Deep Learning Weekly: Issue 417

This week in deep learning, we bring you How to Build Reliable AI Agent Architecture for Production, How Much Power will Frontier AI Training Demand in 2030?, and a paper on TextQuests: How Good are LLMs at Text-Based Video Games?.

You may also enjoy GLM-4.5: Reasoning, Coding, and Agentic Abilities, From GPT-2 to gpt-oss: Analyzing the Architectural Advances, a paper on OpenCUA: Open Foundations for Computer-Use Agents, and more!

As always, happy reading and hacking. If you have something you think should be in next week's issue, find us on Twitter: @dl_weekly.

Until next week!

Industry

Introducing GPT-5 | OpenAI

OpenAI released GPT-5, a significant leap in intelligence over all previous models, featuring state-of-the-art performance across coding, math, writing, health, visual perception, and more.

GLM-4.5: Reasoning, Coding, and Agentic Abililties

The team at Z.ai introduced two new GLM family members called GLM-4.5 and GLM-4.5-Air – designed to unify reasoning, coding, and agentic capabilities into a single model.

Claude Sonnet 4 now supports 1M tokens of context \ Anthropic

Claude Sonnet 4 now supports up to 1 million tokens of context on the Anthropic API.

Squint gets $40M in funding to accelerate human-to-machine collaboration in manufacturing

An industrial automation startup called Squint has raised $40 million as it bids to build on a vision of “agentic manufacturing,” where humans collaborate with artificial intelligence agents.

MLOps & LLMOps

AI Agent Design Patterns: How to Build Reliable AI Agent Architecture for Production

A technical blog post discussing the practical breakdown of the design principles for AI agent architecture that help to ship and scale real-world AI agents.

Four places where you can put LLM monitoring

A strategic blog post outlining four crucial locations for implementing LLM monitoring to effectively identify and mitigate dangerous or malicious AI actions.

Elysia: Building an end-to-end agentic RAG app

An innovative blog post presenting Elysia, an open-source, agentic RAG framework built on a decision-tree architecture that features dynamic data display types, AI data analysis, and more.

Learning

From GPT-2 to gpt-oss: Analyzing the Architectural Advances

An analytical article providing a detailed comparison and evolution of large language model architectures from GPT-2 to OpenAI's new open-weight gpt-oss models.

How Much Power will Frontier AI Training Demand in 2030?

A white paper summary forecasting that the electrical power demand for training frontier AI models will grow exponentially, potentially reaching 4-16 gigawatts by 2030 ...