Darktable is the best photography editing and classifying software I've ever used. New users should be warned it's quite power-user centric though. Adopting isn't like opening up Lightroom and messing around. You may have to read a bit of documentation regarding the modules and different workflows (i.e. Sigmoid vs FilmicRGB).
I think the Ansel developer has a YouTube series with tutorials on how to edit in Darktable using the Filmic RGB workflow. Not sure if that's where I'd get started nowadays (I've just adopted Sigmoid and it's way quicker to edit with it), but it gave me a solid base in how to use this software.
Darktable is another piece of software that really shows warts in UX.
For example, you can right click empty space to "remove" or "update path to file..." in the left hand rail. There shouldn't be a right click option in empty space.
Everything lower case makes it very difficult to quickly parse, especially the settings menu.
Inkscape is also full of these sins. For that matter, so is KDE.
I tested literally all alternatives to Lightroom, including even other commercial Windows only software, just because I want to get off the subscription. I tested two specific functions I use and need, and zero of all other programs do them well. It's frustrating.
to add on: alternatives I am seriously considering is buying a macbook m2 purely for lightroom and run linux on intel, and macbook m2 + asahi + dumping the intel.
Is there an established way to take into account that people rate different types of restaurants on very different scales? "5 stars great burgers and cheap beer" vs "the avocado soufflé on the 6th course was such a letdown, no creativity at all, ruined my evening, 3 stars". I've seen particularly harsh reviews on fine dining restaurants that speak more about the entitlement of the customer than on the quality of the place.
Firstly: restaurants need to be divided into categories and ranked within the category.
5 star McDonald’s is worse food than 1 star fine dining. (Regardless of number of visitors.)
In the fast food burger category that Mickey D might be the beating a 3 star Burger King hands down in cleanliness, friendly customer service and timely orders.
Secondly: You need to filter out the average Joe’s.
Especially with regards to ethnic food, they try an amazing Pho and be like ”meh, too bland - the mexican one was better”. They’re two different kitchens! You can’t compare them.
Maybe I'm missing something obvious, but what is the idea behind quantizing and tokenizing time series? We tokenize text because text isn't numbers. In the case of time series, we're... turning numbers into less precise numbers? The benefit of scaling and centering is trivial and i guess all timeseries ML does it, but I don't see why we need a token after that.
I'm building upon insights from this paper (https://arxiv.org/pdf/2403.03950.pdf) and believe that classification can sometimes outperform regression, even when dealing with continuous output values. This is particularly true in scenarios where the output is noisy and may assume various values (multi modal). By treating the problem as classification over discrete bins, we can obtain an approximate distribution over these bins, rather than settling for a single, averaged value as regression would yield. This approach not only facilitates sampling but may also lead to more favorable loss landscapes. The linked paper in this comment provides more details of this idea.
Isn't it a given that classification would "outperform" regression, assuming n_classes < n_possible_continuous_labels?
Turning a regression problem into a classification problem bins the data, offers more examples per label, simplifying the problem, with a tradeoff in what granularity you can predict.
(It depends on what you mean by "outperform" since metrics for classification and regression aren't always comparable, but I think I'm following the meaning of your comment overall)
Tokenisation turns a continuous signal into a normalized discrete vocabulary: stock "went up a lot", "went up a little", "stayed flat". This smooths out noise and simplifies matching up similar but not identical signals.
> We tokenize text because text isn't numbers.
Text is actually numbers. People tried inputting UTF8 directly into transformers, but it doesn't work that well. Karpathy explains why:
Text can be represented by numbers but they aren't the same datatype. They don't support the same operations (addition, subtraction, multiplication, etc).
Interesting. Can you explain how this is superior and/or different from traditional DSP filters or other non-tokenization tricks in the signal processing field?
Traditional DSP filters still output a continuous signal. And it's a well-explored domain, hard to imagine any low-hanging fruit there.
My intuition is the following: transformers work really well for text, so we could try turning a time series into a "story" (limited vocabulary) and see what happens.
I think it could also have a connection with symbolic AI: The discrete tokens could be the symbols that many believe is useful or necessary for reasoning.
It is also useful for compression, reducing memory requirements by the quantization and small integer representations.
My primitive understanding is that we approximate a Markovian approach and indirectly model the transition probabilities just by working through tokens.
I really wonder if this is
1) Nostalgia.Obviously, the first tracks you heard were your introduction to the genre, so they will have emotional value.
2) First-mover advantage, for the lack of a better word. Obviously you perceive every new track t_i in the context of t_{0..i-1} that you already know. Maybe track t_443, which sounds cheap and commercial now, would have sounded innovative if you heard it as t_33?
3) Genre saturation. How many truly new tracks can be done in psytrance (or any other specific genre) while still being psytrance, before every track sounds like one that was done before? I don't envy the poor souls trying to express individuality in established, specialized genres.
I also wonder whether we could use deep learning to learn representations unbiased by our own history with the genres, and analyze this in a more unbiased way. But I'm also sure you can make the model say whatever you want if you bend the data the right way :)
The first time I heard goa trance it was in 90s, Astral Projection. It was too weird for me. I couldn't stand it!
Then in the early 00s I got introduced to psytrance. I started listening to it. Then I refound my love for goa trance (see my other post for examples) and I fell in love. It was an exciting time for me. I fell in love with two women, I used my first time MDMA and mushrooms later on, all of this while I learned to know this fantastic genre.
One could argue it was related to my first listening of AP. But I don't think I heard MWNN before, and I instantly loved it, and he has a very distinctive style. If you take Shakta, he's also made psytance, same with Miranda, MWNN, AP, etc etc but I don't like it.
Also, back in the 90s I listened to a lot of progressive trance as well (the usual names: Paul Oakenfold, Paul van Dyk, stuff mixed by Tiesto, Ferry Corsten, etc etc). How would you explain that?
Plus I like a lot more electronic music, mostly (but not solely) from 90s. I even like some psydub/psybient. But psytance? Nope. Well, barely.
I think the bar has been raised by what even casual producers can do with DAWs. I gave a listen to a classic drum and bass album recently (Wormhole by Ed Rush and Optical) and while I really enjoyed it, by the end of the album I can kind of tell how many of the various samples were used over and over again. The Wikipedia article[1] lists the hardware/software they used, and it looks like the recording computer was an Apple LCII. So a lot of what they were able to make seems like it would have been limited by what hardware they were able to finagle.
Compare that to a similarly layered and voiced project in a modern DAW. While the compute requirements have gone up drastically, the complexity the DAW has compressed into a single screen would make it trivial to make a similar album now (from a purely technical standpoint - the album is still the work of an artist).
I don't get the point the author is trying to make:
* The article starts off by arguing that Gantt charts are for factories not for information work; therefore, software providing "bygone ways" of project management (such as Gantt charts) is bad.
* Then it criticizes that software that masters multiple project management methodologies is also bad because SaaS companies are trying to make money by adding features.
* Then, we find that the issue is that project management software is a simplistic UI in front of a relational database. They don't work because project management is not a problem that can be put into databases.
* Then the problem is that the smartphone generation will
never understand relational databases because they are used to smooth UIs, and the problem is that there is no undo button.
* Next, the problem is that we aren't all thinking like managers, so PM methodologies... don't apply at all to us?
* However the supporting example instead says that the problem is that we aren't all thinking like developers, and developers dogfooding their PM software is not necessarily a great selling point for everyone else (this is a take I find somewhat convincing).
* Finally, the issue is poor planning, and no software will
ever solve it.
Maybe I am not commenting on the "most charitable reading" of this article, but it leaves me confused. The author is echoing a broadly-felt frustration with the world of PM tools. They further various criticisms of PM tools, but it feels like they struggle to find what their own criticism actually is.
The article hit all the main points I would use for a wholesale take-down of Taylorism and all the MBA BS that grew in its wake, but it's presented in a way that you already have to know where the author is coming from to know what they're saying.
The key point the article missed is that "management doesn't know jack". The author should have opened with that. Everyone thinks they're boss is an idiot, but why is that? This is important because Taylorism was based on bifurcating the thinkers and the doers into managers and workers respectively. Managers would create the schedule based on their unique knowledge and insight, and workers would carry out their directives without skepticism or doubt. In reality, employees have unique knowledge and insight that management does not possess. Gantt charts, PERT, were all created based on the assumption of management expertise. Knowledge work turned this dynamic around explicitly. Knowledge workers, by definition, know more than their managers. How is a manager going to create a work breakdown structure and a schedule for work they don't understand?
MBAs have the answer. The MBA perspective is that management its self is a discrete discipline that can be done in isolation from the work. A good manager, with the management skill set can just as easily manage a hospital, as they can an airport, a shoe factory, a fast food franchise, or a nuclear power plant. This is done through metrics. By measuring performance, the manager can know where and when to conduct corrective actions to satisfy metrics. The metrics are ultimately tied to progress on schedules, WBSs, Gantt charts etc. If these don't map to your projects, the metrics they generate are useless. They will guide managers to make misapplied corrective actions that miss the mark and prevent work from happening, rather than correct any under-performance.
For project management tools, doesn't matter that the interfaces are nicer. It doesn't matter how they're implemented on the back end. It doesn't matter if the companies that make them also use their own products. The fact that your project doesn't map well to a prescribed schedule, or that you're laser focused on KPIs that don't matter. None of that changes anything. The core assumptions these tools are based on are wrong. They don't care, they just want your money.
During my masters thesis in a chemistry lab, I got a side task to look at a data analysis script and make it run faster. It was a "C/C++" code (i.e. procedural C-style code using C++ stdlib for convenience) that read a file line by line and then fed it to a slow processing function, then aggregated the results. It took over a day to run.
Without even looking at the processing function, which I considered some sciency science, I set up pthreads and mutexes on the result array and such to reap almost perfectly linear scaling. So far, so good.
Then I ran a profiler to see what was actually taking so long.
... Uh, why are you spending all this time copying strings back and forth?
Turns out they passed all strings by value. Sprinkling in a few const & here and there got a 1000-fold speedup or such. I felt pretty stupid for my multithreading antics after that.
Great achievement. The title is of course misleading: The Orbitrap itself (the part that "fits in your hand") was hardly miniaturized, it's about the same size as a regular one [0]. The achievement is to miniaturize "mass, volume and" (IMO especially!) "power requirements" of the box around it (which, even miniaturized, does not fit in your hand). This runs at 41W and weighs 8kg. A commercial instrument runs at a total of ~2kW and weighs >100kg.
(Though in space they have the convenient advantage that no extra vacuum system is needed, which makes up a lot of space and energy consumption of these instruments here on Earth. The atmosphere on Europa is conveniently just about the "natural" operating conditions of an Orbitrap, which is required for its high accuracy.)
There have been a few new developments in "open-source Retool alternatives" recently. I tried out Openblocks, Illa Builder, Tooljet, Budibase, and Appsmith, with a toy project to test the suitability for a frontend for scientific software.
