I’ve had the idea for a while to use an LLM to gather metadata about books for me as well as generate tag lists for themes, plot, writing style, etc for everything in my ebook library. You could also generate non spoiler plot summaries and produce recommendations for similar books.

Look at MIT and UC Berkeley’s CS curricula and start tackling things that you haven’t covered. They’re both available freely online and you might still be able to find video recordings of Cal’s lectures somewhere (they recorded every class for students who weren’t present or had difficulty understanding in real time until 2015 or so but were hit with an ADA accessibility lawsuit because they weren’t captioned or something.)

I liked the whole “send cheaters to their own server shard with other cheaters” approach that showed up a while back

Interesting that the one has such large capacitors in it. I imagine that is as last-ditch effort to keep the board powered long enough to finish flushing all of its caches in the event of a power failure.

That’s exactly the point of power loss protection (aka PLP.) As a side effect of not needing to wait for a flush after a write synchronous write workloads are dramatically faster on enterprise drives with PLP.

Edit: To add a bit of detail - you don’t need to wait for a flush after a synchronous write with PLP because the drive firmware can lie and immediately return from a flush call because there’s enough backup power to complete that flush if the power were cut.

Ruckus APs and Opnsense have been solidly reliable for me for 5y now. No random fucking with unifi bugs (like having my WPA enterprise SSID punting users out onto the management vlan at random instead of the Kerberos assigned VLAN for that user, thanks unifi) and fantastic wireless coverage has me completely satisfied with my infra choices. Also, Ruckus unleashed handles controller duty on the primary AP rather than requiring a management container, that’s also a plus.

You can access Gmail over IMAP and pull down messages locally. If you do this; Back. Up. Your. Mbox.

Also, fun fact, you can move messages from a local mbox to Gmail while preserving read status and original dates if you want to add old email to Gmail for some reason.

+1, your list of browser extensions, list of plugins and list of available fonts are also available to anyone trying to fingerprint you. This idea that NAT will somehow obscure you enough to be anonymous online is security voodoo.

Yes, the machine that stays off 363 days of the year is such a security risk to my home network 🙄

I assume it’s a way to build a training dataset to fine tune a model over time. If they pay creatives for 2y while feeding the AI at some point they’ll no longer need humans in the loop (outside of spot checking or go/no-go judgements.) That’s how I’d get this off the ground anyway.

I’ve been waiting for this to happen. We’ve seen an absolutely huge increase in the number of readers who expect books to be a collection of specific tropes showing up in various scifi and fantasy book communities asking for book recommendations that match a chosen hyper specific trope collection. I’m not sure when this started but I’ve noticed more of it every year for a while.

I’ve been expecting someone to start churning out AI produced content (books, short videos, generated game plots) in the same collection of tropes style since generative AI really caught on back in ~2022. It’s amusing as hell that trashy-short-soap-opera-on-tiktok is where it really gained a hold first. I think we’re going to see a whole lot more of this across every form of media soon and I hope the flood doesn’t draw too many users away from legit content creators who work hard to produce their own content.

https://www.reddit.com/r/StableDiffusion/comments/1eij95i/running_flux_on_an_apple_m1_using_comfyui/

Some info in the thread above

I mean, the horror of having to tick a box to use rotating v6 addresses. These are all solved problems, they’re not a flaw worth ignoring the entire ipv6 protocol over. Most major operating systems have moved to stable privacy preserving addresses by default, that’s true, but it’s not all that difficult to turn on address randomization and rotation either. And, hell, if you’re that married to NAT as security just use NAT66 and call it a day, nothing about NAT is exclusive to ipv4.

Your firewall should take care of that, it’s pretty rare to be connected directly without one and by default any decent routing package will filter incoming traffic that’s not in the state tracking table. NAT isn’t designed for security, any security benefit it provides is a side effect rather than the intended purpose.

Edit: check out ipv6 privacy extensions too, there are solutions there that can reduce info disclosure if that’s a concern. You can accomplish many of the same benefits of NAT with v6 features without the downsides that NAT brings.

Ipv6 is fantastic, it has less overhead than v4 and removes the need for NAT or other translation. Support can be spotty in cheaper and older devices but there’s no reason not to learn and adopt it where possible.

The only windows machine on my home network is the backup Windows laptop that I only boot when I need to run something like Odin to flash a tablet or some niche Nintendo switch management software.

Most enterprise drives are TLC these days, MLC just doesn’t provide the storage density that enterprises require anymore. I only mentioned MLC because you’ll occasionally find mSATA drives in the <=256GB range that use MLC. You have to check the datasheet for each model, look for endurance rated at 5DWPD or higher, those will typically be MLC or heavily over provisioned TLC. If you want enterprise drives with greater endurance than the usual 0.5 or 1 DWPD look for the over provisioned models with capacities like 400GB, 800GB, 1.6T or 3.2T. those are 512GB, 1TB, 2TB and 4TB raw capacity drives with a bunch of flash set aside for wear leveling purposes. You don’t often see 300GB, 600GB, 1.2T or 2.4T drives anymore but those are often very high endurance (write intensive, 10 DWPD or so) models.

Check the datasheets for drives when you’re shopping and you can get a pretty good idea of what their durability is like, I usually buy 1 DWPD drives for write occasional bulk storage and 3+ DWPD for anything with a serious write workload. You can also help the drive controller a bit by running blkdiscard against the entire device before partitioning, then only partition and use ~80% of available space. The drive controller will typically grab free unused blocks and use them for wear leveling but only if they’ve been marked free (TRIMmed) and never allocated after. If you can’t find or can’t afford high endurance drives you can usually buy a larger lower endurance drive and over provision it in this way to extend its lifespan.

(The last time MLC flash was really common was back in maybe 2014-2015, some of the older Samsung pro drives like the 850/860 pro were built using MLC. Those had legendary real world endurance, I think they’d get up to 10+PB written before actually failing. It’s a shame they didn’t have PLP because they would have made good budget array storage if they did.)

My approach to this has always been to buy one enclosure and validate it, then go buy like 8 more after thorough testing. Obviously don’t place an order for 10 units of an unknown tech item from AliExpress or you’re looking at a bad time. Look for enclosures that use known good chipsets and there’s not as much risk as you’re expecting. I have something like 8 msata enclosures here that work flawlessly and another half dozen sata+nvme rtl9210b enclosures that also work well.

Buy used Samsung mSata or m.2 2230 drives on fleaBay. Stick with Samsung and other well known brands with decent spec sheets and warranties, that’s the cheapest way to handle durable storage on a pi. USB enclosures are like $5-7 on AliExpress or fleaBay.

Buy MLC drives if you need higher endurance (check the model no and look up the datasheet.) TLC will usually be fine for a few years, MLC will last a bit longer. If you’re killing drives faster than you expect buy larger (512 instead of 256GB), blkdiscard the entire device once it’s installed and then only partition 60-80% of it. Never touch the rest of the freed storage and the drive controller should be able to use those blocks for wear levelling to reduce the NAND wear rate.

Edit: One heads up, I usually buy used drives from eBay because their buyer protection is top tier, if there’s anything wrong with the drive when it’s delivered or when I test it it goes right back for a refund. This makes buying blind viable thanks to an easy return policy.

If you’re sourcing used drives somewhere else insist on seeing SMART data before purchasing and don’t buy heavily worn drives. Look at the drive model datasheet, find the warranted endurance of the drive (if it’s a 512GB drive rated for 1 DWPD over 3y that means the rated endurance is ~ 0.5T * 365 * 3 or roughly ~550TB written over 3y. Pass on buying drives approaching their rated endurance, try to buy lightly used drives wherever possible and you shouldn’t have problems with reliability.

Buy external drives. Don’t run them in RAID, use one to store backups and plug it in once or twice a week to copy data to it.

The secret to RAID is that it doesn’t buy you data protection, it buys you uptime to access data while a device in the array is failed. This is most valuable to businesses that can’t afford the downtime that recovery from a backup incurs. The most paranoid RAID will still fail sooner or later, due to hardware or software failure, and as a home user with a limited budget you’re far better off having one offline backup that you can use to recover data from once that happens.

Backup only data you can’t afford to lose (eg: don’t backup downloaded data that can be replaced easily, like a game or movie collection) and your backups will be much more manageably sized and you won’t need to spend as much on your backup drive. If a backup disk is too much for your budget you can always exploit cloud backup plans, backblaze PC backup has no limit on the size of your backups and only charges something like ~$60/yr.

Edit: It’s also worth thinking about what kind of data you’re storing and splitting that data across multiple devices if possible. If you’re storing bulk data where performance isn’t critical, like backups from other machines or a movie collection, you can pay a much lower price by buying a hard drive instead of flash. Even if only some of your data requires fast flash you can still use a cheaper HDD to store bulk data and buy a smaller flash drive for performance sensitive tasks. When I build NAS I split my data two pools, one bulk pool of HDDs and one much smaller fast pool comprised of flash storage. Put performance critical data on flash, put bulk storage on HDDs, this will allow you to spend less on bulk and still have fast storage performance for tasks that require it. A 512GB or 1TB SSD alongside a 4TB, 6TB or 8TB HDD is significantly cheaper than spending on a 4TB or 8TB SSD.

Shop eBay for refurbished storage, it’ll be significantly cheaper than spending on brand new drives.

I don’t encourage people to buy anything older than ~2016 or Skylake era. Older chips tend to eat enough power that they’re more expensive over time (usually less than a year after purchase) than newer more power efficient parts. Run the math on power consumption with the chip’s TDP for a year as an estimate and you’ll often be surprised by just how expensive chips from <2016 end up being to run. Cpubenchmark.net will do that for you if you use the comparator, just remember to set your average kWh cost.