TLDR AI 2025-07-02

AI is pretty good at creating bugs. Can it also resolve them?

Seer is Sentry's new AI agent — and the only AI agent that uses all of the context that Sentry provides: errors, stack traces, commit history, logs, traces, spans, breadcrumbs.

Seer sifts through the data, finds the root cause of an issue (with 94.5% accuracy!), and can open a PR with a fix. Ship it, close the tab, and move on with your life. It's private by default, supports ±any stack, and it actually works.

Watch Seer work its magic and learn how the Sentry engineering team uses Seer to automatically debug issues. Join the workshop →

Amazon introduced a new AI foundation model to optimize warehouse robotics and celebrated the milestone of deploying one million robots. The model improves efficiency by 10% and supports faster deliveries.

Grammarly is acquiring Superhuman to enhance its AI productivity platform by integrating AI-native email capabilities. Superhuman users send and respond to 72% more emails per hour with AI, showcasing increased efficiency. This acquisition supports Grammarly's plan to create applications as intelligent agents, positioning email as a key platform for multi-agent assistance.

xAI has raised $5 billion in debt through secured notes and term loans and $5 billion in equity secured through a strategic equity investment. The debt offering was oversubscribed and included prominent global debt investors. The funding will go towards the startup's efforts to build out its AI infrastructure and develop its Grok chatbot. It will continue buying Nvidia and AMD chips for a 1-million-GPU facility outside of Memphis.

The latest update to Sentence Transformers introduces support for training sparse embedding models, ideal for hybrid search and reranking. The blog walks through model components and training steps and highlights available SPLADE-based models.

Nick Turley and Mark Chen discuss ChatGPT's viral launch, evolving safety strategies, breakthroughs in multimodal models, and the rise of agentic coding with Codex. They also share insights on internal culture and the vision for AI-powered assistants.

This post discusses how to build an AI factory that improves itself. The aim of the project was to build a factory that can produce itself. It aimed to create a factory of agents that can produce code, verify it, and improve themselves over time. The post covers the idea behind the AI factory, how to build the factory, scaling the factory, and ideas for future improvement.

Arguing against case-by-case analysis of concepts like grokking or double descent, these authors propose shifting focus toward broader theories that offer more practical value for real-world deep learning applications.

Researchers compared raw AI models with their chat-trained versions to understand what safety training actually changes inside the model—like finding the "off-limits" neurons that make a model refuse harmful requests. They discovered that Anthropic's standard comparison technique was hallucinating thousands of fake differences, but their BatchTopK fix revealed real interpretable safety circuits like refusal detection and stereotype filtering, capturing 88% of behavioral changes versus just 60% for the broken method.

Finding antibodies that bind to disease targets usually means screening massive libraries for months—most pharmaceutical discovery is basically a numbers game. Chai Discovery's AI cuts straight to viable drug candidates by designing them computationally, hitting 16% success rates and finding working antibodies for 50% of targets with just 20 attempts each.

X launched a pilot allowing LLMs like Grok or ChatGPT to submit Community Notes. These AI-generated notes will be vetted by humans and subject to the same trust mechanisms as existing user contributions.

Key AI safety paths include international cooperation, domestic initiatives, and societal resilience. Strategies involve forming alliances like CERN for AI, using deterrence to prevent unsafe development, and fostering responsible actors within countries. A diversified approach across these strategies offers insurance against uncertainties and leverages various expertise.

You'll probably like our flagship newsletter. It's all about tech, science, and programming.

Same quick format. Still free.

Subscribe now.

Huawei will open source two of its PanguAI models as well as some of its reasoning technologies.

Cloudflare launched Pay per Crawl, a marketplace enabling sites to charge AI bots for content scraping.

Boris Cherny and Cat Wu, product managers who worked on Claude Code, have joined Cursor.