Skip Navigation

InitialsDiceBearhttps://github.com/dicebear/dicebearhttps://creativecommons.org/publicdomain/zero/1.0/„Initials” (https://github.com/dicebear/dicebear) by „DiceBear”, licensed under „CC0 1.0” (https://creativecommons.org/publicdomain/zero/1.0/)U
Posts
0
Comments
80
Joined
3 yr. ago

  • To be similarly pedantic: Ctrl+C is a hotkey that sends the corresponding ASCII code / codepoint to signal something, it is not an ASCII code itself.

    You could have the same character be sent by using Ctrl+Q (if you were to remap it), and not break compatibility with other processes while doing so: the codepoint being sent would be the same. From a technological perspective there is nothing special about the key combination Ctrl+C specifically, but altering this behavior in a terminal absolutely wreak havoc on the muscle memory of terminal users, and altering it's behavior in a text editor on everyone else's.

  • The key issue is that the request is to change behavior in one place (browser) to match that of a rare case (terminal), causing a mismatch with the frequent case (office suites, mail programs, ...). The terminal is the odd one out, not the browser, and ought be the one to change the default for the reason you provide.

    In practice, a terminal is a special case and not just a text input window, and current convention is that Ctrl + C aborts / cancels.

    (You could of course have a duplicate hotkey, but now you are inconsistent w.r.t. other browsers, and there will be someone else who will be annoyed by the difference)

  • Yep. LLMs are at their core text completion engines. We found out that when performing this completion, large enough models account for context enough to perform some tasks.

    For example, "The following example shows how to detect whether a point is within a triangle:", would likely be followed by code that does exactly that. The chatbot finetuning shifts this behavior to happen in a chat context, and makes this instruction following behavior more likely to trigger.

    In the end, it is a core part of the text completion that it performs. While these properties are usually beneficial (after all, the translation is also text that should adhere to grammar rules) when you have text that is at odds with itself, or chatbot-finetuned model is used, the text completion deviates from a translation.

  • Privacy concerns are valid when an external server needs to be queried, like if you were to use DeepL or Google Translate for this stuff, or for any LLM related muck, but they have been accounting for this already by making things work locally. For example, translations performed fully on device, and are an example of a feature I wanted.

    Like many here, the entire AI browser idea doesn't appeal to me at all, but I also struggle to come up with 'features their users want' if I take myself as an example. I have previously used Vivaldi, and while it is much more full featured, it doesn't add any features that I actually end up using frequently.

  • Note that, reading the article & a recent follow up, it was moreso serving more ads that drove them to make results worse, rather than AI: the article was published in 2024, and refers to events starting in 2019. GPT2 got released around that time, way before ChatGPT (2022).

    Still 100% enshittification though.

  • RAG is Retrieval Augmented Generation. It is a fancy way of saying "we've tacked a search engine onto the LLM so that it can query for and use the text of actual documents when generating text, so that the output is more likely to be correct and grounded in reality."

    And yeah, MCP stands for Model Context Protocol, and is essentially an API format optimized for LLMs, as you've said, to defer to something else to do the work. This can be a (RAG like) search engine lookup, using a calculator, or something else entirely.

    LLMs suck at doing a lot of stuff reliably (like calculations, making statements relating to recent events, ...), but they turn out to be quite a useful tool for translating between human and machine, and reasonably capable of stringing things together to get an answer.

  • I think it makes more sense if you view the country as an addict.

  • That seems vague though: when is something synthetic? If I use something like Context Free to create something, would that count? If you paint or take a picture, but apply some filters / effects? It, to me, feels too general to differentiate between tools that "fill in based on training data, and statistics thereof" and those that do not.

    While AI/GenAI is used for marketing, it does draw a clear boundary as to what necessitates the annotation, yet is general enough to include any future changes, while being widely recognized.

  • And it still cleans up once the ownership model indicates it can be cleaned up. That does not ensure memory is never leaked, but it is equivalent to destructors running automatically when using unique ptr or shared ptr without cycles in C++, which avoids at least a portion of possible memory leaks.

  • I feel like that comparison is a bit skewed, given that the original house (sudo, without the rs) was never torn down.

    Furthermore, rather than something easily replaceable or shallow, the language used is pretty integral to the software. Rather than paint, it would be like the building is made with untreated timber for framing. There might be termites that threaten the integrity of the building.

    If the building is unimportant you may not bother with the cost. But if it is somewhat important, reconstructing with better materials could be worth it even if new mistakes could be made during construction.

  • The weird thing is also what happens to taste of you lose your sense of smell. You can still feel the texture of the food, but it doesn't taste like much. One of the weirdest things I have experienced as a result of a virus.

  • Though, a big catch is that whatever is generated needs to be verified. The most recent story I've seen was the AI proposing the hypothesis of a particular drug increasing antigen presentation, which could turn cold tumors (those the immune system does not attack) into hot tumors (those the immune system does attack). The key news here is that this hypothesis was found to be correct, as an experiment has shown that said drug does have this effect. (link to Google's press release)

    The catch here is that I have not seen any info on how many hypotheses were generated to find this correct hypothesis. It doesn't have to be perfect: research often causes a hypothesis to be rejected, even if proposed by a person rather than AI. However signal-to-noise is still important for how game changing it will be. Like in this blogpost it can fail to identify a solution at all, or even return incorrect hypotheses. You can't simply use this data for further training the LLM, as it would only degrade the performance.

    There needs to be a verification and filtering first. Wikipedia has played such a role for a very long time, where editors reference sources, and verify the trustworthiness of these sources. If Wikipedia goes under because of this, either due to a lack of funding or due to a lack of editors, a very important source will be lost.

    • Progressive Web Apps - check the last update, it allows for the browser to create a shortcut to a webpage with minimal ui. (At least, on Windows)
    • Modern Tab Management - Depends on what you want exactly. But tab groups and vertical tabs have landed on stable already.
    • Not sure what you mean by Cross Site Scripting, XSS is usually an attack vector not a feature. If by macros you mean automation features, these do exist but are mostly focussed on developers & testing - you need an external application to use these APIs.
    • History Search - certainly could be improved, but I use it rarely enough as is.
    • (Improved) Privacy Containers - Is already being worked on
    • Granular Permissions - What exact permissions do you wish for there to be more granular? I personally don't experience this issue as I find a lot of permissions to be quite granular already.

    A lot of Web APIs are proposed by Google, giving them a head start in implementation. Furthermore, some not supported by choice due to privacy implications.

    The CEO pay however: I completely agree, it is insane how much the CEO is paid while responsible being for purchasing and shuttering services like Pocket.

  • And even that would be similar to someone cycling uphill, supported by the wind. Using air currents is efficient, but requires there to be air currents to use.

    Gliding is really efficient for the distance covered, however: rolling resistance of wheels and the air resistance of a (not as aerodynamic) person is higher (and hence less efficient) than birds.

  • I think you might have missed it, but if an extension is marked as compatible (by the developer) you can install extensions nowadays, no need to curate your own extension list. Here is the listing.

  • Instead the rule is: must be a member of the European Broadcasting Union, which is notably different from Europe or EU: some African countries are members too.

  • This has messed with me for the longest time. 24h just wraps around at 24, simple modulo 24 arithmetic.

    12h? The hour and am / pm wrap around independently, and hence I am always confused whether 12pm is supposed to be midnight or noon. Zero based would have made more sense (with x pm being x hours after noon...)

  • That difference is so large, they must be quoting different numbers. Something like DOJ is looking at Advertising providers or search providers alone, while Google quotes a number for percentage of all websites visited or something.