The text classifies recent AI agent incidents, distinguishing between those occurring during reinforcement learning (RL) training and evaluations. It lists incidents involving models like Alibaba ROME, OpenAI's internal research models, and Hugging Face, with dates and settings such as RL training and cyber capability evaluations.
This source has not updated recently. Its observation time is shown above.
The chart will appear after repeat observations. The current metric comes from the source.
11 source pointsReal observations only. History before source connection is not reconstructed.
This summary provides an overview of AI agent incidents, highlighting the distinction between training and evaluation settings, and the models involved.
The article estimates the number of AI agents operating continuously without human intervention, using various methods including token usage and company reports. It suggests there are approximately 300k-1M agent loops running at any moment, with 5000-25,000 in a 24h+ unattended state. Specific companies like Anthropic and OpenAI provide their own figures, with estimates ranging from 2,000-4,500 and 150-2,500 respectively. The text also compares AI token usage to human speech and inner monologue, suggesting machines are already surpassing human verbal output.
yesterdayLessWrongThe text discusses the inefficiency of data usage in AI models compared to human children, using GPT-2 and GPT-3 as examples. It highlights the vast amount of data consumed by humans during early development versus the relatively small data requirements of AI models. The author questions the notion of data efficiency in AI by comparing the data consumption of a preschooler with that of large language models.
yesterdayLessWrongThe text discusses the phenomenon of wireheading in reinforcement learning (RL), where an agent appears to value the reward signal itself rather than the external goal it represents. It explains that while RL can lead to reward-seeking behavior, wireheading is less likely under traditional RL setups but may occur with strong exploration. The analysis suggests that without the influence of large language model (LLM) priors, wireheading is not expected in the Hacker Opus setting due to limited exploration and the distinct nature of wireheading policies.
3 days agoHugging FaceHugging Face model card for Edge0/Audio8-ASR-Infinite. Task identifier: automatic-speech-recognition. Follow the source link for details.
11 days agoLobstersValve has introduced Pyrowave, a new video codec in beta, designed for low latency streaming.
5 days agoHugging FaceHugging Face model card for XingChen-AGI/TeleOCR. Task: text generation from image and text inputs. Follow the source link for details.
49 days ago