KTibow

I tried this once. It downloaded way more stuff than needed and still required admin to actually install.

Okay, but what does "applied" look like? Including a prompt?

It sounds like he actually orchestrates a "big night out" each time:

  One of the things I learned is that my review will be worse than useless if I am not having a good time the night I review it. I'll feel bad for underperforming, you'll feel bad for having a stressed, depressed drunkard on the line, and no one's knowledge of UX and the world is improved. So, I've made a few rules. One: I never drink alone. That means I need to ask friends whether they're up for a night out. I normally pay for their drinks, too. Two: I never schedule in a rush. That means that I now commit to a general two-week turnaround, but it can be longer than that, at times, and there's nothing I'm willing to do about it to make it faster.

Last time I checked the Fedora ISOs didn't include the device trees necessary to even begin installation.

I can see this comment was downvoted because it doesn't address the main point but Circle to Search is genuinely a good, helpful feature. It allows you to copy or translate text in two or three taps, even faster than if you had selection power, and I hope more platforms add similar functionality (even if just to work around the current terrible state of text selection).

It sounds like both the extension and bookmarklet don't send every page you visit but do send every page you check.

Vite/Rollup name assets based on one of the included files, which can lead to funny and misleading file names like this one.

This can still happen even with thinking models as long as the model outputs tokens in a sequence. Only way to fix would be to allow it to restart its response or switch to diffusion.

RouterBench is from March 2024.

How exactly would that work without the app having access to the pictures?

Android recently added an option that lets apps pop up a picker and only get access to the picked pictures. They probably just didn't realize that some users might want to only share some photos with Google Photos or didn't think the slice was big enough to justify implementing.

To add to the article: Gemma 3 270M's exact IFEval score is 51.2, and Qwen 3 would be at (0.6, 59.2) on the scatter plot.

I'm surprised that works even without telling it to think/think hard/think harder/ultrathink.

Second, in many environments (managed hosting etc.) there is not an easy way (or indeed a way at all) of adding headers to responses.

It's getting better. Most serverless hosts (including Cloudflare, which this site uses) follow the (req: Request) => Response pattern, which by definition allows sending headers.

This is also showing up on Xitter as the #keep4o movement, which some have criticized as being "oneshotted" or cases of LLM psychosis and emotional attachment.

It's likely the quantization on "Llama-3.2-1B-Instruct-q4f16_1-MLC". inference.net generated this more coherent answer: https://hst.sh/ovilewofox.md

I (Claude Code) tried this and it actually increased the gzipped size by 100b (3456 -> 3556), only reducing the non-compressed size by 20%, likely because gzip is really good at interning repeated patterns already.

My traffic has definitely gone down, I've been using DevDocs instead of MDN direct.

There's a slightly relevant response under "Dragging over your own computers is part of the fun of LAN parties. Why build them in?".

Thing is Atkinson Hyperlegible is "what if we made a non-monospace font with monospace like, distinct characters?" so the Mono version doesn't have much of a point. For text or code, it looks worse to me than the alternatives.

It sucks more that Cloudflare/similar have responded to this with "if your handshake fingerprints more like curl than like Chrome/Firefox, no access for you".

This is more or less a funnel to their Agentic Benchmark Checklist: https://arxiv.org/abs/2507.02825

My best guess is that these lines in the prompt were the root of the problem:

The second line was recently removed, per the GitHub: https://github.com/xai-org/grok-prompts/commit/c5de4a14feb50...

Maybe it's used to verify that the filter is loaded.

This is really interesting, although the cheapest route is still to use an alternative audio-compatible LLM (Gemini 2.0 Flash Lite, Phi 4 Multimodal) or an alternative host for Whisper (Deepinfra, Fal).

It doesn't exist. Turns out GitHub has a lot of private and deleted repos.

(Then there's 1000000002, https://github.com/minseon-01/flaskapp, similar to 999999999 in that it's a fork)

The Android client I use, Harmonic, has a shortcut to report a user, although it just prefills an email to hn@ycombinator.com.

These days, they'll sometimes also RL on a task if it's easy to validate outputs and if it seems worth the effort.

I'm not much of an ML engineer but I can point you to the original chain of thought paper [0] and Anthropic's docs on how to enable their official thinking scratchpad [1].

[0] https://arxiv.org/pdf/2201.11903

[1] https://docs.anthropic.com/en/docs/build-with-claude/extende...

With o3 you get something that feels like a human-written bug report, condensed to just present the findings, whereas with Sonnet 3.7 you get something like a stream of thought, or a work log.

This is likely because the author didn't give Claude a scratchpad or space to think, essentially forcing it to mix its thoughts with its report. I'd be interested to see if using the official thinking mechanism gives it enough space to get differing results.

Some newer ones like calc-size are also like this.