Meet Joi: A native Mac app for End-to-End Audio-to-Audio chat, running 100% locally! 🍎🎙️
Running Qwen3omni-30b-a3b-instruct (4-bit) on M3 Max (36GB): ⚡️ "Thinking" Speed: ~30 tok/s🔊 TTFA (Time to First Audio): 3-5s
Private, immersive, and uncensored.
@_akhaliq@awnihannun
@thsottiaux Codex Desktop keeps silently overwriting my explicit model_reasoning_effort = "ultra" setting and downgrading it to low, while the official Codex Claude plugin rejects both max and ultra—please stop mutating user config and support the reasoning levels the CLI already accepts.
hey openai, this privacy boundary needs to be painfully obvious
Codex works on a local codebase or workspace. that workspace should not be bound to, uploaded to, or treated as part of an account workspace just because Codex can access it. 1/5
ChatGPT Projects on the web and Codex projects are easy to confuse. they are absolutely not the same thing: a Codex project is rooted in a local workspace, and that workspace must never be fully synced into the account. 4/5
New Anthropic research: A global workspace in language models.
Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with.
We found a strikingly similar divide inside Claude.