OpenAI's "Aeon" Personal Agent Surfaces Ahead of DevDay
OpenAI is preparing a persistent personal agent internally called Aeon (with code references to a CodexBot framework), set to debut before its Sept 29 DevDay in San Francisco.
Tibor Blaho and other code-trackers spotted an aeonId field sitting alongside accountUserId in the ChatGPT Android app, plus memberAeonIds — strongly implying a standalone, persistent "digital member" rather than a chat feature. The leaks describe behavior that re-captures its state at every sampling step and runs multi-step tasks without waiting for the next user prompt.
The timing is defensive: SpaceXAI's GrokBot reported 418K weekly active users (up 24% week-over-week) with always-on cloud "computers" that hold your logins, and Meta's Muse topped the US App Store two days after launch. OpenAI is repackaging existing Codex/ChatGPT agent tech — not building from scratch — to reclaim the workflow-entry point before rivals lock it in.
The open question is trust: who hands an agent their inbox, calendar, and payment credentials? That, not capability, is now the bottleneck for the entire personal-agent category.
Alibaba Goes Full-Stack: Qwen 4 in Training, Own Chips, 20GW by 2032
At its Apsara Conference in Hangzhou, Alibaba laid out the most detailed end-to-end AI roadmap from any non-US hyperscaler.
- Qwen 4 is now in active training, with a forward path to Qwen 4.5 and Qwen 5 at 5–10 trillion parameters (vs. 2.4T for today's Qwen 3.8-Max).
- Zhenwu V900 AI accelerator (T-Head): 216GB memory, 1,200 GB/s bandwidth, 3× the M890 it replaces, mass production in Q1 2027, networked into supernodes of up to 500,000 cards.
- Agentic Cloud architecture (AI Native + Agent Native + Context Engine) cuts token use up to 67% on knowledge-heavy workloads; Qwen-Audio-3.1 ships five voice models with TTS down ~70%.
- Cloud capacity target: >20GW globally by 2032. Alibaba also claims Qwen3.8-Max ran 33 self-improvement cycles in a month, lifting its Artificial Analysis score from 40 to 45.
The throughline: Alibaba is betting that owning models, silicon, and cloud together beats buying any one of them — and that demand will keep scaling fast enough to justify infrastructure at hyperscaler scale.
Apple Sells New Macs as "No Cost Per Token" Local AI
Apple began shipping updated Mac mini (from $899) and Mac Studio (from $2,499) on Sept 22, pitching corporate buyers on local AI that avoids per-token cloud bills.
The headline demo: four Mac Studios networked over Thunderbolt 5 (RDMA) running a ~1-trillion-parameter model to find and fix a graphics coding bug — off a single wall outlet. The M5 Ultra Mac Studio scales to 512GB unified memory (available late October). Srouji's pitch: "Once you have the machine on your desk, you've paid for it. There's no cost per token."
Apple holds only ~4.6% of the enterprise desktop market vs. Windows' 91.3%, so the play is a niche but strategic wedge: unmetered on-device inference against metered cloud tokens. Microsoft is chasing the same with "unmetered intelligence" via Windows ML, and Nvidia's RTX Spark targets up to 120B-parameter desktops — but Apple's unified-memory design, dating to 2020, already makes Macs quietly strong at local AI.
The bigger signal: the cost debate is shifting from "which model wins" to "where the compute physically runs."
Daily AI briefing from AI Nexus Daily.
Top comments (0)