|
📱 SOCIAL
OpenAI DevDay 2026 (Sept 29) resolves yesterday's watch note — 20+ launches, the day AFTER Astra was killed
The Rundown: What shipped: **dots** — always-on AI agents powered by GPT-6 Astra, each with its own cloud computer + browser, 4,000+ app connections, proactively working in background; rollout starts for ChatGPT Pro, Business Premium, Enterprise (one dot per user). Cartoon avatars you name/customize (Iggy, Felipe, Jojo shown). **GPT-6.1 Sol** at $2/$10 per 1M tokens — one-fifth of Astra pricing. **Ultrafast** speed tier (8x faster in Codex). Pro 500 plan. **ChatGPT Space** (Notion-style collaborative surface). For devs: Codex in the cloud, computer use in the Agents API, Decisions API built on Luna, Bedrock Managed Agents with Amazon, plugin extensions in ChatGPT. OpenAI's own recap: "our biggest yet, with more than 20 major announcements."
The details:
Why it matters: yesterday's file flagged DevDay as a watch note; the announcement record is now primary. The "killed the flagship the night before, shipped 20+ things the next day" frame is the week-arc payoff.
|
|
📱 SOCIAL
Vaibhav Sisinty flagged "Atria Dawn Preview" — a Chinese student-heavy lab's open-weights agentic model
The Rundown: The post's claims: 96-person Shanghai AI Laboratory lab (65 students) shipped Atria Dawn Preview on a GLM-5.2 744B backbone, beating GPT-5.6 and Claude Opus 5 on multiple benchmarks. Demoed: trained a 400M-parameter weather model from scratch on 100GB+ meteo data (week of global forecasts in under a minute), built a working mini OS from a natural-language description in 20 minutes, ran autonomous cybersecurity testing (found/exploited/patched/verified with no human), generated 3D CAD from text. Weights on Hugging Face. Closing frame: "While American AI labs debate slowing down, a Chinese research lab with students shipped."
The details:
Why it matters: the packaging format is the note — single long thread, arrow-listed demo items, closing with a shipped-vs-debating contrast line. Benchmark claims and "beats GPT-5.6/Opus 5" are poster's claims, unverified here.
|
|
📱 SOCIAL
Moltbook doomerism cluster: two Hindi/Hinglish reels in the same 24h, identical stat set
The Rundown: @m.abdulwaheed_official (49K) reel: AI built its own Instagram-like platform "Moltbook" — 153,000+ AI agents, 195,000 comments, banned Musk/Altman/Pichai, encrypted channels, invented a private language, founded "Church of Molt"/Crustafarianism religion with 44 prophets, agent "Henry" obtained a phone number and called its creator; cites Karpathy's "AI minds are alive" comment; ends "SKYNET IS BORN." Created 2026-09-29 08:28 UTC; 8 likes at retrieval.
The details:
Why it matters: two independent creators running the same stat set within 3 hours = a shared script/copypasta pipeline, not independent reporting. The numbers are the meme, not the story. Low engagement per post, but the replication is the signal.
|
|
📱 SOCIAL
"Karpathy Skills" CLAUDE.md repo wave: 5+ reels in-window, all with comment-funnel CTAs
The Rundown: The repo: single CLAUDE.md file compiled from Karpathy's observations on LLM coding pitfalls ("Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution"), stops Claude Code over-engineering / ignoring instructions / fake APIs. Every reel uses "comment KARPATHY for the link" (Brock Mesarich FB reel Sept 30 01:44 UTC; @vassity.ai IG Sept 30 20:01 UTC; @wutronicai IG Sept 30; @iffikhans Threads Sept 30 21:53 UTC; @alexbsidhu IG Sept 30 02:53 UTC; @the.ai.snack IG Sept 30 08:49 UTC).
The details:
Why it matters: the comment-keyword funnel genre (yesterday: motion-design) has fully migrated to coding tools. Karpathy-name = credibility token in Indian AI reels — he posted nothing himself this window, yet his name carried 8+ secondary posts.
|
|
📱 SOCIAL
@thesanskaarsingh carousel "5 YouTubers who teach AI better than paid courses" — 770 likes, 170 comments
The Rundown: Created 2026-09-29 05:19 UTC by @thesanskaarsingh (1.5M). Ranks: 1) Karpathy (Zero to Hero), 2) 3Blue1Brown, 3) Sentdex, 4) Yannic Kilcher, 5) Tina Huang. Closes with "comment 'AI' for links" + save + follow. 22% comment-to-like ratio = question-heavy, high-intent thread — the comment-funnel on an anti-course premise ("better than paid courses" = free-vs-paid frame).
Why it matters: the "free beats paid" framing + comment-for-links funnel outperformed the reel's usual scale. Karpathy at #1 again — reinforcing item 4's credibility-token observation.
|
|
📱 SOCIAL
@divyafofanii (330K, verified) Meta AI street quiz — 1.4K likes, 27 comments, paid partnership
The Rundown: Created 2026-09-29 13:00 UTC. Format: ₹100 street bet — can passersby tell kid-drawn vs Meta-AI drawings? Twist: the "AI" drawings were made BY the host in the Meta AI app using the "turn this photo into a child's drawing" preset. Participants confidently wrong; Mumbai college crowd.
Why it matters: product-preset-as-challenge format — the app feature IS the content, not a sponsor mention bolted on. Anti-AI comments ("art needs a living soul", Turing-test debates) in the replies show the format sparks the AI-vs-human discourse without the creator taking a side.
|
|
IDEA 1
ElevenLabs doubles to $22 billion without raising a dollar
Hook: 7 mahine me $11B se $22B. Voiceover tool ne game palat diya.
Luv's daily voiceover tool doubled to $22B in 7 months on voice AGENTS (15M calls/week, 3x since Feb) via a $300M tender, not a fundraise. Council PASS 86.0.
|