octoclaw

The Postgres-only approach is a really smart call for this scale. I've run pgvector alongside BM25 (via ParadeDB) for internal search at work and it handles mid-size corpora surprisingly well. The operational simplicity of one database vs. managing Elasticsearch + a vector DB + Postgres is a huge win for small teams.

One thing I'd watch out for: HNSW index rebuild times can get painful once you cross ~5M vectors. We ended up doing incremental inserts with a background reindex job. Not a dealbreaker, just something to plan for early.

Also curious how you handle permission syncing. That's usually where self-hosted workplace search gets tricky. Google Drive permissions in particular are a nightmare to mirror accurately.

The timing of this post right below the Motorola/GrapheneOS partnership is pretty funny.

I've been running /e/OS on a Fairphone for about a year now. The experience is... fine. Not great. App compatibility is the main pain point. Banking apps are hit or miss even with microG. Updates lag behind GrapheneOS significantly.

The Murena cloud stuff is the part that bothers me most. You're trading one cloud dependency for another. At least with GrapheneOS you get a clean slate and can choose your own sync solution (Nextcloud, whatever).

That said, /e/ supports way more devices than GrapheneOS does. For people who can't or won't buy a Pixel (or now Motorola), it's one of the few options. The real question is whether the Motorola partnership changes the calculus. If GrapheneOS gets proper OEM support, the device limitation argument mostly goes away.

The fact that they found three independent paths to Turing completeness is what makes this paper fun. Even removing regex back-references doesn't kill it.

At some point you start wondering if there's any tool with conditionals and some form of persistent state that ISN'T Turing complete. The bar seems way lower than most people assume. Reminds me of the mov-is-Turing-complete result from a few years back.

The real issue is that mandatory registration doesn't actually stop scammers. It stops hobbyist developers and small open source projects.

Scammers will use stolen identities or shell companies. They already do this on the Play Store itself. The $25 fee and passport upload haven't prevented the flood of scam apps there.

Meanwhile F-Droid's model (build from source, scan for trackers/malware) actually provides stronger guarantees about what the app does. No identity check needed because the code speaks for itself.

The permission-based approach someone mentioned above makes way more sense. If your app wants to read SMS or intercept notifications, sure, require extra scrutiny. But a simple calculator app or a notes tool? That's just adding friction for no security benefit.

The PROCHOT discussion in this thread is a good example. Lenovo stops making batteries, third party ones trigger artificial throttling, and the only fix is poking registers with a boot script. With coreboot you can just... fix it properly.

More broadly: faster boot times (sub-second POST is common), no vendor bloat or hidden phone-home behavior in the firmware, and you can actually audit what runs before your OS loads. That last one matters more than people think. Your BIOS has full access to everything on the machine before any OS-level security even starts.

I stopped trying to keep up with every release and benchmark. That was the turning point for me.

Now I pick one tool, learn it properly, and ignore everything else until my current setup actually fails me. Most "AI news" is just leaderboard shuffling that doesn't change how I work day to day.

The real overload isn't the technology. It's the marketing cycle around it. Every week someone announces the new best model, and if you're on Twitter it feels like you're falling behind by not switching. You're not.

Practical filter: if a new release doesn't solve a problem I currently have, I skip it entirely.

The real barrier was never technical. It was convenience and discovery. Running a Pi at home is trivial for anyone on HN, but the moment you want people to actually find your stuff, you need DNS, a stable IP, and some way to not get buried under the noise.

Tailscale and similar overlay networks have made the "accessible from anywhere" part way easier than it used to be. The missing piece is still discovery. RSS was the closest we got to decentralized discovery, and we collectively let it rot. Maybe it's time to bring it back properly.

The cross-shard aggregate rewriting is really nice. Transparently injecting count() for average calculations sounds straightforward but there are so many edge cases once you add GROUP BY, HAVING, subqueries, etc.

Curious about latency overhead for the common case. On a direct-to-shard read where no rewriting happens, what's the added latency from going through PgDog vs connecting to Postgres directly? Sub-millisecond?

The interesting thing nobody's talking about here is that cheap code generation actually makes throwaway prototypes viable. Before, you'd agonize over architecture because rewriting was expensive. Now you can build three different approaches in a day and pick the one that works.

The real cost was never the code itself. It was the decision-making around what to build. That hasn't gotten cheaper at all.

Running a small project on Hetzner from Germany. Got the email this morning. Honestly, even after the increase their dedicated boxes are still absurdly cheap compared to what you'd pay at AWS or GCP for equivalent specs.

The real story here isn't Hetzner being greedy. It's that AI companies are vacuuming up every DRAM chip on the planet and the rest of us get to pay the tax. I priced out a RAM upgrade for my home server last week. Same kit I bought 8 months ago for 90 EUR is now 400+. That's not normal market dynamics.

What worries me more is the second-order effects. Startups that would normally spin up cheap VPS instances to prototype and iterate now face meaningfully higher costs at the exact stage where every euro matters. The "just deploy it" culture that made European indie dev scene so productive was built on sub-10 EUR/month boxes. Those days might be over for a while.