Explore Appleās research on LLM failures, Codexās repo-editing powers, Scale AIās big move, and a 2-week AI crash course
At WWDC 2025, Apple introduced upgraded on-device (~3B param) and server-based foundation language models powering Apple Intelligence, plus a new Foundation Models framework granting third-party developers swift, privacy-preserving access to onādevice generative AI capabilitiesāeven offlineāall using simple Swift integration.
Generative AI enables āvibe hackingā: novices can now prompt LLMs to create malware, while seasoned hackers can scale sophisticated, polymorphic attacksālike fast zeroāday exploits. As AIāarmed adversaries rise, those defending against it must leverage AI too.
Meta has agreed to buy a 49% stake in Scale AI for $14.8 billion, bringing Scale CEO Alexandr Wang into Meta to lead a "superintelligence" lab aimed at boosting its AI capabilities after Llama 4 underperformed. Scale, valued at $13.8 billion, generated $870 million in 2024 and projects over $2 billion in 2025.
Anthropic quietly launched āClaude Explains,ā an AIāgenerated, humanāedited blog aimed at showcasing its Claude model, but shut it down after a month amid criticism over partial automation and concerns the pilot overstated Claudeās accuracy and marketing intent.
OpenAI's cloud-based software engineering agent first introduced in May is now available to ChatGPT Plus users. It can read, edit, test, and commit code in isolated repos to write features, fix bugs, and propose PRs.
This helpful repo contains a large collection of reference implementations for MCP. It also contains references to community built servers and additional MCP resources. Every reference implementation uses either the Typescript MCP SDK or Python MCP SDK.
OpenAI released o3-pro for ChatGPT Pro users and for their API. It is a version of o3 meant to think for longer periods of time and be more reliable in terms of responses provided.
Built by Henry Shi, AI Crash Course is a GitHub repo with resources and a learning plan for busy developers wanting to get up to speed with every AI-related in 2 weeks.
Apple researchers analyze when LLMsā chain-of-thought reasoning succeeds or fails based on problem complexity, revealing that performance varies significantly with task structure and highlighting both strengths and hidden limitations.
Get the latest AI tools, frameworks, and insights delivered to your inbox. Join hundreds of developers staying ahead of the curve.
100% free weekly newsletter. No spam. Unsubscribe anytime.