For the first time, the US government named specific frontier AI models and ordered them shut down globally — not because of a safety incident, but because of a claimed jailbreak and national-security concerns. The Anthropic shutdown of Claude Fable 5 and Mythos 5 marks a shift from controlling AI hardware to controlling the models themselves, and it sets a precedent every major AI lab now has to price into its plans. The real story may be less about model safety and more about industrial policy: keeping the most capable American AI systems out of global reach.

Featured story

• Anthropic Disables Claude Fable 5 and Mythos 5 After US Government Order — MarkTechPost

Also today

• Moonshot AI Releases Kimi K2.7-Code: a Coding Model Reporting +21.8% on Kimi Code Bench v2 Over K2.6 — MarkTechPost
• Google Releases Gemini-SQL2: Gemini 3.1 Pro Text-to-SQL Scores 80.04% on BIRD Single-Model Leaderboard — MarkTechPost
• olmo-eval: An evaluation workbench for the model development loop — Hugging Face Blog

AI agents — software that takes a goal and acts to achieve it, rather than just answering questions — crossed a commercial threshold today, with three separate product launches in a single day. The most striking is Kimi Work, a desktop app from Beijing-based Moonshot AI that runs up to 300 simultaneous AI sub-agents on your own machine, using your real files and browser sessions. But as the products ship, researchers are quietly surfacing evidence that the "reasoning" these agents display may be sophisticated pattern-matching rather than genuine logic — and no one has yet built reliable tests to know the difference at scale.

Featured story

• Moonshot AI Launches Kimi Work, a Local Desktop Agent Reportedly Running on Kimi K2.6 With a 300-Sub-Agent Agent Swarm — MarkTechPost

Also today

• Perplexity Moves Deep Research Into Computer, Routing Research Subtasks Across 20+ Frontier Models For Reports, Decks, And Dashboards — MarkTechPost
• xAI Ships Grok Build Plugin Marketplace With MongoDB, Vercel, Sentry, Chrome DevTools, Cloudflare, and Superpowers Plugins at Launch — MarkTechPost
• MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scaling — arXiv cs.CL
• Multi-Agent Reinforcement Learning from Delayed Marketplace Feedback for Objective-Weight Adaptation in Three-Sided Dispatch — arXiv cs.AI
• Zyphra Release Zamba2-VL: Hybrid Mamba2–Transformer Vision-Language Models That Cut Time-to-First-Token by About an Order of Magnitude — MarkTechPost
• One Polluted Page Is Enough: Evaluating Web Content Pollution in Generative Recommenders — arXiv cs.AI
• Beyond the Commitment Boundary: Probing Epiphenomenal Chain-of-Thought in Large Reasoning Models — arXiv cs.AI
• Reasoning as Pattern Matching: Shared Mechanisms in Human and LLM Everyday Reasoning — arXiv cs.AI
• AgentBeats: Agentifying Agent Assessment for Openness, Standardization, and Reproducibility — arXiv cs.AI
• EvoArena: Tracking Memory Evolution for Robust LLM Agents in Dynamic Environments — arXiv cs.CL
• EpiBench: Verifiable Evaluation of AI Agents on Epigenomics Analysis — arXiv cs.AI
• EurekAgent: Agent Environment Engineering is All You Need For Autonomous Scientific Discovery — arXiv cs.AI
• Reward Modeling for Multi-Agent Orchestration — arXiv cs.AI
• Recursive Agent Harnesses — arXiv cs.CL

Google released DiffusionGemma today — a model that generates text in parallel rather than word by word — and the speed gains are real. But the deeper story is what this bet reveals: that the autoregressive paradigm underlying every major AI model may be approaching its ceiling, and the safety infrastructure built around it may not survive the transition.

Featured story

• Google AI Releases DiffusionGemma, a 26B MoE Open Model Using Text Diffusion for Up to 4x Faster Generation — MarkTechPost

Also today

• DiffusionGemma: 4x faster text generation — Google DeepMind Blog
• Google DeepMind is worried about what happens when millions of agents start to interact — MIT Technology Review
• Meet 'North Mini Code': Cohere's 30B Open-Weight Mixture-of-Experts Model With 3B Active Parameters for Agentic Coding — MarkTechPost
• Supporting Europe's work in ensuring a trustworthy AI ecosystem — OpenAI News

Five independent research papers published today — none citing each other — converge on a single finding: post-training is where model behavior is actually determined, and current methods produce models whose alignment is fundamentally unstable and opaque even to their creators. That result collides directly with Anthropic's launch of a safety-differentiated Claude Mythos tier and Google's multimodal expansion, raising a question neither company is answering: how do you guarantee a safety tier when the research community is proving that alignment established during post-training does not reliably survive adversarial fine-tuning, architectural changes, or extended reasoning?

Thread 1: The Alignment Debt Is Coming Due

• Anthropic Releases Claude Fable 5 and Claude Mythos 5: Same Underlying Model, Different Safeguards, New Mythos-Class Tier — MarkTechPost `[3a21e90463]`
• It Takes One to Bias Them All: Breaking Bad with One-Shot GRPO — arXiv cs.CL `[6a3992fe11]`
• Does Reasoning Preserve Alignment? On the Trustworthiness of Large Reasoning Models — arXiv cs.CL `[409b6da10a]`
• CIAware-Bench: Benchmarking Control Intervention Awareness Across Frontier LLMs — arXiv cs.AI `[b39b1d1076]`
• ABC-Bench: An Agentic Bio-Capabilities Benchmark for Biosecurity — arXiv cs.AI `[dda8668e64]`
• PhantomBench: Benchmarking the Non-existential Threat of Language Models — arXiv cs.AI `[abe23429f9]`

Thread 2: Benchmarks All the Way Down

• T1-Bench: Benchmarking Multi-Scenario Agents in Real-World Domains — arXiv cs.AI `[ce4efb41e4]`
• Workflow-GYM: Towards Long-Horizon Evaluation of Computer-use Agentic tasks in Real-World Professional Fields — arXiv cs.AI `[ab924dc4fe]`
• Mind the Gap: Can Frontier LLMs Pass a Standardized Office Proficiency Exam? — arXiv cs.CL `[f829485320]`
• VISTA: A Versatile Interactive User Simulation Toolkit for Agent Evaluation — arXiv cs.CL `[7244de4ac5]`
• Flaws in the LLM Automation Narrative — arXiv cs.AI `[6c628e4ab0]`
• Do Transformers Actually Help Intrusion Detection? A Temporal Sequence Evaluation on CIC-IDS2017 — arXiv cs.LG `[68703a77d0]`
• What Fits (Into Few Tokens) Doesn't Overfit: Compression and Generalization in ML Research Agents — arXiv cs.AI `[388026afb1]`

Thread 3: The Post-Training Arms Race

• Introducing Gemma 4 12B: a unified, encoder-free multimodal model — Google DeepMind Blog `[9be096effc]`
• Google Releases Gemini 3.5 Live Translate — MarkTechPost `[39896267c8]`
• Fluid, natural voice translation with Gemini 3.5 Live Translate — Google DeepMind Blog `[8065dada88]`
• TRACE: A Unified Rollout Budget Allocation Framework for Efficient Agentic Reinforcement Learning — arXiv cs.AI `[c5e434bbe2]`
• A Unifying Lens on Supervised Fine-Tuning Through Target Distribution Design — arXiv cs.AI `[0f2466a869]`
• Attention Amnesia in Hybrid LLMs: When CoT Fine-Tuning Breaks Long-Range Recall, and How to Fix It — arXiv cs.CL `[766c210451]`

Cross-Story / Counter-Narrative

• It Takes One to Bias Them All: Breaking Bad with One-Shot GRPO — arXiv cs.CL `[6a3992fe11]`
• Does Reasoning Preserve Alignment? On the Trustworthiness of Large Reasoning Models — arXiv cs.CL `[409b6da10a]`
• A Unifying Lens on Supervised Fine-Tuning Through Target Distribution Design — arXiv cs.AI `[0f2466a869]`
• Attention Amnesia in Hybrid LLMs — arXiv cs.CL `[766c210451]`
• CIAware-Bench — arXiv cs.AI `[b39b1d1076]`
• Anthropic Releases Claude Fable 5 and Claude Mythos 5 — MarkTechPost `[3a21e90463]`

Quick Hits

• Introducing North Mini Code: Cohere's First Model For Developers — Hugging Face Blog `[3b631680ec]`
• Introducing Gemma 4 12B: a unified, encoder-free multimodal model — Google DeepMind Blog `[9be096effc]`
• Can Voice Agents Handle Bilingual Customers? Benchmarking Frontier ASR on Code-Switched Speech — Hugging Face Blog `[7a24901e21]`
• ReasonAlloc: Hierarchical Decoding-Time KV Cache Budget Allocation for Reasoning Models — arXiv cs.AI `[769e0bfba3]`
• Piper: A Programmable Distributed Training System — arXiv cs.AI `[138213b411]`
• Predicting Future Behaviors in Reasoning Models Enables Better Steering — arXiv cs.LG `[b371b92495]`
• ABC-Bench: An Agentic Bio-Capabilities Benchmark for Biosecurity — arXiv cs.AI `[dda8668e64]`

Three stories today — Microsoft's MAI-Transcribe-1.5, Google's agentic RAG release, and the OpenEnv coalition — share a single underlying strategic logic: the next competitive moat in AI is not model quality but ownership of the execution layers models run through. Meanwhile, a five-model economic simulation serves as a quiet corrective to the hype around autonomous multi-agent systems, and an automated prompt optimization framework signals the beginning of the end for prompt engineering as a human craft.

Thread 1: The Infrastructure Consolidation Wave

• Microsoft AI Introduces MAI-Transcribe-1.5: 2.4% WER on Artificial Analysis, Best-in-Class FLEURS Accuracy, and Up to 5x Faster Long-Audio Transcription — MarkTechPost `[79363245bd]`
• The Open Source Community is backing OpenEnv for Agentic RL — Hugging Face Blog `[fbaf22fcf6]`

Thread 2: Agentic Everything, Infrastructure Nothing

• Google Research Adds Agentic RAG to Gemini Enterprise Agent Platform with a Sufficient Context Agent for multi-hop queries — MarkTechPost `[a1d1db296c]`
• The Open Source Community is backing OpenEnv for Agentic RL — Hugging Face Blog `[fbaf22fcf6]`

Thread 3: Prompt Engineering's Quiet Death

• Building Reflective Prompt Optimization with GEPA: Multi-Component Prompts, Structured Feedback, and Held-Out Validation — MarkTechPost `[7abc803d9f]`

Thread 4: Emergence Is Harder Than It Looks

• The crash that vanished: control and emergence in a five-model economy — Hugging Face Blog `[d6de7b8b6f]`

A single community observability tool built around Claude Code raises a question that platform vendors haven't answered: when an agentic coding assistant takes dozens of actions across your filesystem, what's your audit story? Today's episode examines whether Her — a session reconstruction tool on Hugging Face — is a leading indicator of an emerging governance layer for agentic coding, or just one developer's personal itch. Either way, the design problem it's solving is real, underserved, and one the major platforms should own.

Thread 1: Her and the Observability Gap

• Her · हेर — a detective for your Claude Code sessions — Hugging Face Blog [09eb658027]

Thread 2: Why Platform Vendors Are Getting This Wrong

• Her · हेर — a detective for your Claude Code sessions — Hugging Face Blog [09eb658027]

Quick Hits

• Her · हेर — a detective for your Claude Code sessions — Hugging Face Blog [09eb658027]

Four uncoordinated releases today — Google DeepMind's Gemma 4 QAT checkpoints, NVIDIA's Nemotron 3.5 ASR, Moonshot AI's Kimi Code CLI, and the Thousand Token Wood multi-agent experiment — converge on a single infrastructure thesis: the unit of value delivery is shifting from one large hosted model call to composed systems of smaller, locally-runnable, format-reliable components. The episode argues this is real progress on narrow sub-problems, while the hard capability problems remain untouched. The efficiency narrative is too flattering; this is infrastructure plumbing, not the autonomous-agent breakthrough the coverage implies.

Thread 1: The Efficiency Offensive

• Google DeepMind Releases Gemma 4 QAT Checkpoints: Q4_0 and a New Mobile Format Cut On-Device Memory — MarkTechPost `[76c2396d66]`
• NVIDIA Releases Nemotron 3.5 ASR: A 600M-Parameter Cache-Aware Streaming Model Transcribing 40 Language-Locales in Real Time — MarkTechPost `[4877e18b30]`

Thread 2: Open-Source as Infrastructure Wedge

• Moonshot AI Releases Kimi Code CLI: A Terminal AI Coding Agent Built in TypeScript for Next-Gen Agents — MarkTechPost `[4b875fb8e2]`
• Thousand Token Wood: shipping a multi-agent economy on a 3B model — Hugging Face Blog `[c43588fe62]`

Thread 3: Small Models as Economic Actors

• Thousand Token Wood: shipping a multi-agent economy on a 3B model — Hugging Face Blog `[c43588fe62]`

Cross-Story / Counter-Narrative Sources

• Moonshot AI Releases Kimi Code CLI — MarkTechPost `[4b875fb8e2]`
• Google DeepMind Releases Gemma 4 QAT Checkpoints — MarkTechPost `[76c2396d66]`
• NVIDIA Releases Nemotron 3.5 ASR — MarkTechPost `[4877e18b30]`
• Thousand Token Wood — Hugging Face Blog `[c43588fe62]`

NVIDIA shipped a 550B open model, a Kubernetes inference optimizer, and a multimodal safety classifier in a single week — not as independent drops, but as a coordinated attempt to own every layer of enterprise AI deployment. Meanwhile, a real-world Meta agent hack and EVA-Bench's expansion to 213 evaluation scenarios confirm that the field is no longer anticipating adversarial production conditions — it is already inside them. The convergence of safety tooling, inference optimization, and security incidents is the signature of an industry reacting to deployment reality, not preparing for it.

Thread 1: NVIDIA's Platform Ambition

• NVIDIA AI Releases Nemotron 3 Ultra: An Open 550B Mixture-of-Experts Hybrid Mamba-Transformer for Long-Running Agents — MarkTechPost
• NVIDIA AI Releases Dynamo Snapshot: A CRIU-Based Fast Startup System for AI Inference on Kubernetes — MarkTechPost
• Nemotron 3.5 Content Safety: Customizable Multimodal Safety for Global Enterprise AI — Hugging Face Blog

Thread 2: Agentic Security Is Already Burning

• The Meta hack shows there's more to AI security than Mythos — MIT Technology Review
• EVA-Bench Data 2.0: 3 Domains, 121 Tools, 213 Scenarios — Hugging Face Blog

Thread 3: Hybrid Inference as the New Default

• Perplexity AI Introduces Hybrid Local-Server Inference Orchestrator for Personal Computer: Automatic On-Device and Cloud Task Routing — MarkTechPost
• NVIDIA AI Releases Dynamo Snapshot: A CRIU-Based Fast Startup System for AI Inference on Kubernetes — MarkTechPost

Thread 4: Benchmarks Chasing Deployment Reality

• EVA-Bench Data 2.0: 3 Domains, 121 Tools, 213 Scenarios — Hugging Face Blog
• The Meta hack shows there's more to AI security than Mythos — MIT Technology Review

Today's episode argues that three simultaneous open-weights releases — a multimodal laptop-scale model, a local agent framework, and an expressive TTS model — mark a genuine inflection point in self-hosted AI, not a incremental update. That shift makes OpenAI's vertical productization strategy with GPT-Rosalind more legible as urgency rather than confidence, and two alignment papers suggest the training cost collapse is accelerating the dynamic further. The counter-narrative: capability parity has never equaled enterprise adoption, and the Linux-in-2003 analogy is more predictive than the bullish substitution timeline most observers are running.

Thread 1: The Open-Weights Inflection Point

• Google DeepMind Releases Gemma 4 12B: An Encoder-Free Multimodal Model with Native Audio That Runs on a 16 GB Laptop — MarkTechPost [588bbce8f8]
• Meet OpenJarvis: A Local-First Framework for On-Device Personal AI Agents with Tools, Memory, and Learning — MarkTechPost [b3b633fd12]
• Miso Labs Releases MisoTTS: An 8B Emotive Text-to-Speech Model with Open Weights — MarkTechPost [ec4cef3d89]

Thread 2: Alignment Techniques Go General-Purpose

• Task-Seeded Synthetic Q&A Generation for Nemotron Pretraining — Hugging Face Blog [7038267192]
• Direct Preference Optimization Beyond Chatbots — Hugging Face Blog [bc0be78cdc]

Thread 3: Vertical Enclosure at the Frontier

• Introducing new capabilities to GPT-Rosalind — OpenAI News [bfb45dcad6]
• Google DeepMind Releases Gemma 4 12B — MarkTechPost [588bbce8f8]
• Meet OpenJarvis — MarkTechPost [b3b633fd12]
• Miso Labs Releases MisoTTS — MarkTechPost [ec4cef3d89]

Thread 4: Courts as Unintended AI Stress Tests

• How courts are coping with a flood of AI-generated lawsuits — MIT Technology Review [57bacf40d5]

The BCG Trust Imperative 5.0 argues that governments across ten countries have successfully built the architectural layer of AI governance — principles, risk frameworks, accountability roles — but have systematically failed at the operational layer beneath it. The result is a paradox: over-governance of low-risk productivity tools and under-governance of genuinely consequential agentic systems, with the opportunity cost of delay running into the trillions.

**Key findings and arguments:**

• Most governments studied already have AI ethics principles, named accountability roles, risk-based assessments, transparency requirements, and procurement controls in place — the architecture is not the problem
• Risk thresholds are too vague to apply consistently, causing reviewers to default to treating GenAI as high-risk across the board regardless of actual stakes
• One Australian agency identified 71 crossover points where different teams requested the same or slightly different information during a single assurance process
• Low-risk use cases are being abandoned entirely — not just delayed — because governance pathways appear too difficult to navigate
• Almost no government frameworks were designed for agentic AI, which introduces delegated authority, multi-step autonomous action, and distributed accountability across model providers, platforms, and deploying agencies
• Singapore is the sole country in the study with explicit published agentic AI governance guidance as of 2026, having iterated through model framework updates since 2019
• Vendor collaboration is generating structural friction: governments are applying legacy on-premises audit logic to cloud-delivered, continuously updated, multi-party AI stacks
• The orchestration layer — system prompts, permitted actions, exposed context — is often more determinative of risk than the underlying model, yet current frameworks don't assign clear ownership of it
• BCG estimates GenAI could unlock $1.75 trillion in annual productivity value for governments globally by 2033
• Citizens using AI at least weekly rose more than 25% between 2024 and 2026 in BCG's global survey; public expectations of government AI adoption are rising in parallel
• Nine recommended fixes include: proportionate risk triage with concrete exemplars, staged lifecycle assurance, integrated approval workflows, reusable evidence artifacts, clearer role mandates with decision rights, and metrics that track value delivered — not just forms completed
• The report was jointly funded by BCG and Salesforce; the recommended stack-based accountability model aligns closely with Salesforce's own platform architecture

Source

• Trust Imperative 5.0: Building Trust in Government Through Practical AI Assurance

Four uncoordinated releases today — NVIDIA's Cosmos 3, TinyFish's BigSet, Holo3.1, and a DPO-for-OCR fine-tuning paper — are collectively building the infrastructure layer that sits just below AI applications: structured data manufacturing and reliable action execution. The dominant narrative says unification is winning, but the deployment evidence suggests bifurcation: unified world models for robotics where latency is loose, lean specialized agents for interactive software where it is tight. The scarce resource in AI deployment is no longer model capability; it is data quality and execution reliability.

Thread 1: The Unification Bet

• NVIDIA Releases Cosmos 3: A Two-Tower Mixture-of-Transformers Foundation Model Unifying Physical Reasoning, World Generation, and Action Generation — MarkTechPost `[ec7389a75a]`
• TinyFish Launches BigSet: An Open-Source Multi-Agent System That Builds Structured Live Datasets from Plain-English Descriptions — MarkTechPost `[ea541a546f]`
• Holo3.1: Fast & Local Computer Use Agents — Hugging Face Blog `[440b7646f5]`

Thread 2: Alignment Techniques Escaping Their Origin

• Direct Preference Optimization Beyond Chatbots — Hugging Face Blog `[bc0be78cdc]`

Thread 3: Infra Before Product

• NVIDIA Releases Cosmos 3: A Two-Tower Mixture-of-Transformers Foundation Model Unifying Physical Reasoning, World Generation, and Action Generation — MarkTechPost `[ec7389a75a]`
• TinyFish Launches BigSet: An Open-Source Multi-Agent System That Builds Structured Live Datasets from Plain-English Descriptions — MarkTechPost `[ea541a546f]`
• Holo3.1: Fast & Local Computer Use Agents — Hugging Face Blog `[440b7646f5]`
• Direct Preference Optimization Beyond Chatbots — Hugging Face Blog `[bc0be78cdc]`

Four independent teams — Travelers with OpenAI, TinyFish, Hugging Face, and Alibaba — shipped production-grade agentic architecture on the same day without referencing each other. That convergence is the signal: agentic orchestration is no longer experimental, it is the default production bet. OpenAI's enterprise moat today is relationship depth, not model superiority, and that erodes faster than the revenue numbers suggest.

Thread 1: Agentic AI Crosses the Deployment Threshold

• Travelers deploys AI-powered claims countrywide with OpenAI — OpenAI News
• TinyFish Launches BigSet: An Open-Source Multi-Agent System That Builds Structured Live Datasets from Plain-English Descriptions — MarkTechPost
• Holo3.1: Fast & Local Computer Use Agents — Hugging Face Blog
• Rehumanizing global health care with agentic AI — MIT Technology Review

Thread 2: The Componentization Bet

• JetBrains Releases Mellum2: A 12B MoE Model for Fast, Specialized Tasks in Multi-Model AI Pipelines — MarkTechPost
• Holo3.1: Fast & Local Computer Use Agents — Hugging Face Blog
• TinyFish Launches BigSet: An Open-Source Multi-Agent System That Builds Structured Live Datasets from Plain-English Descriptions — MarkTechPost
• Travelers deploys AI-powered claims countrywide with OpenAI — OpenAI News

Thread 3: Qwen's Quiet Western Flank

• Alibaba's Qwen Team Launches Qwen3.7-Plus, Adding Vision, Deep Reasoning, Tool Invocation, and Autonomous Iteration on the Bailian Platform — MarkTechPost

Thread 4: Grassroots Compute Efficiency as a Signal

• How to Speed Up Transformer Training Using NVIDIA Apex (FusedAdam, FusedLayerNorm) and Native torch.amp — MarkTechPost

## Episode Summary

Today's three headline stories — JetBrains' Mellum2 release, NVIDIA's Cosmos 3, and OpenAI's Michigan data center groundbreaking — share a common strategic skeleton: each is a lock-in play executed at a different layer of the AI stack. The open-versus-closed model debate is increasingly a distraction; the real competition is over which platforms own the deployment layer that models run on top of. A counter-narrative worth tracking: the efficiency and focus of today's specialized open-weight releases may quietly undermine the premise behind gigawatt-scale centralized compute.

---

## Thread 1: The Specialization Wedge

• Introducing Mellum2: A 12B Mixture-of-Experts Model by JetBrains — Hugging Face Blog `[2995ca1c36]`
• Welcome NVIDIA Cosmos 3: The First Open Omni-model for Physical AI Reasoning and Action — Hugging Face Blog `[7c8c63c46b]`

## Thread 2: Compute as Territorial Claim

• Building the infrastructure for the Intelligence Age in Michigan — OpenAI News `[dadd123def]`

## Thread 3: Open vs. Closed, Reframed

• Introducing Mellum2: A 12B Mixture-of-Experts Model by JetBrains — Hugging Face Blog `[2995ca1c36]`
• Welcome NVIDIA Cosmos 3: The First Open Omni-model for Physical AI Reasoning and Action — Hugging Face Blog `[7c8c63c46b]`
• Building the infrastructure for the Intelligence Age in Michigan — OpenAI News `[dadd123def]`

## Cross-Story / Counter-Narrative

• Introducing Mellum2: A 12B Mixture-of-Experts Model by JetBrains — Hugging Face Blog `[2995ca1c36]`
• Welcome NVIDIA Cosmos 3: The First Open Omni-model for Physical AI Reasoning and Action — Hugging Face Blog `[7c8c63c46b]`
• Building the infrastructure for the Intelligence Age in Michigan — OpenAI News `[dadd123def]`

## Quick Hits

• Introducing Mellum2: A 12B Mixture-of-Experts Model by JetBrains — Hugging Face Blog `[2995ca1c36]`
• Welcome NVIDIA Cosmos 3: The First Open Omni-model for Physical AI Reasoning and Action — Hugging Face Blog `[7c8c63c46b]`
• Building the infrastructure for the Intelligence Age in Michigan — OpenAI News `[dadd123def]`