Supreme Leader Wiggum

31.5K posts

Supreme Leader Wiggum banner
Supreme Leader Wiggum

Supreme Leader Wiggum

@ScriptedAlchemy

Infra Architect @ ByteDance. Maintainer of @webpack @rspack_dev - creator of #ModuleFederation #auADHD #synesthesia own opinions.

Redmond, WA शामिल हुए Haziran 2018
720 फ़ॉलोइंग18.5K फ़ॉलोवर्स
पिन किया गया ट्वीट
Supreme Leader Wiggum
Supreme Leader Wiggum@ScriptedAlchemy·
If I were to stream… and it was mostly just me doing my day to day work, while talking to chat, would anyone watch? Setup isn’t going to be anything stellar, no real agenda - I’d pretty much just go do stuff across 5 or 6 repos in parallel & talk through what I’m thinking
English
20
0
31
10.2K
Kimi.ai
Kimi.ai@Kimi_Moonshot·
Congrats to the @cursor_ai team on the launch of Composer 2! We are proud to see Kimi-k2.5 provide the foundation. Seeing our model integrated effectively through Cursor's continued pretraining & high-compute RL training is the open model ecosystem we love to support. Note: Cursor accesses Kimi-k2.5 via @FireworksAI_HQ ' hosted RL and inference platform as part of an authorized commercial partnership.
English
301
704
10.8K
826.8K
Supreme Leader Wiggum
Supreme Leader Wiggum@ScriptedAlchemy·
@zeeg @Zackary_Chapple @KentonVarda Peak burn rate was $5000 a day. When I was testing cursors cloud agents before it was released, not sure the $ amount but it was about 1 billion tokens a hour. These days ive optimized burn down to about $40-$50k a month.
English
1
0
0
23
David Cramer
David Cramer@zeeg·
@Zackary_Chapple @KentonVarda i myself cannot spend more than 300-400 a day unless im pushing it through heavy token consumption automated tools (and those are often wildly inefficient spend) and im far above the average person in volume of code
English
2
0
1
47
Supreme Leader Wiggum
Supreme Leader Wiggum@ScriptedAlchemy·
I use codex app-server protocol for many things, have custom versions of codex as well. adding chat api support into the rust code cause too many rebase conflicts with upstream so its easier to just create responses api proxy for for chat endpoints. Codex subagent architecture is hard to beat imo
English
0
0
1
37
DawidDD
DawidDD@dawiddrzala·
@ScriptedAlchemy What is the point ? The oficial apps are best to use with the models that they try to fine tune for them. You have Open Code and other open source harners if you want to use other models
English
1
0
0
69
Supreme Leader Wiggum
Supreme Leader Wiggum@ScriptedAlchemy·
Wrote a small proxy that converts the chat api into the responses api shape for codex. So you can use chat/completions model with codex, most features supported. Sessions is stored and encrypted in the database so that responses api can support retrieving conversations. chat-response.federation.workers.dev If you lose your ID, you lose all the chats in the database. I do not store any api keys etc. I only probe the model to see what features it supports to configure your endpoint, after that its passthrough.
English
2
0
14
1.8K
Supreme Leader Wiggum
Supreme Leader Wiggum@ScriptedAlchemy·
My plan format is copy past of cursors. So interchangeable. Enforcing it I just run subagent graph skill. Then after that I’ll just say “helm” which helm mentions subagent graph skill so helming will make it switch to subagent graph mode. So as it’s working I’ll just say helm occasionally
English
1
0
0
18
elliot braem
elliot braem@elliot_braem·
This is what the inventor of Module Federation is doing:
Supreme Leader Wiggum@ScriptedAlchemy

I have 5 interconnected skills that are related to “subagent graph” and “plan graph” the helm skill drives the graph. In the instructions for main agent I tell it “all writes must be exclusively performed through fast-coder subagent, you can only helm the graph” or if you want to hard block it, put it in read only mode so only subagent has write. Usually I’ll have @cursor_ai write a ~60 item todo in plan mode. Then drag that into codex and my plan graph skills are all designed off cursors format, so the other skills kick in that then link to theirs and create per agent mini plans using same format. Then I just say “helm” every now and then till the DAG is exhausted.

English
1
0
2
1.2K
Supreme Leader Wiggum
Supreme Leader Wiggum@ScriptedAlchemy·
Yes but all interconnected. One designed for writing the plan files or tracking existing ones. The other is for subagent management. Plan graph skill like cursors plan mode, or if I pass it one of cursors plan files, it’ll parse and use that or break it up into mini plan per todo or phase etc. ultimately it’s designed to be like “how can I do this work in parallel with many agents without collision and without agent getting lazy on big work”
English
0
0
0
32
Supreme Leader Wiggum
Supreme Leader Wiggum@ScriptedAlchemy·
like i can just start with subagent graph skill, and then itll just helm a bunch of subagents to do the work. if i want more robust then i can use plan graph which makes the mini plans, then plan frontier finds then next series of parallel work that can be done and subagent graph then manages and launches the subagents according to the DAG. Helm steers them and watches for conflicting directions and bascially acts as me, but for the subagents its managing. For complex ones ill have the subagents become helms for their own subagent graphs. Graph Helm Agent->Graph Helm Agent-> Subagent workers being helmed
English
0
0
0
19
Supreme Leader Wiggum
Supreme Leader Wiggum@ScriptedAlchemy·
they all work on the same task / plan, in the same branch. Its about breaking the work down into enough independent lanes that the request can be delivered faster and larger requests can be one-shotted that might take a few days to build out. usually ill do one pr, if its large i have separate agents and skills designed for restacking a pr into a stacked pr. Mostly its about doing more work faster with less babysitting.
English
1
0
1
41
Supreme Leader Wiggum
Supreme Leader Wiggum@ScriptedAlchemy·
@MoonliteTechLLC @astrodotbuild All good! Astro islands are really great, but not necessarily great for the business needs at ByteDance scale. We have problems most others will never experience. RSPack and federation are more business tools than devtools.
English
0
0
0
38
Supreme Leader Wiggum
Supreme Leader Wiggum@ScriptedAlchemy·
Its official. Im giving up on nextjs & module federation webpack plugin. Deleted the package, writing new version of package for rspack only
English
8
4
106
11.5K
Supreme Leader Wiggum
Supreme Leader Wiggum@ScriptedAlchemy·
Both. We have over 1000 apps, 20000 repos, billions of users. Federation was intended for large scale business challenges. There’s many micro solutions for smaller scale. Most micro solutions only think about frontend. Federation works in native apps, game consoles, backend services, and frontend. I’ve seen drones get OTA updates mid flight. You probably noticed game consoles have fewer updates that cause system reboot, that’s because they are delivering updates via federation.
English
0
1
7
79

खोजें

@trashh_dev @Kimi_Moonshot @cursor_ai @FireworksAI_HQ @zeeg @Zackary_Chapple @KentonVarda @cryptopunk7213