A hybrid approach would work well too in some situations. When I worked in wetland restoration, one thing we did to reduce the ecological impact of herbicides was to cut the weeds a few inches above the water or soil, and then to paint herbicide directly on the cut face with a foam brush. This was a lot more labor intensive than spraying, but reduced the amount of herbicide used by 80-90%, and prevented any from landing in the water via over-spray. Even without the cutting step, applying herbicide directly to the weeds with a brush would be a big improvement over spraying.
The first software I made that had users other than myself was a Lemonade Stand clone for the TI-83 Plus. I shared it with people at my school using a link cable and, as there wasn't much competition in the calculator game market, it was a big hit. The next school year, I attended a drivers ed class at our crosstown rival high school and found that my game had somehow spread to that school as well! I have no idea how it happened, but it was very satisfying seeing something I made have a life of its own. Ben Brode was on the Search Engine podcast recently and shared a very similar story.[1] I wonder how many other programmers of our generation had this experience.
One thing that my game had that the original did not (and I don't know if I invented this or copied it from the version I played) was a hidden demand parameter that persisted between days. Advertising and word of mouth from satisfied customers would cause it increase, while customers going away empty-handed would cause it to decrease. That turned the game into a kind of clicker where the numbers could continue to go up and up. I also added the ability to save your game by serializing the game state to an image file, since that was the only way to save data on the TI-83 Plus that wasn't likely to be clobbered by another program.
I have a story from my high school physics teacher as well. Before he was a teacher, he worked at Kodak on digital imaging. His group had made a discovery, and after clearing it with management, they submitted a paper to a scientific journal. It was accepted, but just before the issue went to print, government censors decided that the idea had military applications and the paper was classified.* The issue had already been typeset and there was no time to make a new layout, so the journal published the issue with nine 'intentionally left blank' pages in the middle without explanation. He had the issue on his bookshelf and would show it off every year as the most significant achievement in his career. It was still classified at the time, so he couldn't say anything more about it.
* Kodak worked on spy satellites for the DoD, among other things, so the work could have been funded by the military as well.
The Library of Congress has an archive of every tweet from 2006 to 2017. In 2018, they began selectively archiving tweets. I can't find how selective they are exactly, but as far as I can tell, the project is still ongoing.
Chromatic aberration is a type of optical distortion and is a separate issue to color rendition. Color rendition refers to the lens's ability to transmit light equally across the color spectrum. If the lens is more transparent to red wavelengths than blue, images will look warmer, for example. Chromatic aberration, on the other hand, is a type of distortion in which a lens fails to focus all colors to the same convergence point. It will negatively affect image quality even if you use the lens to take black and white photos, since the result is a blurrier image.
> Color rendition is also impacted by the sensor, assuming digital cameras.
Particularly since the old lens was designed for film, perhaps even black and white film. The choice of film has a much larger impact on color rendition than the lens would have. Also, unlike sharpness, color rendition is highly subjective and easily corrected. If you're shooting to JPG and you don't like how the camera is interpreting the colors from a lens, most cameras allow you to customize the white balance.
Nick Carver, a photographer, had a really interesting video recently where he explained the progression of light at sunset. Prior to watching the video, I understood the primary phenomenon of Rayleigh scattering, so I didn't really expect to learn anything. However, there's a lot more going on as the sun sets than I realized. The properties of sunlight (color, hardness, direction) shift in interesting and surprising ways even after the sun falls below the horizon.
The segment begins at 7:10, but if you already have a basic familiarity with the science, you can probably skip to 12:00.
I've heard this called the curb cut effect. (It's a subject right in 99% Invisible's wheelhouse and there is a good episode about it that mostly focuses on the history of literal curb cuts.)
Spotify is the largest music streamer by far, being over twice as large as its next biggest competitor, Apple Music. It's natural that the company would be the biggest target for criticism, both for Spotify-specific problems and as a stand-in for the whole industry. Also, speaking as a mostly-satisfied subscriber, Spotify still has a lot of issues. The quality of the desktop client in particular is poor and it never seems to improve.
One thing that jumps out to me is that the white fur on the animals has a strong green tint due to the reflected light from the green surfaces. I wonder if the model learned this effect from behind the scenes photos of green screen film sets.
It is a challenge for these models to generate images of counterintuitive or unusual situations that aren't depicted in the training set. For example, if you ask for a small cube sitting on top of a large cube, you'll likely get the correct result on the first attempt. Ask for a large cube on a small cube and you'll probably get an image of them side-by-side or with the small cube on top instead. The models can generalize in impressive ways, but it's still limited.
I think that's 99% true right now, but it's not something that will be true forever. Almost all AI images that you see today are the result of a person typing a description into a prompt and getting an image back. If you're lucky, the person played with the prompt a bit until they got something that more closely matched their intention. However, there are tons of ways to integrate AI into a creative process in ways that allow an artist to iterate on an image and to be intentional about composition, color, style, and so on. Currently, it's only highly technical people who are able to do this, so the result is a lot of stuff that is technically impressive, but of low artistic value. (Quite similar to the output of the stereotypical camera nerd who has the latest camera and has its specs memorized, but takes uninteresting photos.) Eventually, these tools will find their way into the hands of artists and they will use it to make art.
I believe that when thinking about AI art, it's helpful to keep photography in mind as a source of analogies. For instance, both give anyone the ability to create images easily. Both gave existing artists heart attacks and were dismissed because they didn't require artistic training to use. Both resulted in a flood of images being created without thought or effort. The main difference is that we skipped straight to the smartphone camera age with AI.
I see a lot of negativity, but this is actually one of the first things I tried to do with ChatGPT when it got Bing integration. I was in an unfamiliar town for a wedding and I had forgotten my tie. I asked ChatGPT where I could buy one downtown. (If you're curious, it utterly failed, with two of the three suggestions not selling men's clothing at all, let alone ties.) I would love to be able to be on the highway and ask how far the next rest stop is or if there are any places coming up where I could get a sandwich. Google has that information, including menus in many cases, but it's difficult to access, especially while driving.
My proposed regulation is to introduce a hazardous vehicle license for large trucks and SUVs. Everyone is automatically granted one alongside their normal drivers license, but if you're caught driving recklessly (weaving through traffic on the highway, aggressively tailgating, speeding through residential or urban areas), you lose it and you have to drive a normal-sized car. Perhaps there's a way to earn it back through paying a fine and taking a safe driving course. The classification of vehicles as hazardous should be based on factors known to increase pedestrian and vehicle collision fatalities to encourage safe designs.
> Thinking about it, I wonder if both people picking multiple numbers and adding them all up mod N would improve things.
I bet that our bias against clustering and repetition when creating a pseudo-random list of numbers would mean that a list would be less random than an individual pick. As a simple example, if picking two numbers, people would probably be more likely to pick one even and one odd number than two even or odd numbers, since a mix of even and odd feels more authentically random, resulting in an excess of odd sums.
As an alternative, what if each person created a short list and then blindly chose one number from the other person's list?
> I suppose it's more appealing to our TikTok-conditioned minds.
Tik Tok directly influences the charts as well as video views are counted as streams. As a result, songs that are fun or funny to dance to for 20 seconds can quickly end up in the Top 10. Tik Tok also has had a significant indirect influence recently with songs and artists that go viral there finding mainstream success afterward, e.g. Lil Nas X, Doja Cat, and Dreams by Fleetwood Mac.
Some additional context: the editor of the manual was Sam Leffler, who was a coworker of Lasseter at Lucasfilm at the time.
> The later, more popular versions of the BSD Daemon were drawn by animation director John Lasseter beginning with an early greyscale drawing on the cover of the Unix System Manager's Manual published in 1984 by USENIX for 4.2BSD.[7] Its author/editor Sam Leffler (who had been a technical staff member at CSRG) and Lasseter were both employees of Lucasfilm at the time. About four years after this Lasseter drew his widely known take on the BSD Daemon for the cover of McKusick's co-authored 1988 book, The Design and Implementation of the 4.3BSD Operating System.[8] Lasseter drew a somewhat lesser-known running BSD Daemon for the 4.4BSD version of the book in 1994.
I would think that car rentals would be one of the slowest areas for electric adoption. You're immediately limited to people only intending to drive short distances, which is less common for car rentals than day-to-day car ownership. Of those, many will be people far from home where they're not familiar with the charging infrastructure. Maybe they can stay at a hotel with chargers, but what if there are only a couple and they don't get one one night? It just sounds like a way to make planning a trip more complicated.
I wish this would be addressed at the OS level. If a target popped into existence less than ~0.25 seconds before it was touched, a touch event shouldn't be generated. Humans reaction times aren't fast enough to hit a button that quickly anyway.
To borrow a phrase from Microsoft's history, "Embrace, Extend, Extinguish." AI proves to be incredibly useful and we welcome it like we welcomed the internet. It becomes deeply embedded in our lives and eventually in our bodies. One day, a generation is born that never experiences a thought that is not augmented by AI. Sometime later a generation is born that is more AI than human. Sometime later, there are no humans.
The Babylonians had similar ideas. Code of Hammurabi, Law 229: If a builder builds a house for someone, and does not construct it properly, and the house which he built falls in and kills its owner, then that builder shall be put to death.
Since they've gone to the trouble of protecting it with an insecure hash, couldn't they also have designed this hardware feature so that it could be completely disabled until the device is rebooted? This vulnerability doesn't persist through reboots, so it would be sufficient to have the firmware lock the feature out during startup outside of development or manufacturing contexts.
For 35mm, flatbeds just do not have comparable resolution to a digital camera with a good macro lens. The most you'll get out of the best flatbed is about 2300dpi (measured from resolved detail, not pixels). Generally, you end up with about 6-8MP of usable detail.
If you own a digital camera already and are comfortable tinkering with your setup and workflow at first, camera scanning is the way to go. You'll get almost the full detail that 35mm film has to offer in less time overall for about the same cost.
Something to consider is if you might have children and grandchildren who will stumble on them someday. One of my favorite things to do when I visited my grandparents was to go down into their basement and flip through their collections of Life, Popular Mechanics and National Geographic. It gave me a window into life in the decades before I was born in a way that history classes never could. As much as I love the Internet Archive, I would not have found clicking through a digital magazine nearly as absorbing.
> [W]e’re crawling up to a million pages a day, as you can see on our stats page.
> Given that Mwmbl is still relatively unknown, it seems plausible that we can reach our target of crawling three billion pages a day, to refresh the entire index in one month.
I think this is supposed to read "it seems plausible that we can reach our target of crawling three million pages a day."
That happened naturally as early Windows developed alongside rapidly improving hardware. The situation today is due to an arbitrary policy decision by Microsoft. Microsoft could easily create a SKU of Windows 11 that runs as well on ten year old hardware as Windows 10 does.