@mindreframer.bsky.social

One thing I notice more and more is that people think verbal abuse and personal attacks are acceptable, if they are directed at people who don't have a dismissive attitude towards AI and LLMs.

I wrote this earlier today on lobsters, but the thread was deleted. I feel like I had a conversation of that sort multiple times, so I figured I leave my thoughts on that issue for others to read. It's on the question of not having kids because of the state of the world.

Bild

Giving an LLM a code sandbox massively increases its capabilities, but managing VM lifecycles is a pain. Introducing foreverVM, a Python REPL-as-a-service. Under the hood, we swap between memory and durable storage. To you (and your agent) it acts like an always-on REPL. See more at forevervm.com

Updated my post with notes on this: for Ollama you need to set num_ctx to a higher value than the default 2048, plus I got tips on running large prompts using the Apple MLX framework

Sync engines enable web apps with: - instant (< 1ms) interactions - full-stack reactivity - dramatically simpler dx Despite this, sync hasn't taken off because it doesn't generalize to more complex apps. I'll be speaking about how @zero.rocicorp.dev solves this next week. Come check it out!

Yonz@yonz.bsky.social · 2y ago

Aaron Boodman @aaronboodman.com will introduce Zero: A Sync Engine for the Whole Web for the final session. The much-anticipated launch from the folks that brought you Replicache.

Some of you know that I lead many Core infrastructure teams at Google, and one of those teams is Multiplatform development. As such, I've been extremely interested in WASM lately- you've probably heard about WASM's benefits for performance, but did you also know WASM is great at *portability*?

Drawing of how WASM's workflow can work from a variety of languages, to many outputs

Single-writer systems are 🤩 - they transform traditional transactions into a much simpler problem: fencing zombies. This allows read-modify-write without coordination, instead relying on commits to discard data if its fenced (rare). SlateDB enforces single-writer. Let the use cases "sync" in 😉

My philosophy at Bluesky: * The cloud for smaller services, that will remain relatively cheap even w/scale. * Dedicated servers for services that need to exist in many regions, without being wildly expensive. * On-prem for distributed systems that would be dramatically more expensive otherwise.