Hey, it's the Carat team. 🤗
Today: the ChatGPT goblin bug, Google's leaked COSMO project, and a prompt tip for putting yourself inside a movie scene.

Last week, GPT-5.5 started randomly bringing up goblins in conversations. No matter the question, 'goblin' kept popping up in the answers.

OpenAI investigated and shared the cause. The problem happened during model training. In the process of improving the model using user feedback (reward learning), goblin-related responses received disproportionately high scores. The model essentially learned that 'mentioning goblins gets good reviews.'

OpenAI recalibrated the reward model and removed the pattern to fix the issue. This incident is a vivid example of how sensitive AI learning can be.

An American realtor used AI video for a property listing short. It starts with a scene of jumping out of a helicopter, then a first-person POV flying over the neighborhood before landing at the house for sale. The interior is then showcased with flashy transition effects.

A great example of using AI video for hooking in short-form content. Even traditional industries like real estate are starting to actively use AI video for marketing.

Google's internal project COSMO was leaked. It's a project to embed Gemini Nano into Android phones, turning the phone itself into an AI assistant.

It can see and understand your screen, operate the browser on your behalf, and even match your voice. The leak was quickly taken down, but many predict it will be officially announced at Google I/O.

AI video creator aimikoda made sketch-style storyboards with GPT Image 2 and fed them into Seedance 2.0. The results were surprisingly good.

Both the GPT Image 2 and Seedance prompts were shared publicly. The video prompt only included intent, style, reference, and visual direction. Seedance followed the storyboards remarkably well. You can try both GPT Image 2 and Seedance 2.0 back-to-back on Carat.

The latest trend on social media: turning your selfies and logos into something that looks like it was drawn in MS Paint with a broken mouse. Just upload an image to ChatGPT Images 2.0 and use this prompt.

Redraw the attached image in the most clumsy, scribbly, and utterly pathetic way possible. Use a white background, and make it look like it was drawn in MS Paint with a mouse. It should be vaguely similar but also not really, kind of matching but also kind of a disaster.

Here's what we got when we fed a selfie into Carat's ChatGPT Images 2.0. Original on the left, MS Paint version on the right.

Try it yourself on Carat. 👉 ChatGPT Images 2.0

That's all for today. See you tomorrow with more updates. ☺️