AI
TLDR: NEW YORK—Axios CEO Jim VandeHei spent a year feeding AI personal and business data, then testing OpenAI Codex and Anthropic Claude Code across Axios. He says AI can outperform humans in complex diagnosis and research, but current agent work breaks on security, access, and systems integration.
Key Takeaways:
- Axios tested AI across every department and trained staff for agent to agent work, led by VandeHei using a phone controlled AI personal operating system.
- He logs 1 to 2 hours daily on ChatGPT or Claude, including uploading MRI and blood work and running clinical checks with his doctor.
- He forecasts job change over time plus new revenue lines, while warning that scaled deployment stalls on human workflows, security gates, and agent collaboration.
VandeHei’s experiment reads like an origin story for AI as a personal co worker, except the boss is also the intern. The real bottleneck is not intelligence, it is getting AI to act safely in messy human systems.
VandeHei’s experiment reads like an origin story for AI as a personal co worker, except the boss is also the intern. The real bottleneck is not intelligence, it is getting AI to act safely in messy human systems.
Q&A
What does it mean that the biggest blockers are security and access rather than model quality?
It suggests companies will have to treat AI deployment like identity and permissions engineering. Even smarter models stall when they cannot reliably see, verify, and act on the right internal data.
If AI can outperform many people most of the time, why do users still get unimpressive results?
Because performance depends on feedback and iteration. VandeHei emphasizes daily training inputs and explicit guidance, which casual users often skip.
How could uploading personal medical records change the clinician conversation inside a company?
It can shift AI from a general assistant toward an evidence backed advisor, but it also raises governance questions. Companies may need medical validation workflows and strict data handling rules.
Why is agent to agent coordination harder than using one chatbot well?
Agents multiply edge cases, data sharing constraints, and timing issues. If one agent cannot confirm what another knows or is allowed to access, the collaboration collapses into operational chaos.
What kind of training could help rank and file workers become effective AI accelerators without becoming AI specialists?
Firms can focus on prompt discipline, task decomposition, and review habits that create a feedback loop. The goal is not coding skill, it is consistent evaluation and iterative improvement.
Top in AI
OpenAI IPO filing tests private speed versus public scrutiny
Bristol Myers Shrinks Procurement Timelines Using AI Readiness Doubts
Gulf funds double down on AI as IPO clock ticks
NotebookLM turns prompts into documentary grade art decks
No comments yet. Be the first to share your thoughts!