Frontier Capability Developments

Nvidia announced DLSS 5 at its GTC conference, introducing generative AI directly into real-time graphics rendering rather than using it purely for upscaling. CEO Jensen Huang framed this as the 'GPT moment for graphics' — blending traditional hand-crafted rendering with generative AI to deliver visual output. Early reactions are sharply divided, with some developers calling the result 'slop' that unacceptably alters artistic intent, as reported by The Verge. The technology appears to function as a real-time generative AI filter applied to game visuals, fundamentally different from previous DLSS versions that focused on resolution upscaling through neural networks trained on high-quality reference images.

The controversy centres on whether generative AI should modify the visual output of games beyond what developers explicitly designed. Previous DLSS iterations aimed to reconstruct detail that was downsampled for performance — a technically sophisticated interpolation but one that preserved authorial intent. DLSS 5's generative approach may introduce visual elements not present in the original render, raising questions about whether players are experiencing the game as intended or a generatively altered version. This mirrors broader tensions in AI deployment: when does assistance become substitution, and who decides what constitutes acceptable modification of creative work.

Encyclopedia Britannica and Merriam-Webster filed suit against OpenAI alleging that GPT-4 'memorized' their copyrighted content and generates responses 'substantially similar' to their original work, as reported by The Verge. The complaint states that OpenAI repeatedly copied Britannica's content without permission for training purposes. This lawsuit differs from earlier copyright challenges by focusing explicitly on memorisation — the model's ability to reproduce training data verbatim or near-verbatim — rather than broader questions about whether training constitutes fair use.

The memorisation angle is strategically sharper than general fair use arguments. If plaintiffs can demonstrate that GPT-4 reliably reproduces substantial portions of Britannica or Merriam-Webster content when prompted, they may sidestep the transformative use defence that has protected some AI training practices. This approach aligns with recent research showing that large language models do memorise training data, particularly when that data appears multiple times or consists of distinctive factual sequences. The outcome could force model developers to implement more aggressive memorisation mitigation during training, potentially degrading model performance on knowledge-intensive tasks where verbatim recall of facts is valuable.

OpenAI's delayed 'adult mode' for ChatGPT will reportedly support text-based erotica at launch but not image, voice, or video generation, according to The Verge. An unnamed OpenAI spokesperson described the content as 'smut rather than pornography,' suggesting a distinction between written sexual content and explicit visual or audio material. This incremental approach contrasts with the company's initial positioning of adult mode as a comprehensive policy shift, and the continued delay since announcement indicates internal debate or external pressure over implementation.

The text-only limitation reveals OpenAI's risk calculation: written erotica has longer precedent as a content category (published literature, fanfiction platforms) with clearer legal boundaries than AI-generated visual pornography, which raises questions about consent, deepfakes, and CSAM generation risk. The limitation also reduces technical attack surface — text generation is easier to filter for prohibited content than image generation, where adversarial prompts can bypass safeguards more effectively. However, limiting adult mode to text while competitors like Mistral and open-weight models offer fewer content restrictions may put OpenAI at a competitive disadvantage among users seeking unrestricted AI tools.