AI Daily — 2026-05-13

English 中文

Token Superposition Training Boosts LLM Pretraining Speed 2–3× · LeCun: World Models Essential fo...

Covering 17 AI news items

🔥 Top Stories

1. Token Superposition Training Boosts LLM Pretraining Speed 2–3×

Token Superposition Training (TST) speeds up standard LLM pretraining by 2-3× at matched FLOPs without altering architecture, optimizer, tokenizer, or data. In the first third of training, it processes contiguous token bags with input embeddings averaged and a modified cross-entropy; the remainder follows normal next-token prediction. The inference-time model is identical to conventional pretraining, and the method has been validated on 270M, 600M, 3B dense scales and 10B-A1B MoE; led by bloc97, gigant_theo, and theemozilla at Nous Research. Source-twitter

2. LeCun: World Models Essential for Reliable Agentic AI

Yann LeCun argues that a reliable agentic AI system requires a world model, which current LLMs lack. He contends that LLMs cannot predict the consequences of their actions before acting and thus do not demonstrate true intelligence. Without a world model, these models are limited in agency and foresight. Source-twitter

📰 Featured

LLM

Claude Plans Offer Monthly Credits for Programmatic Use — Starting June 15, paid Claude plans will include a dedicated monthly credit for programmatic usage, covering Claude Agent SDK, claude -p, Claude Code GitHub Actions, and third-party apps built on the Agent SDK. Separately, social chatter warns that using certain tools with Claude can trigger a 25x reduction in usage, disguised as free credits. Source-twitter
AI Ends LeetCode Interviews, Proponent Praises It — An AI-powered claim argues that modern models can solve LeetCode-style interview puzzles in one shot, ending a decade of memory-based coding questions. The post praises AI progress and foresees a drastic shift in how technical hiring is conducted, as interview puzzles become obsolete. Source-twitter
Qwen 3.6 Plus Free on Nous Portal for Limited Time — Alibaba’s Qwen 3.6 Plus is now available on Nous Portal for free for a limited time. Nous Portal provides access to 300+ models and bundles tokens and paid tools for easier setup and billing, highlighting the Hermes Agent integration. Source-twitter
MCP-Cosmos Injects World Models into MCP for Predictive Task Automation — MCP-Cosmos extends the Model Context Protocol by integrating generative World Models into MCP-based agents, addressing the gap between task planning and execution-time dynamics. The framework aims to provide long-horizon foresight for complex tasks in MCP environments, balancing planning with predictive execution. This open research aligns LLMs, external tools, and environment modeling to enable more capable autonomous agents. Source-huggingface

AI Tools

Codex tests apps across viewports using in-app browser — Codex can now use the in-app browser to test apps at multiple viewport sizes, controlling the device toolbar and simulating clicks across breakpoints. It captures key screenshots during long tests and presents them at the end, and can speed up testing by hiding animations and running 1-2x faster. Annotations were optimized to send faster and use fewer tokens. Source-twitter
AiToEarn: AI-driven content marketing for solo creators — AiToEarn is an AI-powered content marketing platform for OPCs (one-person companies), creators, and brands to build, publish, distribute, and monetize content across major platforms. It supports five deployment methods, including Docker-based private deployment, and integrates with OpenClaw, Claude, Cursor, and other agents via MCP for cross-platform use. Recent updates include a 2026 content marketplace, MCP support across agents, OpenClaw integration, and offline merchant promotion features. Source-github

Embodied AI

World Action Models: Next Frontier in Embodied AI — Vision-Language-Action models generalize well but lack explicit world dynamics under intervention. World Action Models (WAMs) integrate predictive environment dynamics into the action-generation process, forming embodied foundation models. This emerging paradigm seeks to unify perception, language, and action in embodied AI. Source-huggingface

Multimodal

AlphaGRPO Enables Self-Reflective Multimodal Reasoning in UMMs — AlphaGRPO proposes a novel framework that applies Group Relative Policy Optimization to AR-Diffusion Unified Multimodal Models to boost multimodal generation without a cold-start stage. It enables Reasoning Text-to-Image Generation by inferring implicit user intents and Self-Reflective Refinement where the model autonomously diagnoses and corrects its outputs. Source-huggingface

⚡ Quick Bites

Anthropic’s $1T vs Google’s $4.5T Valuations Spark Debate — A tweet compares Anthropic’s trillion-dollar valuation with Google’s $4.5 trillion valuation, suggesting two possible conclusions: Anthropic is overpriced or Google is underpriced. The post frames this as a puzzle about relative AI market value and invites analysis. Source-twitter
OpenAI Codex Enterprise Promo: 2 Free Months for Switching — OpenAI is offering an enterprise promotion for Codex: eligible customers who switch within 30 days receive two free months of Codex usage for new users. The post encourages sharing with a CTO to bring teams onto Codex. Source-twitter
Jensen: Need More NVDA GPUs, Enable HLS Playback — Jensen Huang tweeted that Nvidia needs many more GPUs and referenced enabling HLS playback. The post highlights demand for Nvidia hardware to support AI and video workloads, reflecting GPU supply considerations in the AI industry. Source-twitter
ToolCUA Advances GUI-Tool Path Orchestration for CUAs — ToolCUA examines when Computer Use Agents should stick with GUI actions or switch to tool calls to optimize task execution. It highlights that scarce high-quality interleaved GUI-tool trajectories and the cost and brittleness of collecting real tool data impede optimal planning. The work outlines steps toward achieving optimal GUI-Tool path orchestration for CUAs. Source-huggingface
Tradeoffs: speed, price, and intelligence in AI models — The author expresses anxiety about not using the smartest model, even if it means slower performance. They propose focusing on price/speed versus price/intelligence tradeoffs when deploying AI. The note highlights practical cost-performance considerations in AI systems. Source-twitter
Claude Code weekly limits up 50% through July 13 for all plans — Claude Code is increasing weekly usage limits by 50% through July 13. The uplift applies to Pro, Max, Team, and seat-based Enterprise users. Source-twitter
AI progress often goes unnoticed as models improve — An observation from a tweet argues that claims about AI models being intelligent enough and the pace of improvements are hard to notice. It suggests this perception is a ‘take that ages badly’ and reflects on how public perception lags behind underlying progress in AI. Source-twitter

Generated by AI News Agent | 2026-05-13