Perplexity AI Is Lying about Their User Agent

Perplexity AI Is Lying about Their User Agent

Robb Knight discusses blocking AI bots, specifically criticizing Perplexity AI for not respecting robots.txt or using the correct user agent to access and summarize his blog posts. Despite implementation of server-side blocking and attempts to restrict access, Perplexity AI appears to bypass these measures, causing Robb to question the ethics and methods of such AI companies. Read more here.

Announcing ARC Prize

The ARC Prize is a $1,000,000+ competition aimed at advancing open artificial general intelligence (AGI) by solving the ARC-AGI eval. The competition is hosted by Mike Knoop and François Chollet. Current AI models excel at memorization but fail at general intelligence. The ARC-AGI measures true general intelligence, and a solution could significantly advance AGI research. The competition invites global participation to discover innovative ideas and counter the trend of closed-source AI research. More details and sign-up options are available on their website. Read more here

How Meta trains large language models at scale

Meta discusses scaling efforts and technical challenges in training large language models (LLMs). The transition to generative AI requires vast computing resources. Key considerations are hardware reliability, fast recovery, efficient state preservation, and optimal GPU connectivity. Innovations span software, scheduling, hardware, data center deployment, reliability, networking, and storage. Meta addresses these complexities, aiming to continue leading AI advancements. Read more here

Getting 50% (SoTA) on ARC-AGI with GPT-4o

Redwood Research achieved a 50% accuracy rate on the ARC-AGI test set using GPT-4 by generating approximately 8,000 Python programs per problem and choosing the correct ones from these samples. They improved performance through meticulous few-shot prompts, program revisions, and specialized grid representations. Their method surpassed the previous state of the art of 34% accuracy. The research highlighted potential for further scalability and challenged assertions that LLMs cannot effectively learn in context. Read more here

Update on the Recall preview feature for Copilot+ PCs

Microsoft announced an update on the Recall feature for Copilot+ PCs, initially set to be broadly available on June 18, 2024, but now starting in the Windows Insider Program for further testing. This decision reflects Microsoft's commitment to security and quality, aligning with their Secure Future Initiative (SFI). Recall, a feature that creates a visual timeline of screen snapshots for easy retrieval of past activities, will incorporate robust privacy controls and encryption, emphasizing user consent and security. The feature, designed for Copilot+ PCs, underscores Microsoft's approach to integrating AI with stringent security measures to enhance user productivity and safety. Read more here