
AI Unraveled: Latest AI News, ChatGPT, Gemini, Claude, DeepSeek, Gen AI, LLMs, Agents, Ethics, Bias [FULL RUNDOWN] GPT-5.4’s Computer Use, Anthropic’s "Safety Theater" Memo, and the Death of Hallucinations (March 5th Rundown)
Mar 5, 2026
They unpack GPT-5.4’s new native computer use, huge context window, and model tiers. The conversation covers a leaked memo accusing safety posturing over a Pentagon deal and the tensions fracturing the AI industry. They highlight token-level hallucination detection, agent governance as agents gain OS access, and global moves from China to Apple shaping AI deployment and provenance for synthetic media.
AI Snips
Chapters
Transcript
Episode notes
GPT 5.4 Enables Native Computer Use
- GPT 5.4 introduces native computer use via a Windows Codex sandbox, turning the model from a chat oracle into an agent that can click and type within an isolated environment.
- The sandbox gives structured access to file system and accessibility APIs while virtualizing kernel access to prevent destructive actions like registry rewrites.
Reduced Hallucinations Matter More With Agents
- OpenAI claims GPT 5.4 is more factual: 18% fewer errors and 33% fewer hallucinations versus GPT 5.2, crucial when agents act on your files.
- The Thinking variant supervises subagents and simulates outcomes before executing to compound safety improvements.
Apple's Token Level Hallucination Detection
- Apple published Reinforcement Learning for Hallucination Span Detection to detect hallucinations at the token/span level instead of a binary flag.
- The RL detector is trained to match human annotators and outperformed prior methods on the RAG Truth benchmark, enabling targeted auditing.
