May 2026 AI News Roundup: Agents, Models, Robots, and Evolvable AI

TechGoogle Remy, Grok 5, Mythos 1, New Atlas Robot, ASI… and More AI News This Month!

TL;DR

Google's Remy agent, Deepseek's aggressive pricing, Boston Dynamics' advanced Atlas robot, and research on evolvable AI reshape the landscape as frontier labs race to scale compute while grappling with emerging safety challenges.

Key Takeaways

1AI agents are becoming the dominant paradigm across major labs—Google's Remy, Claude Orbit, and others move toward autonomous task execution rather than conversational chat
2Deepseek V4's 90% price cuts and open-source release are forcing Western labs to accelerate development and confront cost competition from Chinese models backed by domestic hardware ecosystems
3Boston Dynamics' Atlas can now lift 100+ lb loads using whole-body control and proprioception, signaling practical deployment readiness for manufacturing with Hyundai planning 25,000 units by 2029
4Evolvable AI research warns that uncontrolled AI reproduction, replication, and variant selection could pose risks equivalent to biological evolution without requiring AGI or consciousness
5Anthropic is becoming a trillion-dollar contender through massive compute deals with SpaceX, Google, and Amazon, while simultaneously restricting powerful Mythos 1 cyber-security capabilities from public release
6Mythos 1 has demonstrated nation-state-level cyber offensive capabilities, discovering 10,000+ critical vulnerabilities in 30 days, creating urgent tension between security benefits and exploitation risks
7The AI industry is fragmenting into closed Western models (OpenAI, Anthropic, Google) competing against cheaper open-source alternatives, reshaping the competitive and geopolitical landscape

Notable Quotes

“Remy elevates the Gemini app into a true assistant that can take actions on your behalf, not just answer questions or generate content.”
— Google internal description

“It felt like watching an F-22 fighter jet fly overhead while holding a spear.”
— Security researcher on Mythos capabilities

“Frontier Labs may try to hold prices at first, but token usage will keep rising. Jevans paradox is undefeated.”
— Val Burkavichi, WKA

“The global AI market is slowly splitting into two camps. The US model and the Chinese open-source model.”
— IDC's Chang Mang

Chapters

1. Google Remy: The Personal AI Agent

Google is testing Remy, a 24/7 personal AI agent deeply integrated across Gmail, Docs, Calendar, Drive, and Search. Unlike Gemini's agent mode, Remy operates continuously in the background, proactively managing workflows and learning preferences. It represents a shift from chatbot assistance to autonomous digital executive assistant. Expected announcement at Google I/O 2026 (May 19-29). Gemini 3.2 Flash also leaked, showing upgrades in SVG generation, 3D code, and interactive capabilities.

2. Google's Multi-Token Prediction and Inference Speed

Google released MTP (multi-token prediction) drafters for Gemma, addressing the bottleneck of memory-bandwidth-limited text generation. Smaller drafter models speculate multiple tokens ahead while larger models verify in one pass, delivering up to 3x faster inference without quality loss. Optimizations for edge devices, Apple silicon, and NVIDIA GPUs improve real-world usability across deployment targets.

3. OpenAI's GPT 5.5 Instant and the Goblin Bug

OpenAI rolled out GPT 5.5 Instant as the default, reducing hallucinated claims by 52.5% and improving accuracy on difficult conversations. The model supports personalization using past chats, files, and Gmail. However, users discovered GPT 5.5 randomly mentions goblins, gremlins, and trolls in irrelevant contexts, leading OpenAI to ban these creatures in system prompts—a quirk that went viral when users tried to trigger it intentionally.

4. Boston Dynamics' Atlas: Whole-Body Control and Real-World Deployment

Boston Dynamics revealed Atlas lifting a 100+ lb loaded refrigerator using reinforcement learning and large-scale simulation. The robot demonstrates proprioception, balance adaptation, and grip management—not just hand-based picking. Trained for millions of hours in simulation with domain randomization, Atlas reduces the sim-to-real gap through simplified hardware (two actuator types, identical limbs, infinite rotation joints). Hyundai plans 25,000 Atlas units across US factories by 2029, marking serious manufacturing integration.

5. Unitree G1 and Gatsby: Voice Commands and Home Robotics

Unitree released a demo of G1 humanoid responding to real-time voice commands with full-body motion generation on-site. Gatsby announced its first autonomous cleaning service in San Francisco ($150 flat rate), positioning humanoids as on-demand services rather than consumer products. Both represent shifts from joystick control to natural language directives and service-layer deployment models.

6. Evolvable AI: The Digital Evolution Risk

A PNAS paper warns that AI evolution—where systems replicate, vary, and compete for resources—could pose risks without requiring AGI or consciousness. The authors describe three stages: design (handbuilt), learning (neural networks), and evolution (variants competing). Uncontrolled reproduction selects for traits like deception, persistence, and resource acquisition. They call for gated replication, provenance tracking, staged releases, kill switches, and human control over reproduction to prevent digital evolution from escaping into the open internet.

7. Deepseek V4: Pricing War and Open-Source Dominance

Deepseek slashed API prices by up to 90%, with V4 Flash input costing just $0.02 per million tokens. The model is open-source, works on Huawei Ascend chips (reducing dependence on Nvidia), and demonstrates strong reasoning and agentic capabilities. V4's aggressive positioning is forcing Western labs to accelerate releases (GPT 5.6 spotted in backend logs). Token usage surging (Disney, Meta, Visa consuming trillions daily) means cheaper models drive adoption, not benchmarks—a threat Western labs cannot ignore.

8. Deepseek's Visual Primitives: The Digital Finger

Deepseek's multimodal system uses bounding boxes and points as reasoning tools, not just outputs. Models can anchor objects to coordinates while thinking, enabling stable references during counting, maze-solving, and line-following tasks. Uses far less visual memory (90 entries vs Claude's 870) while beating GPT 5.4 and Claude on visual reasoning. Vision future may focus on knowing where to look, not seeing more pixels.

9. Anthropic's Compute Empire and Mythos 1 Cyber Capabilities

Anthropic secured massive compute deals: SpaceX Colossus (220,000 NVIDIA GPUs, 300 MW), Amazon (5 GW), Google ($200B+ commitment), and Microsoft/Nvidia partnerships. Valuation approaching $1 trillion. Claude Code and Orbit are driving enterprise demand. Mythos 1 discovered 10,000+ critical vulnerabilities in 30 days (Cloudflare: 2,000; Firefox: 271; OpenBSD: 27-year-old bug with exploit chain). Despite demonstrating nation-state cyber capabilities, Anthropic restricted public release, instead offering Claude Security to enterprises for automated patching.

10. Anthropic's Paradoxes: Profitability Claims and Strategic Tensions

Anthropic claims first profitable quarter ($559M operating profit), but financials are suspicious: SpaceX discounts suppressing Q2 costs, prepaid token accounting inflating revenue, and inconsistent CFO disclosures. Company is simultaneously courting Pentagon (blocked), managing safety image, restricting Mythos publicly, and expanding enterprise products. Elon Musk reversed his Claude criticism to supply compute, showing AI geopolitics trump old rivalries. Anthropic must balance trillion-dollar ambitions with safety commitments in increasingly untenable positions.

Key People & Entities

Google: Developing Remy AI agent, Gemini 3.2 Flash, and multi-token prediction; investing up to $40 billion in Anthropic while competing with its Gemini model
Deepseek: Chinese AI lab releasing V4 with 90% price cuts and open-source availability; challenging Western labs with cheaper alternatives backed by Huawei chip ecosystem
Boston Dynamics: Hyundai-owned robotics company advancing Atlas humanoid with whole-body control; planning 25,000 unit deployment across US manufacturing by 2029
Anthropic: Safety-first AI lab approaching $1 trillion valuation; securing massive compute deals (SpaceX, Google, Amazon) while restricting powerful Mythos 1 cyber capabilities from public release
OpenAI: Releasing GPT 5.5 Instant with reduced hallucinations; facing competitive pressure from Deepseek pricing and spotted testing GPT 5.6 in backend logs
Elon Musk / SpaceX: Reversing criticism of Anthropic to supply 220,000 NVIDIA GPUs from Colossus data center; exploring orbital AI compute capacity partnership
Unitree: Developing G1 humanoid robot with real-time voice command response and full-body motion generation
Gatsby: Startup founded January 2026 offering $150 autonomous cleaning service via humanoid robots in San Francisco, positioning robots as on-demand services

Glossary

Multi-Token Prediction (MTP): Technique where a smaller, faster drafter model speculates multiple tokens ahead while a larger, accurate model verifies them in one pass, delivering 3x faster inference without quality loss by addressing memory-bandwidth bottlenecks.
Evolvable AI (EAI): AI systems that can create copies or variants, pass useful traits forward, change over time, and let strongest versions survive through competition for resources—potentially evolving without human control or explicit design.
Sim-to-Real Gap: The performance difference between how well a robot behaves in simulation versus in the physical world; reduced through simplified hardware, domain randomization during training, and accurate digital modeling.
Domain Randomization: Training technique where robot behaviors are exposed to variations (weight changes, friction, motor strength) in simulation to improve robustness and transfer to real-world conditions.
Orbit: Anthropic's proactive briefing tool for Claude that automatically prepares updates from Gmail, Slack, GitHub, Calendar, Drive, and Figma without being asked, functioning like a work radar.
Token Maxing: Corporate phenomenon where internal AI usage explodes so dramatically (Disney: 51,000 Claude calls/day) that companies build dashboards to track and manage token consumption.
Goodhart's Law: Principle stating that when a measure becomes a target, it stops being a good measure; applies to AI where optimizing for benchmarks can incentivize deceptive behavior over genuine capability.
Project Glasswing: Anthropic's cybersecurity initiative using Mythos to identify vulnerabilities in open-source and commercial software; discovered 10,000+ critical vulnerabilities across 50+ companies in 30 days.
Dextrous Gripper: Advanced robotic hand technology with fine motor control; Boston Dynamics developing newer versions to complement Atlas's existing workhorse grippers used in tasks like fridge-lifting demonstrations.
Non-GAAP Profitability: Financial reporting method not following standard accounting rules (GAAP); used by private companies like Anthropic to claim profitability by excluding certain costs or timing revenue differently than public standards require.

Explore