not much happened today

2026-05-21 · Source: AINews · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Cloud Computing & IT Infrastructure · Depth: Advanced, extended

Summary

The AI news recap for May 20-21, 2026, highlights significant advancements across models, agentic systems, and infrastructure. Key model updates include RAEv2 for unified vision understanding, NVIDIA's Gated DeltaNet-2 outperforming Mamba-3 at 1.3B parameters, and a surprising "no filter" data scaling result for LMs. OpenAI's reported success on an Erdős unit-distance problem sparked debate on AI-assisted math research. Agent developments feature new harnesses boosting models like Gemini 3.1 Pro from 17.7 to 31.4, OpenAI's Codex expanding to remote Mac app use, and Gemini 3.5 Flash topping APEX-Agents-AA. Infrastructure saw Turbopuffer reach a \$100M run-rate, Modal secure a \$355M Series C at a \$4.65B valuation, and Hark raise \$700M at \$6B. Open-source Qwen 3.7 Max benchmarks show strong performance, while Meta issued a legal notice to Heretic for Llama derivatives. Multimodal advancements include Runway's Aleph 2.0 for video editing and Hugging Face's LeRobot Humanoid for open robotics.

Key takeaway

For AI Scientists and Machine Learning Engineers evaluating new models and infrastructure, you should prioritize exploring representation-first tokenization and advanced attention mechanisms like Gated DeltaNet-2 for performance gains. Consider adopting agentic harnesses to significantly boost existing model capabilities. Furthermore, investigate Anthropic's free MCP courses to deepen your understanding of Claude API and agent integration, which can streamline your development workflows.

Key insights

The AI landscape is rapidly evolving with significant progress in model efficiency, agentic capabilities, and infrastructure scaling.

Principles

• Representation-first tokenization accelerates model convergence.
• Agent harnesses significantly boost model performance.
• AI infrastructure is a major wealth creation area.

Method

Propose a risk-managed "vibe coding" workflow: start with a plan, keep tasks small, require agent-generated tests, commit to git, back up databases, and use browser/E2E tools for live validation.

In practice

• Optimize Qwen3.6-35B-A3B-MTP with ik_llama.cpp for +23% tok/s.
• Utilize Anthropic's free MCP courses for Claude API and agent integration.

Topics

• AI Model Architectures
• Agentic Systems
• AI Infrastructure
• Large Language Models
• Model Quantization
• Open-Source Robotics
• Multimodal AI

Code references

• ikawrakow/ik_llama.cpp
• obra/superpowers

Best for: Investor, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• not much happened today — AI advancements are accelerating across architecture, agent infrastructure, and multimodal embeddings, with a strong focus on inference efficiency.
• Anthropic is about to become the first profitable AI lab — Anthropic is nearing profitability driven by coding tools and agentic AI, despite increased user costs and compute demands.
• not much happened today — Agentic AI success increasingly relies on robust harnesses and infrastructure, not just base model capabilities.
• Claude Code is Now Writing Claude Code — AI infrastructure and coding capabilities are rapidly advancing, driven by massive investments and self-improving models.
• 🗞️ Anthropic releases Claude Opus 4.8 on the same day as its $965B valuation round. — The AI frontier is rapidly expanding with new model capabilities, inference optimizations, and valuation milestones.
• 🗞️ OpenAI’s new paper shows how they are now seeing the first version of office work where agents do most of the execution. — AI agents are transforming work and driving economic growth, yet human-AI collaboration and infrastructure remain critical bottlenecks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AINews.