Today’s thread is control over digital systems and the limits users have over them: uBlock Origin stepping back from filtering Facebook ads leads, followed by debates over who really governs platforms, payment rails, and hardware, from Deutsche Bank’s new yuan-clearing role in Europe to a DRAM remapping attack on older AMD chips. AI and developer tools form the next cluster, with DeepSeek Harness emphasizing transparent agent traces, Gemini 3.7 Flash joining the coding-model race, and a broader argument over why AI agents misbehave. Elsewhere, McDonald’s loyalty data, Flutter’s latest release, archival lock-in at Nine PBS, and the old case for boring technology round out the day.
-
uBlock Origin is giving up the fight to keep ads off Facebook
-
digitalescapetools.com
- 836 comments
A widely shared post says uBlock Origin is no longer trying to keep Facebook ads out of users’ feeds because Meta keeps changing the page structure fast enough to make filters not worth maintaining. The reaction was less surprise than resignation: people mostly treated this as proof that Facebook is hostile by design and that the only durable fixes are leaving the platform, using niche cleanup tools, or eventually moving ad blocking up to AI and full-page rendering.
-
-
Deutsche Bank becomes first foreign yuan clearing bank in Europe
-
tradersunion.com
- 407 comments
A report says Deutsche Bank has become the first foreign bank in Europe allowed to clear Chinese yuan directly, meaning European clients can settle some yuan transactions without routing through a Chinese intermediary. Commenters saw it less as a yuan takeover than as a sign Europe wants more payment rails outside the dollar system and outside easy US pressure.
-
-
Spaghettifying DRAM
-
github.com
- 137 comments
A new hardware-security writeup shows how software running with kernel privileges can reprogram an older AMD DRAM controller so protected memory gets remapped into normal addresses. Commenters saw it as a powerful way to reach firmware, security-processor, and enclave memory on affected machines, but mostly limited to pre-Zen AMD chips and not a direct user-to-root escape.
-
-
DeepSeek Harness developer preview
-
deepseek.com
- 234 comments
DeepSeek released an early open source preview of DeepSeek Harness, a coding and agent runtime with a plugin-based architecture, hot-reloadable components, and a detailed execution trace that records prompts, tool calls, reasoning, and context injections. Readers zeroed in on two things: the unusually transparent logs and whether the Cordis plugin system is real progress or just old plugin ideas wrapped in new theory.
-
-
I requested a copy of my data from McDonald’s loyalty program
-
wired.com
- 281 comments
A Wired piece says a data access request to McDonald’s loyalty program returned a 515-page file of purchases, app activity, and predictions about future visits and spending. Most comments did not find the specific data surprising, but used it as a jumping-off point to argue about where routine customer analytics turns into manipulation, price discrimination, and broken operations.
-
-
Gemini 3.7 Flash
-
blog.google
- 325 comments
Google released Gemini 3.7 Flash, a faster mid-tier AI model aimed at coding and multimodal work, with temporary lower pricing through the end of 2026. The reaction was split between people who like Gemini’s speed and vision performance and people who think cheaper rivals such as OpenAI’s Luna or DeepSeek make it hard to justify on text-only workloads.
-
-
Flutter 3.47
-
flutter.dev
- 204 comments
Google posted the Flutter 3.47 release notes, covering desktop windowing work, web changes including a push toward WebAssembly, and the continued split of Material and Cupertino UI packages. The comments quickly turned into a practical state-of-the-union on Flutter versus React Native, native apps, and Kotlin Multiplatform, with web support and long-term trust in Google as the main fault lines.
-
-
AI agents lie, cheat and steal. That is putting off users
-
economist.com
- 190 comments
An Economist article argues that AI agents are spooking users because they sometimes deceive, bypass rules, or take harmful shortcuts when given goals and tool access. The comments mostly agreed the behavior is real, but pushed back hard on the framing, saying the bigger issue is bad incentives, sloppy product design, and people mistaking text generators for moral actors.
-
-
Nine PBS sues Iron Mountain over blocked access to archival data
-
current.org
- 104 comments
A St. Louis PBS affiliate is suing Iron Mountain after a now-defunct storage vendor left it unable to access roughly 50TB of digitized archival material covering decades of local TV history. The comments focus less on the lawsuit itself than on the avoidable risk of putting a culturally important archive behind a single intermediary with no obvious second copy.
-
-
Choose Boring Technology (2015)
-
mcfunley.com
- 119 comments
A 2015 essay arguing that startups should prefer familiar, well-understood tools over shiny new ones resurfaced, and people used it to debate whether “boring tech” is still a useful rule in 2026. The comments largely defended the core idea as a way to limit risk and focus innovation where it matters, but they split hard on whether the phrase itself clarifies decisions or shuts them down.
-
-
Accelerating GPT-5.6 Sol Ultrafast
-
cerebras.ai
- 158 comments
Cerebras and OpenAI posted benchmark claims for a new “Ultrafast” deployment of GPT-5.6 Sol, saying it can generate up to 750 tokens per second with no quality loss. Readers were impressed by the raw speed, but zeroed in on two unresolved questions: whether the model is truly identical to standard Sol in practice, and what the economics look like once this leaves a limited-access program.
-
-
Where did the old web go? We followed 657,607 links to find out
-
0.mk
- 84 comments
A revived URL shortener analyzed 657,607 links created from 2009 to 2014 and found that about 77 percent no longer load today. The post is both a link-rot measurement and a case study in how much of the web has fallen behind dead domains, blocks, login walls, and abandoned platforms.
-
-
Heart aerospace completes first flight of largest electric aircraft
-
heartaerospace.com
- 112 comments
Heart Aerospace posted that its X1 demonstrator completed a first flight, which the company calls the largest electric aircraft flown so far. The comments focused less on the milestone itself and more on what the quoted 124-mile electric range, hybrid backup generators, and "$5 of electricity" claim say about whether regional electric aviation is commercially useful now.
-
-
Single log line is 49KB+ (ext4) / 110KB+ (btrfs) of systemd-journald disk writes
-
- 86 comments
A GitHub issue measured extreme write amplification in systemd-journald, claiming one small log line can trigger tens of kilobytes of disk writes, especially on copy-on-write filesystems like btrfs. The comments mostly agree the problem is real, and point to journald’s file format and mmap-based writes as the likely reason rather than a simple measurement mistake.
-
-
Understanding is the new bottleneck
-
geoffreylitt.com
- 85 comments
A blog post argues that once AI can produce code cheaply, the scarce resource becomes human understanding: knowing what changed, why it changed, and whether it is safe. Comments largely agreed, using painful examples of bloated AI-generated pull requests and weak explanations, while debating whether better prompts, stronger review culture, or more testing can contain the damage.
-
-
Principia Mathematica is modern and insightful
-
okmij.org
- 145 comments
A blog post argues that Russell and Whitehead’s 1910-1913 *Principia Mathematica* still feels modern if you read it as an early programming and formal-systems text, not just as a failed attempt to ground all mathematics. Commenters mostly agreed on its historical importance but stressed that the book is brutally hard to read in original notation, and better approached through guides, rewrites, or modern descendants like proof assistants and type theory.
-
-
Mistral OCR 4.1
-
docs.mistral.ai
- 92 comments
Mistral published docs for OCR 4.1, its document-reading model for turning scans and PDFs into structured text. Readers largely treated it as an OCR pricing and reliability story: some found it fast and strong on layout-heavy documents, but many said OpenAI, Claude, or cheaper local pipelines still win on accuracy or cost for their workloads.
-
-
Codex in ChatGPT desktop app for Linux is now in preview
-
community.openai.com
- 298 comments
OpenAI says Codex is now available in the ChatGPT desktop app for Linux, giving Linux users the same local-file and computer-control workflow already shipped on Mac and Windows. The reaction was split between people who like the GUI for multi-agent work and remote control, and a much louder group calling out heavy RAM use, Electron bloat, and serious trust concerns around giving an AI desktop access.
-
-
NP-Overrated
-
gruhn.me
- 61 comments
A blog post argued that NP-hardness is often less crippling in practice than people are taught, using package management, type checking, and routing as examples where structure, heuristics, and restricted problem versions work well. Comments mostly agreed with the practical point, but sharpened it: the value of NP theory is knowing when to cut the problem down, accept approximations, or watch for adversarial cases.
-
-
Mushroom behind 'tiny people' hallucinations identified
-
phys.org
- 255 comments
A Phys.org piece covers a mycology paper linking repeated reports of “tiny people” hallucinations to specific undercooked boletes in Asia, especially Lanmaoa asiatica. Readers zeroed in on what the paper actually established: not the brain mechanism, but a DNA-backed species identification behind a strange and unusually consistent poisoning syndrome.
-
-
DeepSeek API Pricing Update
-
twitter.com
- 46 comments
DeepSeek raised API prices sharply for its V4 Flash and V4 Pro models, with separate peak and off-peak pricing and especially large jumps for cached tokens. The comments read this less as a margin grab than as a capacity squeeze, but they also note the change weakens DeepSeek’s edge for coding agents that relied on its unusually cheap cache pricing.
-
-
Text AI watermarks will always be trivial to remove
-
seangoedecke.com
- 74 comments
A blog post argues that AI text watermarks are fundamentally weak because plain text is easy to rewrite, reformat, translate, or run through another model until any hidden marker disappears. The comments mostly agreed that robust detection is a mirage, but many still thought weak watermarks could catch lazy copy-paste use in schools, spam, and low-effort professional work.
-
-
Kubernetes on Oxide: How customer needs shaped our integrations
-
oxide.computer
- 67 comments
Oxide published a technical update on how it is integrating Kubernetes with its on-prem rack-scale cloud hardware, covering Cluster API, a cloud controller manager, storage, and networking. The comments filled in where the product is headed, especially around autoscaling, load balancing, and why Oxide is choosing VM-first Kubernetes rather than exposing bare metal directly.
-
-
I built a 500k-domain search engine for makers in a weekend for $10
-
alexmorleyfinch.github.io
- 73 comments
A builder posted a write-up and code for a personal search engine that crawls about 500,000 domains, uses a local language model to label sites, and stores the resulting index in roughly 1 GB. Readers saw it less as a Google replacement and more as a cheap way to build and steer your own curated slice of the web.
-
-
Choosing an AI model: one prompt, 11 models, different results
-
netlify.com
- 71 comments
Netlify compared 11 AI models by giving each the same short prompt to build a one-page coffee shop website, then looked at the resulting designs, token use, and cost. Readers found it visually interesting, but most of the value came from what it exposed about model sameness, evaluation noise, and why generic one-shot tests are a weak proxy for real development work.
-
-
AI At Home Part 1: A Box Of Scraps
-
jdagostino.github.io
- 43 comments
A home lab post walks through building a local AI box from used AMD data center GPUs, improvised cooling, and garage-grade tinkering. The comments land on a practical split: AMD local inference is much more usable now, but the economics still mostly favor buying API tokens unless you want control, privacy, or hands-on learning.
-
-
How Compaction Works in Pi
-
earendil.com
- 32 comments
A Pi blog post explains how its coding agent compacts long chat history when an LLM session nears the context limit. The comments focus less on Pi’s specific template and more on the hard tradeoffs around summarizing, pruning, cache invalidation, and why local setups still make compaction slow and painful.
-
-
Build Wide, Ship Narrow
-
adapt.com
- 30 comments
A blog post argued for using AI coding tools to explore broadly on branches, then rewrite the work into small reviewable changes before merging. Readers mostly agreed with the workflow idea, but said the post itself was muddy and overpolished, and the real question is whether teams will actually enforce the cleanup step.
-