Nico Baier
15.9K posts
Nico Baier
@nbbaier
Developer and linguist based in Chicago. PhD from Berkeley. I've also taught at McGill and UBC. I pretend to hate puns but I really don't
Chicago เข้าร่วม Ekim 2011
4.3K กำลังติดตาม1.1K ผู้ติดตาม
I am finally using the tool for some real work!
Pekka Enberg@penberg
I am pretty tired of drowning in Ghostty tabs while running coding agents, so throwing some tokens at it to build a tool in Rust that runs on Linux. It's essentially just Ghostty with a nice workflow to managing coding agent workspaces.
English
English
@nbbaier here's an explanation ;)
x.com/vaibcode/statu…
Vaibhav Gupta@vaibcode
apparently more people like order-n planning. Here's how i do it in a bit more detail. tldr: 1. define a shitty ticket 2. research relevant parts of the codebase + prior art 3. understand and select from the landscape of design choices <-- this is time intensive 4. define a better ticket 5. research relevant parts of the codebase + prior art 6. understand and select from the landscape of design choices <-- this is time intensive 7. assemble a "plan" 8. assemble a plan with code snippets 9. let it rip 10. meticulously read test cases and compiler snapshots 11. go back to 1 with whatever I and agent got wrong. i repeat steps 4-6 until it really works. we're building BAML, a new programming language, fully in rust. and have had high success with 10-20k line PRs that "just work". X wouldn't upload so here's a video that shows more details: youtu.be/8wQBvT6L8Xk
English
plan mode only works for order-1 difficulty tasks.
if you task is order-0, one shot it
order-1, plan mode
order-n, you need a recursive plan mode where you grill into increasing levels of depth at each layer.
for implementing test syntax, i had 4 times i had to create a "plan" just in this run (and 2 phases of defining the ticket)
Matt Pocock@mattpocockuk
I have also stopped using plan mode It creates a plan FAR too eagerly and usually asks you zero questions en route The whole point of planning is to get on the same wavelength with the LLM, not to generate an asset you don't read /grill-me all the way
English
@penberg @nbbaier Its hard to keep up with all the very cool and interesting LLM orchestrators, Im not even sure you can call swarm that. The way it keeps global and project-specific state looks very robust. repo-explainer.com/penberg/swarm
English
Nico Baier รีทวีตแล้ว
I can't read code, but I use open-source repos all the time. So I made a Github explainer: the @github for humans.
Simple human-facing explanations of the best repos out there. Because humans don't read code anymore, and vibe coders never did.
repo-explainer.com
English
me and Claude are too busy trying to vibecode our way out of the permanent underclass
Genius Tech@Geniustechw
Guys, what's stopping you from obtaining this physique?
English
Nico Baier รีทวีตแล้ว
NASA’s launches aren’t an advertisement for their upcoming IPO
Elias Makos@eliasmakos
Can someone explain why the Artemis launch video feeds didn’t look as good as this?
English
@nbbaier @elithrar @mattzcarey @thomas_ankcorn y'all just have to wait for us to publish the artifacts @elithrar mentioned 😁
English
ive been working w/ @elithrar, @mattzcarey, and @thomas_ankcorn on this and it's possibly the most exciting/cool thing ive worked on in my career
also finally had the chance to work on a project w/ these 3.
some of the most talented and fun colleagues to work with
Matt Silverlock 🐀@elithrar
working on the announcement blog, dev docs & API for something coming very soon 📦
English
Searching my library is now fast enough for live searches as I type!
Querying 100k paragraphs and 20k figures takes ~10-30ms.
The bottleneck is now 600-800ms for multi-modal reranking...
English
@dillon_mulroy Would be cool if you could actually have a chat interface for this
English
working on the announcement blog, dev docs & API for something coming very soon 📦
English
this visualization shows how agentic logic can work alongside deterministic tracking of what is happening, who should engage when, and when one agent should wait for another.
it’s to understand what happens inside without needing to navigate the code.
at this point, I’m pretty certain I’ll go all in on this.
English