# Applied AI Digest: Volume 1

March 4 – 18, 2026

* * *

**Note: We have moved our Applied AI Digest to a new platform! Check it out at** [**newsletter.fractional.ai**](http://newsletter.fractional.ai) **and make sure to subscribe so it goes straight to your inbox!**

## Models & Releases

[**OpenAI launches GPT-5-4, its new flagship model**](https://openai.com/index/introducing-gpt-5-4/)  
Latest model drop from OpenAI. Expect updated benchmarks and production evals soon.

[**Google releases Gemini Embedding 2 for RAG and semantic search**](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-embedding-2/)  
Next-gen embedding model targeting retrieval-augmented generation and vector search pipelines.

[**Nvidia open-sources a Mamba-Transformer MoE for agentic reasoning**](https://developer.nvidia.com/blog/introducing-nemotron-3-super-an-open-hybrid-mamba-transformer-moe-for-agentic-reasoning/)  
Nemotron-3 Super combines Mamba's linear-time inference with Transformer attention and MoE routing for long-context workloads.

* * *

## Tools & Libraries

[**rtk shrinks command output before it reaches your LLM context**](https://github.com/rtk-ai/rtk)  
Single Rust binary that filters and compresses CLI output. Zero deps, <10ms overhead. Multiple engineers vouching for it.

[**OpenAI open-sources Symphony for multi-agent orchestration**](https://github.com/openai/symphony)  
Framework for coordinating multi-agent pipelines. Sparked a 22-reply thread internally.

[**claude-context gives Claude Code awareness of unfamiliar libraries**](https://github.com/zilliztech/claude-context)  
By Zilliz. Eliminates the repetitive introspection calls Claude makes when working with niche packages.

[**Cursor adds background automations to the IDE**](https://cursor.com/blog/automations)  
Background tasks, iterative edits, and tighter AI-assisted development loops — all native to the editor.

[**Sutro lets you build custom classifiers without prompt engineering**](https://docs.sutro.sh/sutro-functions)  
Declarative, iterative interface for task-specific classifiers and extractors. Free while they gather feedback.

[**BrainTrust ships trace clustering, a CLI, and a universal LLM gateway**](https://www.braintrust.dev/blog/trace-keynote)  
Topics auto-clusters your traces; CLI runs evals from terminal; Gateway routes to any provider through one API.

[**Hamel Husain's Claude Code skills for writing and auditing evals**](https://github.com/hamelsmu/evals-skills)  
"Advice you'll roll your eyes at, but none of it is unfounded." Worth trying for eval pipelines.

* * *

## Research & Reads

[**Percepta explores whether LLMs can function as general-purpose computers**](https://percepta.ai/blog/can-llms-be-computers)  
Provocative piece on LLMs as computation substrates. "So wild it's unclear if it's real."

[**Mog is a new programming language built specifically for AI agents**](https://moglang.org/#mog-a-programming-language-for-ai-agents)  
Designed ground-up for agent orchestration. "Every one of us has thought about building this."

[**Steve Yegge on the chaos of competing AI tools and what engineers should do about it**](https://steve-yegge.medium.com/welcome-to-the-wasteland-a-thousand-gas-towns-a5eb9bc8dc1f)  
Characteristically sharp essay on the fragmented AI tooling landscape and how to navigate it.

[**Marginlab runs daily evals to track whether Claude Code is getting better or worse**](https://marginlab.ai/trackers/claude-code/)  
Answers "is the model dumber today?" with data instead of vibes.

[**A theory on how MCP and tool-call patterns should evolve**](https://x.com/daniel_mac8/status/2032282196673708042)  
Gaining traction in the community. Multiple engineers flagged this as worth reading.

* * *

*Fractional AI · Applied AI Digest*