Good morning, AI enthusiast.

A significant breakthrough in algorithmic design comes from Google DeepMind with their new AI system, AlphaEvolve. This innovative agent has already shattered a 56-year-old optimization record, showcasing its power.

By leveraging Google's Gemini models within an evolutionary framework, AlphaEvolve is not just theorizing but actively improving real-world systems, even speeding up Gemini model training. With its ability to discover novel solutions, how profoundly will such AI-driven algorithmic breakthroughs reshape scientific research and technological development?

In today’s AI recap:

• DeepMind’s AlphaEvolve shatters 56-year algorithm record

• Meta's suite of new AI tools for scientific research

• Google Gemini expands to Android devices and GitHub

• US revamps AI chip export regulations

AlphaEvolve: DeepMind's New AI Forges Superior Algorithms

The Recap: Google DeepMind introduced AlphaEvolve, an AI agent that designs and refines algorithms for complex mathematical and scientific challenges. This system already helps optimize Google's own infrastructure.

Unpacked:

• AlphaEvolve combines Google's Gemini models with an evolutionary process; it generates, critiques, and evaluates potential solutions to enhance accuracy and reduce hallucinations.

• The system rediscovered top solutions for 75% of benchmark math problems and even improved known algorithms, including one for matrix multiplication that surpasses a 56-year-old record and DeepMind's specialized AlphaTensor tool.

• Beyond theory, AlphaEvolve generated optimizations that recover 0.7% of Google's global compute resources and sped up Gemini model training by 1%.

• AlphaEvolve builds upon concepts from earlier systems like FunSearch and targets problems with "machine-gradable" solutions, with DeepMind planning an early access program for academics.

Bottom line: AlphaEvolve showcases AI's growing capability to not just apply knowledge, but to discover and create novel, more efficient solutions to fundamental problems. This could significantly accelerate scientific discovery and technological advancement by augmenting human expertise. Professionals can look forward to tools that tackle increasingly complex optimization challenges.

Meta Pushes Scientific Frontiers with New AI Tools

The Recap: Meta FAIR recently unveiled a suite of significant AI advancements, releasing new datasets, models, and research aimed at accelerating molecular discovery, enhancing generative modeling, and deepening our understanding of language acquisition in the human brain.

Unpacked:

• Meta released the Open Molecules 2025 (OMol25) dataset, its largest and most diverse collection of high-accuracy quantum chemistry calculations, set to speed up atomic-scale design in healthcare and energy. You can explore the OMol25 dataset and its associated models on Hugging Face.

• The new Universal Model for Atoms (UMA) offers a foundational machine learning potential trained on over 30 billion atoms, providing robust predictions of molecular behavior across varied materials. Details are available in the UMA paper.

• Researchers introduced Adjoint Sampling, a novel technique that trains generative models using only a reward signal, eliminating the need for traditional training data, with promising results in molecular generation. The methodology is detailed in the Adjoint Sampling paper.

• A collaborative study with the Rothschild Foundation Hospital mapped language representation development in the human brain, revealing striking parallels with LLMs and offering new insights into efficient learning. You can read their research paper for an in-depth look.

Bottom line: These releases from Meta underscore a commitment to open science and aim to empower the global research community. By providing powerful new tools and foundational models, Meta is helping to unlock complex scientific challenges and foster innovation across multiple critical domains.

Gemini Goes Wide: Android Expansion & GitHub Integration

The Recap: Google is significantly broadening Gemini's reach, bringing its AI capabilities to a wider range of Android devices and integrating Gemini Advanced directly with GitHub. These updates, unveiled as part of Google's latest announcements, aim to embed AI more deeply into daily and professional tasks.

Unpacked:

• Gemini is now expanding to enhance user experiences across more Android devices, including Wear OS, Android Auto, Google TV, and upcoming Android XR platforms.

• Android users everywhere can now access Gemini Live's advanced camera and screen sharing features at no cost, making powerful AI tools more accessible.

• Gemini Advanced now integrates with GitHub, enabling developers to interact directly with public or private repositories for coding tasks like generation, explanation, and debugging.

Bottom line: Google is embedding Gemini deeper into its ecosystem, aiming to make AI an indispensable part of daily digital life and professional workflows. This expansion signals a strong push to bring powerful, practical AI tools directly into the hands of users and developers, enhancing productivity across the board.

US Revamps AI Chip Export Rules

The Recap: The U.S. Department of Commerce has rescinded a Biden-era rule that would have limited AI chip exports to some EU nations, while also strengthening other chip-related export controls. This decision, welcomed by the European Commission, arrived just before the original rule's May 15 enforcement date.

Unpacked:

• The "AI Diffusion Rule," issued January 15, 2025, would have restricted access to critical AI chips for certain EU countries like Poland, potentially hindering their AI development plans.

• European Commission officials actively pushed back against the planned caps, arguing U.S. interests align with the EU purchasing American chips without limitations.

• The rescission averts potential diplomatic friction and removes uncertainty for EU nations previously designated as "second-tier" under the rule.

• Alongside this reversal, the Commerce Department announced new guidance to bolster controls on overseas AI chips, specifically targeting risks associated with PRC advanced computing ICs and the use of U.S. AI chips in Chinese AI models.

Bottom line: This policy shift eases transatlantic tech tensions and provides a boost for EU countries reliant on U.S. AI chip technology. However, the strengthened controls targeting China underscore a continued focus on strategic competition in the global AI landscape.

The Shortlist

OpenAI expanded its data residency program to Asia, allowing organizations in Japan, India, Singapore, and South Korea to store ChatGPT Enterprise, Edu, and API data locally.

Audible announced plans to use its own AI production technology for audiobook narration and will launch an AI-powered translation tool beta, aiming to expand its catalog.

Reports suggest Anthropic is safety testing a new model called 'claude-neptune,' fueling speculation about an impending Claude 4 release.

TikTok launched 'AI Alive,' an in-app image-to-video tool that transforms users' static photos into short videos using AI and text prompts, available via the Story Camera.

Signing off,

David, Lucas, Mitchell — The Recap editorial team