This past week’s AI news roundup saw Anthropic knock OpenAI from the top spot, AI audio generators landing in court, and Language Learning Models (LLMs) struggle with a simple puzzle. After multiple AI creators touted models nearly comparable to OpenAI’s GPT-4, Anthropic finally released an updated version of its ‘Claude Sonnet 3.5,’ surpassing GPT-4o and Gemini 1.5 Pro on almost all MMLU benchmark tests. Another upgrade, ‘Claude Opus 3.5,’ is due soon, posing a challenge to OpenAI’s rank dominance.
Apple delayed its Intelligence features rollout in the EU following Meta’s canceled launch of Meta AI, both due to the region’s stricter regulations. AI companies Suno and Udio faced a lawsuit by the Recording Industry Association of America for allegedly training their AI using stolen copyrighted sound recordings.
A DeepMind study found deep fakes for opinion manipulation the leading type of AI misuse, with LLMs notably generating plausible yet incorrect information under a condition referred to as ‘semantic entropy.’ Even advanced LLMs are found failing to solve a simple river-crossing puzzle, implying limitations to their logical reasoning abilities.
AI has also been shown to predict human anxiety levels from reactions to photos, a potential asset for understanding human emotions. However, the accessibility of these models is being questioned. While marketed as ‘open,’ companies often restrict certain aspects. Sam’s investigation into the ‘openness’ of AI models unearths these discrepancies.
An exciting development in the open model space is EvolutionaryScale’s ESM3, a biology-focused generative model transforming prompts into proteins, allowing scientists to make advancements beyond previous natural constraints.
Some additional AI news includes Meta marking real photos mistakenly as ‘Made by AI,’ SoftBank’s CEO predicting AI will be 10,000 times smarter than humans within a decade, restrictions planned by OpenAI on its API, and the hesitant launch of OpenAI’s GPT-4o voice assistant due to safety concerns.
Lastly, noteworthy stories also included Toys “R” Us stirring online controversy for an ad created using OpenAI’s Sora, and OpenAI’s acquisition of the collaborative screen sharing tool creator ‘Multi.’