nowahe

Ah I was wondering why I couldn't get past 600 nits on my M5 why it worked great on my M1. Guess I'll just have to live without it for now

By any chance, do you know what Claude Code's sandbox feature uses under the hood and how that relates to your solution ? From what I remember it also uses the native MacOS sandbox framework, but I haven't looked too deep into it and don't trust it fully

IIRC from a comment in another thread, it's marked as deprecated to stop people from using it directly and to use the offical macOS tools directly. But it's still used internally by macOS.

And I think that what CC's /sandbox uses on a Mac

I'm not a native English speaker, but I read a lot of books in their original English. Being able to quickly look up a word you've never encountered is a god send

I find myself really missing this feature when I occasionally read a paper book, thinking about clicking the word on the page to get a definition.

I've downloaded a few from yggtorrent, but there might be some more niche/less public sites I'm not aware of

The weird thing is that when I tried to tell it to distill it to a much smaller message it had no problem outputting it without any followup questions. But when I edited my message to ask it to generate a larger response, then I got stuck in the loop of it asking if I was really sure or telling me that `I apologize, but I noticed this request would result in a very large response.`

It sparks me as odd, because I've had quite a few times where it would generate me a response over multiple messages (since it was hitting its max message length) without any second-guessing or issue.

I've had it refuse to generate a long text response (I was trying to concise a 300kb documentation to 20-30kb to be able to put it in the project's context), and every time I asked it replied "How should structure the results ?", "Shall I go ahead with writing the artifacts now ?", etc.

It wasn't even during the over-capacity event I don't think, and I'm a pro user.

Do you happen to have the name of the place you found it ? I'm in France, and I wasn't really able to find anything with a quick google search

I also found that passing all the codebase in a single file in the context works really well. I've tried Cursor et al, but I found that it not having the full context of the codebase (and having to do some back and forth for it to requests files) was slower and didn't really yield any better results. Granted I work on projects where the codebase fit in 100-200kb text files, but I'm still only at 20% of the context limit in Claude.

Also I found the UX of Claude to be better for this, especially their Projects feature. I can just put the codebase in the Project's context, and start a new conversation to ask different questions/solve different problems.

The only pain point I have is that it seems to be pretty optimized to only show changes in existing files, not rewriting them in full, which is a bit of a pain to copy-paste into my IDE. I'll see if I can write out a system prompt to force it to generate diff or a similar format that could more easily be applied automatically to my code.

I switched to an Ergodox-EZ about 3 years ago, and I don't regret doing it at all. I didn't switch layout tho, stayed on AZERTY.

My experience is that at first it was very frustrating. I went from ~80WPM on my macbook keyboard to ~20WPM on the ergodox. After a couple of weeks I was able to write text at a comfortable speed again, but any special character was painfully slow, as I had to consciously think where each character was, and often look it up on my layout. After about 3 months I was back up to 80WPM.

What took a long time as well, was configuring my layout to fit my programming needs, it took me about 6 months to come up with a layout that had everything I needed (you can see it here if you're curious[1]). My recommendation is to do it incrementally, trying with something general at first, then with use seeing what feels right and doesn't.

In the end it was a really good idea, a lot of back pain I had has gone away, and after long typing session I have way less pain in my wrists and hands.

[1]: https://configure.zsa.io/ergodox-ez/layouts/BOLz0/ybXMx/0

If you haven't seen it already, you might enjoy this story from gwern : https://gwern.net/fiction/clippy

Auth0 seems to be down as well, we can't use the login forms on our websites nor login onto the management console itself. Getting timeouts everywhere

I found a tool to do just that, iCloud Photo Downloader[1] (tho you do need to have your photos backed up to iCloud).

I run it as a cron on my NAS, and the only intereaction I need to have with it is to give it a new 2FA code every ~2-3 months. Highly recommend it

[1]: https://github.com/icloud-photos-downloader/icloud_photos_do...

Out of curiosity, how does your Image Similarity Search works ? Are you also using some feature of Apple's Vision framework, or running some ML model on your linode instance ?

Highly recommend this as well, it has been a godsend for me for digging into multi-GB Cloudfront log files and extracting stats and patterns.

There is a good (albeit a bit old now, as the interface has been overhauled since) quick demo video from VD's creator : https://www.youtube.com/watch?v=N1CBDTgGtOU

Looks like the image in the article is a crop of this[1] view, which can be accessed in high res.

You can find all the JWT images in full res on here[2].

[1]: https://stsci-opo.org/STScI-01GCVNZ68YTC7FPTBSNA3QDGYW.png

[2]: https://webbtelescope.org/resource-gallery/images

Basically, to validate a block, you need to bruteforce a SHA-256 hash of the block by incrementing a nonce in it. And by essence, bruteforceing is wasteful (and by extension, expensive).

And this is by design, as the only way to mint a new bitcoin is to throw away computational power (ie energy -> money). And the amount of power needing to be wasted is constantly adjusted by the network (it's targeting a certain amount of blocks / hour, adjusting the difficulty of the sha bruteforce, compensating for technological improvement).

Now, to create a bitcoin you need to mine a block (solving the bruteforce), inherently requiring a set (on average) amount of real world value (mostly energy) to be irrevocably wasted. For the miners to recoup those losses, they MUST sell the bitcoin they just created for at least their lost value. Which in turn, guarantee the minimum value of each bitcoin.

And with this system, the minimum value of each bitcoin is inversely equal to the amount of value "wasted".

I have a feeling that this breach is older than what they're letting on.

On July 27th, I received ~7 emails, about 10 minutes apart, warning me of a new device logging in my Plex account. It didn't correlate with any activity on my part, and the IPs were all over the place (for context I'm in France). Here is some of the IPs that were used :

  - 191.101.41.35 (US)
  - 185.199.103.40 (US)
  - 103.43.200.58 (India)
  - 2001:16a2:def3:200:40cf:530f:ff72:1747 (Saudi Arabia)

Fortunately the password is only used on Plex, and I just generated a new one and signed out my devices, and that was it.

(Semi)surprisingly, Strasbourg looks like a good contender as well, as it covers a good portion of the blue banana, most of Germany, and about half of France

A couple of years ago I started architecting transcoding/streaming video pipelines on AWS for my work. It sparked a general interest/curiosity into the internal workings of video compression algorithms, which then allowed me to revisit a subject I've always wanted to try out, datamoshing. Something about purposefully breaking codecs to generate artifacts that lets you peak at their inner workings really nerd sniped me.

The most advanced video I was able to create is this datamoshed version of Carpenter Brut - Turbo Killer clip : https://streamable.com/7d9h4b. I'm particularly proud/pleased with the transition at the 2:00 minute mark. It still a long way from what I would want to be able to accomplish (as the moshing is still pretty random and looks good only by "chance"), but I'm pretty happy with my progress overall.

My long term goal for it would be able to create moshing of music video clips, and sync the glitching to the beat/structure of the music (using Spotify Audio Analysis feature [0][1]).

As a sidenote, a tool I discovered and has been a godsend into tinkering with video codecs is FFglitch[2]. It's a fork of FFmpeg with the ability to directly access and modify internal MPEG2/4 values (such as the Motion Vectors, macroblocks, DCT coefficient, etc). Compared to other datamoshing techniques (such as corrupting AVI files with random data, manual key frames deletion, etc) it's order of magnitudes more advanced and precise, allowing you to precisely and purposefully influence the codec, instead of randomly breaking it and seeing what sticks. However its downside is that you have to have a solid grasp of the inner workings of those codecs to achieve anything.

Finally, for anyone looking for a good resource to get an understanding of Video Encoding, I highly recommend this[3] github repo as a starting point. It contains a good description of the workings, as well as an amazing references list.

[0]: https://developer.spotify.com/documentation/web-api/referenc...

[1]: https://spotify-audio-analysis.glitch.me/

[2]: https://ffglitch.org/

[3]: https://github.com/leandromoreira/digital_video_introduction

From personal experience, I feel like the crash following a multi-day binge isn't necessarily due to amphetamine withdrawls, but more so the consequences of multi-day sleep deprivation, dehydration, under-nutrition; all the while having over-exerted yourself.

On the comedown of a multi-day binge, I don't feel depressed per se. I'm constantly tired while awake, which I can only keep at bay if I'm physically active. Also, my ability to focus and get motivated dips below baseline, inversely proportional to how high over baseline I was. (And I would speculate that it is mostly due to a depletion of dopamine & its precursors rather than a withdrawl, as it can be alleviated through a decent diet).

From what I've read as well, most of the long term side effects of amphetamine abuse stem from chronic lack of sleep, food & water rather than damage from the molecule itself. Although repeated overstimulation of your reward/dopaminergic system definitely can't be good for you, be it through adaptation or damage of the dopaminergic neurons (I remember reading somewhere that long term meth users are more prone to Parkinson's).

As a semi-regular user of speed, I can almost guarantee you that a gram of street amphetamine is nowhere near 100% concentrated. You would be lucky to find it at 20% (cut with half caffeine half random white powder). I've done acetone washes of amphetamine paste, and it's almost unbelievable how little amph there is in it.

A few times I was able to get my hands on really pure dextroamphetamine (>90% validated by a lab test), and even with a tolerance, there is no way you're consuming more than 100mg over the course of a day. 10-15mg was my sweetspot to get enough focus without getting locked into tasks and feeling wired. Although speed like that is pretty hard to find and quite pricey (~60-70€/g).

I've also realized that most of the unpleasant physical side effects of speed (palpitations, jittering, elevated cardiac rythm, anxiety, etc) are actually due to the massive amount of caffeine in it and not the amphetamine themselves (tho I'm sure there is a synergy between them). Even while feeling pretty high on pure amphetamine, my bpm only rose by about 4-5 bpms (compared to 15+ on street speed). For anyone in the same case, L-theanine works wonders to calm down the caffeine side effects, highly recommend you keep some around just in case.

You might be interested by the book "Consciousness and the Brain" by Stanislas Dehaene. In it, Dehaene tries to define Consciousness, based on a series of reproducible experiments designed to discriminate between conscious and unconscious processes.

For example, he is able to show that our unconscious processes that are responsable for parsing text and numbers are able to assign conceptual meaning to words/numbers (ie "five" and "5" lead to the same meaning), before any conscious awareness.

Although my summary hardly does it justice (as I'm still half-way through reading it), I highly recommend Scott Alexender's book review on Astral Codex Ten[0], it gives an exhaustive overview of the book and the subjects it covers.

[0]: https://astralcodexten.substack.com/p/your-book-review-consc...

On macos, whenever I have a high memory pressure, coupled to a large swap usage, I find that the system becomes really sensitive to I/O intensive tasks.

If I start an IO intensive task (eg a recursive grep or a backup), any task that involves pulling data from the swap (eg switching desktops) gets significantly slower, and I often see the dreaded beach ball

I've tried, but for me limiting myself to the default/whole integer scaling is more of a pain than the bugs I encounter with it.

With 4k, at 1:1 I basically have two choices of resolution : either 3840x2160 native (which results in tiny text, even on a 32") or 1920x1080 HiDPI (which is way too small of an effective resolution, especially at 32").

With 1:1.5 scaling, I can get 2560x1440p HiDPI, which seems to be more stable (and renders text more crisply), but still a tad too small for what I need.

A 5k display would fix those issues, but the price difference between 4k & 5k is way too steep to me for the marginal improvement I would get out of it.

However, I've never had an issue with 4k video playback, even on non integer scaled resolutions. What's more, they virtually never drop frames, even with the system under load. My bet is that their hardware video decoding pipeline is significantly better that what's present on Intel CPUs, and is able to better delegate video decoding to the hardware decoders.

I've never pushed it enough for it to completely lock up, but I've had some weird bugs/crash happen to me. Like all webviews crashing at the same time (all electron apps + all browsers), with an impossibility to run anything (commands would hang, trying to open apps didn't do anything, etc). But what was uncanny was that the UI would happily run smoothly at 120hz, with all the menus still working and being responsive.

However I've had a lot of issues with WindowServer. I have an external 4k display connected to my mac (Macbook Pro 2021 14"), which I usually scale down to 3200x1800 HiDPI. And 2-3 days after a fresh boot, I'll start to have issues such as the WindowServer process jumping to around 140-180% CPU, and staying there until I reboot (or restart the process). I'll also randomly have my cursor start "glitching" and jump around in a 20px radius around the place I want to point; but only on my external screen.

My feeling is that it has something to do with the non-integer scaling I'm doing , and WindowServer entering at some point a state which introduces rounding errors (which would explain the glitching cursor) or something like that. Also, switching resolutions (or even refresh rate) of the external display will temporarily fix the cursor glitching issue.

But my biggest complain, by far, is that there is virtually zero way for me to troubleshoot/debug the problem. As far as I can tell, there is no way for a debugger to attach itself to the WindowServer process, no documentation on internal workings of the rendering pipeline, virtually no logs, etc. I've been trying to narrow down those issues for 6-8 months, and I've stilled not been able to track it down

On a (sad) related note, a few month ago[1] the devs announced the end of active development on RES, and from now on will only receive bugfixes, if any at all.

Which is understandable considering that since the redesign, they have been more and more working against a black box, with the Reddit staff actively trying to prevent access to certain functionalities outside of their clients.

I'm still worried about Reddit turning off old.reddit.com one day, or making some major breaking change crippling RES. Tho it could be a good way to quit using it, as the friction of the new design overpowering by far my desire to go on it.

[1]: https://old.reddit.com/r/RESAnnouncements/comments/sh83gx/an...

I spend a lot of time remoting into fresh *nix systems, so I also almost don't have any aliases, with one notable exception : ll (aliased to ls -lah).

It's just so engrained into my muscle memory that I do it without thinking about it most of the time.

And the workaround I found for it is adding a macro on my keyboard (through QMK but can be done with anything) that just types out 'alias ll="ls -lah"\n'.

Having bought a few artist compilation on it, after clicking 5 links manually I resorted to writing a script to parse the webpage and get the links there.

But what's even worse is that their CDN throttles your IP after ~15 downloads and you have to wait 1-2 hours before you can download again (which I've already hit a few times, even while clicking on the links manually)

If you have a kobo, I highly recommend checking out koreader[1], which is an open-source ebook reader and is highly configurable (to the point where you can even write custom css rules from withing the e-reader), integrates with calibre natively (you can even connect to your Calibre library over Wifi), has a solid support for CBZ (with features like auto-cropping the border around your comic, automatically or manually), and more.

Having used it, I don't think I could go back to Kobo's reader (and even less to Amazon's, which is a joke in term of configuration options).

[1] : https://github.com/koreader/koreader