Deep Learning Weekly: Issue 433
Deep Dives
Explore related topics with these Wikipedia articles, rewritten for enjoyable reading:
-
Mixture of experts
12 min read
The article mentions Mistral Large 3 as a 'sparse mixture-of-experts model' - understanding this neural network architecture technique would help readers grasp why this is significant and how it differs from dense transformer models
-
Theory of mind
15 min read
The SoMi-ToM paper directly evaluates Theory of Mind capabilities in AI models - this cognitive science concept about understanding others' mental states is foundational to understanding why this benchmark matters for AI development
This week in deep learning, we bring you Introducing Mistral 3, MCP Explorer, and a paper on Agentic Bridge Framework: Closing the Gap Between Agentic Capability and Performance Benchmarks.
You may also enjoy Laying the Foundations for Visual Intelligence, 8 learnings from 1 year of agents, a paper on SoMi-ToM: Evaluating Multi-Perspective Theory of Mind in Embodied Social Interactions, 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
The Mistral team announced Mistral 3, which includes three state-of-the-art small models and Mistral Large 3 — a sparse mixture-of-experts model.
Laying the Foundations for Visual Intelligence—$300M Series B | Black Forest Labs
Black Forest Labs raises $300M Series B at $3.25B valuation to advance visual intelligence models beyond its popular FLUX image generation platform.
New training method boosts AI multimodal reasoning with smaller, smarter datasets
Researchers at MiroMind AI and several Chinese universities have released OpenMMReasoner, a training framework that improves the capabilities of models in multimodal reasoning.
Skill Learning: Bringing Continual Learning to CLI Agents
The Letta team released Skill Learning, a way for Letta Code to dynamically learn skills over time.
MLOps & LLMOps.
An educational project for learning Anthropic’s Model Context Protocol through a narrative-driven and interactive learning experience.
8 learnings from 1 year of agents
A detailed retrospective blog post sharing 8 key learnings from a year of developing PostHog AI, focusing on architectural choices like using a single LLM loop and the power of continuous model improvements.
OpenSearch as an agentic memory solution: Building context-aware agents using persistent memory
A blog post that explores the memory challenges facing AI agents, introduces agentic memory’s core concepts, and demonstrates how to integrate it with your agent frameworks.
Build and Run Secure, Data-Driven AI Agents
A technical guide detailing the deployment of NVIDIA’s AI-Q Research Assistant and Enterprise RAG Blueprints, which use Nemotron NIMs and an agentic Plan-Refine-Reflect workflow on Amazon EKS.
Learning
Evaluating honesty and lie detection techniques on a diverse suite of dishonest models
An alignment report evaluating techniques like fine-tuning and prompting to improve AI honesty and detect lies across five specialized testbed models.
A Rosetta Stone for AI benchmarks
A statistical paper introducing a “Rosetta Stone” framework that stitches together around 40 ...
This excerpt is provided for preview purposes. Full article content is available on the original publication.