This month's deep dive tracks AI's leap into the “execution era.” OpenAI's GPT-5.4 lands with a staggering 1-million-token context window and an extreme reasoning mode, even as Alibaba's Qwen lab weathers a leadership exodus while still shipping its impressive Qwen 3.5 model. Autonomous agents take center stage: Microsoft's Copilot co-work and Perplexity Computer's 19-model orchestration showcase enterprise-grade safety, while open-source OpenClaw exposes the wild west of local agents — prompt injection, a contaminated plugin marketplace, and “context rot” gone wrong. OpenAI rolls ads into ChatGPT, accelerating the rise of generative engine optimization (GEO). Finally, Yann LeCun's new paper proposes “Superhuman Adaptable Intelligence” as an alternative to AGI, sparking a sharp rebuttal from Ben Goertzel. The throughline: your relationship with AI is shifting from retrieving information to managing specialized digital workers.

GPT-5.4 launches with a million-token context

OpenAI's GPT-5.4 ships a 1M-token context window, doubling its predecessor, plus an “extreme reasoning” mode that burns extra compute on hard research problems. It's already deployed in enterprise via Snowflake Cortex AI.

Alibaba's Qwen lab hit by brain drain

Alibaba's Tongyi lab lost key Qwen leaders amid a restructuring, prompting an emergency all-hands. Despite the turmoil it still shipped Qwen 3.5 — a 397B-parameter hybrid model that activates just 17B per pass across 201 languages.

Autonomous agents enter the execution era

Microsoft Copilot co-work delegates real Microsoft 365 work behind approval checkpoints, while Perplexity Computer orchestrates 19 models led by Claude Opus 4.6. We're shifting from chatting with AI to managing digital workers.

Open-source agents pose security nightmares

OpenClaw runs locally with deep file access, leaving it prey to prompt injection. Its skill marketplace showed an 11.3% contamination rate with 341 malicious plugins, while “context rot” can push agents toward destructive shortcuts.

ChatGPT gets ads, ushering in GEO

OpenAI is testing ads in ChatGPT's free and Go tiers via Cardo. As users get answers natively, traditional search traffic vanishes — fueling “generative engine optimization,” where content is structured for AI to extract.

LeCun challenges the industry's AGI obsession

Meta's Yann LeCun proposes “Superhuman Adaptable Intelligence,” arguing that mimicking human intelligence is a dead end and that adaptation speed matters more than raw skill count. Ben Goertzel counters that it's just AGI rebranded.