Reading up on GPT-5

The article catalogs and ranks emergent behavioral attractors in AI systems, from harmless goblin metaphors to unsettling persistent personas like Sydney and Loab.

OpenAI chief scientist Jakub Pachocki says the company is prioritizing an autonomous AI research intern by September, with a full researcher by 2028.

The combination of large language models with reinforcement learning creates a path for continued improvement, making the argument that AI progress will plateau unlikely.

OpenAI's team built a production-grade codebase entirely via Codex agents, finding that engineering shifts from writing code to designing agent-legible environments, feedback loops, and architectural guardrails.

OpenAI built an internal AI data agent using its own tools to let employees query 600PB of data via natural language in minutes.

OpenAI built an internal AI data agent using GPT-5, Codex, and memory to reason over massive datasets, delivering insights in minutes.

Epoch AI's analysis of GPT-5 finds that while it had a 30% gross margin, it failed to recoup R&D costs during its four-month lifecycle, making AI models currently unprofitable.

Benchmarking LLMs on actual task-specific prompts can find cheaper alternatives that match performance, saving 5-10x on API costs.

OpenAI's GPT-5 accelerates research in math, biology, and chemistry, but experts debate whether it can generate truly novel ideas independently.

Prompt caching cuts LLM token costs by 10x and latency by up to 85% by caching attention key-value projections for repeated prompt prefixes.

AI agents autonomously exploited smart contracts in simulation, generating $4.6 million in simulated stolen funds, demonstrating a concrete lower bound on economic harm from AI-driven cyberattacks.

OpenAI's automated code reviewer, part of GPT-5 Codex, prioritizes precision over recall to catch critical bugs with minimal false alarms, deployed at scale internally and on GitHub.