I don’t know. Quite the opposite experience here. I’ve always only heard stupid ignorant shit about unions from my colleagues. Things only strated shifting over the last coupleof years.

FOSS communities probably won’t form around most of these projects, as you said. Assuming the software can be managed by a few people with LLMs, likely local in the very near future, it’s quite possible that plenty end users who didn’t use FOSS before would jump on these instead of paying money to MS or Adobe. That in itself is a huge win for the majority. Depriving MS from revenue that is. Then for community FOSS, this has potentially significant positive implications about our biggest challenge - dealing with proprietary formats and proprietary blobs. This could mean perfect .doc/.docx compatibility in LibreOffice. It could mean a perfectly working nouveau driver for NVIDIA GPUs. How about fixing the tons of bugs we workaround in motherboard firmware. Or fixing that corner case in Wine so it can run your fav obscure game. Or reverse-engineering communication protocol of cloud-connected IoT to integrate it in Home Assistant. Etc. Etc.

In short: FOSS is a social activity. Assuming this method can produce very good source from binaries, that itself won’t create FOSS communities, but FOSS communities can use it to significantly improve the FOSS they care about.

2x R9700, Ryzen 7700, 64GB RAM. The absolute best I’ve seen from llama.cpp was 60tps but the typical was 45-55tps. Radiance gave me 150tps on the first try. It can do parallel output while keeping 80-120tps with 2-4 streams. It can do more in parallel but I heven’t tested what the numbers look like. As far as I understand these numbers are possible on a single R9700 with the MXFP4 model. Haven’t tried it.

Radiance. Was using the latest llama.cpp. This thread caught my eye as simple enough to try.

If you do it at 3AM anyway… may as well eliminate the need to even think about mismatches. 😁 I think mine also used to do it at 3AM before I switched to ZFS snapshots + send/recv.

Immich uses user/YYYY/MM folders on my setup, and I exclude any thumbnail directories from the backup too. So not many dirs/files to check for the backup software and it’s very quick.

That makes sense. I copy the whole set of data dirs so I can trivially start it after restore or start it elsewhere without extra steps. Also because this strategy works with all services so I don’t have to consider how to backup/restore each one. Makes adding new services less work.

pgdump with live file copy while the service is in active use can result in files the db doesn’t know about, or files it thinks are there that were actuall deleted. Probably can be fixed after the fact.

More generally, as a someone who’s done software for a very long time, I’ve learned that the further away I go from the happy path of a software program, the less tested it is, the more bugs there are and the poorer the edge case handling is. It’s inherent to software development with limited labour. So for backups I lean on the process kill/recovery edge case that they all must handle. Snapshot + backup from that snapshot looks like a process death to the service upon restore.

Thats kinda how the whole Chinese competition started - optimization to make due with limited hw.

Also I just changed the inference engine I use for local models and Qwen 3.8 27B went from 50tps to 150tps. The new engine is optimized for my hw.

Immich has so many dirs and files that it will take hour/hours for any file-based program to walk that tree (on an HDD) in order to do a backup. I used to use rsync, syncthing, duplicity and had gotten it to about under an hour for 1TB library. The service has to be shutdown to preserve consistency of the backup. So that downtime was okay once per day. Still a heavy operation. Not great, not terrible.

Because of the above you might want to do one of the following:

• Use snapshotting filesystem like ZFS or Btrfs. That allows to avoid the downtime as you can backup a frozen-in-time snapshot while Immich is running.
• The above but with ZFS om both the Immich end and the backup end. That allows transferring just the delta between the new Immich snapshot and the last snapshot the backup has. This happens entirely within the filesystem so there are no file operations needed. At every point in time ZFS knows what blocks have changed between now and the previously backed up snapshot. When you ask it for that, it starts spitting data without spensing an hour figuring out what changed. That allows online backups every hour or even more frequently. The backup of my 1TB Immich instance (HDD) takes 5-10 seconds to initiate the network transfer, then it takes whatever time it takes to get to the backup machine, which is usually equivalent to transferring whatever new was stored in Immich.

So depending on how painful migrations are for you, I’d highly recommend the latter. Otherwise the former. The ZFS snapshot send strategy works for any service you may run.

As is tradition.

This appears to be a Jetpack Compose UI for Krita.

No issues upgrading from 12.1.

They can still re-rank Google’s results to surface more relevant information.

But it feels it got much worse much faster over the last 2 years.

Some companies already run Chairman Xi’s big models on-prem and do not limit token usage on those.

I don’t want it take shit photos, it keeps doing it despite that. :D

Joke aside the FP6 has a decent cam. It’s inconsistent indoors but it’s decent overall.

I know for a fact Ford ships a modem with a SIM and they use it sparingly because it’s expensive. Don’t know abt the rest.

It’s this setting: https://github.com/aosp-mirror/platform_frameworks_base/blob/1cdfff555f4a21f71ccc978290e2e212e2f8b168/packages/SettingsProvider/src/android/provider/settings/backup/SecureSettings.java#L68. You can trace it in the source code. The general name is Reduce Bright Colors. Check Graphene’s src for that. Woulda linked Google’s AOSP web page but it’s giving me 503 because that’s how Google do these days.

Stop hitting them, they already dead.

Def is