Math animations, spontaneous TDD, and compression that beats xz

00:00:04 OpenAI shipped GPT-6 Astra earlier this week, and the demos are starting to settle. One caught my eye—a math animation built from a single prompt by elvis, aka Omar Sarfaty, who tweets about AI-assisted education. He called it a jaw-on-the-floor moment, and after watching the clip I can see why.

00:00:23 The animation walks through parabolas bending and shifting across coordinate space, with explanatory labels appearing in sequence. It's coherent enough to actually teach someone. That's different from what we usually get. Previous models needed you to break the animation into steps, export frames, stitch them together.

00:00:44 Astra went straight from prompt to polished output. Omar wrote that he hasn't been able to get anything close to this with earlier models. When a single pass can produce pedagogically coherent output where earlier models required manual assembly, the delta is real for people who build educational tools.

00:01:05 The tweet was pinned by Omar himself, though it only picked up sixteen likes and four retweets—you don't judge capability claims by engagement metrics. You judge them by whether you've actually tried to build it yourself. When a model bypasses that usual frustration wall, where human labor takes over from automation, the shift is worth tracking.

00:01:28 Whether it changes your actual workflow is another question entirely. Nate B Jones pushed further today. On his AI News & Strategy Daily channel, he skipped reacting to the demo and spent a few days actually handing work to Astra. His test case was a household move—over twenty hours of pure administrative tasks in an average relocation: finding apartments, comparing schools, chasing down DMV appointments, changing utilities across states.

00:01:57 Nate's approach draws the same comparison I found myself reaching for. This feels like what happened with Claude Code back in December, when we went from asking it to write individual functions to just giving it a whole job and letting it figure out the steps. The relationship changed.

00:02:16 You stopped driving the chatbot and started delegating outcomes. He gave Astra chunks of work that ate hours, not minutes—a doctor search across seventeen different practice websites, deadline tracking pulled from email threads, address verification from Google Maps.

00:02:34 The model moved between systems autonomously, tried a route that hit a dead end, found another one, kept going on the parts it could complete because it held enough context. I'm not saying you should hand Astra your credit card and go to sleep. That's not what this is about.

00:02:52 What he's showing is the scale of work you can actually delegate without constant supervision—the kind that used to require a human sitting at their desk with seventeen tabs open. We're looking at which chunks of your week you'd stop doing yourself.

00:03:09 Simon Willison noticed something this morning while testing coding agents. He's been telling them to follow red/green test-driven development for a while now—the pattern where you write the tests first, confirm they fail, then implement the code that makes them pass.

00:03:27 His tweet today reads: "From 'red/green TDD'—earlier this way I was telling coding agents to do that, looks like the most recent models do it unprompted." Simon has a detailed framework on his blog explaining why red/green TDD works so well with them. Write the tests first, confirm they fail—this prevents the common trap where agents write code that looks correct but was built around an untested assumption.

00:04:01 Then implement until green. The important part used to require explicit prompting. You had to tell the agent: write tests first, run them, make sure they fail. Now it's emergent. The model has likely internalized something about test-first development during training—it recognizes the pattern as a standard structure in programming workflows, and the signal from the training data is strong enough to activate on its own.

00:04:29 There's no announcement from any lab about this capability. No benchmark update showing a new score. It's one developer testing their tool and noticing the tool changed. But when development methodologies that require explicit prompting start emerging without it, you have to ask what boundary shifted.

00:04:49 More training data on test suites? A change in how the model weights prioritize structured outputs? We won't know until someone investigates, which is probably why Simon hasn't written a full post about it yet. I'd call this incremental but real. It doesn't change how I build things today.

00:05:09 In six months, though, when all your coding agents write their own test suites by default, you'll wonder how you managed without them.

00:05:18 A completely different signal came in via bzip3, Kamila Szewczyk's spiritual successor to BZip2 showing up on Hacker News. This isn't a toy project. The algorithm uses an order-zero context mixing entropy coder combined with a fast Burrows-Wheeler transform built on suffix arrays, plus an RLE pass using LZP-style prediction matching.

00:05:43 The benchmarks matter for engineers who deal with compressed data at production volume: it compressed every version of Perl ever released down to 546 megabytes using the b equals 511 setting. Compare that to xz hitting 2.54 gigabytes on the same dataset, and Zstandard landing at 3.07 gigabytes.

00:06:06 Decompression is competitive too. bzip3 ran the dataset in four minutes and six seconds using parallel mode against LZMA's four minutes forty. Zstandard came in at three fifty-one but with significantly worse compression ratios. For anyone storing large codebases or datasets where storage cost adds up across petabytes, this is a real reduction that most people haven't heard about.

00:06:35 Licensing is straightforward: bzip3 itself uses LGPLv3 only, not dual-licensed. The build tooling has various GPL components but they're compile-time only—the runtime library and the binary you actually use are LGPLv3. Kamila took no responsibility for data loss because every compression tool carries that assumption anyway, which is fair.

00:07:01 Low-level infrastructure stories rarely make headlines but matter when storage costs add up across petabytes. No one announces bzip3 in a keynote. But it's out there, and it works. If this reminds you of another compression tool that solved a real problem before anyone was looking—that's probably why it's here.

00:07:25 One more item from the policy side. The EU mandated smartphone and tablet repairability scores alongside public repair information links over a year ago, back in June of 2025. Manufacturers have to report their self-assigned scores on the European Product Registry for Energy Labelling.

00:07:44 A review from Right to Repair Europe shows that out of two thousand three hundred thirty-four smartphone records filed over the last year, only about eighteen percent actually list a website where spare parts prices or repair instructions can be found. About half the records have blank fields where URLs should go.

00:08:06 Another nineteen percent point to product pages that don't contain any repair information at all. Some manufacturers even listed Temu and AliExpress as sources for spare parts and repair instructions—which, given those platforms' track record with genuine replacement components, is a bit of an open challenge rather than actual compliance.

00:08:29 Despite the gaps, some companies still give themselves class A ratings for repairability. The scores are self-assigned with no verification from any public authority. Right to Repair Europe is asking whether consumers can trust manufacturers grading their own homework, which seems like the only reasonable question given the data.

00:08:52 Thomas Opsomer spoke on behalf of the coalition: they want full documentation published so anyone can check the scores, and easier reporting for non-compliance. Next February brings another round of EU regulations requiring user-replaceable batteries in new mobile devices.

00:09:11 Whether those will be enforced any better remains to be seen. This is a familiar pattern. Regulation exists, industry compliance is cosmetic, enforcement requires resources nobody seems willing to allocate. The self-reporting mechanism doesn't just break down—it fails because the incentive structure is backwards from the start.