Kaggle Introduces Community Benchmarks

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...

print("News & Trends")

Kaggle Introduces Community Benchmarks (3 min. read)

Kaggle has introduced Community Benchmarks, a new product that lets you build, run, and share your own custom benchmarks for evaluating AI models at no cost. Powered by the kaggle-benchmarks SDK, you can now create your own AI evaluations (“tasks”) and put them together into a collection (“benchmark”). This launch is just the beginning. They have a rich feature roadmap planned out for Community Benchmarks, including support for more AI models, Task & Benchmark versioning, multiple task runs (pass@k), and more.

PaperBanana: Automating Academic Illustration for AI Scientists (6 min. read)

Creating publication-ready illustrations can be a tedious task for AI researchers. Enter PaperBanana, a framework that automates this process by coordinating five specialized agents (Retriever, Planner, Stylist, Visualizer, and Critic) to transform raw scientific content into polished diagrams and plots. Evaluated on the PaperBananaBench dataset, which includes 292 methodology diagrams from NeurIPS 2025 papers, PaperBanana consistently outperforms existing methods in faithfulness, conciseness, readability, and aesthetics. This innovation streamlines the illustration process, allowing researchers to focus more on their core work.

Introducing the Codex App (9 min. read)

OpenAI has unveiled the Codex app for macOS, a robust interface designed to manage multiple coding agents simultaneously, enabling parallel task execution and collaboration on extended projects. This app allows developers to delegate substantial tasks to agents, streamlining the software development lifecycle from design to maintenance. For a limited period, Codex is available to ChatGPT Free and Go users, with doubled rate limits for Plus, Pro, Business, Enterprise, and Edu plans, enhancing accessibility and performance across various platforms.

print("Applications & Insights")

The "LLM-as-Analyst" Trap: A Technical Deep-Dive into Agentic Data Systems (8 min. read)
This article delves into the "Simple Agentic" pattern, where large language models (LLMs) are equipped with tools to fetch data but handle the final analysis themselves. While this approach is easy to implement and demo, it poses significant business risks, including accuracy issues, increased costs, and challenges in verifiability. The author illustrates these pitfalls through a case study of a locally run financial assistant, highlighting the hidden dangers of relying solely on LLMs for data analysis.

Inside OpenAI’s In-House Data Agent (5 min. read)
OpenAI has developed a bespoke AI data agent, powered by GPT-5.2, to streamline data analysis across its vast datasets. Integrated into tools like Slack and IDEs, the agent handles complex queries end-to-end, from understanding questions to executing analyses. It employs a multi-layered context system—incorporating metadata, human annotations, code insights, institutional knowledge, memory, and real-time data—to ensure accurate and efficient results. This self-improving agent reduces manual effort, enabling teams to derive insights swiftly and accurately.

LLM Observability Best Practices Guide
Datadog's guide delves into the critical aspects of monitoring Large Language Models (LLMs), highlighting challenges like hallucinations, performance issues, and security vulnerabilities. It emphasizes the importance of real-time monitoring to enhance model performance, ensure explainability, and bolster security. The guide also outlines key features to seek in an observability solution, such as comprehensive application stack visibility and robust anomaly detection, to effectively manage and optimize LLM applications.

Ads Candidate Generation Using Behavioral Sequence Modeling (8 min. read)
Pinterest's engineering team has developed a transformer-based sequence model to enhance ad relevance by analyzing users' offsite behaviors. This two-tower architecture predicts future advertiser interactions, leading to a significant boost in conversion rates and a reduction in cost per action. The model's success underscores the power of behavioral sequence modeling in delivering personalized ad experiences.

Context Graphs, One Month In (5 min. read)
A month after introducing context graphs (structured records of organizational decision-making), Foundation Capital reflects on the widespread adoption and discourse surrounding the concept. Industry leaders like Dharmesh Shah and Aaron Levie have embraced context graphs as essential for capturing the 'why' behind decisions, not just the 'what.' The article underscores the transformative potential of context graphs in enterprise AI, emphasizing their role in creating a living, queryable map of decision processes that traditional systems overlook.

print("Tools & Resources")

TRENDING MODELS

Image-Text-to-Text
moonshotai/Kimi-K2.5
⇧ 203k Downloads
Kimi-K2.5 is a 171B parameter model designed for advanced image-to-text tasks, offering high accuracy in generating textual descriptions from images.

Image-to-Text
zai-org/GLM-OCR
⇧ 96.3k Downloads
GLM-OCR is optimized for optical character recognition, providing efficient and accurate text extraction from images.

Text Generation
stepfun-ai/Step-3.5-Flash
⇧ 8.69k Downloads
Step-3.5-Flash is a 199B parameter model tailored for high-speed text generation, suitable for various natural language processing applications.

Text Generation
Qwen/Qwen3-Coder-Next
⇧ 18.7k Downloads
Qwen3-Coder-Next is an 80B parameter model focused on code generation, assisting developers with efficient and accurate code completion.

Automatic Speech Recognition
Qwen/Qwen3-ASR-1.7B
⇧ 104k Downloads
Qwen3-ASR-1.7B is a 2B parameter model designed for automatic speech recognition, delivering high-quality transcriptions across various audio inputs.

TRENDING AI TOOLS

• 🖥️ Qwen3-Coder-Next: Advanced coding assistant for efficient software development and debugging.

• 🐘 pg_tracing: PostgreSQL tracing for performance insights and debugging.

• 🔍 Unwrap: AI-powered customer intelligence

print("Everything else")

• Gemini CLI introduces extensions, enabling users to integrate tools and customize their AI-powered command line.

• ShareChat successfully scaled their ML feature store using ScyllaDB to handle increased data demands efficiently.

• Amazon now offers Alexa Plus free to US Prime members, enhancing smart home capabilities.

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