In Its Own Window

With ChatGPT Images 2.0, the model maker flexes on powerful renderings that illustrate not just “pretty picture skills” but also reasoning. Remember, it was only a few years ago that image genAI spat out incoherent text universally. Now, the image maker can show you a world that conforms to yours, because it’s endowed with a much bigger context window. My colleague Alex Wissner-Gross wrote about this on X this way:

“OpenAI released ChatGPT Images 2.0, its first image model with thinking capabilities, able to search the web, generate multiple distinct images from one prompt, and audit its own outputs. GPT-Image-2 promptly swept every Image Arena leaderboard with a record +242 point Text-to-Image lead, and in a dizzying bit of recursion, OpenAI demoed the model generating photorealistic screenshots of ChatGPT conversations.”

That’s pretty impressive, but it’s not the only type of recursive AI going on over there. Increasingly, OpenAI systems are self-referential. They contribute to the work that goes on at the company, which is focused on new models that will blow the prior ones out of the water.

AI Builds AI

It turns out there’s a startup partially funded by OpenAI, among others, working on this kind of reality, where AI can design its own successors in a profoundly strange kind of regeneration, without too much human oversight at all. Think of the available footage of humanoid robots building humanoid arms and hands for their brethren.

Recursive Superintelligence is brand new. Internal documentation explains its origins this way:

“Co-founder Richard Socher brings deep natural language processing experience from Salesforce, while Tim Rocktäschel, a University College London professor and former Google DeepMind senior researcher, brings experience from the Genie world model project. The team also includes former OpenAI researchers Josh Tobin, Jeff Clune, and Tim Si, alongside talent from Meta and Google.”

That’s an impressive roster. What is the company for?

Recursive Superintelligence, which hasn't officially launched yet, wants to build an AI system that keeps improving itself without any human involvement,” writes Matthias Bastian at The Decoder. That’s quite a heady goal.

But there’s more: in this wild context, with so many working on advancing AI platforms, autonomously or otherwise, there are also rumors of new OpenAI models about to drop, and they’re not just based on nothing.

Codex Displays Unknown Models

There’s a model picker inside of OpenAI’s Codex, a drop-down list showing available options. Somebody “leaked” a screenshot of this, with model titles including “arcanine” and “glacier-alpha.”

Importantly, both of these are unknown models, with no publicly available footprint. GPT, for its part, swears that any mention of them relies only on circumstantial evidence. But some of the humans close to these projects expect something to drop within a week. What do you expect from a model named “arcanine”? It’s hard to say. We don’t have any specs.

What’s clear, though, is that AI is, as a whole, still heating up. Companies are getting involved in this massive sea change, to be integrating AI into systems, or to be left behind in the shuffle. To that end, legacy migration is paramount.

“Legacy systems weigh companies down with costs, risks, and inefficiencies,” writes Serena Clifford at Fullstack. “AI-augmented modernization offers a faster, more practical path forward. Unfortunately, updating these systems can feel like a catch-22: Although traditional modernization is costly, time-consuming, and disruptive, delaying modernization efforts is high-risk.”

So – Clifford suggests – use AI:

“AI-augmented legacy modernization, or AI-assisted legacy modernization, is the use of AI throughout the software modernization process,” Clifford writes. “Software engineers employ techniques such as agentic swarm coding and AI-assisted code refactoring to update and migrate older enterprise systems into modern, scalable environments. Instead of relying only on manual work, tools powered by machine learning and generative AI help developers analyze large codebases, identify dependencies, and translate outdated programming languages into modern ones.”

That’s another example of recursive AI: using AI to drive systems to the new AI they need, while also using AI to create AI, and modeling AI behavior with AI. Now that truly does sound recursive, doesn’t it?

Stay tuned.