timlod

FWIW, the performance considerations section is a little simplistic, and probably assumes that exact dataset/problem.

For GIN for example, perfomance depends a lot on the size of the search input (the fewer characters, the more rows to compare) as well as the number of rows/size of the index.

It also mentions GiST (another type of index which isn't mentioned anywhere else in the article)..

So, I checked, and this life cycle assessment (first link I could find, seems very comprehensive) states a 27x return on energy: https://www.vestas.com/en/sustainability/environment/lifecyc...

This as opposed to a tweet about someone who 'read a life cycle analysis article in some engineering journal like 10 years ago'.

Please don't spread misinformation.

I've used it before for realtime-uses (not production though where you'd need 100% guarantees for no drop-outs), latency has not been an issue. I think you essentially get the latency of the plugins you're using since this is a JUCE wrapper.

Ultimately it depends on how much work you do and how efficient an audio thread you built. pedalboard is not a library which does audio playback itself, it just effects buffers you give it. I used python-sounddevice, which are bindings for PortAudio - if you don't use much CPU you can comfortably run plugins in realtime.

Obviously you're still beholden to the GIL in Python (until further notice) so if worse comes to worst you might experience the unlucky dropout.

It's not virtual analog tech scientists building a better musical instruments :)

I had no clue what a fog harp is, turns out it's used for harvesting water - interesting tech!

Wow, I didn't see that. Great to see this level of interest early on!

Audio Source Separation I think is the general term used in research. It is often applied to musical audio though, where you want to do stem separation - that's source separation where you want to isolate audio stems, a term referring to audio from related groups of signals, e.g. drums (which can contain multiple individual signals, like one for each drum/cymbal).

The title is a bit confusing as open-source separation of ... reads like source separation, which this is not. Rather, it is a pitch detection algorithm which also classifies the instrument the pitch originated with.

I think it's really neat, but the results look like it could take more time to fix the output than using a manual approach (if really accurate results are required).

I'd say it depends on what your other options are. I don't think the technology will replace me soon, even at the rate I see it improving. At this point it's still a tool we can use to deliver faster, if we use it wisely.

Especially about ChatGPT et al. - I use it daily, but having the proper foundation to discern and verify its output shows me that it's still very far from being a competent programmer for any but the 'easy' tasks which have been solved hundreds of times over.

Like I hinted, I also view all of this hype sceptically. I dislike the 'we need AI in our org now!' types and am not planning on taking on projects if I don't see their viability. But there's obviously still a lot of demand and people offering services like those in TFA who're just looking to cash in, and that seems to work.

If you can find projects you believe you can make a difference in with your background, why not give it a shot?

Time will tell, but I would cautiously say say yes.

Note that I don't usually work in that particular space (I prefer simple solutions and don't follow the hype), didn't sell myself using 'AI' (I was referred), and also would always tell a client if I believe there isn't much sense in a particular ask.

This particular project really uniquely benefits from this technology and would be much harder, if possible at all, otherwise.

I'm a Data Scientist currently consulting for a project in the Real Estate space (utilizing LLMs). I understand the article is hyperbole for perhaps comedic purposes, and actually do perhaps 80% align with a lot of the authors views, but it's a bit much.

There is industry-changing tech which has become available, and many orgs are starting to grasp it. I won't deny that there's probably a large percentage of projects which fall under what the author describes, but these claims are doing a bit of a disservice to the legitimately amazing projects being worked on (and the competent people performing that work).

Pretty cool! Tried it with RHCPs Dani California (https://lamucal.ai/songs/red-hot-chili-peppers/dani-californ...) and there's a lot of wrong chords still. Impressive nonetheless, and already quite useful in the song-part recognition (assuming it's all the ML)! Lyrics seem right too.

The source separation only seems to be available when downloading their app, which I didn't do, so I can't comment on that.

I do like the Task Energy plot towards the end (second to last plot).

It shows how wasteful these new Intel processors are compared to their direct competition.

I’ve been using Windows throughout my childhood and start of my CS career - now I use Windows for specific software (audio/music) and Linux for developing (about 8 years I guess). I had a 1-year stint with macOS because I was developing an iOS app, and have been the troubleshooter for people with macs at my previous job, so I consider myself somewhat ‘multilingual’ when it concerns OSs.

As a power user, Linux is just so much nicer. I constantly get frustrated, especially with macOS, about stuff that I can’t easily. In Linux my stuff works and if it doesn’t it can be made to work (usually). In Windows/Mac it’ll often take considerable effort to make the system work the way I want, or it’s just not possible.

I think with proprietary software ‘it just works’ is only a thing if you’re happy with the basic experience that is tuned to the average person. If you have more complex needs, you should be using Linux (and if you know your stuff or use the right distro, things will likely also ‘just work’).

I see, I did misread it!

But I don’t think I can agree with what you meant then - why would our brain mostly ignore it in real life but not in video games? Where is the transition from something feeling real (in an immersive way) to us not liking it because it feels awkward, and why does it happen? I’m imagining a “perfect simulation” game which is like real life in ways that matter/don’t get in your way in terms of gameplay - I think everyone would be awed (of course this can be argued though). What would need to degrade in terms of realism for it to seem awkward/not be immersive anymore?

I agree with the movie example, but in a game you don’t have to watch the mundane - it’s just background “noise” to make the world believable.

Well, weird movements in games should be a thing of the past in the near future, as we can begin to extract motion capture data from videos of normal people acting normally.

I think it depends on the type of game you have, but I wouldn't underestimate this type of technology for say, open world games where it might make the game more immersive due to convincing realism.

That already exists in some forms, most notably perhaps in Dubler 2. I’m also working on an open source version of something similar, albeit quite involved and aiming at professional real-time performance and drumkit augmentation (if all goes well to share on HN this month for the first time :))

Recently released free short course (1 hour-ish minus the project) on getting structured output from LLMs. I found it quite useful to quickly get an idea of how this can be done!

They do make the tinnitus more noticeable for sure! By blocking out sounds which would otherwise “mask” the tinnitus, it can become noticeable if it’s otherwise pretty quiet. But in conditions where you really need earplugs I think you still won’t really notice it that much when wearing earplugs (very loud situations).

I've had Tinnitus since I was 14 (when I went to a concert and stood in front of the speakers).

A couple of pieces of advice to people who might be struggling with their tinnitus:

1. You need to learn to cope with it - once you're used to it, it will mostly fade into the background and be manageable. Accepting that it'll never be silent again was very difficult, but that's the only thing hat helped me feel better in the end.

2. Wear ear plugs when it gets too loud! It's too easy to get irreversible damage to your hearing, and that's the only thing you can really do - prevent it.

Curiously, yesterday I woke up at night because the tinnitus had gotten louder again - stupidly, I played drums the other day at a jam session without earplugs. I could punch myself for that one, and see it as (yet another) wakeup call to be more careful.

I actually started recording with S1 v1 in my teens because I got it for cheap. After some longer breaks, I kept coming back to it (and buying updates) - it really is good!

Compared with the Ableton workflow my bandmates are using nowadays, I still prefer S1.

Requires Ubuntu 23.04 and Wayland, so not necessarily applicable to many Linux users, but great to see more professional software being ported to Linux in any case!

Note: Studio One is a DAW (digital audio workstation). Other offerings that exist on Linux include Ardour, Bitwig and Reaper.

I use emacs-jupyter, which has async execution. Sorry for the late reply, only just now noticed this!

Do you have any source for scene reconstruction being hollywood-quality given 50+ images 10 years ago (assuming the images don't come from a controlled environment)? I can find for example this from 2016: https://substance3d.adobe.com/magazine/go-scan-the-world-pho... However, there's still a lot of manual work involved, and you don't easily get near-perfect PBR textures (which would be what I'd consider hollywood-quality). I'd say the devil is in the details - if you want the best quality, you'd still need to control a lot of the environment.

The more assumptions you can bake into the parameters of some model, the more degrees of freedom you get in the actual measurement process (e.g. reducing the amount of actual data necessary).

I've long switched to Emacs/Org, but used JupyterLab extensively before (as a data scientist). It's way more powerful than vanilla notebooks since you can open notebooks/code/related side-by-side, easier to extend (with lab extensions), etc.

I always thought people only still used vanilla notebooks because that's what people say they use, e.g. "I work with Jupyter notebooks" (even though that may well be in JupyterLab). So most regular users wouldn't necessarily know about JupyterLab.

SEEKING WORK | REMOTE or TOKYO

------------------------------

Data Science - freelance and consulting both possible.

https://loderhose.com (contains contact info)

https://www.linkedin.com/in/tim-loderhose/

Technologies: most tech relevant to Data Science/Engineering, e.g. PyTorch, pandas, scikit-learn, FastAPI, etc.

I have both BSc/MSc in Data Science and over 4 years of professional experience working as a Data Scientist with all kinds of data. I don't shy away from any data problem, and value reproducibility and code quality in my work. Feel free to ask for a quick chat!

I don't think it is, not for a business such as the one described in the article. I'm currently living in Japan, and in a relatively short amount of time I've personally met several business owners who might have acted similarly, sincerely.

There is a large amount of small and niche businesses here which are not profit-driven.

In case OP doesn't respond, I could probably help you with this - feel free to send an email!

I would recommend constructing some spectrograms from specific sounds, especially simulated ones, to help you connect the visual with the audible.

For example:

- Sine sweeps (a sine wave that starts at a low frequency and sweeps up to a high one) - to learn associate the frequencies you hear with the Y-axis

- Sine pulses at various frequencies - to better understand the time axis

- different types of noise (e.g. white)

Perhaps move on to your own voice as well, and try different scales (log or mel spectrograms, which are commonly used).

With this, I think you can develop a familiarity quickly!

This is a very clickbaity title. Over 6 years ago I worked as a student on reconstruction of 50 micron human brain (part, occipital lobe) MRI images, also captured in 9.4T. Obviously a 10x higher resolution is still impressive, but it’s a mouse brain, and it’s certainly nowhere close to 64M times the resolution of comparable prior work…

Some of those graphs are rather hard to pass, as the same domain is presented twice side-by-side, but with different X- and Y-axes. For example, comparing NN and linear regression models we see orders of magnitude difference on the Y-axis. Would really take the point of 'no change in RV model' home if they had the same scale.

Apart from that, many years ago, at a previous job, I prototyped automatic recognition of model degradation. I think one of the approaches I tried was comparing KL divergence between predictions on the test set and new data. This way you don't need to have the ground truth available (in which case you can make measures based on the residuals). Worked quite well iirc to signal when one should retrain and/or investigate context drift. Nowadays, with relative maturity in explainability methods, you can go much further and can much better find out what exactly causes the drift.

In general, I'm not a big fan of just saying 'error's up, let's retrain'. Why did the error go up? Understanding the data here better is usually quite valuable in the business context.