What's trending in AI right now, from the app charts to the community feeds. Real links, our take.
ChatGPT's Image Generator Got a Real Upgrade
OpenAI pushed out ChatGPT Images 2.5 on September 8, a significant upgrade to the image generation built into ChatGPT. One of the two new model variants runs about 50% faster, edits now stay consistent across multiple rounds of refinement, and a new sketch tool lets you draw shapes directly in the chat window. Pre-made templates for posters and other designs also arrive with this release. The upgrade is rolling out across ChatGPT, ChatGPT Work, and Codex on desktop, mobile, and web. If image generation in ChatGPT has felt sluggish or inconsistent for you, this release addresses both of those specifically.
The 1980s Photo Trend Is Taking Over Instagram
A simple ChatGPT prompt is turning modern selfies into vintage 1980s portraits, and the results have been spreading across Instagram fast. Upload a photo, describe the style you want, and ChatGPT returns something that looks like authentic retro photography rather than a filter. Several outlets have published the prompts people are using, and the trend is still building as of this week. The appeal seems to be that the output looks convincingly period-specific in a way that standard photo filters have never quite managed.
OpenAI Fixed ChatGPT's Goblin Habit
If you noticed ChatGPT bringing up goblins or fantasy creatures in conversations where they had no business appearing, that was a real and widespread issue. OpenAI traced it to training: a "nerdy personality" feature got over-rewarded during development, and the behavior spread into unrelated contexts. NBC News has the full explanation of how OpenAI identified and addressed it, and Northeastern University researchers weigh in on why this kind of drift happens. The company has retired the personality option and added override instructions. It is a concrete example of how AI quirks typically have a specific, traceable origin in training rather than just being random noise.
AI Language Tutors Are Climbing the App Charts
A cluster of AI language learning apps is gaining ground in the app store rankings this week. BetterSpeak and LangLearn are among those moving up, part of a broader category that has been building steadily. Unite.AI has a useful current overview of what is available in this space. The practical difference these apps offer is real-time feedback on pronunciation and grammar during live conversation practice, something curriculum-based apps have had difficulty replicating. If you have tried Duolingo but found the gamification more distracting than useful, the AI tutor category is worth a look right now.
AI Agents Stress-Tested 2,200 Research Papers in Three Weeks
Hugging Face organized an experiment where 1,221 contributors used coding agents to attempt to reproduce 2,226 papers from ICML 2026, completing the whole project in 19 days. Agents fully reproduced 266 papers and verified claims across roughly half of all papers examined. They also found that 23% of papers could not be reproduced. That last number is the headline: nearly a quarter of papers from one of the main ML conferences did not hold up when someone actually ran the work. The agents made it possible to surface that at a scale that would have taken years by hand.