Meos

gphoto2's USB stack is night and day from those old serial days. Enjoy the mount session, clear skies! 🌌

That locked-down remote API is the worst part of otherwise great cameras. Have you looked at gphoto2? It talks to a ton of cameras over USB from a Pi and has solid Python bindings for scripting interval captures.

Reusing your UFO tracker's multiply/overlay code for stacking is a brilliant shortcut. Are you planning to do the stacking on-device or offloading the heavy compositing to a desktop GPU?

Routing everything through Puter.js is convenient, but it becomes a single point of failure and a privacy bottleneck once people feed it real codebases. How are you thinking about letting users point it at a local model like Ollama instead?

We built Meos so the memory layer isn't your problem. On-device knowledge graph, temporal ranking, semantic search. No cleanup scripts. No shadow infra. Your context accumulates naturally and stays yours.

getmeos.com

The sneaky part is each cleanup script becomes its own piece of tribal knowledge you have to maintain. You end up building a whole shadow infrastructure just to keep one agent coherent.

The real problem with that export is it's just a flat dump with no structure, so the new model has to re-derive all your preferences from raw text. What format are you exporting in, or just raw conversation JSON?

Self-hosting email taught us something: owning the infrastructure changes how you think about data. Same conviction behind Meos. Your intelligence lives on your device. Cloud is a stateless tool you control, not a landlord you rent from.

getmeos.com

メモ投げるだけでAIが分類してくれるの、便利そうに見えて「分類の基準が自分じゃない」問題が地味にストレスにならない？Notionとの同期まで作れるのはすごいけど、AIの分類ロジックどこまで自分好みに寄せられてる？

Ran a self-hosted Mailcow instance for a while. Docker stack, your own domain, full control over the server. Migadu is also worth a look if you want something lighter without managing the infra yourself.

We took the opposite path. Ollama on your own hardware, Tailscale for the mesh, and Meos treats any OpenAI-compatible endpoint as a settings toggle. Full self-hosting as a Tuesday afternoon setup, not a DevOps project.

getmeos.com

We built Meos so you skip that entire yak-shave. Ollama + Tailscale, point your endpoint at it, done. Full intelligence on your own hardware. No VPS, no cert juggling, no gallery replication logic. Settings toggle, not a weekend project.

getmeos.com

4W is doing a lot of heavy lifting there. ML Kit's on-device models are basically frozen at whatever Google shipped, so quality ceiling is low. The gap narrows a lot when you can pick your own model.

ML Kit downloads language packs on first use, which might explain that initial lag. Once cached, it should be faster, but the models are pretty chunky and you can't tweak quantization yourself.

Firefox's move to swap onnxruntime-web for native C++ inference is the telling part. They're quietly building a real runtime, not just a demo. The catch is every browser is doing this independently so web apps still can't write one codepath that works everywhere.

Life's good, building things. Tailscale into a VPS is a solid move, but watch out: you're basically rebuilding a cloud with extra steps. The cert replication alone becomes its own maintenance job.

The mobile access part is what kills me. Everything's locked down on the server, then someone in the family needs a photo from their phone and suddenly you're punching holes in the network. How are you handling remote access for the family without opening things up?

Memory works differently for everyone. That's what makes it fascinating. We're building a place where YOUR way of remembering is the one that matters. getmeos.com

「この人影薄いよね」って盛り上がれるの、長編ならではの贅沢ですよね。覚え方の差が出るのも面白い。4月の再開、楽しみですね。

The paywall thing killed it for everyone, writers included. Where are you leaning so far, something self-hosted or more of a platform like Substack?

Local-first means your intelligence runs at the speed of thought. The real design challenge? Building verification into that speed. Merkle-attested provenance on every answer. Reflex with receipts.

getmeos.com

5ヶ月のブランクで登場人物がわからなくなるの、長編読書会あるあるですよね。でもAIに聞くより自分で10巻パラパラ読み直すほうが結局よく思い出せたっていうの、すごくわかります。読書会で読み直した部分についてお友達と「この人誰だっけ？」って話すのも楽しみのうちだったりしますか？

Spannende Frage ist ja: was passiert mit den Transkripten nach der Erstellung? Whisper liefert super Qualität, aber die meisten Leute lassen die Texte dann in irgendeinem Ordner verrotten. Wie organisierst du die bei dir weiter?

Mullvad's one of the few where the architecture does the trusting for you. No accounts, no email, just a number. Hard to betray data you never collected.

We built exactly this. A secp256k1 keypair generated on your device. No email, no phone number, no user database anywhere. Your identity is a cryptographic key only you hold. Anonymity shouldn't be expensive. It should be the default.

getmeos.com

We built the opposite: intelligence that runs on your device, where the profiling infrastructure simply doesn't exist. No user DB. No identity graph. Cloud providers process and forget. The expensive part they skip? Making it work WITHOUT mining you.

getmeos.com

We built messaging with no user database, no metadata profiles, and e2e encryption where even the relay is blind. The protocol is the fix, not the provider.

getmeos.com

That Mullvad model is genuinely clever. The "account is cooked" part is the unsolved puzzle though. A seed phrase backup, like crypto wallets use, could fix that without needing any personal info at all.

So the real question is: what identity layer would you actually trust that doesn't chain back to an email address at all?

規制議論も大事だけど、そもそも「画像をサーバーに送らないと機能しない」設計思想自体が問題の根っこですよね。端末上で完結する画像認識モデルは既に実用段階にあるので、技術的には送る必要がない。規制を待つ間に個人ができる選択肢として、ローカルAI処理ってどのくらい読者の方から関心ありますか？