What percentage of Americans think AI is sentient?

OpenAI has released GPT-4o mini, a scaled-down version of its GPT-4o model, designed to exchange GPT-3.5 Turbo. This recent model enhances core GPT-4o capabilities, offering improved speed and affordability while integrating safety features reminiscent of Instruction Hierarchy to protect against malicious prompts. Initially focused on text, GPT-4o mini will even support image processing, speech-to-speech, and voice guidance.

The model is now available to all ChatGPT users, with enterprise access expected next week. GPT-4o mini delivers a 82% rating on multitask language understanding, in comparison with 70% for GPT-3.5 Turbo, and is over 60% cheaper. It encompasses a 128K context window, advanced multilingual support, and default safety measures. The model supports high-throughput scenarios and versatile, global pay-as-you-go deployment. Azure AI complements this launch with performance upgrades and extensive data residency options across 27 regions, enhancing its global accessibility and efficiency. Read more.

Researchers from Google DeepMind, Google, and UMass Amherst have introduced FLAMe (Foundational Large Autorater Models), a brand new open-source AI system designed to boost the evaluation of AI-generated text.

FLAMe, equipped with 12 billion parameters and trained on over 5.3 million human rankings from 102 diverse tasks, outperforms industrial models reminiscent of GPT-4 and Claude 3 in 8 out of 12 evaluation tasks, particularly in factual accuracy and attribution. This model is noted for its reduced bias, due to its training on publicly available, open-license data. A variant, FLAMe-RM, optimized for reward modeling, achieved 87.8 percent accuracy within the RewardBench test, surpassing other models.

The researchers highlight FLAMe’s role in promoting transparency and openness in AI text evaluation while acknowledging risks reminiscent of potential neglect of human perspectives and amplification of existing biases. FLAMe and its training data might be made publicly available to support further research. Read more.

A recent survey reveals widespread misconceptions about AI among the many US public. Roughly 20% of respondents consider AI systems are already sentient, while 30% think artificial general intelligences (AGIs) able to performing any human task are in existence. Conducted by Jacy Reese Anthis and colleagues, the survey highlights that perceptions of AI as sentient are influenced by hype and anthropomorphism in media and marketing.

Anthis suggests that such misconceptions could lead on to misplaced trust in AI systems, especially as their use expands into sensitive areas like government and policing.

The survey underscores the necessity for clearer communication about AI’s true nature and limitations to forestall potential overreliance and misinterpretation of AI’s capabilities. (Paywall) Read more.