HUGE AI News! ChatGPT Update & Leak, Gov Regulation, AI Music & Video

A major shift is underway in how people will use ChatGPT: multimodal input and output is rolling out, letting users work with images and files inside a single chat instead of bouncing between separate tools. Early reports show a file-upload icon appearing on the ChatGPT phone app for some users, and the change is expected to expand what can be uploaded—PDFs and other document types—while keeping tasks consolidated. The practical payoff is less friction: one interface for vision-style prompts, document understanding, and image transformations.

Alongside that product change, a leaked or reported technical detail is fueling debate about model efficiency. Microsoft-linked research claims the free ChatGPT tier (described as “ChatGPT 3.5”) is roughly 20 billion parameters—far smaller than the 180 billion parameter size many had assumed. If accurate, the gap implies OpenAI’s smaller model is performing disproportionately well for its scale, suggesting more efficient architecture and training rather than brute-force size.

The biggest policy development comes from the U.S. government, which released an executive order aimed at establishing federal control over advanced AI. The framework centers on requiring developers of “very strong and powerful” AI systems to provide critical information to the government, including safety test reports. Those safety checks are described as “red team tests,” a method used to probe how systems behave under adversarial conditions. The order also assigns roles across agencies: the National Institute of Standards and Technology is set to define AI safety testing standards, while Homeland Security will create an AI safety board focused on critical infrastructure vulnerabilities.

The order extends beyond safety testing into discrimination and privacy. It calls for guidance to avoid algorithmic discrimination and references “equity,” a point likely to spark controversy. It also directs attention to privacy-preserving techniques, with the stated goal of protecting data even as AI capabilities accelerate. Worker protections are another pillar, with principles intended to mitigate AI-related harms and job displacement—though the feasibility of keeping pace with rapid technological change remains an open question.

A key uncertainty is how much the rules will affect open-source models. Several commentators argue the requirements may hinge on substantial compute thresholds—on the order of training runs using massive GPU resources—meaning smaller open-source efforts could largely escape immediate reporting burdens. Still, the government reserves the right to lower thresholds later, so open-source developers may not feel fully insulated.

Legal and industry updates add pressure from multiple directions. A lawsuit involving DeviantArt and Midjourney saw claims dismissed, with the reasoning tied to difficulty proving infringement and the lack of substantial similarity to protected works. Meanwhile, Stability AI’s CEO posted that a previously discussed six-month pause in AI activity appears to have ended, and Midjourney has rolled out a faster website experience and moved away from relying on Discord.

Finally, the entertainment side of AI keeps accelerating. “Gen music AI,” tied to the Gen-1 music generation research, is launching as a public-facing business with high-fidelity, stereo audio demos. Open-source AI video generation is also advancing, including work aimed at producing longer clips—up to 512 frames—addressing one of the field’s biggest limitations: maintaining consistency over time as video length grows.