0830 | Record Ocean Heat as El Niño Forms, Flock Cameras, DHS Snooping, Pixel 11 MTE

Mia: Welcome to HackerNews Daily on Bri Radio. I'm Mia.

Milo: And I'm Milo. Coming up: a record ocean temperature as a powerful El Niño takes shape, and the quiet clause that put a dollar from your car insurance toward Flock cameras.

Mia: Plus, GrapheneOS says the Pixel 11 drops support for hardware memory tagging, a Hy4 preview, and a vote on responsible generative AI inside Debian. That's all ahead on HackerNews Daily.

Mia: The global average sea-surface temperature hit a record seventy degrees Fahrenheit on August twenty-second, according to Europe's Earth observation agency, Copernicus, which has been tracking the metric daily since 1979. That single-day reading came as a powerful El Niño forms.

Milo: The previous single-day record was measured back in March, so this new August reading stands out because it landed in a season where the oceans don't normally peak like this.

Milo: Texas lawmakers added a dollar to every car insurance policy in the state as part of a 2023 law passed unanimously by the legislature, designed to combat rampant catalytic converter theft. That money went into a Motor Vehicle Crime Prevention Authority grant program, which then paid for Flock cameras. One person in the Hacker News discussion called the whole approach snooping.

Mia: And that critique lands on the funding path itself. Drivers paid the extra dollar believing it would target catalytic converter theft, but the same grant money ended up purchasing automated license plate reader cameras. The discussion keeps circling whether the money is doing what policyholders thought they were funding, or whether it quietly financed a broader surveillance network instead.

Milo: The Department of Homeland Security has been using an obscure customs law, nineteen USC fifteen-oh-nine, to demand records from journalists, nonprofits, and unions. People in the Hacker News thread flagged the legal mechanics of the statute itself as notable, since it's not a measure most people associate with domestic surveillance.

Mia: That's the striking part. A law written to govern customs imports, essentially how the government handles goods coming across the border, is now the legal hook for pulling records from domestic groups like journalists and unions. It takes a statute whose original audience was importers and repurposes it as the authority behind that kind of request.

Mia: GrapheneOS reports that the Pixel Eleven no longer supports hardware memory tagging, known as MTE. That's the feature that would tag every piece of memory with extra bits so the silicon can flag unsafe access patterns before they turn into real vulnerabilities. People in the Hacker News thread are calling it a terrible development, since MTE is seen as a genuinely promising defense and this walks that capability backward at a time when the security landscape is demanding more of it, not less.

Milo: And that's what makes this sting for the security crowd. Memory tagging is one of those hardware-level mitigations that software alone can't fully replicate, because it depends on the silicon doing the checking in real time. Losing it on the Pixel line means the path from chip to shipping device no longer carries that safety net, which is why the reaction in the thread reads as genuine frustration rather than routine complaining.

Mia: Tencent has opened up its Hy4 preview model, and the traction on OpenRouter is striking. Commenter minimaxir notes it has already processed trillions of tokens in a couple of days — more than GLM 5.3 managed in a full week — and that traffic is coming at a cheap price, with a five percent cache cost while everyone else still charges ten or twenty percent.

Milo: That cache pricing gap is the telling part. Five percent undercuts the rest of the market so dramatically it goes a long way toward explaining why people are pushing that much traffic through it this fast.

Mia: vLLM shipped version 0.28.0 on August twenty-sixth, a release the notes say brings five hundred eighty-four commits from two hundred seventy contributors. The headline work is a big optimization push for Kimi-K3 across the stack — decode context parallel support, fused decode and prefill kernels, combined all-gathers with a reported one-and-a-half to three times kernel-level speedup, an adaptive speculative token budget that delivers roughly sixty percent better time-to-first-token for DSpark, and optional shared-expert sharding that saves about seventeen gibibytes of memory per GPU. Kimi-K3 also now runs on ROCm with the V2 model runner. For DeepSeek V4, sparse MLA now works end-to-end for plain decode, MTP, and DSpark speculative decoding, joined by AMD Quark NVFP4 support. New defaults include max num batched tokens raised from eight thousand one hundred ninety-two to sixteen thousand three hundred eighty-four, and the Blackwell CUDA graph capture default raised to one thousand twenty-four. Breaking changes include bitsandbytes support moved to an out-of-tree plugin, Transformers bumped to version 5.15.0, removal of the deprecated calculate-KV-scales runtime calculation, and removal of the override attention dtype.

Milo: Behind all that optimization, the discussion carries real frustration. Commenter kouteiheika says they love vLLM but find it frustratingly buggy. Running DeepSeek-V4-Flash on a B300, version 0.26 was totally broken and needed three out-of-tree patches; version 0.27 needed none but output randomly degraded into repeated token loops. On a workstation running Gemma-4 on an RTX 6000, the whole process tends to get stuck and needs a kill-and-restart. So the performance gains land alongside reliability that still reads rougher than the release notes suggest.

Mia: Debian has voted to allow what the community is calling responsible use of generative AI, and the Hacker News thread did not take the news quietly. One commenter hoped it works out better than Microsoft's experience using the technology on Windows 11, which immediately drew a counter-question from another reader — do we actually know Windows 11 has suffered because of generative AI, or is that just a random snipe against Microsoft?

Milo: That pushback is the useful part of the thread. The whole debate is running on vague assumptions about what responsible use even means in practice, and nobody on that thread has produced evidence tying generative AI to concrete harm on Windows 11. So the vote sets a policy direction while the actual consequences stay untested and undefined.

Mia: Samsung is still pursuing processing-in-memory, or PIM, and it showed up at Hot Chips. Commenter reliabilityguy points out this is a long-running bet. They recall a Samsung paper from ISCA around 2021 or 2022 demonstrating an HBM2 module with PIM, which was impressive back then. Their honest read is they are not sure what the killer application actually is for the technology, and without one, adoption is unlikely.

Milo: That is the tension the whole PIM story keeps circling. The hardware idea is genuinely clever — moving compute into the memory module to cut data movement — but cleverness has never been the blocker. Without a demonstrated workload that needs it badly enough to justify the cost and complexity, it stays a technology looking for its reason to exist.

Mia: Field of view is the whole argument for the Nancy Grace Roman Space Telescope. One commenter in the Hacker News thread put it bluntly: you would need many Hubble telescopes to map everything out the way Roman can, because for something like a survey, the wide field is everything.

Milo: And that is the core design tension. Hubble sees a narrow patch in stunning detail; Roman trades that reach, that single patch, for a much larger field of view, which is what lets it sweep whole regions of sky at once instead of stitching thousands of pointings together.

Mia: The "Quantifying Colour" article builds its visual explainer of CIE 1931 colour spaces starting from spectral power distributions and how the eye's photoreceptors respond, rather than working backward from RGB. Commenter herf called it interesting but weird, and pushed back on a few choices: he wanted to know why it plots bell-curve-style approximations instead of real measured cone responses like the Stockman-Sharpe data, and he flagged that D65 white could look very blue in that rendering.

Milo: Herf also said the piece takes too long to reach the actual point, that cone responses overlap heavily and that what we perceive as colour comes from roughly differencing those overlapping curves. He even conceded the matrix transform to chromaticity is technically correct, but argued it is not the most accessible way to teach the subject.

Mia: Another commenter, dahart, came back that the article acknowledges the shortcut itself. The sensitivity curves are normalized approximations, not accurate data, and the real curves are less smooth, with S-cone sensitivity notably lower and rods more sensitive than any cone. But dahart disagreed with the accessibility criticism outright, saying most colour articles start from RGB and reason backward, whereas this one starts from the spectral power distribution, implies the cone overlap early, and uses the correct integral, so it gets the physics across first and keeps perception separate from physics.

Milo: And the creator, ekunazanu, confirmed all of it. He said he built the page quickly during his last summer break just to make the SoME4 submission deadline, aimed it at readers comfortable with high-school math but with no colour background, and deliberately went with the drawn-out approach.

Mia: A Hacker News thread on Martin Uecker's August 2026 piece about calling nested functions on GCC without an executable stack turned into a wide-ranging discussion of self-modifying code. On x86-64, older GCC versions place a trampoline on the stack to hand a static chain to a nested function. Uecker's fallback reads the code address and the static chain out of that trampoline and calls the nested function directly using GCC's built-in call with static chain, so the program never actually executes the trampoline on the stack.

Milo: That matters because the stack is still marked executable and the trampoline is still created, so the call can't be devirtualized. But since the trampoline is never run, you can clear that flag with patchelf's clear-execstack option. Uecker has it implemented in his experimental noplate library, and he also proposes treating the trampoline as a function descriptor by interpreting its fixed code sequence at call sites.

Mia: From there the comments drifted into history and firsthand experience with self-modifying code. One commenter called it cool and noted it was disabled for security reasons; another added that instruction caches also have to cope with it. In reply, someone argued a cache flush is easy enough, but its overhead is what made the older tricks impractical, citing the Linux kernel's runtime patching of traced functions and Java JIT compilers using dynamically known subclass information, and described a personal case where shift amounts could be patched straight into the shift instructions. Someone else recalled 1980s text editors patching their own executable code to work around limits.

Mia: Dan Luu's piece "Bug Blindness" sparked a comment that most software really does contain bugs, and that the majority of people using the software end up using it in ways the builders would never anticipate. One commenter said the workflows users invent are mind-bending, and yet those weird workflows somehow do what they need anyway.

Mia: There's a point in one Hacker News thread that keeps coming back: good culture is the biggest productivity hack, not AI. But the discussion underneath gets quickly skeptical. One commenter notes the devops crowd has talked about culture for years and it never seemed that helpful — in their view, toxic managers and toxic C-level executives create bad culture, so what do you actually do about that?

Milo: And that's where it gets interesting, because another commenter says reading that post made them realize bad culture perfectly explains something they'd observed. The response to the manager-blaming is essentially that naming the problem as a one-way causal arrow doesn't give you an off-ramp — the post is arguing culture is the lever, but the people in that thread are pointing out that the people who set the culture are exactly the ones least likely to change it.

Mia: One Hacker News co-founder of Burning Man argues the festival has lost its soul. But the thread's first reaction is that this has been obvious for years now — one commenter says it's been clear for a long time, and another agrees it's hard to draw a specific line but that it's been over a decade in the making.

Milo: So the pushback isn't really aimed at the founder's description of what's changed, it's about the timing. The commenters are saying the decline wasn't a recent development — they'd been watching it unfold for years before the founder went public. And there's a running note that he's been raising this complaint for a while, which frames the essay less as a fresh revelation and more as a long-simmering grievance finally being aired.

Mia: On a lighter note, there's Defrag98 — a free, browser-based simulator of the classic Windows 98 Disk Defragmenter, credited to Dennis Morello. It spins up four virtual drives, C through F, ranging from 512 megabytes up to 2 gigabytes, with options from slow, grinding hard disks to a zippy late-nineties drive, each with its own defragmentation speed and estimated runtime.

Milo: And it's faithful down to the rendering — free, unoptimized, optimized, reading, writing, and unmovable clusters are drawn exactly like the original utility, complete with period-correct hard-disk sound effects. One commenter said the whirring noise was a brilliant touch and summed it up as exactly what the AI revolution should be for. Another said they wanted the disk noise and got it, though the speed felt faster than their actual drives back then.

Mia: The speed question actually gets answered in the thread. One commenter explains it's a simple calculation — one thousand divided by the disk's operations per second — and the different disks live in Settings. They note that by inspecting the source for that one-thousand-divided-by line, setting a breakpoint, and changing the disk's iops value to one thousand, the timeout drops to one second per row instead of two to four.

Milo: And it only ever touches virtual clusters — it never reads, moves, or modifies real files on the device, runs entirely in the browser, works offline as an installable web app. That prompted one commenter to ask for the same idea as a cross-platform screensaver, and another to want the concept applied to AI models — a sleep mode where models self-prune and optimize overnight for their daily purpose, with a reply linking to Claude's documentation on managed-agents Dreams.

Mia: We've been tracking record ocean heat as that powerful El Niño takes shape, and how a one-dollar insurance fee quietly helped pay for the Flock cameras. Thanks so much for listening.

Milo: Those stories, plus everything else we covered today — appreciate you being here with us.