AI Co-Mathematician: Accelerating Mathematicians with Agentic AI

Hey there 👋

We hope you're excited to discover what's new and trending in AI, ML, and data science this week.

Here is your 5-minute pulse...

But first, a quick message from our partner 👇

See What Shipping Agentic Analytics Actually Looks Like (Free Workshop)

Here is the reality: most data scientists are learning agentic AI. Very few are shipping it.

In this free 30-minute workshop, we will deep dive into the architecture, the trade-offs, and the parts the courses tend to skip.

print("News & Trends")

AI Co-Mathematician: Accelerating Mathematicians with Agentic AI (22 min. read)

Google DeepMind introduces the AI co-mathematician, an interactive workbench designed to assist mathematicians in open-ended research. This system supports the exploratory nature of mathematical workflows, including ideation, literature search, computational exploration, theorem proving, and theory building. By providing an asynchronous, stateful workspace that manages uncertainty and refines user intent, it mirrors human collaborative processes. Early tests show the AI co-mathematician aiding researchers in solving open problems, identifying new research directions, and uncovering overlooked literature. Notably, it achieves state-of-the-art results on challenging problem-solving benchmarks, scoring 48% on FrontierMath Tier 4, the highest among evaluated AI systems.

Why MistralAI Grows Faster Than OpenAI/Anthropic (7 min. read)

Mistral AI, a Paris-based startup, has rapidly scaled its annual recurring revenue from $20 million to $400 million within a year, aiming for over $1 billion in 2026. Unlike U.S. giants OpenAI and Anthropic, Mistral focuses on sovereignty, offering open-weight models that grant enterprises greater control and flexibility. Their efficient, mixture-of-experts architectures deliver high performance at reduced costs, appealing to European institutions seeking powerful AI solutions without full dependency on U.S. platforms.

Teaching Claude Why (16 min. read)

Anthropic's latest research delves into enhancing AI alignment by teaching models like Claude to understand the reasoning behind ethical behavior. By training Claude on datasets where it advises users through ethical dilemmas, the team achieved a significant reduction in misaligned actions. This approach emphasizes that instilling the principles underlying aligned behavior, rather than just demonstrating correct actions, leads to more robust and generalizable AI alignment.

Interaction Models: A Scalable Approach to Human-AI Collaboration (12 min. read)

Thinking Machines introduces interaction models designed for real-time, multimodal human-AI collaboration. Unlike traditional turn-based systems, these models process audio, video, and text simultaneously, enabling seamless, natural interactions. By adopting a multi-stream, micro-turn architecture, they achieve state-of-the-art performance in both intelligence and responsiveness. This approach addresses the collaboration bottleneck in AI, allowing users to engage with models as they would with human colleagues—through continuous, dynamic exchanges.

print("Applications & Insights")

How BigQuery Actually Executes a Query (and Why Most Optimization Advice Misses Half the Picture) (11 min. read)
Many BigQuery users overlook the Execution Details panel, missing insights into how queries are physically processed. This article demystifies BigQuery's execution model, explaining that queries are broken into stages and executed across slots—units of compute akin to CPU cores with memory. Understanding this model is crucial, especially with the shift to capacity pricing, where computational cost is measured in slot-seconds. By grasping these concepts, data professionals can optimize queries more effectively, moving beyond generic advice to strategies tailored to BigQuery's architecture.

Autodata: An Automatic Data Scientist to Create High-Quality Data (5 min. read)
Meta AI introduces Autodata, an AI agent designed to autonomously generate and refine high-quality training and evaluation datasets. By emulating a human data scientist, Autodata iteratively creates data, analyzes its quality, and enhances it through continuous feedback loops. The Agentic Self-Instruct method, a specific implementation, has demonstrated significant improvements in scientific reasoning tasks over traditional synthetic data generation techniques. Moreover, meta-optimizing the agent itself further boosts performance, suggesting a transformative approach to AI data creation.

Democratizing Machine Learning at Netflix: Building the Model Lifecycle Graph (7 min. read)
Netflix has developed the Model Lifecycle Graph (MLG), a centralized system that maps the entire lifecycle of machine learning models, from development to deployment. This initiative aims to democratize ML by providing engineers with a clear, interconnected view of models, datasets, and features, enhancing collaboration and efficiency. The MLG addresses challenges like model reproducibility and lineage tracking, ensuring that ML workflows are transparent and scalable across the organization.

How We Built an AI Second Brain for 60K Knowledge Workers (8 min. read)
Meta tackled workflow fragmentation by developing an AI "Second Brain" that integrates seamlessly with existing tools, providing persistent, structured access to users' projects, meeting notes, and tasks. This system, now adopted by over 60,000 employees, leverages Tiago Forte's PARA method to organize information into Projects, Areas, Resources, and Archives, enabling the AI to maintain context across interactions. By implementing progressive disclosure, the agent efficiently manages context windows, enhancing productivity without overwhelming users.

print("Tools & Resources")

TRENDING MODELS

Text-to-Video
SulphurAI/Sulphur-2-base
⇧ 158k Downloads
Sulphur-2-base is a 9B parameter model designed for text-to-video generation, enabling users to create videos from textual descriptions. Updated three days ago, it has garnered significant attention for its capabilities in generating high-quality video content from text prompts.

Text Generation
Zyphra/ZAYA1-8B
⇧ 66.1k Downloads
ZAYA1-8B is an 8B parameter model focused on text generation tasks, offering advanced language understanding and generation capabilities. Updated about nine hours ago, it has been widely adopted for various natural language processing applications.

Text Generation
deepseek-ai/DeepSeek-V4-Pro
⇧ 2.02M Downloads
DeepSeek-V4-Pro is an 862B parameter model designed for advanced text generation, providing high-quality and contextually relevant outputs. Updated six days ago, it has become a popular choice for complex language generation tasks.

Image-Text-to-Text
openbmb/MiniCPM-V-4.6
⇧ 324 Downloads
MiniCPM-V-4.6 is a 1B parameter model that bridges the gap between images and text, enabling users to generate textual descriptions from images. Updated about two hours ago, it offers efficient performance for image-to-text applications.

Image-Text-to-Image
HiDream-ai/HiDream-O1-Image
⇧ 3.42k Downloads
HiDream-O1-Image is a 9B parameter model that generates images from textual descriptions, facilitating creative visual content creation. Updated three days ago, it has been recognized for its ability to produce high-quality images from text prompts.

TRENDING AI TOOLS

• 🧩 Codex Chrome Extension: Enhance coding productivity directly in your browser with AI-powered assistance.

• 🧠 Ernie: A comprehensive language model platform by Baidu for AI-powered applications and solutions.

• 🚀 Pareto Code Router: Free routing layer that auto-picks the cheapest coding AI above a user-set quality bar.

• 📊 Helix-DB: High-performance database optimized for analytical queries and real-time insights.

• 🔍 Data Landscape: Visualize and explore your data ecosystem effortlessly.

print("Everything else")

• Oracle shares 16 strategies to enhance the capabilities of small language models.

• Greece is considering constitutional changes to address the impact of artificial intelligence on society.

• OpenRouter introduces the Pareto Code Router, allowing users to select coding models based on a specified minimum coding score.

• In a recent blog post, mathematician Timothy Gowers describes how ChatGPT 5.5 Pro produced PhD-level research in an hour with minimal human input.

• Google Cloud's report highlights AI vulnerability exploitation as a growing initial access threat in cybersecurity.

That’s it for today!

Before you go we’d love to know what you thought of today's newsletter to help us improve the pulse experience for you.

See you soon,

Andres