OpenAI Develops Recursive AI Features and Supports Self-Improving Systems Startup

OpenAI has released ChatGPT Images 2.0, an image model with capabilities for web searching, generating multiple images from one prompt, and auditing its outputs. The model can produce photorealistic screenshots of simulated ChatGPT conversations, demonstrating improved reasoning and context handling compared to earlier image generation systems.

Alex Wissner-Gross stated on X that the model achieved a record lead in text-to-image benchmarks and swept image arena leaderboards.

systems are increasingly self-referential, contributing to the company's work on new models. A startup partially funded by OpenAI, along with other investors, is developing AI systems designed to improve themselves without human involvement. The startup's team includes co-founder Richard Socher, with experience in natural language processing from Salesforce, and Tim Rocktäschel, a professor at University College London and former senior researcher at Google DeepMind.

The team also features former OpenAI researchers Josh Tobin, Jeff Clune, and Tim Si, as well as talent from Meta and Google. Matthias Bastian reported at The Decoder that the startup aims to build an AI system that keeps improving itself without human involvement.

and Broader AI Applications

A leaked screenshot from OpenAI's Codex tool shows a model picker with unknown options, including 'arcanine' and 'glacier-alpha'. These models have no publicly available information, and expectations exist for potential releases within a week. In related developments, AI is being applied to modernize legacy systems.

Serena Clifford wrote at Fullstack that AI-augmented modernization uses techniques like agentic swarm coding and AI-assisted code refactoring to update older enterprise systems. This approach helps analyze codebases, identify dependencies, and translate outdated languages into modern ones, addressing costs and inefficiencies of legacy systems.