Hey there—David Kemmerer here, founder of AITools.inc and now The Recap.

First off, thanks for signing up. When you originally subscribed to AITools.inc, the goal was simple: cut through the noise and help you stay ahead of the AI wave. That mission hasn't changed—it's just gotten sharper.

The Recap is your new daily cheat code. In each edition, we scan the internet, synthesize what actually matters in AI, and package it up into a quick, no-BS read you can skim in under five minutes. No hype. No fluff. Just signal.

This is the first one. I hope you dig it—and if you don’t, reply and tell me why. Seriously.

Let’s get into it.

In today’s AI recap:

• Prime Intellect’s decentralized INTELLECT-2 model

• Google launches AI Futures Fund for startups

• OpenAI unveils HealthBench for healthcare AI

• US Copyright Office head fired after AI report

Decentralized AI Training Arrives

The Recap: Prime Intellect has released INTELLECT-2, the first 32B parameter language model trained using globally distributed reinforcement learning on a permissionless network. This marks a potential shift away from centralized supercomputer training.

Unpacked:

• INTELLECT-2 showcases training a reasoning LLM using fully asynchronous RL across a dynamic, permissionless swarm of compute contributors, differing from traditional centralized methods.

• The training utilized 285k verifiable math and coding tasks and modified the GRPO recipe with techniques like two-sided clipping and advanced data filtering for stability and learning efficiency.

• Prime Intellect has open-sourced INTELLECT-2 along with the associated code and data, aiming to foster more open research in decentralized AI training.

Bottom line: This release demonstrates that complex RL training can be effectively decentralized, potentially lowering barriers to entry for large-scale model training. As RL consumes more compute for sampling, leveraging distributed, even consumer-grade, hardware could become increasingly viable, challenging the dominance of massive, centralized clusters.

Google Launches AI Futures Fund

The Recap: Google unveiled the AI Futures Fund, a new initiative designed to invest in and partner with ambitious AI startups. The program aims to accelerate AI development by providing funding, resources, and expert collaboration.

Unpacked:

• Participants get early access to advanced Google DeepMind models, including Gemini for multimodal tasks, Imagen for image generation, and Veo for video creation.

• The fund provides hands-on support from Google researchers and engineers, plus dedicated technical help and Google Cloud credits to build and scale products.

• This initiative places Google alongside competitors like Amazon and Microsoft, who are also significantly backing AI startups through investments and partnerships.

Bottom line: For AI startups, this fund offers a valuable pathway to access Google's cutting-edge models, expertise, and potential funding. It highlights Google's commitment to fostering the AI ecosystem and staying at the forefront by nurturing emerging innovation.

OpenAI Unveils HealthBench

The Recap: OpenAI has released HealthBench, a new benchmark designed to rigorously evaluate AI models in realistic healthcare scenarios. This aims to standardize how we measure AI performance and safety in health settings.

Unpacked:

• Developed with input from over 250 physicians practiced in dozens of countries, ensuring clinical relevance.

• Features 5,000 realistic health conversations evaluated across nearly 50,000 rubric items to measure performance and safety.

• The benchmark is now open-source and available in OpenAI's GitHub repository for community review and use.

• It enables meaningful open-ended evaluation beyond narrow tests, revealing significant performance improvements in newer models over predecessors like GPT-3.5.

Bottom line: HealthBench provides a crucial tool for developers and researchers building AI for healthcare. Establishing robust, clinically informed evaluation standards is essential for fostering trust and ensuring AI systems are safe and helpful when deployed in sensitive medical contexts. We're excited to see how this shapes future AI development in the health sector.

Copyright Office Head Fired After AI Report

The Recap: Shira Perlmutter, head of the U.S. Copyright Office, was abruptly fired by the Trump administration just one day after the office released a significant draft report questioning whether AI training on copyrighted works always qualifies as 'fair use'.

Unpacked:

• The report suggests that making commercial use of copyrighted works to create AI content that competes in existing markets likely goes beyond established fair use boundaries.

• Perlmutter's dismissal occurred Saturday, May 10th, immediately following the report's May 9th release and just days after Librarian of Congress Carla Hayden was also fired.

• Rep. Joe Morelle (D-NY) slammed the firing as a brazen, unprecedented power grab potentially linked to tech interests mining copyrighted data for AI.

• The administration quickly named new acting leadership, including Deputy Attorney General Todd Blanche as acting Librarian of Congress and Paul Perkins as acting Register of Copyrights.

Bottom line: This sudden leadership change throws significant uncertainty into U.S. AI policy, particularly concerning copyright and fair use. AI companies relying on scraped data face a potentially shifting legal landscape.

The Shortlist

Apple plans to introduce an AI-powered battery management feature in iOS 19, analyzing user habits to optimize energy consumption and potentially display charging time estimates on the Lock Screen.

Perplexity AI is reportedly raising around $500 million in a new funding round led by Accel, pushing its valuation to $14 billion as its ARR approaches $120 million.

Economists found in a recent Danish study that generative AI chatbots have so far had no significant impact on earnings or recorded work hours across 11 analyzed occupations, suggesting real-world productivity gains are lower than some trials indicate.

Cybercriminals are luring users with convincing fake AI video and image editing tools, often promoted on Facebook, to distribute Noodlophile malware designed to steal browser credentials and cryptocurrency wallet details.

Signing off,

David, Lucas, Mitchell — The Recap editorial team