Today’s thread is AI tools getting more capable while becoming harder to trust or live with: developers are frustrated with Opus 5’s coding behavior and writing style, Z.ai’s GLM-5.3 is drawing attention for offensive security willingness and a planned weight release, and Alibaba’s Qwen 3.8 shows how much open local models can now do despite practical tradeoffs. Around that, the web and browser stack loom large, with Firefox’s uBlock Origin support standing out beside link rot and renewed complaints about anti-user website design. Also in view: encrypted AI inference, law-enforcement pressure on secure software, home batteries cutting power prices, and a critique of AI-lab overconfidence.
-
Why does Opus 5 feel worse to work with?
-
mun-logadan.github.io
- 687 comments
A blog post argues that Anthropic’s Opus 5 coding model feels worse mainly because it no longer stops to clarify intent and instead barrels ahead. The comments mostly said the bigger practical problem is different: Opus 5’s writing has become dense, jargon-heavy, and exhausting to read, while many users also report more overreach, worse comments, and more switching to GPT/Codex or older Claude models.
-
-
GLM-5.3: Frontier coding with emergent cyber capabilities
-
z.ai
- 505 comments
Z.ai announced GLM-5.3, a coding-focused language model it says reaches near-frontier performance on software engineering and cybersecurity tasks while using the same base model as GLM-5.2 and getting the gains from post-training. Readers cared less about the benchmark charts than about two practical facts: it appears unusually willing to do offensive security work, and Z.ai says it will release the weights in two weeks.
-
-
Qwen 3.8 27B
-
huggingface.co
- 552 comments
Alibaba’s new Qwen 3.8 27B is an open-weight language model small enough to run locally on high-end consumer hardware, and its published benchmarks put it near much larger frontier models on coding and agent-style tasks. Early users say it is genuinely strong for a 27B model, but often slow, memory-hungry, and prone to spending huge numbers of tokens on reasoning unless you tune it down.
-
-
Firefox is now the last major browser that still supports uBlock Origin
-
pcworld.com
- 106 comments
A PCWorld piece says Firefox is now the only mainstream browser that still runs the full uBlock Origin ad blocker, while Chrome and other Chromium browsers keep moving to the more limited Manifest V3 model. The comments mostly treat that as a real strategic edge for Firefox, but also point out that Brave, Vivaldi, Edge, and uBlock Origin Lite muddy the headline.
-
-
Where did the old web go? We followed 657,607 links to find out
-
0.mk
- 207 comments
A revived URL shortener analyzed 657,607 links shared from 2009 to 2014 and found that about three quarters no longer load, a concrete snapshot of how much of the web has vanished or become unreachable. The comments turned that into a broader argument about what people mean by the “old web” and whether today’s AI-heavy, platform-controlled internet can ever recover its older feel.
-
-
In Australia, a home battery boom has helped cut wholesale power prices
-
e360.yale.edu
- 240 comments
A Yale Environment 360 piece says Australia’s surge in rooftop batteries is soaking up daytime solar oversupply and helping push down wholesale electricity prices. Commenters broadly agreed the effect is real, but argued hard over who captures the savings, whether home batteries are the best use of subsidy money, and how much this model depends on Australia’s unusually strong rooftop solar adoption and dynamic pricing.
-
-
When Genius Fails: The Intellectual Arrogance of the AI Labs
-
weightythoughts.com
- 184 comments
A blog post argues that AI lab leaders overestimate how easily intelligence in one field transfers to every other field, using recent AI hype and a young AI-focused hedge fund manager’s blowup as examples. Commenters largely agreed that the same arrogance shows up whenever outsiders claim they can automate jobs they do not understand, though some pushed back that the prediction can still be directionally right even if the current messengers are not.
-
-
Every Fucking Website (2020)
-
lxe.github.io
- 424 comments
A satirical 2020 webpage that recreates the worst habits of modern sites hit the front page and sparked a long venting session about what has made the web miserable to use. The useful signal was not just nostalgia. People zeroed in on which anti-user patterns are now standard, why companies keep shipping them, and where regulation and browsers have helped or failed.
-
-
Google is making private AI practical with homomorphic encryption
-
blog.google
- 159 comments
Google posted about HEIR, an open source toolkit for running AI inference on encrypted data using homomorphic encryption, so a cloud provider can compute on your prompt or records without seeing the plaintext. Readers mostly focused on whether this is anywhere near usable today, with benchmarks and prior work suggesting it is still dramatically slower than normal inference and mainly suited to narrow, high-value cases.
-
-
Going Dark, and the era of law enforcement hacking
-
blog.cryptographyengineering.com
- 98 comments
A cryptography blog post argues that AI-driven bug finding could soon make mainstream software and phones much harder for police and intelligence agencies to hack, pushing them to demand legal backdoors instead. Readers debated whether software will really get that secure, but mostly agreed the bigger risk is governments trying to turn covert hacking into explicit vendor cooperation.
-
-
France's top court blocks social media ban for under-15s
-
reuters.com
- 149 comments
Reuters reports that France’s top court blocked a law that would have barred children under 15 from using social media, ruling that the measure went too far in restricting freedom of expression and privacy. The comments mostly agreed with the outcome, but split hard on what should replace it: device-level parental controls, broader platform regulation, or simply better parenting.
-
-
Bluesky Protocol Services
-
atproto.com
- 64 comments
Bluesky posted a new set of managed infrastructure components for the AT Protocol, including a new Jetstream firehose service with history and replay, plus supporting relay and account-hosting services. Readers focused less on the announcement copy and more on what it says about ATProto becoming real backend infrastructure, along with sharp complaints about rough developer experience and operational transparency.
-
-
The TEMU-Fication of Software, Digital Goods and Services
-
xn--gckvb8fzb.com
- 101 comments
A blog post argues that AI is doing to software and digital content what Temu did to consumer goods: making it vastly cheaper and more abundant while pushing quality down and flooding markets with disposable output. Commenters largely agreed with the framing for media, but split on whether software will actually get worse or whether buyers simply will not care if the cheaper version works.
-
-
RustDesk now supports true unattended remote access on Wayland
-
rustdesk.com
- 91 comments
RustDesk posted that it now supports unattended remote access on Wayland, meaning you can reach a Linux desktop at the login screen and control it remotely without someone present. The comments were mostly positive on the feature itself, but quickly shifted to RustDesk’s security model, licensing status, and how it stacks up against VNC, RDP, Sunshine, and VPN-based setups.
-
-
How AI text watermarking works
-
declaude.org
- 97 comments
A post explains AI text watermarking as a hidden bias in word selection that lets a model provider later test whether a passage likely came from its system. The comments mostly treated it as easy to bypass in practice, but useful for platform operators who want to spot their own AI output and keep it out of training data.
-
-
Ruby 4.0 Universal RCE Deserialization Gadget Chain
-
elttam.com
- 22 comments
A security post shows a new Ruby 4.0 gadget chain that turns unsafe `Marshal.load` deserialization into remote code execution using classes that ship broadly with Ruby. The comments mostly treat it as a real exploit path for a long-known unsafe primitive, and focus on why removing one gadget at a time is not enough.
-
-
Maximizing the value of your Claude Code sessions
-
claude.com
- 83 comments
Anthropic published a guide on how to use Claude Code more cheaply and effectively, with advice on prompt caching, clearing context, file attachment, and session management. Readers mostly treated it as a sign that the product still makes users manage too many low-level implementation details by hand.
-
-
Don't classify, hallucinate
-
softwaredoug.com
- 84 comments
A blog post describes a product-classification trick: have an LLM invent a plausible taxonomy label for a query, embed that invented label, then match it to the closest real category. Readers found the idea familiar from retrieval methods like HyDE, useful as a cheap shortcut in large taxonomies, but shaky unless it is benchmarked against simpler embedding and reranking baselines.
-
-
I turned my RSS feeds into an e-ink newspaper to stop reading on my phone
-
heyjonny.dev
- 58 comments
A developer described turning RSS feeds into a daily e-ink “newspaper” to make casual reading feel more like print and less like doomscrolling on a phone. The conversation quickly turned into a practical guide to the current e-ink RSS stack, and a reality check on where the approach still breaks down.
-
-
Show HN: We Implemented the IPv8 Internet-Draft in Linux, Libc, and BGP
-
goonhost.rocks
- 59 comments
A hosting company posted a real lab implementation of a joke-like “IPv8” Internet-Draft, patching Linux, libc, routing tools, FRRouting, Nginx, and curl to see how the spec behaves in practice. The value is not the protocol proposal itself, which commenters largely dismissed, but the concrete failure report showing where a “just make IP look more like IPv4” redesign breaks on MTU, routing policy, hardware, and Internet governance.
-
-
WhatCable: Know what your USB-C cable can do
-
whatcable.uk
- 63 comments
WhatCable is a small utility site and app that asks your computer what a USB-C cable claims it can do, so you can sort mystery cables without specialized gear. The useful signal is practical but limited: commenters liked it for identifying charge-only vs data-capable cables, while repeatedly warning that software and cheap testers cannot verify high-speed performance claims.
-
-
an ambiguity in C89 which will never be fixed
-
sebsite.pw
- 45 comments
A blog post digs into a corner-case ambiguity in the 1989 C standard around old-style function declarations and implicit int behavior, arguing it can never be definitively resolved because the rule was removed in later standards. Commenters mostly treated it as a fossil from pre-prototype C, useful today mainly as a reminder of how legacy code and old toolchains still surface long-dead language traps.
-
-
I want extern "fil-C"
-
domenkozar.com
- 36 comments
A blog post argues Rust should be able to call Fil-C, a memory-safe C variant, through an `extern "fil-C"` interface so teams could use Rust for safe code and still wrap legacy C with stronger runtime protection than normal FFI. Comments focused less on whether that sounds nice in theory and more on the hard part: Fil-C is not C ABI compatible by design, and its runtime model makes a clean Rust bridge much harder than the post suggests.
-
-
Count Binface receives over a quarter of votes in Clacton by-election
-
bbc.com
- 342 comments
A satirical candidate called Count Binface took 26.9% of the vote in a UK by-election triggered by Nigel Farage, a high-profile right-wing politician under investigation over an undeclared £5 million gift. Commenters treated the result less as a serious near-upset and more as a public humiliation for Farage, because mainstream parties stayed out and Binface still kept the story focused on corruption rather than victory.
-