is, quote, "so horribly written that it’s impossible to read it without AI help." She said it has "lots of name dropping" and "confusing citations." And my favorite part, she said, "It feels like something written by someone who’s on psychedelics.6:54 So much unclear and doesn’t make sense." So let that sink in.6:58 The AI may have solved the problem.7:01 The answer might be CORRECT.7:03 But the explanation it gave is functionally useless to the very humans it's supposed to be helping.7:10 It’s like finding a book that contains the secrets of the universe, but it’s written in a language that no one, not even a machine, can translate back into a human thought.7:22 What good is it?7:23 This is where we get to the pattern.7:26 Where have we seen this before?7:28 Well, we're seeing it right now, everywhere.7:31 It's a problem of translation.7:33 And as if on cue, a new paper just hit the arXiv server this week titled "Navier-Stokes lost in translation." Again, don't worry about Navier-Stokes—it's another huge, messy math problem.7:47 The paper is about the process of AI autoformalisation.7:51 That's the fancy term for an AI taking a proof written in normal human language and translating it into a formal, machine-verifiable language like Lean.8:01 The idea is that the computer can then check the formal proof with one hundred percent certainty.8:08 But here’s the devastating point the authors make.8:12 Even if the AI produces a formally PERFECT proof in Lean, that gives you ZERO guarantee that the original human-language proof was actually correct.8:23 The AI might have misunderstood the original argument.8:26 It might have misinterpreted a subtle semantic point, or filled in a logical gap with its own reasoning that wasn't in the original text.8:36 The translation itself is the weak link.8:39 And how hard is it to create a perfectly faithful translation from messy human language to precise formal logic?8:47 The authors have an answer for that.8:50 They say, quote, "Providing semantically faithful AI autoformalisation is harder than any computational problem including the Halting problem." For anyone who doesn't live and breathe computer science, the Halting problem is the classic example of something that is PROVABLY impossible for a computer to solve.9:12 So what they're saying is that the very thing we need to do to trust these AI proofs—ensuring the translation is perfect—is literally, provably, impossible.9:23 So what does it all add up to?9:25 You have AIs producing "solutions" that are unreadable to human experts.9:30 And you have a formal argument that verifying the link between human ideas and AI proofs is an impossible task.9:38 The whole enterprise starts to look...9:41 a little shaky.9:42 This is where the synthesis comes from, in a post by one of the greatest living mathematicians, Terence Tao.9:50 He’s been watching all of this, and he just wrote a post that I think will define this entire debate for the next few years.9:59 He argues that we need to start thinking about "Math 2.0." He says that for centuries, "Math 1.0" measured progress by one metric: did you solve the problem?10:10 But he argues that was never the whole story.10:13 A real breakthrough wasn't just an answer.10:16 It was an answer that generated a huge amount of community activity.10:21 It spawned talks, workshops, new papers, follow-up collaborations.10:26 It gave dozens of other researchers new tools to work with, new avenues to explore.10:32 A solution was valuable because it nourished the entire ecosystem of mathematics.10:37 And what he sees happening now with AI is the exact opposite.10:42 He says, quote, "solutions to open problems are now being harvested at large scale in an unsustainable fashion." He describes the process as being driven by, quote, "AI prompters who have no interest in the broader field itself once their initial target is ‘solved.’" The AI provides an "answer," a box is ticked, the prompter moves on.11:06 But the community gets nothing.11:08 There's no new technique to learn, no beautiful new structure to understand, no foundation to build on.11:15 There’s just...11:16 a formally verified answer that no one can read, to a question that has now been stripped of all its context and community.11:25 It’s the intellectual equivalent of strip-mining.11:29 You get the resource, but you destroy the landscape in the process.11:33 And this brings me all the way back to Margaret Hamilton.11:38 Her work on the Apollo program was the absolute epitome of Tao's "Math 1.0." It was a massive, collaborative human effort.11:46 It wasn't just about solving one problem; it was about inventing the entire discipline of software engineering as they went along.11:56 The code she and her team wrote was so robust, so clear, so well-designed that it became the foundation for decades of computer science.12:05 It created a universe of follow-up work.12:08 It nourished the ecosystem.12:10 Years later, reflecting on the Apollo program, she said, "Looking back, we were the luckiest people in the world." That’s not something you say when you’ve just strip-mined an answer.12:23 That’s something you say when you’ve been part of building a cathedral.12:28 It's about the meaning, the community, the shared struggle.12:32 So the question this week sets up isn't really about whether AI can solve hard problems.12:39 It's becoming clear that it can.12:41 The real question is what kind of progress we actually want.12:45 Are we building a library of unreadable answers, an oracle that hands down solutions from on high without explanation?12:54 Or are we trying to build a community of understanding, where the goal isn't just the answer, but the beautiful, intricate, and deeply human process of getting there?13:06 This week, with the passing of a giant like Margaret Hamilton juxtaposed against these unreadable AI proofs, that question feels more urgent than ever.13:16 We're not just deciding what our tools can do; we're deciding what we want to use them for.13:23 And what kind of world we want to build with them.
