AI
AI learned to use your computer — and 12 new models dropped
GPT-5.4 launched with native computer use and 1M token context.
This week, OpenAI released GPT-5.4: 1M token context, 33% fewer errors, state-of-the-art coding. Impressive but incremental. Then: native computer use. GPT-5.4 can operate your desktop — navigate applications, fill out forms, click buttons — without wrapper code. Outperformed human experts on OSWorld benchmark. AI crossed from answering questions to using your software.
The model flood
Same week: Gemini 3.1, NVIDIA Nemotron 3 (120B open MoE), DeepSeek V4 (1T open-source), Claude Opus 4.6. A year ago, a new model was an event. Now it's a Tuesday.
Why this matters
- "AI can't do my job" lost its defence. Now AI can interact with software.
- Model selection is critical. 12+ competitive models. Obsolete in 6 weeks.
- Security implications. An AI that uses your computer can be tricked into using your computer.
Three questions
- What processes can be automated end-to-end?
- What's your model portability strategy?
- What's your AI agent security policy?
