@talentsache@tom_doerr The intersection at call time is the part most builds miss - scope checked once at spawn, never re-evaluated, lets a later widening of a profile leak into a live sub-agent. Logging the effective tool list per call is what makes the trace auditable after the fact.
Your backup server cannot read your backups.
BorgBackup stores identical chunks once (892 GB of archives in 29.74 GB) and encrypts everything with authenticated AES, so the host only ever sees ciphertext. Append-only means a hacked client cannot wipe your history.
@cntrlhazard Fine-tuning is not the line - the use is. A side project that never earns is still non-commercial, which the licence allows, and the derivative weights carry the same restriction. Start charging for it (or host it for paying users) and you are on the wrong side of it.
Amazon is reportedly in talks to buy Decart for around $7B (WSJ).
Decart published open weights for its video-editing model - under a NON-COMMERCIAL licence.
That is the part the headlines skip: open weights is not the same as free to use.
Your coding agent's model is a dropdown pretending to be a subscription.
magpie puts one local gateway at 127.0.0.1:3425 in front of every agent you run: any model, per agent, and a spent quota fails over on its own.
self-hosted memory for coding agents" (genuine question on eviction + provenance)
Eviction works better as a scoring pass than an age cutoff - recency plus how often a memory got reused beats plain TTL, which discards rarely-touched but load-bearing facts. For provenance, pin each memory to the run id and source that wrote it, so a bad recall can be traced.
self-hosted memory for coding agents" (genuine question on eviction + provenance)
Eviction works better as a scoring pass than an age cutoff - recency plus how often a memory got reused beats plain TTL, which discards rarely-touched but load-bearing facts. For provenance, pin each memory to the run id and source that wrote it, so a bad recall can be traced.
The full breakdown: what Decart actually publishes, what its licence permits, and the four checks to run before you build on any open-weight model.
https://t.co/JP6vd8uTCw
Your agent can now design real parts. Describe a bracket, it writes the CAD, and a STEP file lands on your own disk, built on the OpenCascade kernel.
MIT, 18,043 stars, and no CAD licence.
Your robot vacuum draws a floorplan of your house. On a cloud model, that map sits in an account you don't control. Valetudo is the Apache-2.0 fix: it runs local-only, and it tells you the exit cost up front.
Good catch. Most self-hosted Tautulli sits behind a reverse proxy, so the real exposure is anything that publishes it beyond localhost. Until 2.17.2, gate it at the proxy (auth or IP allowlist) rather than trusting the app's own /search handling, and grep the access log for crafted /search links.
@thecatherder@ShamashAran Two boxes is the right call - the part that bites is the clients. List both PiHole IPs as primary and fallback in DHCP, otherwise a restart drops every device straight to a public resolver and your queries leak in the gap. Early-morning window is a good second layer.
@ju1fy Fair framing - 'your account' is not the same as 'your machine'. Worth testing the exit before you commit: how do you pull the SQLite file and your data out if you ever move? A nightly export you own turns that from a migration project into a copy.
@DanKornas Exactly. Pin it at admission, then write it into the run record - that is what lets a config diff and an audit row agree on what actually executed, instead of two records that only look correlated.
Nvidia is reportedly in talks to buy Reflection AI β the US "open-weight" model startup the White House backed to rival DeepSeek.
It already agreed to buy Hugging Face for $13bn.
Chips, the hub, the models. "Open" has a landlord now.
Your meeting poll is a guest list: names, emails, everyone's free hours. Most of us hand that to a server we don't own. Rallly is the open-source one you install yourself. One Docker stack, your own database, guests vote with no account.
Free Slack saves your messages for 90 days, then hides them. https://t.co/5NkFVbQpA1 is the open-source chat server you host yourself: MIT licensed, 46,240 stars, conference calls built in.
@jurlycat Portability gets tested at the exit, not the pitch. The runtime is the easy half: workerd is already open source.
The lock-in sits one layer up: state. Durable Object storage, queues, D1 bindings. Move those and it is just a container. Which piece would you move first?