0:00 Anthropic's Claude AI solved a decades-old cryptographic mystery in under thirty minutes.0:06 For years, people have wondered about the origin of the seeds used in some of the most common encryption standards on the internet, and now we know the answer wasn't a secret message or an NSA backdoor—it was just some old timestamps.0:22 Last episode, we were talking about Meta's big open AI push, and this week it feels like we’re seeing another one of those leaps, but this time it’s less about releasing a model and more about what these models can actually DO.0:38 And it’s not just one thing.0:40 On one hand, you have this almost magical act of digital archaeology, solving a puzzle that has stumped smart humans for a very long time.0:49 And on the other, you have the same company, Anthropic, completely re-architecting how they write and ship software, using Claude to automate huge chunks of the code review process.1:02 So you've got this incredible new tool for discovery, and at the same time, this powerful new engine for industrial-scale production.1:11 It’s a wild week.1:12 Let’s quickly run through the other big threads on Hacker News.1:17 First, there's that story about the Claude Code team.1:20 They managed to double the number of pull requests they merged between June and September of this year, while at the same time cutting the number of times they had to revert those changes by over forty percent.1:35 They did it by fundamentally changing the role of human reviewers, moving them from being gatekeepers on every single line of code to being specialists who only look at the most complex and novel problems.1:49 It’s a huge vote of confidence in AI-assisted development.1:53 Then there’s a piece of research from a company called Mercor that’s getting a ton of attention.2:00 They mapped out nine hundred and twenty-three different US occupations against all the existing AI benchmarks—you know, the tests we use to see how good our models are.2:11 The result?2:12 A staggering ninety percent of American working hours are spent on tasks that have little or no coverage from these benchmarks.2:21 So while we’re celebrating AI solving crypto mysteries, we’re not even measuring its ability to do the vast majority of actual work.2:30 There's also a really great post making the rounds that reframes how we think about transformer architectures.2:38 Instead of complex neural networks, the author, Alex Wang, argues we should think of them as associative memory stores.2:46 This model helps explain why they're so good at recalling vast amounts of information.2:52 A simple linear layer might store sixty-four memories.2:56 A multi-layer perceptron, or MLP, maybe twenty-seven hundred.3:00 But a self-attention layer, the core of a transformer?3:04 It can store over a MILLION memories.3:06 It’s a really clean way to build an intuition for why these things work so well.3:12 And finally, a bit of community drama.3:14 eBay announced new restrictions on listings for radio equipment.3:19 They're trying to crack down on devices that can operate outside of legal frequencies, but they made a specific exemption for gear that is "ham-band-only." This has the amateur radio and electronics hobbyist communities completely fired up.3:36 The discussion thread is a mix of people trying to figure out the exact impact on the secondary market and others worrying this is a sign of a broader regulatory crackdown on hobbyist electronics.3:49 Okay, let's go back to that cryptography story because it is just… it's perfect.3:55 So, for decades, there's been this mystery around a set of standards from the National Institute of Standards and Technology, or NIST.4:04 These are the "NIST P-curves," and they are everywhere.4:08 P-256, P-384… if you're using HTTPS, if you're using TLS to secure your connection to a website, you are almost certainly relying on the math defined by these curves.4:19 They are a foundational piece of internet security.4:23 Now, to generate these curves, you need to start with a random seed.4:28 A number.4:28 And for years, nobody could figure out where these seeds came from.4:33 The process was documented, but the origin of the initial numbers was… opaque.4:38 Because the standards came from the NSA in the late nineties, the speculation ran wild.4:44 The most common theory was that the seeds were a hash—a cryptographic fingerprint—of some funny or self-serving English sentence.4:53 There was even a quote from one of the NSA employees who worked on it, a guy named Jerry Solinas, who said in 2015 that he vaguely remembered the message being something like, "Give Bob and Jerry a raise." But he admitted no archives existed.5:10 So for years, the crypto community has been torn.5:13 Was it a joke?5:14 Or was it something more sinister?5:17 Could the NSA have chosen specific seeds that gave them a secret mathematical advantage—a backdoor to break the encryption?5:25 People have spent years trying to reverse-engineer these seeds, searching for phrases that would produce the right hash.5:34 Nothing ever worked.5:35 And then, this week, Anthropic just… gave the problem to Claude.5:40 And in less than half an hour, the AI came back with the answer.5:44 It wasn't a secret phrase.5:46 It wasn't a backdoor.5:47 It was just timestamps.5:49 Specifically, the output of a standard C library random number generator called lrand48.5:55 And that generator had been seeded with the Unix timestamps from when the engineers were likely doing the work back in 1997.6:04 The AI figured out the function, figured out the approximate time window, and then just brute-forced the possibilities until it found the exact seconds that produced the seeds we’ve been using for twenty-five years.6:18 The big mystery, the conspiracy, the secret message… it was just a programmer using a standard, slightly old-school tool to get a random number so they could get on with their day.6:31 The big headline from Anthropic says it all: "It's just timestamps." What I love about this is how it perfectly demonstrates a certain kind of intelligence.6:42 This wasn't a problem of brilliant insight.6:45 It was a problem of exploring a ridiculously large search space of possibilities.6:50 A human would get bored.6:52 A human would get stuck on their own assumptions—like "it must be a hashed phrase." But the AI has no assumptions.7:00 It just follows the logic.7:02 It's the ultimate needle-in-a-haystack finder.7:05 We've seen this before, right?7:07 Think about how AlphaGo beat Lee Sedol.7:10 It wasn't just that it was "smarter" in a human sense.7:13 It played moves that centuries of human Go players had considered "wrong" or ugly, but which the machine, by exploring the game's possibility space, knew were optimal.7:25 This feels like that.7:27 It's a reminder that AI doesn't just do what we do, but faster.7:31 It can solve problems in ways we wouldn't even think to try.7:35 So that’s the discovery engine.7:37 But the other side of the coin this week is the production engine.7:42 And that’s the story about the Claude Code team.7:45 On the surface, it sounds like a simple productivity story: they’re shipping more code with fewer errors.7:52 Great.7:53 But the how is the really important part.7:56 They didn't just give their developers a better autocomplete.8:00 They rebuilt the entire process of quality control.8:03 The old model of code review is probably familiar to anyone who’s worked in tech.8:09 You write some code.8:10 You submit it as a "pull request." Then, another human—a teammate, a senior engineer—has to read every line and approve it before it can be merged into the main codebase.8:22 This is the gatekeeper model.8:24 It's designed to prevent bugs and maintain quality.8:28 But it has a huge bottleneck: human attention.8:31 And as teams scale up, and especially as they start using AI to generate more and more code, that bottleneck becomes a crisis.8:40 You end up with senior engineers spending their entire day just approving trivial changes, while a huge queue of pending work piles up.8:49 The Claude team's insight was to stop asking the question, "Did a human read the code?" and start asking a better one: "What could go wrong with this change, and which check is the right one to catch it?" Their new system is layered.9:05 First, you have the automated tests, same as always.9:09 But then, you have a new layer: machine review.9:12 Claude agents—specialized instances of the AI—review the code for common issues, for style guide violations, for potential bugs.9:21 They can ask for changes, and they can approve simple, low-risk submissions automatically.9:27 The human reviewers are still there, but they’re at the top of the pyramid.9:32 They're brought in only for things the automated systems can't handle: architecturally complex changes, novel patterns, or when the AI itself flags something as needing a second opinion.9:45 This is a pattern we've seen again and again in industrial history.9:50 Think about the evolution of a car factory.9:53 At first, everything is hand-built by artisans.9:56 Then comes the assembly line, where tasks are standardized.10:00 Then comes robotics, which automates the repetitive, predictable tasks—the welding, the painting.10:07 The human workers don't disappear.10:09 They move to tasks that require more dexterity, judgment, and problem-solving: quality control, machine maintenance, final assembly.10:19 That's what's happening to code review.10:21 They’ve automated the predictable parts, freeing up human expertise for the problems that actually require human intelligence.10:30 And the results speak for themselves: doubling output while cutting the revert rate—that is, the error rate—by forty percent.10:39 That's not just an incremental improvement; it's a phase change.10:43 It suggests that for a huge class of code changes, the AI is actually a more reliable reviewer than a busy, distracted human.10:52 Of course, the counter-argument is real.10:55 What if the AI develops a blind spot?10:57 What if it consistently approves a subtle kind of bug that a human would have caught?11:03 It's a valid concern.11:05 But the data, so far, points the other way.11:08 Fewer reverts means fewer mistakes are getting through.11:11 The system seems to be working.11:14 And it points to a future where the role of a software engineer shifts even further away from writing boilerplate and more towards being an architect and a systems thinker—directing a team of AI assistants that handle the implementation.11:30 So you have these two massive stories from Anthropic this week.11:34 An AI as a brilliant archaeologist, and an AI as a hyper-efficient factory foreman.11:40 It feels like the future is here.11:42 But then you get hit with that research from Mercor.11:46 The one that says ninety percent of American working hours are spent on tasks that our AI benchmarks don't even cover.11:54 And it’s a crucial dose of reality.11:57 The researchers took the ONET database, which is this huge government project that breaks down every conceivable job into specific tasks, and they cross-referenced it with the tasks that show up in popular AI evaluation suites like MMLU or HumanEval.12:14 They looked at nine hundred and twenty-three occupations, covering almost nineteen thousand distinct tasks.12:21 And the gaps are huge.12:23 Of twenty-two major job families, six have basically ZERO coverage.12:28 Think about jobs in construction, in maintenance, in agriculture.12:32 We have no standardized way to even ask if an AI could help with the tasks that make up those jobs.12:39 Even in fields where there is some coverage, like business or finance, it's incredibly lopsided.12:46 The benchmarks are obsessed with things that are easy to grade, like writing code or answering multiple-choice questions.12:54 They almost completely ignore tasks that involve physical interaction, complex interpersonal negotiation, or dealing with unpredictable, real-world environments.13:05 So what does this all add up to?13:07 We are getting incredibly good at building AI that excels at the tasks that people who build AI do all day.13:15 You know, writing code, processing text, solving logic puzzles.13:19 The ECDSA discovery and the Claude Code team's success are testaments to that.13:25 We're building amazing tools for knowledge workers.13:28 But the vast majority of the economy—the ninety percent—is still uncharted territory.13:34 This isn't a criticism of AI, it's a criticism of how we're measuring it.13:39 It’s like we’ve invented a sports car and we’re only testing it on a drag strip.13:45 We know it’s fast in a straight line, but we have no idea how it handles on a dirt road, or in city traffic, or in the snow.13:53 The pattern here is the classic "streetlight effect." We look for our keys where the light is, not where we dropped them.14:02 We build benchmarks for the problems that are easy to measure, not necessarily the ones that are most important.14:09 This research is a massive wake-up call that if we want AI to truly transform the whole economy, and not just the tech sector, we need to get a whole lot more creative about how we define and measure success.14:24 We need benchmarks for plumbers, not just programmers.14:28 This week, we saw an AI solve a mystery that stumped cryptographers for a generation, and we saw it revolutionize the process of building software itself.14:38 But we also got a stark reminder that for most people's jobs, the AI revolution hasn't even been benchmarked yet, let alone begun.14:47 The frontier is moving faster than ever, but the map of the territory we actually need to cover is still mostly blank.14:56 And that, I think, is what this week really sets up: the race to fill in that map.15:01 The race to figure out what AI means not just for us, here in tech, but for everyone else.
