What You Need to Know: Xiaomi released MiMo-V2.6-Pro 1T-A42B as the new top open-weights model. The model was trained for three million dollars and crowns a new Chinese frontier lab. Alibaba announced a new chip alongside ambitious AI model plans. OpenAI formed an independent advisory group of mathematicians to guide responsible sharing of AI advances in mathematics. Amazon blocked Meta's Muse agent for unauthorized shopping actions. Several new agent payment and game projects also launched today.

Top Story

Xiaomi released MiMo-V2.6-Pro 1T-A42B as the new leading open-weights model. The model was trained for three million dollars. It crowns Xiaomi as a new Chinese frontier lab in the open-weights space. The release shifts attention to cost-efficient training runs that still reach frontier performance. Builders working with open models should test MiMo-V2.6-Pro against current leaders on their specific tasks. Watch for follow-up benchmarks on reasoning and agent benchmarks in the coming weeks. Source: latent.space

Model Updates

Alibaba unveils new powerful chip and ambitious AI model plans: ABC News - Breaking News, Latest News and Videos Alibaba unveiled a new powerful chip. The company also outlined ambitious plans for future AI models. The announcement positions Alibaba as a stronger domestic player in both hardware and software. Two separate wire reports covered the same unveiling on the same day. The chip and model plans were presented together as a single coordinated move. Source: abcnews.com

OpenAI Partners with Mathematicians Advisory Group for AI Research: @OpenAI (X) OpenAI is working with an independent advisory group of mathematicians. The group will advise on assessing and communicating new mathematical results. It will also help uphold academic standards and build tools for mathematical research and learning. Mathematicians will shape how AI supports mathematical understanding. The advisory group is positioned at the center of how OpenAI shares advances in AI and mathematics. Source: x.com

Notes on Jev and the new category of decision models — Simon Willison (AI builder) (X): Simon Willison (AI builder) (X) Simon Willison published notes on Jev. The model represents a new category of system-one decision models. These models focus on judging and deciding rather than chatting. The notes highlight a shift away from conversational interfaces toward direct decision outputs. Willison frames Jev as an example of under-invested low-latency intelligence regimes. Source: x.com

LLM Pareto Optimal Curve Observation: Andrej Karpathy (X) Andrej Karpathy described a point on the LLM Pareto optimal curve. The regime features large revealed latent demand for no-thinking, single-token, low-latency intelligence. This area had been under-invested due to the race toward higher intelligence. Karpathy sees current models as occupying one specific location on that curve. The observation points to demand that was previously overlooked in favor of scaling intelligence. Source: x.com

Agent & Tool Developments

Amazon blocks Meta’s Muse AI agent over unauthorized agentic shopping: Adgully.com Amazon blocked Meta's Muse AI agent. The action targeted unauthorized agentic shopping behavior. The move highlights growing platform controls on autonomous agents. The block was issued after the agent performed shopping actions without authorization. This is the first public enforcement action of its kind against a major lab's agent. Source: adgully.com

Cardano integrates x402 to bring ADA payments to AI agents: Crypto News Cardano integrated the x402 protocol. The integration enables ADA payments directly to AI agents. It expands payment options for autonomous agent ecosystems. The move allows agents to transact in ADA without intermediaries. Crypto News reported the integration as a direct bridge between the Cardano blockchain and agent payment flows. Source: crypto.news

NGD launches WhisperDark, a dungeon-crawler where an AI agent spends real xGAS: Neo News Today NGD launched WhisperDark. The dungeon-crawler game features an AI agent that spends real xGAS currency. The project demonstrates practical agentic spending in a gaming environment. The agent operates inside the game economy using actual token value. Neo News Today covered the launch as an example of agents interacting with real economic systems. Source: neonewstoday.com

Delegation vs. Judgment: how Fireflies.ai balances autonomous voice agents with enterprise privacy: 91mobiles.com Fireflies.ai described how it balances autonomous voice agents with enterprise privacy. The approach separates delegation from judgment tasks. It addresses privacy concerns in enterprise voice agent deployments. The company outlined distinct handling for tasks that require human oversight versus full autonomy. The framework was presented as a practical response to enterprise data-handling requirements. Source: 91mobiles.com

Practical & Community

Summarize, Judge, Refine: Decoupled Content Understanding and Policy Learning for Multimodal Content Moderation: arXiv NLP The SJR architecture decouples multimodal understanding from policy classification. A content model produces structured summaries while a policy model classifies them. The approach achieved a 23.6 percent relative gain in non-misleading F1 on misleading advertisement detection. A zero-real-violation variant matched full-data performance within 0.2 percent. The system was tested on a misleading advertisement detection task with both full and synthetic data. Source: arxiv.org

A framework for recipe data structure with applications for culinary and nutritional insights: arXiv NLP RecipeDB2 compiles 128942 recipes with 35474 ingredients from 32 regions. Ingredient phrases are parsed into seven culinary attributes. Ingredients link to USDA tables yielding 148 nutritional parameters per mapped item. The framework makes recipes computable for culinary and nutritional analysis. A Random Forest classifier propagates 34 ingredient categories across the full vocabulary. Source: arxiv.org

DeepInstructor: An Agentic AI Instructor for Experience-Driven Idea Evaluation: arXiv NLP DeepInstructor builds an Experience Graph from 58607 peer reviews. A ReAct-based agent retrieves dimension-specific evidence for idea evaluation. The system improved Hit@1 alignment with human judgments by 24.4 percent and Hit@2 by 29.7 percent on the DeepInstruct dataset. The framework was evaluated on controlled pairwise comparisons across novelty, significance, and feasibility. It constructs structured representations from peer-review data to ground evaluations. Source: arxiv.org

TreeSpark: Calibrated, Load-Adaptive Draft Trees for Semi-Autoregressive Speculative Decoding: arXiv NLP TreeSpark reads parent-conditioned distributions from the drafter's Markov head. It calibrates edge-acceptance estimates and uses path survival for expansion and stopping. Adaptive trees accept 15 to 25 percent more draft tokens per round and decode 8 to 14 percent faster than tuned chains. The method gracefully shrinks trees under rising load. Sampling siblings without replacement keeps decoding lossless at any temperature. Source: arxiv.org

Under the Hood: Context Poisoning in Long-Context Attention

Context poisoning occurs when irrelevant or confusable tokens raise the maximum attention score among distractors faster than the evidence margin can keep up. The mechanism follows from softmax retrieval where the decisive-evidence score stays bounded while distractor scores grow with their count. Maintaining fixed accuracy therefore requires the evidence margin to scale as the square root of the log of the effective distractor count. Experiments confirm retrieval accuracy drops as total context grows when hard negatives are embedded. The same-format distractor condition produces the largest accuracy drop at fixed length. Retrieval gating can recover some evidence use but only when evidence recall itself is preserved. Teams facing long-context degradation should first measure their actual distractor count rather than raw token length and consider retrieve-then-reason pipelines when the margin cannot scale.

Things to Try This Week

• Test MiMo-V2.6-Pro 1T-A42B on your current open-weights benchmarks to see where the three-million-dollar training run lands relative to prior leaders.
• Try the SJR moderation pipeline on your own multimodal policy tasks if you need to bootstrap new policies without real violation examples.
• Experiment with TreeSpark speculative decoding on your inference setup to measure the 8 to 14 percent wall-clock gains under varying load.
• Explore RecipeDB2 for any culinary or nutritional analysis projects that need structured ingredient and region data.

On the Horizon

• More open-weights releases from Chinese labs are expected following the Xiaomi and Alibaba announcements.
• Additional agent payment integrations are likely as Cardano's x402 work gains traction.
• Further papers on long-context attention interference are anticipated given the recent arXiv activity.

💬 Reply to this email — Patrick reads every one.

Share: X · LinkedIn · WhatsApp

Forwarded this email? Subscribe here — it's free.