Greg's recent activity

  1. Comment on Hackers can read private AI-assistant chats even though they’re encrypted in ~tech

    Greg
    Link
    I love this kind of research: it’s impressive that they managed it at all, it exposes a flaw that a lot of people just wouldn’t have thought of, and as far as I can see it gives a very...

    I love this kind of research: it’s impressive that they managed it at all, it exposes a flaw that a lot of people just wouldn’t have thought of, and as far as I can see it gives a very straightforward mitigation in just padding or chunking tokens to disguise their length.

    Haven’t had a chance to read in detail yet, but it’s interesting that Google’s platform is already immune - they do love their protobufs, so I’m thinking maybe that’s inherently obscuring the individual tokens?

    7 votes
  2. Comment on Fun programming challenge: figure out which sets of passports grant visa-free access to the whole world in ~comp

  3. Comment on Unique things to do in Las Vegas? (and Los Angeles) in ~travel

    Greg
    Link Parent
    I wandered into the Neon Museum pretty much just on the basis that I'd seen one or two cool and vaguely apocalyptic looking photos from there, and ended up having an amazing hour long chat with...

    I wandered into the Neon Museum pretty much just on the basis that I'd seen one or two cool and vaguely apocalyptic looking photos from there, and ended up having an amazing hour long chat with one of the staff. My answer to "let me know if you have any questions" was something like "I honestly don't even have enough context to know what to ask - what do you think is most interesting here?" and she just lit up, talked me through the historical context, the engineering, the weird social circumstances that led to huge fucking neon signs being a key part of the local culture, everything. Pure good luck to catch the right person in the right mood on a quiet afternoon, but it was a great experience!

    1 vote
  4. Comment on Bitcoin tops $72,000 for the first time in ~finance

    Greg
    Link Parent
    I guess it depends on what you’re counting as an advocate; the loudest and most numerous generally fall into one or more of those buckets, but there are plenty of serious financial types gambling...

    I guess it depends on what you’re counting as an advocate; the loudest and most numerous generally fall into one or more of those buckets, but there are plenty of serious financial types gambling on crypto just as they do on any other asset class.

    I see it as interesting tech that could’ve found a niche, and instead became a microcosm of our insane financial system as a whole. The numbers are made up, but not substantially more so than the ones underlying some similarly volatile derivatives.

    9 votes
  5. Comment on Recommendations for wireless earbuds for extended PC use? in ~tech

    Greg
    Link Parent
    I use mine for gaming sometimes because apparently the mic in them is much clearer for everyone else than the one in my bigger headphones, and they’ve been pretty much fine with Windows once...

    I use mine for gaming sometimes because apparently the mic in them is much clearer for everyone else than the one in my bigger headphones, and they’ve been pretty much fine with Windows once they’re connected.

    There seems to be some kind of sync issue about half the time when they power on - took a while to work out but it seems like they connect to Windows individually as you take each one out of the case, and then the audio comes through with each ear a few hundred milliseconds off from the other, which is enough to be unusably distracting. Good news is that just hitting disconnect/reconnect from the Windows Bluetooth settings once they’re in your ears seems to reliably sync them up again. I’ve had a million issues with Windows audio over the years in general, across a whole range of hardware, so that counts as fine based on the expectations I’ve got here!

    4 votes
  6. Comment on Recommendations for wireless earbuds for extended PC use? in ~tech

    Greg
    Link
    Possibly a stupid question, but if latency, charging, and cost are all important would wired be an option? That’ll likely get you the best combination of those three things if it’s a possibility.

    Possibly a stupid question, but if latency, charging, and cost are all important would wired be an option? That’ll likely get you the best combination of those three things if it’s a possibility.

    2 votes
  7. Comment on Unique things to do in Las Vegas? (and Los Angeles) in ~travel

    Greg
    Link Parent
    The pinball museum is less a traditional museum, more several hundred playable tables in a warehouse - which as far as I’m concerned is goddamn perfect and by far the best entertainment you can...

    The pinball museum is less a traditional museum, more several hundred playable tables in a warehouse - which as far as I’m concerned is goddamn perfect and by far the best entertainment you can get for $10 of quarters in Vegas, but how much that appeals will definitely depend on your group!

    At the other end of the spectrum, taking a helicopter tour into the grand canyon was truly incredible and fully worth splashing out on in my opinion, but $400 is a lot to spend whichever way you slice it.

    6 votes
  8. Comment on British pubs keep getting demolished and rebuilt in ~design

    Greg
    (edited )
    Link Parent
    Exactly where and how to draw the lines around preservation vs new development will always be at least a little contentious, but I think the balance here is actually pretty good. The planning...

    Exactly where and how to draw the lines around preservation vs new development will always be at least a little contentious, but I think the balance here is actually pretty good. The planning system as a whole is definitely far from perfect, and the broader incentives around housing are an absolute mess in a lot of ways, but I don’t generally see historic buildings as a significant part of the issue.

    The rulings around these pubs aren’t the norm, they’re newsworthy because the local councils did specifically decide to protect them, the developers tried to pull their whole “oops, looks like it burned down by complete unavoidable accident, how terrible, I guess I may as well build my original plans on the site now” bullshit, and rather than rolling over and accepting it as fait accompli the councils have started saying “like fuck you will, you’re paying to put everything back exactly how you found it”.

    It’s pretty much the platonic ideal of justice as far as I’m concerned: person who thinks they’re rich enough to be above the law and/or is trying to treat fines as a cost of doing business gets slapped right back down in a way that costs them enough to discourage others from chancing it and makes the community whole. It’s precisely the right message to be sending: fines aren’t just a cost, the laws aren’t there just to set the price, they’re there because you’re not supposed to do the thing - and while it’s a bit of a shame that enforcing that is even notable, I’m sure as hell glad that it’s happening.


    [Edit] Just to address this bit specifically, because it’s particularly relevant:

    Compromises can be made to maintain the historical facade and build a taller building through it.

    A decent number of these kind of cases were already supposed to be that kind of compromise. The developer applied for permission to build, it was granted with a requirement to preserve historic parts in a way that would make the process slower, trickier, or otherwise more expensive. Developer smiles and agrees, and then (depending on precise level of sketchiness and sophistication) either bulldozes the thing and blames it on a tragic misunderstanding and miscommunication, or just hands a wedge of cash to a low level criminal and claims total shock when they wake up to someone telling them there’s been a fire on the site.

    I can genuinely sympathise with individuals who wanted to make a modification to their homes and got caught up in the bureaucracy, but they aren’t generally the ones who go on to ignore the decision and eventually get ordered by the court to rebuild what they destroyed. There’s nothing noble about the people who find themselves on the wrong end of this, they just thought they could pad their profit margins by forcing the court’s hand.

    6 votes
  9. Comment on WhatsApp announces messaging interoperability in response to Europe's Digital Markets Act (DMA) in ~tech

    Greg
    Link Parent
    Looks like it, based on the developer docs linked from the post. They explicitly require Signal Protocol compatibility, on-device encryption and decryption even when proxying messages, support for...

    Looks like it, based on the developer docs linked from the post. They explicitly require Signal Protocol compatibility, on-device encryption and decryption even when proxying messages, support for WhatsApp’s single device-tied keys, etc.

    I’m very much not a cryptographer and I won’t claim to know what the threat model here is in terms of trusting Meta’s vetting of other closed source clients, or trusting WhatsApp’s handling of client keys to decrypt messages from publicly vetted open source clients for that matter, but at least on the surface it seems like this follows the existing security model.

    4 votes
  10. Comment on WhatsApp announces messaging interoperability in response to Europe's Digital Markets Act (DMA) in ~tech

    Greg
    Link Parent
    Total guesswork, but maybe Meta prefer pushing users to have more of their apps installed, even accounting for the drop off from the extra friction? Messaging apps tend to be pretty sticky, and if...

    Total guesswork, but maybe Meta prefer pushing users to have more of their apps installed, even accounting for the drop off from the extra friction? Messaging apps tend to be pretty sticky, and if someone you need to talk to uses a specific one there’s a good chance you’ll install it if you have to, so I can imagine it being worth it to use them as an on ramp/retention mechanism for their other services.

    Either that or they wanted to tell the EU that interoperability wasn’t feasible, and showing that even their own ecosystem didn’t support it was a strategic decision to support that argument, perhaps?

    3 votes
  11. Comment on Kagi + Wolfram in ~tech

    Greg
    Link
    This is cool! I’m not sure I buy the “semantic data as a bulwark against AI misinformation” angle from the blog post, but having Wolfram’s mathematical functions built in seems like it’ll be...

    This is cool! I’m not sure I buy the “semantic data as a bulwark against AI misinformation” angle from the blog post, but having Wolfram’s mathematical functions built in seems like it’ll be really useful.

    That said, I have noticed recently that the general Kagi search has picked up a mildly annoying habit of returning the lowest fraction rather than a decimal when trying to divide something, so I’m guessing that’s related and hopefully an option in settings somewhere.

    11 votes
  12. Comment on European crash tester says carmakers must bring back physical controls. In 2026, Euro NCAP points will be deducted if some controls aren't physical. in ~transport

    Greg
    Link Parent
    One of the most frustrating things is that in many cases I'd happily pay whatever amount they're making in advertising and tracking for the option to have the product without it - TVs spring to...

    One of the most frustrating things is that in many cases I'd happily pay whatever amount they're making in advertising and tracking for the option to have the product without it - TVs spring to mind in particular here - but that's often not an option at all, and even where it is the companies take that willingness to pay 15-25% extra as a sign of "lost" profit. The company might expect $200 extra in data and ad revenue from you over the lifetime of a purchase, but the willingness to just cover that cost directly means you're classed as a less price sensitive consumer and they feel owed an extra $1,000 or more.

    7 votes
  13. Comment on FastSDXL.AI: Free demo that lets you generate AI images as fast as you can type in ~tech

    Greg
    Link Parent
    Aaand I've just realised that the UI they're giving in the link you mentioned at the top of this comment thread, branded "Copilot Designer" on a Bing URL, is completely different to the "Copilot...

    Aaand I've just realised that the UI they're giving in the link you mentioned at the top of this comment thread, branded "Copilot Designer" on a Bing URL, is completely different to the "Copilot Designer" UI that's presented by selecting "Copilot" on the top menu of the Bing homepage and then "Designer" from the sidebar - even though images I've generated in the latter do show up in the history of the former - that certainly doesn't help when it comes to figuring out the subtleties here!

    What I'm seeing for context.

    2 votes
  14. Comment on What is the most reliable and affordable form of storage medium to use as a backup drive for your computer? in ~tech

    Greg
    Link Parent
    It’s all a question of risks and probabilities, really. Keep a spinning disk powered on, eventually it wears out. Keep it off, eventually the bearings seize. Write too many times to an SSD without...

    It’s all a question of risks and probabilities, really. Keep a spinning disk powered on, eventually it wears out. Keep it off, eventually the bearings seize. Write too many times to an SSD without realising, the NAND degrades. Unplug it for cold storage, eventually the charge in the cells dissipates. You’re generally better off planning for hardware failure, rather than trying to avoid it and ending up screwed if you get unlucky.

    I’ve seen conflicting things on whether keeping drives spinning may actually be safer than storing them powered off (steady state causing consistent wear vs less overall wear but happening in an inconsistent way), but the way I’d look at it is that even if keeping them active marginally increases the odds of failure, it’s worth it to get an immediate notification when that happens rather than only finding out about it after it’s too late.

    There are tens of similar trade offs for pretty much any option, and any single option will fail eventually - so it’ll always come down to your personal balance of complexity, risk tolerance, and cost. The more you care, the more copies you want, across more locations, in more formats.

    As a minimum reasonable balance for the general user, I’d suggest a USB hard drive plus a cloud backup (Backblaze is a good option, but by no means the only good option). The drive will definitely fail at some point, and the chances are you’ll also realise you misconfigured or deleted something important from the cloud at some point, or get locked out, or forget you changed your card number and have it expire, or whatever - but the chances of both happening simultaneously are low enough to be acceptable for most people.

    I put a lot of value on my data, so my own setup is a local NAS with two drive redundancy, backed up to a cloud bucket with point in time snapshots managed from the local side as well as a six month versioned expiry period for anything that does get deleted on the cloud side in case a whole snapshots gets accidentally wiped. That’s expensive, complex overkill for most people, but I generate a ton of code and data that’d cause me way more hassle than it costs to keep safe.

    3 votes
  15. Comment on What is the most reliable and affordable form of storage medium to use as a backup drive for your computer? in ~tech

    Greg
    Link Parent
    Yeah, hard drives eventually just wear out, they’ve got a lot of physical moving parts that can fail over time. They tend to follow a bathtub curve: higher failure rate when nearly new from...

    Yeah, hard drives eventually just wear out, they’ve got a lot of physical moving parts that can fail over time.

    They tend to follow a bathtub curve: higher failure rate when nearly new from manufacturing issues that QA didn’t catch, low chance of failure for a few years after that as the bad ones have already weeded themselves out, and then increasing chance of failure again as motors and bearings age, cumulative vibration damage takes its toll, etc.

    Backblaze publishes quarterly reports on reliability at scale, but on an individual level when you’re buying one or two of the things it’s pretty much luck of the draw. The major manufacturers all make decently reliable products, some will be DOA, and all will fail eventually - often at the most inconvenient possible time - so you’ll want to plan accordingly.

    17 votes
  16. Comment on What is the most reliable and affordable form of storage medium to use as a backup drive for your computer? in ~tech

    Greg
    Link Parent
    Worth noting that those old CDs are going to be stamped, whereas anything writable at home uses dye, which changes the longevity significantly. That said, if you keep them in the fridge it’s...

    Worth noting that those old CDs are going to be stamped, whereas anything writable at home uses dye, which changes the longevity significantly. That said, if you keep them in the fridge it’s reasonable to expect a few decades (the Library of Congress has done some somewhat dated but highly relevant research on optical media ageing).

    Beyond temperature and humidity, the other key factors are storage density (Blu-ray will likely degrade to the point of error more easily as the physical features are much smaller) and precise manufacturing/chemistry (they saw some notable bad batches that almost all degraded far faster than the average).

    Personally I prefer a few copies on actively spinning hard drives - NAS at home backed up to a cloud provider - but if you wanted to go optical I’d imagine at least two copies of everything on media from different manufacturers would be a sensible way to go about it.

    Tape might also be an interesting one: it’s still the way big providers do large, long term cold backup, and the price per TB is hard to beat, but the setup costs are normally prohibitive for home users.

    14 votes
  17. Comment on FastSDXL.AI: Free demo that lets you generate AI images as fast as you can type in ~tech

    Greg
    Link Parent
    Totally makes sense! I will say I've found the chat context handling in Bing's "Designer" mode to be decent - probably better than the compound sentence parsing, as you're seeing - so I had...

    Totally makes sense! I will say I've found the chat context handling in Bing's "Designer" mode to be decent - probably better than the compound sentence parsing, as you're seeing - so I had reasonable luck getting it to build up the desired result step by step, waiting for it to generate an intermediate set of images each time:

    Please create an image of a brown-haired human male paladin in a D&D setting
    ...
    Add a castle to the background
    ...
    Add a teal-haired Drow female rogue standing next to him

    or

    Create a digital art image of a mythical dragon
    ...
    Please make the dragon skeletal and undead

    It's not perfect, not by a long shot, and you have to pretty much consider the intermediate images as throwaways (it'll use them as a rough style guide, but you definitely can't rely on it retaining details you liked from them in the next iteration), but the results did feel a lot better composed to me than doing it in a single shot.

    I'd still say that's not great UX, to be honest: it relies on the user considering the internal state of the system and working around it in a way that is probably less intuitive than giving it as much info as possible up front would be, and the type/wait/iterate loop does feel sluggish and kinda frustrating to me, but with any luck these are the kind of things the product people will be smoothing out as the tech matures.

    2 votes
  18. Comment on FastSDXL.AI: Free demo that lets you generate AI images as fast as you can type in ~tech

    Greg
    Link Parent
    One of the really fascinating bits of watching this whole field develop is that the UI needs to get worked out just as much as the underlying tech. I’m not quite old enough to remember the shift...

    One of the really fascinating bits of watching this whole field develop is that the UI needs to get worked out just as much as the underlying tech. I’m not quite old enough to remember the shift from CLI to GUI, but I do remember video games moving from 2D to 3D, and it took a good few years there for the industry to hit on controls and norms that made sense - this feels kind of similar.

    The thing that particularly brings it to mind is that pretty much exactly what you’re asking for can be done easily enough with multidiffusion - just scribble in the rough area where you want something and attach a specific prompt to that particular region - but turning that into a meaningful, understandable, and usable UI is a goddamn mess right now. We’ll hopefully be getting past the “everything needs to be a text prompt” world in the coming year; sometimes words alone just aren’t the way to communicate something, either to a computer or to a human, but I don’t blame the big guys for being a bit cautious in how they roll that into their products.

  19. Comment on FastSDXL.AI: Free demo that lets you generate AI images as fast as you can type in ~tech

    Greg
    Link Parent
    This is really interesting, I’ve run a few bits and pieces myself now too and the prompt “a man picking his nose” seems to be a surprisingly good stress test even for larger/slower models! I saw a...

    This is really interesting, I’ve run a few bits and pieces myself now too and the prompt “a man picking his nose” seems to be a surprisingly good stress test even for larger/slower models! I saw a lot of extremely bad results there, but they seemed similarly bad in the distilled versions compared to the slower versions, which suggests it’s a more fundamental issue; I imagine there are a solid few papers to be written on reducing those kind of “body horror” error conditions, and I’m betting that improved loss functions are going to be a significant part of it.

    I appreciate you taking the time to play around - I didn’t work on any of these specific models, but I do work in the field, so I feel like I’m in a bit of a “can’t see the forest for the trees” mode a lot of the time! It’s good to see things from an outside perspective.

    1 vote
  20. Comment on The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits in ~comp

    Greg
    Link Parent
    Just what I was thinking! I doubt it’d do much good in a unet, those get unstable even training in bf16, but I’m very tempted to do some tests with a diffusion transformer…

    Just what I was thinking! I doubt it’d do much good in a unet, those get unstable even training in bf16, but I’m very tempted to do some tests with a diffusion transformer…

    2 votes