Skip to content

iFixit says the Switch 2 is even harder to repair than the original

Technology
126 71 3.9k
  • 14 Stimmen
    33 Beiträge
    276 Aufrufe
    I
    Who said they need to retrain? A small modification to their weights in each copy is enough. That's basically training with extra steps.
  • 0 Stimmen
    1 Beiträge
    19 Aufrufe
    Niemand hat geantwortet
  • 1 Stimmen
    1 Beiträge
    21 Aufrufe
    Niemand hat geantwortet
  • 253 Stimmen
    42 Beiträge
    418 Aufrufe
    dojan@pawb.socialD
    Don’t assume evil when stupidity I didn't, though? I think that perhaps you missed the "I don’t think necessarily that people who perpetuate this problem are doing so out of malice" part. Scream racism all you want but you’re cheapening the meaning of the word and you’re not doing anyone a favor. I didn't invent this term. Darker patches on darker skin are harder to detect, just as facial features in the dark, on dark skin are garder to detect because there is literally less light to work with Computers don't see things the way we do. That's why steganography can be imperceptible to the human eye, and why adversarial examples work when the differences cannot be seen by humans. If a model is struggling at doing its job it's because the data is bad, be it the input data, or the training data. Historically one significant contributor has been that the datasets aren't particularly diverse, and white men end up as the default. It's why all the "AI" companies popped in "ethnically ambiguous" and other words into their prompts to coax their image generators into generating people that weren't white, and subsequently why these image generators gave us ethnically ambigaus memes and German nazi soldiers that were black.
  • Biotech uses fermentation to produce milk proteins without cows

    Technology technology
    26
    199 Stimmen
    26 Beiträge
    324 Aufrufe
    M
    Alpro Not Milk comes pretty close for me, oat drink.
  • 310 Stimmen
    37 Beiträge
    366 Aufrufe
    S
    Same, especially when searching technical or niche topics. Since there aren't a ton of results specific to the topic, mostly semi-related results will appear in the first page or two of a regular (non-Gemini) Google search, just due to the higher popularity of those webpages compared to the relevant webpages. Even the relevant webpages will have lots of non-relevant or semi-relevant information surrounding the answer I'm looking for. I don't know enough about it to be sure, but Gemini is probably just scraping a handful of websites on the first page, and since most of those are only semi-related, the resulting summary is a classic example of garbage in, garbage out. I also think there's probably something in the code that looks for information that is shared across multiple sources and prioritizing that over something that's only on one particular page (possibly the sole result with the information you need). Then, it phrases the summary as a direct answer to your query, misrepresenting the actual information on the pages they scraped. At least Gemini gives sources, I guess. The thing that gets on my nerves the most is how often I see people quote the summary as proof of something without checking the sources. It was bad before the rollout of Gemini, but at least back then Google was mostly scraping text and presenting it with little modification, along with a direct link to the webpage. Now, it's an LLM generating text phrased as a direct answer to a question (that was also AI-generated from your search query) using AI-summarized data points scraped from multiple webpages. It's obfuscating the source material further, but I also can't help but feel like it exposes a little of the behind-the-scenes fuckery Google has been doing for years before Gemini. How it bastardizes your query by interpreting it into a question, and then prioritizes homogeneous results that agree on the "answer" to your "question". For years they've been doing this to a certain extent, they just didn't share how they interpreted your query.
  • 157 Stimmen
    12 Beiträge
    109 Aufrufe
    W
    that's not just useless defeatism, but also false. effective end to end encryption exists in multiple forms today. signal, maybe even with a custom server. matrix if the server is being ran on trusted hardware. XMPP too with the right extensions.
  • 203 Stimmen
    6 Beiträge
    64 Aufrufe
    C
    One could say it's their fiduciary duty.