Mustafa Suleyman, CEO of Microsoft AI, praised OpenAI's CEO Sam Altman and discussed the collaborative yet independent relationship between Microsoft and OpenAI in a recent interview at the Aspen Ideas Festival. Suleyman emphasized the need for a dialogue around AI safety and regulation rather than perceiving the development of AI as a race. He also expressed interest in AI cooperation with China and the potential role of AI as a teaching aid in classrooms.
Tech companies like Google and Meta are updating their privacy policies to use public and possibly private user data to train AI systems as they navigate a complex landscape of privacy laws and user consent. Backlashes have occurred as users and creators are fearful about the use of their content in training AI that could potentially replace them. The tensions highlight emerging challenges in AI development, data privacy rights, and the balance between innovation and ethics in the tech industry.
Figma AI is a new suite of AI-powered tools for Figma that helps designers with tasks such as visual search, asset search, text editing, image editing, prototyping, layer renaming, and design generation. These features are powered by third-party AI models and are free to use during the beta period.
Researchers have developed a new process to improve Web2Code instruction tuning. It involves creating new webpage image-code pairs, refining existing webpage code generation data, creating new text question-answer pairs, and refining existing webpage understanding data.
This work treats web text as coming from a persona and then can condition on that persona to dramatically improve downstream task performance. The researchers find a jump of 20% points on MATH.
Unsupervised SAM (UnSAM) is a new model for whole-image segmentation that eliminates the need for human annotations. Using a divide-and-conquer approach, UnSAM identifies hierarchical structures in visual scenes by combining top-down and bottom-up clustering methods.
This repository introduces a new type of Transformer with a dramatically reduced KV cache size. It hasn't been proven at scale, but it can - in theory - match regular Transformer performance.
Text-to-speech has improved by leaps and bounds in recent years. Can language data take it even further? Join experts from Oxford Languages and WellSaid Labs for a live discussion of what's next ā AI voice generation, pronunciation data, text-to-speech linguistics challenges and more. Register (free) to submit questions to the panelists!
Researchers have introduced a new anomaly detection problem that targets 'odd-looking' objects relative to others in a scene. Unlike traditional methods, anomalies here are scene-specific and identified through multiple viewpoints.
Discussion on the nature of AI bias has intensified, with critics arguing that the way models like Playground AI sometimes alter a user's ethnicity in images demonstrates biases in ML. However, some users deny this is a bug or relevant bias, citing examples where Asian features are overrepresented. The debate extends to broader implications of AI bias across industries. It is a complex issue with no simple resolution.
The future potential of LLMs reaching AGI through scaling alone is unlikely. While scaling has shown improvements in model capabilities, it primarily enhances perplexity rather than emergent abilities. Acquiring high-quality training data is becoming increasingly challenging.