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.
By that point I'm pretty sure we'll have an effective compact model that can run locally and transcribe downloaded videos on reasonable hardware. Or you can just sic a paid model like chatgpt on the task. The corporate Internet is entirely focused on subscription service models now, unless you run the model yourself on local hardware you're going to end up paying someone somewhere a service fee.
Edit: y'all need to learn about minified models designed to run on edge hardware, they're a thing and often work shockingly well.
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.
Pro tip, LLMs do an excellent job summarizing YouTube videos now. I've never liked YouTube content, the incentives for creators are perverse and discourage conveying accurate information simply in favor of drawing out every video to maximize ad opportunities. About 95% of the content I might have been interested in could have been better conveyed in a 1-2 page blog post and read in 2 minutes instead of stretched out into a 15 minute video. Having a robot summarize that content is so much less irritating.
YouTube is just on demand TV with extra steps these days. I've stopped watching videos, I have an LLM transcribe and summarize for me now. 99% of the content of a 10-15 minute video can be summarized into 1 or 2 pages and read in under 2 minutes.
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.
Ah the sedimentary filing system. I can tell you exactly when I last touched each layer of each pile and what's there but if I file it all away somewhere I can't tell you shit.
By having long term goals. If you're working toward something bigger in life each day is just progress on that journey.
If you don't have any long term goals start thinking about where you want to be in 5 or 10y and make some. Then you can think about how to get there and start making short and medium term goals to help you along the way.
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.
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.