Everything New in The World of AI!

OpenAI has rolled out DALL·E 3 across ChatGPT Plus, and the biggest practical change is that DALL·E 3 is now available directly inside the ChatGPT mobile app. That means subscribers can generate images on their phones without switching tools—though some users complain the ChatGPT interface is more restrictive, blocking requests involving famous characters unless users bypass safeguards. For beginners, the convenience still makes ChatGPT a low-friction entry point, even if power users often prefer Microsoft Bing’s comparatively looser environment for image generation.

Midjourney is responding on two fronts: higher-resolution output and a push toward a broader web and mobile experience. A new upscaler with 2X and 4X options is live, targeting up to 16 megapixels—described as “photograph size.” Early results look detailed at a distance and natural overall, but close-up inspection shows the upscaling can behave more like a detail enhancer than a perfect recreation of fine textures, and background elements sometimes remain inconsistent. Midjourney is also testing a redesigned website in beta that loads faster and improves browsing, but generation on the site is not yet available in that beta. Access requires at least 1,000 generations on an account, suggesting Midjourney is gating features while it stabilizes the new platform.

The most notable Midjourney-related move may be the launch of a mobile app ecosystem under the name “Nii Journey,” built with Spellbrush and positioned as an AI anime generator. The app supports prompt-based generation, image-to-image, and community-style live feeds where users can watch others’ generations in real time. Midjourney images can also be generated inside the app, and the social, feed-driven design hints at a strategy to reduce reliance on Discord—where Midjourney’s workflow has historically lived.

Beyond image generation, the roundup highlights open-source progress in “audio understanding.” A model called SALMON (Speech Audio Language Music Open Neural network) can analyze audio to produce text descriptions and interpretations, including recognizing background sounds such as gunfire and explosions. The same system can also respond to uploaded music—turning a piano-and-vocal track into an interpretive description of mood and structure. The practical implication is accessibility: audio-to-text systems could help people who are deaf by describing what’s happening around them without needing other people as intermediaries.

A separate AI art development, “PixArt Alpha,” is pitched as a fast-training text-to-image diffusion transformer that can train in about 10% of the time compared with Stable Diffusion-family approaches, with claims of competitive image quality against models like Google’s IM and Midjourney. The efficiency angle matters because it could make home experimentation more feasible, especially since the project plans to co-release code and weights.

On the autonomous-agent front, Hyperight’s assistant (run by CEO Matt Schumer) received an update that improves task completion and adds source citation. A demo shows the assistant searching for restaurants, then making an OpenTable reservation by navigating the booking flow and filling in details automatically. Finally, Meta shared research toward near-real-time decoding of visual perception from brain activity using a system called “MEG,” producing rough reconstructions of what participants viewed after only one second—accurate enough to identify broad categories and colors, while still missing fine details like faces.