AI Watch: Hinton's First RSI Paper, OpenAI's 28-Day Plan, and Image Generation Defaults

AI Watch: Hinton's First RSI Paper, OpenAI's 28-Day Plan, and Image Generation Defaults

A report was headlined “Just now, Hinton published his first RSI paper,” claiming that AI has begun to truly enter the pipeline of “building the next generation of AI.” ^[1]

Another item, dated October 5, showed that OpenAI's head of core products and platform, Thibault Sottiaux, said on X that the team would continue to push forward improvements to Codex and ChatGPT Work over the next 28 days. According to him, OpenAI would release one feature update each day that clearly improves the experience for most Codex / Work users and has practical significance, or otherwise provide a “reset.” The period was expected to run at least from October 5 to November 2. ^[5] A related headline summarized it as a “limited-time 28-day” commitment. ^[3]

OpenAI also announced that it would test visual ads when ChatGPT generates images. When users generate images, the interface may show ads featuring product inspiration, ways to use products, or service experiences; OpenAI said the ads would be clearly labeled and kept separate from the image being generated, would not be directly embedded into or alter the image the user creates, and would not affect answers. The test was expected to launch in the United States in late October, initially only for a group of advertisers. ^[5]

On the image generation side, an analysis article mentioned that in early May, in a broadcast video of the Korean professional baseball league on X, women in the stands were identified as AI-generated, but the video later became a template, with users inserting their faces into similar broadcast footage to generate “themselves caught by the stadium camera.” ^[2] The article recalled that after GPT-4o could draw images in March 2025, Ghibli-animation-style photos briefly flooded social media; after Google's Nano Banana launched in late August, users generated more than 200 million images in two weeks. ^[2] The article further treated “beautiful women” as the default in AI image generation and traced it back to the history of the 1973 Lena Soderberg image being widely used as a test image in the image processing field. ^[2]

From this evidence, the AI industry is reinforcing the narrative of “self-improvement” and next-generation models while advancing commercialization on the product side through high-frequency updates, reset commitments, and ad testing; popular image generation trends continue to be entangled with long-standing default aesthetics and the inertia of test materials. It should be noted that the above cases include news from October 5 and late October as well as time contexts from 2025 and 1973, and they should not be conflated as news from a single point in time.