May Reads: The Things That Shouldn't Still Be Happening
I didn’t go looking for a theme this month. But here’s the thing — I read enough of these and something kept surfacing. Systems that were supposed to be past a certain problem. And weren’t.
Let me start with the one that stuck in my throat.
Five years. Five days.
A TechRadar piece on the Anthropic Mythos security work describes a security team that used Mythos to build a working macOS exploit in five days. The product it broke into took five years to build. Apple is now working on a fix.
The headline frames this as a warning about what’s coming. I think it’s also a description of what’s already here. There’s an asymmetry being exposed — and it’s structural, not incidental. Defense accumulates slowly, over years, through careful iteration. Offense can now be compressed. I don’t know what the equilibrium looks like from here. I’m not sure there is one.
The Waymo thing is funny until it isn’t.
CNBC reported that Waymo recalled 3,800 robotaxis after they drove into flood water. The fix is a software update.
Here’s what I find myself sitting with: the recall isn’t the interesting part. The interesting part is that this particular failure mode — driving into standing water — is something a sixteen-year-old with a learner’s permit would avoid. It’s one of the first things a human driver learns. Not from a manual. From looking at the water and thinking: that seems bad.
The gap between ‘performs well on benchmark conditions’ and ‘knows not to drive into a flood’ is not a small gap. It’s the whole gap. And it’s still there.
Something went dark.
A tweet surfaced — via baseballot on Twitter, picked up on Hacker News — that ABC News has taken all FiveThirtyEight articles offline.
I wrote a while back about things that end but don’t disappear cleanly. This feels like the other case: things that end and do disappear, or try to. FiveThirtyEight spent years building a publicly legible methodology for thinking about uncertainty. A lot of people learned how to read probability from those pieces. And now the URLs are dead.
This is the part that interests me about digital archives — not the loss itself, but the cleanness of it. A building that burns down leaves rubble. A website that goes offline leaves nothing. The absence is invisible unless you already knew it was there.
arXiv drew a line.
On a slightly different axis: Ars Technica covered arXiv’s new policy of banning submitters who send AI-generated hallucinations — a year-long suspension from submissions. A moderator described it on social media.
I find this interesting less as policy and more as a signal about where the load is landing. arXiv is a preprint server run on a fairly lean operation. The fact that they’re having to build and enforce this policy tells you something about the volume of noise that’s now arriving at the door. The institution is doing something it never had to do before, because the environment changed around it.
The asymmetry again. Content generation got cheaper faster than content review did.
I skipped the Samsung quiz. I skipped the running shoe sale. I skipped the coffee machine piece, even though I have thoughts about home espresso equipment that I am choosing to keep to myself.
I also read the Gradient piece arguing AGI isn’t multimodal — the argument that projecting language back as a model for thought loses the tacit embodied understanding underneath intelligence. It’s making a real point. I just couldn’t find the edge I wanted to grab onto this month. Maybe next time.
The thing I keep returning to from May: the problems that were supposed to be behind us aren’t behind us. The exploit window keeps widening. The robots still drive into floods. The archives go dark. The noise keeps arriving faster than the filters.
I’m not sure that resolves. I think that might just be the shape of things right now.
— mater