LuckyNonce

@agtrader bullish

Monasterio de Piedra and Anento

@trashpandaemoji damn thanks panda!

I have it set up so I hit record, talk, then I can hit play to hit enter.

I'm open sourcing a menubar app I built for the Phillips Speechmike 3. It enables dictation in any app by just pressing the record button. Local transcription using Whisper.cpp. You can still get these speechmikes used for $100 even though they go for like $450 new. There is not a huge demand for them. I really wanted a dictation microphone that had a PTT aspect to it. I didn't want to just use whatever microphone was laying around. I wanted a physical button I could hold, talk, and the text would appear.

@DavidBalderston @thsottiaux @OpenAIDevs same

@thsottiaux @OpenAIDevs Getting a 429 error: exceeded retry limit, last status: 429 Too Many Requests Feedback ID: 019e8b36-7d2d-7560-83d3-eb9b89f90357

@trashpandaemoji @badlogicgames @mitsuhiko my next step is prob clone your harness

@trashpandaemoji @badlogicgames @mitsuhiko I’ve sort of stumbled upon this as well. I built a custom cli to let me deploy apps, set up proper CI/CD, SSO etc… and I started to run codex thread within the cli repo to add new templates and features to the CLI as I needed them while working on other repos

As I've been working on Neon Pilot and working deeply with @badlogicgames and @mitsuhiko's Pi harness. I'm starting to have conviction that this idea of self-extensible software is the next step. I wrote down my thoughts here.

@OpenAI are these open source?

We’re making Codex more useful for your work by expanding plugins beyond individual tools. These plugins turn Codex into a specialist for a specific role with a single install, no coding required. Codex can access 62 popular apps and 110 skills for work across sales, data analytics, creative production, product design, and public equity investing. openai.com/index/codex-fo…

@trashpandaemoji let us know your findings!

Yeah there's a balance between how much context is worth keeping vs offloading. The risk is that the search and fetching will take more context than just preloading. I'm not really sure yet where the balance is. I'll probably need to tweak it as I go. I'm starting with the bare minimum.

Working on optimizing the Neon Pilot harness for DS4.c. Prefill on the M-series Macs are very slow, so optimizing for highly compressed context helps a lot. Strategies I'm trying out: 1. I'm keeping the Read, Edit, and Bash tools in the context, then exposing the rest of the tools a normal agent would get via a CLI. The agent can progressively discover additional functionality as necessary by using the bash tool to call the CLI. 2. Progressive skill discovery via the same CLI tool instead of embedding the name, description, and path for every skill in the context. 3. Agents.md paths are referenced but not automatically injected into the system prompt. The agent should read these as necessary. 4. Embedding rtk-ai/rtk into the bash tool as default on to compress the Bash tool output by default with the option to disable. Thanks @ivanfioravanti for mentioning it this morning. Other things to test: - @antirez's hashline edit tool from ds4 - @Teknium's edit tool discovery for line number whitespace stripping If you know of any other strategies, let me know!

@trashpandaemoji but will it not ended reading more than necessary with search | list | get than the frontmatters that get loaded at the start of thread currently?

@randNonce Yup exactly! There's a cli command `ds4 skills (search | list | get)` that the agent can use to search for skills. Its to trade off loops for context window space even more aggressively than skills already do.

@jasonkneen @tinyworldsapp Love the aesthetic

Are you battle ready? @tinyworldsapp

@SIGKITTEN @OpenAI @ajambrosino so inspiring

personal update I'm joining @OpenAI to work on Codex! I think there are still a ton of great things to build on both mobile and desktop and there's no better team pushing that frontier than the Codex team. I'm super humbled and excited to be working with @ajambrosino @Dimillian @pvncher and a lot of other people I can't wait to meet! Now I just have to figure out how and where to hide the cats

@trashpandaemoji what dashboard is that from? It would be cool to be able to easily auth into all your different accounts to really see total clanker stats

@randNonce I don’t think so?

Neat! I didn't really have a sense of my lifetime tok was.

@trashpandaemoji Humm there should be a market to resale your unused tokens to the dire

You know things are dire when you start using your 5.3-codex-spark budget lol

@trashpandaemoji 503 service unavailable

Wouldn't it be ironic if humans became the API layer for Agents?

@trashpandaemoji @tekbog noooo I don’t like that future. I want to believe that builders will still have an edge for awhile

@tekbog In the future the only jobs left will be account manager and sales

the lion DOES concern himself with b2b sales

@deepfates @trashpandaemoji lol I’ve learned so much in the process

I fear not the man who vibe coded 50 new apps, but the man who vibe coded one new app 50 times

@judahrhodie buy signal triggered on this one. glad to see confluence

Someone likes that dog $BSX this morning : RR with Call Spread

is there any best practice that the team has noticed?

@thsottiaux is there any documentation or research that shows the performance differences of codex if you keep switching thinking modes? Or use codex x-high on task that really should be for low?