OpenAI To Launch Its First AI Chip In 2026 With Broadcom

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")

OpenAI To Launch Its First AI Chip In 2026 With Broadcom (1 min. read)

OpenAI will produce its first in-house AI chip with Broadcom by 2026, targeting mass production for internal workloads. The move aims to cut reliance on Nvidia, improve performance control, and secure long-term hardware supply, according to sources cited by the Financial Times. Neither company has publicly confirmed details, but the initiative signals a major step in OpenAI’s vertical integration strategy.

Google Introduces EmbeddingGemma: The Best-in-Class Open Model for On-Device Embeddings (5 min. read)

Google unveils EmbeddingGemma, a 308M parameter open embedding model optimized for on-device AI tasks like Retrieval Augmented Generation and semantic search. Trained on over 100 languages, it ranks highest among sub-500M parameter models on the Massive Text Embedding Benchmark. Its efficiency allows operation on devices with less than 200MB RAM when quantized, enabling private, high-quality embeddings without internet connectivity. EmbeddingGemma integrates seamlessly with tools like sentence-transformers, llama.cpp, and LangChain, facilitating versatile deployment.

Why Language Models Hallucinate (8 min. read)

OpenAI's latest research reveals that language models often "hallucinate"—confidently generating false information—due to training and evaluation methods that reward guessing over expressing uncertainty. This issue arises from models being incentivized to provide answers, even when unsure, to maximize accuracy scores. The study suggests revising evaluation metrics to penalize incorrect answers more than admissions of uncertainty, thereby encouraging models to acknowledge their limitations and reduce hallucinations.

DeepSeek Prepares AI Agent to Rival OpenAI by End-2025 (<1 min. read)

DeepSeek, the Hangzhou-based startup that shook the AI world with its low-cost, open-source R1 model, is gearing up to launch a multi-step autonomous agent by late 2025 designed to learn from its own actions and operate with minimal instruction. This strategic leap signals DeepSeek’s push into the next frontier of AI agents, poised to compete head-to-head with OpenAI in advanced agentic intelligence.

print("Applications & Insights")

Lessons on Building an AI Data Analyst (12 min. read)
Pedro Nascimento shares insights from developing an AI data analyst at Findly, emphasizing that text-to-SQL alone is insufficient for answering complex user queries. He advocates for multi-step workflows incorporating external tools and context, highlighting the importance of a semantic layer to encode business meaning and reduce SQL complexity. Nascimento also discusses the benefits of a multi-agent, research-oriented system, advanced retrieval techniques, and the necessity of continuous model evaluation to meet user expectations for accuracy and latency.

Agentic Design Patterns
"Agentic Design Patterns" is a comprehensive 424-page guide by Antonio Gulli that delves into building intelligent systems through agent-based design. It covers topics like prompt chaining, parallelization, reflection, tool use, planning, multi-agent systems, memory management, learning, and adaptation. The book also explores advanced concepts such as inter-agent communication, resource-aware optimization, reasoning techniques, safety patterns, and evaluation methods, providing practical code examples throughout.

How We Accelerated Secret Protection Engineering with Copilot (7 min. read)
GitHub's Secret Protection team leveraged Copilot to dramatically enhance their secret scanning capabilities. By integrating Copilot into their workflow, they expanded from validating 32 to nearly 90 token types in just a few weeks. This acceleration was achieved by automating framework-driven tasks, allowing engineers to focus on nuanced research and review. The team emphasizes that while Copilot serves as a powerful multiplier, human oversight remains crucial to ensure code quality and accuracy.

Building an Agentic System
This comprehensive guide delves into constructing AI coding assistants for production environments, emphasizing core architectures like reactive UIs with Ink and Yoga, streaming responses, and state management. It explores extensible tool systems for file operations and code execution, robust permission models balancing security and productivity, and strategies for parallel execution to prevent race conditions. The series also covers command systems featuring slash commands and contextual help, drawing lessons from real-world implementations such as Amp and Claude Code.

print("Tools & Resources")

TRENDING MODELS

Text Generation
moonshotai/Kimi-K2-Instruct-0905
⇧ 8.76k Downloads
Kimi K2 is a state-of-the-art language model developed by Moonshot AI, designed to excel in various text generation tasks. It has demonstrated superior performance on multiple benchmarks, surpassing models like GPT-4.1 and Claude Opus.

Image-to-Video
tencent/HunyuanWorld-Voyager
⇧ 4.66k Downloads
HunyuanWorld-Voyager is an advanced model by Tencent that converts images into dynamic video sequences. It leverages cutting-edge techniques to generate high-quality videos from static images, enhancing visual storytelling capabilities.

Text-to-Speech
microsoft/VibeVoice-1.5B
⇧ 245k Downloads
VibeVoice-1.5B is a text-to-speech model developed by Microsoft, offering natural and expressive voice synthesis. With 1.5 billion parameters, it delivers high-fidelity audio outputs suitable for various applications.

Text Generation
swiss-ai/Apertus-8B-Instruct-2509
⇧ 66.4k Downloads
Apertus-8B-Instruct-2509 is an 8-billion parameter language model from Swiss AI, optimized for instruction-based tasks. It provides robust performance in generating coherent and contextually relevant text across diverse prompts.

Translation
tencent/Hunyuan-MT-7B
⇧ 6.56k Downloads
Hunyuan-MT-7B is a 7-billion parameter translation model by Tencent, designed to handle multiple language pairs with high accuracy. It utilizes advanced neural network architectures to deliver precise and fluent translations.

TRENDING AI TOOLS

• 🤖 AI mode: Google’s AI search now in 5 more languages

• 🎙️ Lovable voice mode: Build apps by chatting with AI

• 🚀 Warp code: From prompt to prod inside Warp

• 🌐 Jan-v1-edge: Jan’s powerful new AI search model

print("Everything else")

• Curated list of papers that explore the integration of reinforcement learning techniques with LLMs to enhance their agentic capabilities.

• Framework by Meta on how to navigate product strategy as a data leaders.

• Le Chat integrates with 20+ enterprise platforms and introduces memories

• OpenAI is building an AI jobs platform that could challenge Microsoft’s LinkedIn

• Perplexity introduces Comet, an AI-powered browser designed to be your 24/7 study companion. Students get early access.

• Google’s NotebookLM now lets you customize the tone of its AI podcasts.

• Projects in ChatGPT are now available to Free users.

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