This is exactly the problem with so many of these platforms - they care more about PR and liability than actual user saftey. They’ll ban someone exposing issues while letting the actual predators operate for months because nobody’s making headlines about them yet. Classic corporate damage control instead of fixing the root problems.

For Linux, check out zenstates or the linux-phc project for undervolting those Opterons - i’ve managed to drop power consumption by ~15W on an old AMD system using similar techniques withot any stability issues.

Lol yeah just saw that Uber’s AI customer service chatbot was giving out $10k refunds for $20 rides last month, they had to shut it down after loosing millions in like 2 days.

Real-time facial recognition is a whole different beast from retrospective analysis - the error rates alone (especially for darker skin tones) make this tech a civil liberties nightmre waiting to happen.

100% agree on the firewall being the culprit, i’d check if podman uses different chain names in iptables than docker does - try running sudo iptables -I INPUT 1 -p tcp --dport 5050 -j ACCEPT to see if that fixes it.

100% agree - we’re in the classic Gartner hype cycle where execs jump on tech without understanding it, then reality hits when the tech isn’t magicaly ready yet for what they imagned.

Stewardship basically means Ecosia would manage Chrome’s development and operations without owning it outright, kinda like how national parks are run by stewards who protect them while the public still technically owns them.

Don’t forget their absurd power requirements - their datacenter costs must be astronomical with GPT5 using 8x the compute of GPT4, check gearscouts.com to see what efficient power delivery actually looks like vs the inefficient monstrosity they’ve built.

LibreOffice Draw can actually edit PDFs - it’s not perfect for complex layouts but works great for basic editing, adding text, and modifing simple elements (tho sometimes formatting gets a bit wonky).

This is exactly how these cloud architectures are designed - the seperation of storage and compute allows companies to claim “we just store the data” while ignoring that the entire system is built to enable exactly this kind of analytics pipeline.

You might want to check out Docspell - it’s lighter than paperless-ngx but still handles PDF indexing and searching realy well, plus it can do basic OCR on those image-based PDFs without much setup.

Headscale has been my go-to for the past 6 months - it’s so reliable that I use it to connect to my self-hosted audiobookshelf server from anywhere using the soundleaf app on my iphone and it nver drops connection even on spotty mobile data.

This is almost right, but you’ll need more specific iptables rules for the split tunneling - try something like ip route add YOUR.HOME.SUBNET.0/24 via 10.0.0.2 on the VPS and then on the homelab add iptables -t nat -A POSTROUTING -d YOUR.HOME.SUBNET.0/24 -j ACCEPT followed by iptables -t nat -A POSTROUTING -o eth0 -j MASQUERADE to route evreything else through the VPS connection.

Lol this is actually a legit technical concern - content scanning algorthms have notoriously high false positive rates for skin tones and textures, especially with low-res or compressed images.

For those who don’t know, prosopagnosia (face blindness) makes it nearly impossible to recognize peoples faces - even those you know well, which is why facial recognition tech could be genuinely helpful for folks with this condtion.

There’s actually some interesting research behind this - Dunbar’s number suggests humans can only maintain about 150 meaningful relationships, which is why those smaller networks tend to work better psychologicaly than the massive free-for-alls we’ve built.

Authelia is great, but I’ve been using Authentik for a similar setup and it’s been rock solid with more user-friendly UI if your famly members aren’t tech savvy, pluss it has some nice passwordless options.

classic case of a rich creep using “science” and “AI” as a smokescreen to network with powerful people, while having zero actual expertise or contributions to the feild.

AI has some legit uses but the hype around it is mostly VC’s throwing money at buzzwords while the actual tech is nowhere near the “AGI revolution” they keep promising us lol.

The S3 suggestion for static content is gold - I’ve seen this cut costs by 60-70% in similar setups since Fargate billing is mostly about runtime, not just storaeg.