They/Them, agender-leaning scalie.

ADHD software developer with far too many hobbies/trades: AI, gamedev, webdev, programming language design, audio/video/data compression, software 3D, mass spectrometry, genomics.

Learning German (B2), Chinese (HSK 3-4ish), French (A2).

  • 0 Posts
  • 27 Comments
Joined 1 year ago
cake
Cake day: June 18th, 2023

help-circle

  • Western companies no longer operating in the Russian market, but still producing desirable content. … Western companies have ‘legalized’ piracy in Russia.

    100% this.

    Media is culture, and IMO people have a right to participate in culture. If it’s excessively difficult or impossible to legitimately access culture, one has the moral right to illegitimately access culture, and share it so others also have access.

    It’s inexcusable to refuse to directly sell media. The internet has made it easier than ever to trade access to media for money. Geo-restricted subscription services should be a nice add-on option for power-consumers, not the only way to get access to something.



  • The website does a bad job explaining what its current state actually is. Here’s the GitHub repo’s explanation:

    Memory Cache is a project that allows you to save a webpage while you’re browsing in Firefox as a PDF, and save it to a synchronized folder that can be used in conjunction with privateGPT to augment a local language model.

    So it’s just a way to get data from browser into privateGPT, which is:

    PrivateGPT is a production-ready AI project that allows you to ask questions about your documents using the power of Large Language Models (LLMs), even in scenarios without an Internet connection. The project provides an API offering all the primitives required to build private, context-aware AI applications.

    So basically something you can ask questions like “how much butter is needed for that recipe I saw last week?” and “what are the big trends across the news sites I’ve looked at recently?”. But eventually it’ll automatically summarize and data mine everything you look at to help you learn/explore.

    Neat.


  • I agree that older commercialized battery types aren’t so interesting, but my point was about all the battery types that haven’t had enough R&D yet to be commercially mass-produced.

    Power grids don’t care much about density - they can build batteries where land is cheap, and for fire control they need to artificially space out higher-density batteries anyway. There are heaps of known chemistries that might be cheaper per unit stored (molten salt batteries, flow batteries, and solid state batteries based on cheaper metals), but many only make sense for energy grid applications because they’re too big/heavy for anything portable.

    I’m saying it’s nuts that lithium ion is being used for cases where energy density isn’t important. It’s a bit like using bottled water on a farm because you don’t want to pay to get the nearby river water tested. It’s great that sodium ion could bring new economics to grid energy storage, but weird that the only reason it got developed in the first place was for a completely different industry.






  • I honestly don’t know what that silence would be like. I’ve spent my programming career jumping between domains, becoming an expert then moving on to find a new challenge. Now I’m building AI stuff for medicine.

    In my down time I learn languages, watch videos about physics and math, and play puzzle games.

    My brain actually won’t let me stop. Boredom = pain.




  • I think the big difference between people benefiting at small doses (~0.3mg) and large doses (2+mg) is that the 0.3mg group use it for sleep quality through the night, whereas the 3+mg people just need the sudden shock to get to sleep in the first place.

    The drawback with big doses is that your brain becomes less sensitive so your naturally-produced melatonin might not be enough to keep you asleep for the whole night after the pill wears off. It has a very short half-life in the body (under 1 hour), so there’s no way for a single dose before sleeping to last 8 hours. We naturally produce only 0.06-0.08mg per night, so it’s easy to see how supplementing melatonin could desensitize someone and cause them to wake up after just 4-6 hours of sleep.

    I have ADHD and am in the large-dose category and use 2-3mg of melatonin to help me fall asleep. Without it, I can’t sleep reliably because my brain often won’t shut up. Sleep reliably is so much more important to me than sleep quality.

    Using it only 5 nights a week, I’m not significantly dependent. I can still sleep without melatonin, just less reliably. I’ve tried 0.3mg, but it felt the same as taking nothing.

    For me, 10mg would be excessive and probably harmful in a desensitizing way. The most I’ve taken is 6mg, but it only helped in 2 out of 6 times. The other 4 times my brain just wouldn’t stop. If doubling my usual dose didn’t help, I don’t think doubling it again would be any different.

    There are however studies with higher doses, e.g. this one about kids with ADHD that says:

    two-third of the patients responded to relatively medium doses (2.5–6 mg/d), whereas doses above 6 mg added further benefit only in a small percentage of children.

    so I guess it’s different for everyone.








  • Yeah, I was over-enthusiastic based on their cherry-picked examples. SeamlessExpressive still leaves a lot to be desired.

    It has a limited range of emotions and can’t change emotion in the middle of the clip. It can’t produce the pitch shifts of someone talking excitedly, making the output sound monotonous. Background noise in the input causes a raspy, distorted output voice. Sighs, inter-sentence breaths, etc. aren’t reproduced. Sometimes the sentence pacing is just completely unnatural, with missing pauses or pauses in bad places (e.g. before the sentence-final verb in German).

    IMO their manual dataset creation is holding them back. If I was in this field, I would try to follow the LLM route: Start with a next-token predictor trained indiscriminately on large-scale speech+text data (e.g. TV shows, movies, news radio, all with subtitles even if the subs need to be AI generated), fine-tune it for specific tasks (mainly learning to predict and generate based on “style tokens” (speaker, emotion, accent, pacing)), then generate a massive “textbook” synthetic dataset. The translation aspect could be almost completely outsourced to LLMs or multilingual subtitles.