John Gruber says Claude's new watermarking secretly nudges word choices, and he is furious about it.
What the article says
- Anthropic is now hiding a watermark in everything Claude writes, to comply with a new EU rule.
- The trick tags words as green or red behind the scenes, then quietly favors green ones so the pattern can later prove text came from Claude.
- Gruber argues this can only make writing worse, since it means sometimes picking a weaker word over the model's true best choice.
- Only Anthropic holds the key to detect its own watermark, so nobody else, not universities, not writers, can check text themselves.
- He also digs into Google's similar SynthID system and says Anthropic's own explanation admits the trade off while downplaying it.
What HN is saying
- Most commenters push back hard, saying Gruber misunderstands how LLMs work since there is no single best word, just a probability roll every time.
- Several argue the watermarking method is mathematically clever and provably invisible, just changing where the randomness comes from, not the odds themselves.
- A recurring theory is that Gruber is really venting about EU regulation in general, not the watermark itself.
- Others share his worry from a different angle, noting any proofread text you didn't write yourself could now get falsely flagged as AI generated.
- A practical fear surfaces too, that everyone checking for watermarks has to send private text to every AI company just to find out.
A tiny local AI model matches last year's best cloud models, but it thinks so hard it takes 21 minutes to draw a bicycle.
What the article says
- Alibaba released a new open model small enough to run on a laptop, and it is shockingly capable at coding, vision, and general tasks.
- The catch is its default setting makes it overthink everything, spending many minutes reasoning about even a request to draw a plain circle.
- Turning that reasoning down or off gives answers in a couple of minutes instead, with barely any loss in quality for most tasks.
- It handles harder jobs well too, like spotting objects in photos and running a coding assistant on its own files.
- Even at full speed it feels slow compared to hosted services, though the community is already finding tricks to speed it up.
What HN is saying
- Commenters are floored that a seventeen gigabyte file can do this much on ordinary home hardware, calling it a real turning point for local models.
- Several argue that overthinking is not unique to this model. It is a side effect of how all current models are trained, rewarding thoroughness over efficiency.
- One person shared research suggesting even big proprietary models ramble internally just as much, we simply never see their hidden reasoning.
- The sharpest disagreement is practical. Some say cutting reasoning to low or medium keeps quality intact, others found it still produces broken results without enough thinking.
- A few compare it to rival small models, noting one competitor reasons far less but sometimes cuts corners, like breaking image loading with an unnecessary line of code.
A blogger proposes AI DR, meaning AI, didn't read, and the comments turn into a group therapy session about slop.
What the article says
- A writer says he now skips anything that looks like unedited AI output, even from people he respects.
- His new rule is simple. If you weren't bothered enough to edit it, he isn't bothered enough to read it.
- He carves out an exception for things like customer support, where fully automated replies are expected.
- He calls his approach AI DR, a play on the old too long, didn't read, aimed at AI slop instead of long posts.
What HN is saying
- Commenters compare dumping raw AI text on someone to ignoring a friend to stare at your phone, a new way to be rude.
- Several say the real fix is sharing the prompt instead of the output, since the prompt is the only part that actually carries the person's intent.
- One engineer describes coworkers burying pull requests in AI generated comments explaining code that was never even there.
- A sharper take argues quality should be the only bar. A well written AI piece is fine, a bad human one isn't better just for being human.
- A few point out this same complaint about AI slop got made almost word for word a few years back, so nothing's really new.
Cloudflare quietly adds tracking scripts to your site the moment you switch to their nameservers
What the article says
- A developer switched nameservers to Cloudflare just to serve files from a storage bucket, and found a tracking script had been added to his plain, script free site without asking.
- He had to open the analytics dashboard, turn the feature on, then turn it off again just to get rid of it.
- His complaint is simple. Features like this should require you to opt in, not force you to notice and opt out.
- This post is a warning to others switching to Cloudflare that the same thing will likely happen to their site.
What HN is saying
- A Cloudflare employee showed up and explained the tracking has been on by default for free accounts since last September, framed as a free performance monitoring perk, easy to turn off, and never on by default for paying customers.
- That explanation went over badly. Commenters said sites running on Cloudflare for years got changed with no warning banner, and called the reply corporate spin with no real answer for why free users get tracked by default and paying ones don't.
- Several people explained the mechanism. If your traffic is proxied through Cloudflare rather than just using their name servers, they terminate your encrypted connection and can rewrite your page's html on the way out, which is also how their caching and firewall features work.
- One person found the real trigger wasn't the tracking script at all but the proxy toggle itself, which is also on by default and easy to miss while switching providers in a hurry.
- A few people flagged a second surprise feature working the same way, one that auto obscures email addresses on your site by injecting its own script.
DuckDB is getting a server mode, triggers, and a 40x faster query engine this fall.
What the article says
- DuckDB v2.0 lands this fall with a new SQL parser, storage format, and a redesigned C API.
- A new client server mode called Quack lets any DuckDB talk to another DuckDB over the network, opening the door to long running multi user deployments.
- It finally gets triggers, so you can automatically log changes when data in a table gets updated.
- A faster type for messy JSON like data compresses better and queries much faster without you defining a schema first.
- The team rewrote how it reads and writes files so remote storage queries run dramatically faster, and one benchmark showed a recursive query running forty times quicker.
What HN is saying
- Commenters are genuinely fond of the tool, several say they run whole company platforms or data pipelines on top of it.
- One popular request is a built in way to keep tables in a set order, which could make compression and scanning even faster.
- Someone raised an eyebrow at ten thousand commits in six months and asked if AI coding tools were behind the pace, but another pointed out that messy commit histories with things like typo fixes inflate the count anyway.
- A few people compare it favorably to Postgres and MySQL, saying its query language just feels more thought through.
- Others are asking practical questions, like how it holds up for transactional workloads and whether it is stable enough to trust after past bugginess compared to ClickHouse.