Next Week in The Sequence:

• Our series about world models continues with

• In the AI of the week installment, we will deep dive into Xiaomi’s new

• MiMo-V2-Pro that is killing it on the leaderboAI software ambitions and challenges

Subscribe and don’t miss out:

📝 Editorial: Last Week in AI: NVIDIA’s GTM Releases, Bezos’s $100B Bet, Xiaomi’s Ambush, and the Fracturing of OpenAI

This week in AI feels like a definitive inflection point. We are no longer just talking about chat interfaces and reasoning benchmarks; the industry is aggressively pivoting toward agentic infrastructure and physical world integration. From Silicon Valley to Beijing, the narrative has shifted entirely from “what can AI say?” to “what can AI do?”

Nowhere was this paradigm shift clearer than at NVIDIA GTC 2026. While hardware updates are always a staple, Jensen Huang’s keynote made one thing unequivocally clear: NVIDIA is now a software and infrastructure company. The spotlight was stolen by their agentic software announcements, most notably Dynamo 1.0—an open-source distributed operating system for AI factories—and NemoClaw, an enterprise stack designed to build and orchestrate self-evolving agents. By integrating these with the new Agent Toolkit, NVIDIA is positioning itself as the foundational orchestration layer for the agentic economy. They aren’t just selling the silicon anymore; they are building the entire nervous system for enterprise AI.

Meanwhile, the frontier model race saw a “quiet ambush” from the East. Xiaomi stunned the global AI community with the release of MiMo-V2-Pro. Boasting 1 trillion parameters and a massive 1-million-token context window, the model rivals the performance of GPT-5.2 and Claude Opus 4.6 at a fraction of the inference cost. But what makes MiMo-V2-Pro truly fascinating is its design philosophy. Rooted in Xiaomi’s deep consumer hardware and EV pedigree, the model is engineered specifically for the “action space.” It bypasses the traditional conversational paradigm, built instead to act as the brain for digital workflows and physical robotics. The moat around US-based frontier models is rapidly evaporating.

Further cementing this push into the physical world, Jeff Bezos is reportedly raising a staggering $100 billion fund to acquire and overhaul legacy manufacturing, aerospace, and defense companies. Dubbed a “manufacturing transformation vehicle,” the fund will leverage technology from Bezos’s own spatial AI startup, Project Prometheus. The goal is bold and disruptive: to fundamentally rewire global industrial supply chains by replacing human dependency on the factory floor with spatial and physical AI. It is a massive bet that spatial AI will do to manufacturing what LLMs have done to knowledge work.

Yet, as the technology scales into the real world, early corporate alliances are fracturing under the weight of astronomical compute demands. The foundational partnership of the AI boom is currently on the brink of a courtroom battle. Microsoft is reportedly threatening to sue OpenAI over a multi-billion dollar deal to host OpenAI’s upcoming “Frontier” model on Amazon Web Services (AWS). Microsoft argues this violates the spirit of their Azure exclusivity agreement, while an increasingly independent OpenAI is desperately scrambling for the raw compute capacity required to stay ahead.

From NVIDIA’s agentic operating systems and Xiaomi’s hardware-brained LLMs to Bezos’s industrial buyout fund and the Microsoft-OpenAI fracturing, this week offered a stark glimpse into the next era of artificial intelligence. The models are stepping out of the browser and into the factory, the data center, and the courtroom.

🔎 AI Research

Efficient Exploration at Scale

• AI Lab: Google DeepMind

• Summary: This paper presents an online learning algorithm that drastically improves the data efficiency of reinforcement learning from human feedback (RLHF) by incrementally updating reward and language models as choice data is received. By utilizing an affirmative nudge, an epistemic neural network for reward uncertainty, and information-directed exploration, the algorithm matches the performance of offline RLHF trained on 200K labels using fewer than 20K labels.

EvoClaw: Evaluating AI Agents on Continuous Software Evolution

• AI Lab: USC, UCR, UCSD, Army Research Office, Stanford, Princeton, Haven, OpenHands

• Summary: EvoClaw introduces a novel benchmark and the DeepCommit pipeline to evaluate AI agents on continuous, dependency-driven software evolution rather than isolated, one-off coding tasks. The study reveals that while frontier models excel at independent tasks, their performance drops significantly in continuous environments due to a severe inability to maintain system integrity and prevent error accumulation over an extended development horizon.

MetaClaw: Just Talk – An Agent That Meta-Learns and Evolves in the Wild

• AI Lab: UNC-Chapel Hill, Carnegie Mellon University, UC Santa Cruz, UC Berkeley

• Summary: This paper introduces MetaClaw, a continual meta-learning framework that allows deployed LLM agents to evolve in the wild without service downtime by jointly maintaining a base policy and a dynamic skill library. It achieves this continuous improvement through skill-driven fast adaptation from failure trajectories and opportunistic policy optimization via cloud LoRA fine-tuning during user-inactive periods.

Online Experiential Learning for Language Models

• AI Lab: Microsoft Research

• Summary: This paper introduces Online Experiential Learning (OEL), a framework that enables large language models to continuously improve from real-world deployment experiences. It operates by extracting knowledge from user interactions and consolidating it into model parameters via on-policy context distillation, progressively enhancing task accuracy and inference efficiency over successive iterations.

Efficient Reasoning on the Edge

• AI Lab: Qualcomm AI Research

• Summary: The authors propose a lightweight framework to enable advanced reasoning capabilities on resource-constrained edge devices by combining LoRA adapters with supervised fine-tuning. To address strict latency and memory limits, the system dynamically switches to a specialized reasoning mode only when needed and uses budget-forced reinforcement learning to significantly reduce generation length with minimal accuracy loss.

Recursive Language Models Meet Uncertainty: The Surprising Effectiveness of Self-Reflective Program Search for Long Context

• AI Lab: Apple

• Summary: This study introduces Self-Reflective Program Search for Long Context (SRLM), a framework that guides language models through long contexts using internal uncertainty signals like self-consistency, reasoning trace length, and verbalized confidence. The findings demonstrate that this uncertainty-aware self-reflection method can match or surpass traditional recursive language models across varying context lengths, without requiring explicit self-query mechanisms.

🤖 AI Tech Releases

Agent Toolkit

NVIDIA released its Agent Toolkit, a new agent development platform.

NemoClaw

NVIDIA announced its enterprise-ready version of OpenClaw for semi-autonomous agentic workflows.

Forge

Mistral released Forge, a new platform that allows enterprises to train their own models.

MiniMax M2.7

MimiMax open sourced M2.7, its new model optimized for agentic workflows.

📡AI News You Need to Know About

• NVIDIA AI Announcements At its annual GTC conference, NVIDIA unveiled the Vera Rubin full-stack AI platform (7 chips, 5 rack systems), the NemoClaw enterprise agent stack, DLSS 5, space computing modules, open-source physical AI models, autonomous vehicle deals with Uber and major automakers, and raised its 2025–2027 revenue outlook to $1 trillion — cementing NVIDIA’s push to own every layer of the AI stack from silicon to software.

• Bezos $100B Manufacturing Fund Wall Street Journal Jeff Bezos is reportedly raising $100 billion to acquire and AI-transform legacy manufacturers across aerospace, defense, and chipmaking, tied to his AI startup Project Prometheus.

• Meta AI Content Enforcement Meta announced it will gradually replace third-party content moderation vendors with more advanced in-house AI systems capable of detecting scams, terrorist content, and drug sales at greater speed and accuracy.

• Microsoft Acqui-hires Cove Microsoft hired the entire team behind Cove, a Sequoia-backed AI collaboration startup that built an infinite whiteboard tool, with the product shutting down April 1.

• Fuse Raises $25M AI-native loan origination startup Fuse raised a $25M Series A to modernize the aging lending software used by U.S. credit unions, including a $5M “rescue fund” to help institutions break legacy vendor contracts.

• OpenAI Desktop Super App OpenAI is developing a unified desktop “super app” that will combine ChatGPT, its Codex coding tool, and its Atlas AI browser into a single interface, led by Applications CEO Fidji Simo.

• OpenAI Acquires Astral OpenAI announced plans to acquire Astral, the startup behind popular Python developer tools like uv and Ruff, to bolster its Codex coding platform which now has over 2 million weekly active users.

• Crypto.com Lays Off 12% Crypto.com cut approximately 12% of its workforce (~180 employees), with CEO Kris Marszalek citing enterprise-wide AI integration as the driver and warning that companies failing to pivot immediately “will fail.”

• Xiaomi Share Surge Xiaomi shares jumped up to 5.8% after the company released its MiMo V2 Pro AI model and unveiled the facelifted SU7 EV, which sold 15,000 units in 34 minutes. (No single Xiaomi press release covers both events; Bloomberg is the primary aggregator here.)

• Standard Template Labs Raises $49M Seed Standard Template Labs, an AI-first IT service management startup founded by former Datadog President Amit Agarwal and incubated by ICONIQ, raised a $49M seed round co-led by ICONIQ and CRV to build a platform that autonomously resolves enterprise IT operations end-to-end — replacing legacy ticketing systems like ServiceNow.