Quick tip when using @ChatGPT Codex Codex’s memory system, Dreaming v3 (a memory architecture that automatically analyzes and synthesizes past conversations in the background), is still using GPT-5.6 Luna and Terra. Copy the snippet below into your config.toml to modify it yourself, or have Codex update it for you: Ini, TOML extract_model = "gpt-6-luna" consolidation_model = "gpt-6.1-terra" It’s much smarter and will save on usage.
Always exercise caution when using open-source code from GitHub. Open-source projects promoted as useful—or seemingly useful—skills on social media platforms like Twitter or Threads should always be used with care. Even repositories with thousands of GitHub stars can be manipulated by bots, and it’s common for such projects to conceal wallet or session stealers within their installation processes. Even if they’re harmless now, malicious code can be introduced during future updates. For this reason, I never directly clone or use open-source projects created by individuals I don’t know—even if they have massive star counts. Instead, I study their implementation and architecture, then rebuild them from scratch to suit my own environment. Thanks to AI, reconstructing a project from its structure alone takes very little time. Don’t risk losing your entire local environment by running unknown scripts just to save a few minutes. I strongly recommend using only the logic as inspiration and building your own version from the ground up. This is especially critical if you’re involved in cryptocurrency.
.@arc, let’s build a USDC subscription app Arc hasn’t shown anything since its mainnet launch and has completely flopped, so expectations have largely faded. Still, 500 points are a right—and Arc House has surprisingly been actively updating content, including both online and offline sessions, so I recommend checking in occasionally and participating! (It’s not about making money!) This time, there’s an online session planned where we’ll build a USDC subscription payment app using Arc Studio and Claude Code—I’m joining too, so I wanted to share it with you. 1/ What’s this about? Literally, we’ll build an app that accepts subscription payments in USDC, using AI. - One approach: building it solely with Arc Studio - Another approach: connecting Arc Studio with Claude Code The session will compare these two methods. It’s not a hackathon with submissions or rankings—I’m just participating for the experience. 2/ Schedule & Participation - October 16 (Fri), 00:00–01:00 KST
Video entry for the OpenAI DevDay Side Quests to win prizes. #OpenAISideQuest Used Codex CLI to reorder numbers from 1, 10, 2 to 1, 2, 10 while preserving the original and validating empty inputs. This is a recorded screen showing the actual execution results. Voiceover is synthetic narration. https://t.co/18F0yBwm8M
Muse is great to use like Grok. It seems Muse has officially declared itself completely free and solidified its stance on eliminating fees. From a market positioning perspective, this appears to be an excellent choice. Personally, I’ve often used Muse—well, before that, Meta AI—like Grok whenever I needed it. Because Muse is the only one that can freely access threads and Instagram posts, just like Grok can access all posts on X, it’s very useful for finding relevant information or identifying influencers active on these channels. Of course, the inherent quality of the information is still much better on Twitter.
OpenAI, GPT-5.6 SOL Positioning Change With the imminent launch of Astra and GPT-6 SOL, we have updated the description of GPT-5.6 SOL from “Latest Frontier Agentic Coding Model” to “Reliable Agentic Workhorse for Everyday Tasks.” The catalog priority has also been adjusted from 6 to 4. 1. Before GPT-6 Launch: 1. GPT-5.6 SOL 2. GPT-5.6 Terra 3. GPT-5.6 Luna 2. After Astra Launch (Immediately Prior to This Commit): 1. GPT-6 Astra 2–5. Empty 6. GPT-5.6 SOL 7. GPT-5.6 Terra 8. GPT-5.6 Luna 3. Yesterday’s Commit: 1. GPT-6 Astra 2–3. Empty 4. GPT-5.6 SOL 5–6. Empty 7. GPT-5.6 Terra 8. GPT-5.6 Luna
New benchmark: "StarCraft: Brood War Benchmark" Whenever a new frontier model is released, YouTubers often upload content about it—you may have seen some of these videos. This benchmark directly evaluates AI models’ performance in playing StarCraft: Brood War, and reportedly, none of the tested models surpassed beginner-level play. Distinct playing styles emerged across models: CodeX favored unconventional, trick-based strategies over standard play; Claude played a more classical, textbook style; and Grock simply stopped playing mid-match. All model-inference configuration combinations were pitted against each other in a round-robin format. For full details, please refer to the original report. 🥹 Leaderboard 1. Astra (xhigh) – 18 wins, 0 losses 2. Astra (medium) – 16 wins, 2 losses 3. Fable – 15 wins, 3 losses 4. Astra (low) – 14 wins, 4 losses 5. Sol (medium) – 13 wins, 5 losses ...
Leak of the GPT-6 Astra System Prompt Although it’s impossible to verify whether this actually functions as described, when I asked GPT, it said it seemed largely authentic. According to the leaked version, the most significant update compared to the previous frontier model, GPT-5.6 Sol, is “Ownership over Tasks.” vs SOL While Sol already possessed a high level of autonomous (perhaps even free-form) execution capability, Astra goes one step further: it is designed to understand the user’s ultimate goal or intent, then proactively perform possible actions without pausing at intermediate steps—tracking all the way to actual completion criteria. Notably, once a task has been approved, Astra no longer repeatedly verifies it; even if failures occur, it autonomously resolves recoverable issues and continues progressing—emphasizing Autonomy & Persistence far more than previous versions. https://t.co/y1mzPtJ88E
