[💃 Today's AI News! 💃] ✔️ Today, OpenAI officially released GPT-5.4! This update seems quite significant. It features native computer use, enabling direct mouse and keyboard control by viewing screenshots, and expanding the context window to 1M tokens. (Is Claude watching?) Furthermore, it surpassed human performance (72.4%) in the OSWorld benchmark at 75%, and individual claim errors were reduced by 33% compared to the previous version. GPT-5.4 Thinking and Pro versions are being rolled out sequentially to ChatGPT, API, and Codex. @sama also mentioned the release of the Codex Windows app and the ability to change direction mid-response. ✔️ @AnthropicAI is a bit volatile. Dario Amodei has declared legal action against the Department of Defense's supply chain risk designation. Anthropic had been actively deploying Claude to the Department of Defense, but it drew a red line on autonomous weapons and domestic mass surveillance. The Department of Defense took issue with this and designated it as a supply chain risk. Amodei, however, stated that it was "legally unjustified" and would challenge it in court. They apologized for the leaked internal memo, admitting they were "careless." Meanwhile, reports surfaced that OpenAI had struck a deal with the Department of Defense to replace Anthropic. ✔️ Andrej Karpathy on nanochat reduced GPT-2-class model training to 2 hours on a single 8xH100 node. A month ago, it was 3 hours, but he said the implementation of FP8 and the migration to the NVIDIA ClimbMix dataset were key factors. This experiment involved running eight AI agents (four Claude agents and four Codex agents) simultaneously to automatically optimize nanochat. They reportedly reduced validation loss with 110 changes in just 12 hours. They themselves said, "It's still a mess, but this is the direction." Wow! ✔️ @perplexity_ai also began supporting GPT-5.4 and GPT-5.4 Thinking for Pro and Max subscribers immediately, and added voice mode to Perplexity Computer, enabling voice-controlled AI browsers. Personally, I think it's significant that GPT 5.4's native computer use has been integrated into the general frontier model. Claude was the first to introduce computer use, but OpenAI's integration of this into its flagship model marks the beginning of the agent era.