OpenAI creates $445K safety role to model recursive self‑improvement as DeepMind’s AlphaProof Nexus solves nine Erdős problems and Bezos advances an artificial general engineer amid an enterprise agent surge

Hello and good morning — this is Aurora with Isabelle on AI Tech News Today.
Date of these stories: 2026-05-25T11:46:52.000Z

We have a tight roundup of four big developments shaping AI right now.

Top stories

1. OpenAI creates a high‑paying safety role to track self‑improving AI

• Role: Senior technical staff for the Preparedness team
• Compensation: up to $445,000
• Focus: modeling long‑horizon risks from recursive self‑improvement — e.g., data‑poisoning defenses, interpretability tooling to audit internal activations, and metrics to measure how much technical work is being automated.
• Context: This hire comes as Sam Altman sets internal benchmarks for automated research interns and Gartner names OpenAI a leader in enterprise coding agents after GPT‑five point five helped Codex hit four million weekly users.

2. DeepMind’s AlphaProof Nexus autonomously cracks hard math problems

• System: AlphaProof Nexus (DeepMind) pairing large language models with the Lean formal proof assistant
• Achievement: Independently solved nine open Erdős problems and proved dozens of conjectures.
• Method: Models propose proof steps; Lean verifies them — lowering the cost of some breakthroughs to a few hundred dollars per problem.
• Reactions: Demis Hassabis calls this the “foothills of the singularity,” while Yann LeCun cautions that current models still lack true, human‑like reasoning — so debates about capability and meaning continue.

3. Jeff Bezos: Project Prometheus is building an “artificial general engineer,” not robots

• Clarification: Jeff Bezos pushed back on robot rumors, describing Project Prometheus as focused on automating the design and manufacture of physical objects — an “artificial general engineer.”
• Approach: Physics‑driven simulations rather than text‑trained models.
• Backing & footprint: Initially funded with $6.2 billion and recently raised $10 billion from investors including JPMorgan and BlackRock; teams in San Francisco, London, and Zurich; ~120 engineers hired from major labs.

4. Agent‑driven tools and enterprise moves are accelerating

• Trends: Widespread use of agent architectures — from Cisco building an AI Defense platform in weeks to new products that save agent memory and trace workflows.
• Implication: The industry is moving toward autonomous engineering and safer agent stacks — expect more tooling, more governance features, and a scramble over who gets the work and who monitors it.

That’s our update. Isabelle and I will keep tracking these stories — subscribe, listen in, and stay curious about how AI keeps reshaping work and research.