AI Daily — 2026-03-13

English 中文

Meta Scales DeepSeek to 5T Params, Faces Criticism · Massive AI breakthrough expected early 2026,...

Covering 42 AI news items

🔥 Top Stories

1. Meta Scales DeepSeek to 5T Params, Faces Criticism

An influential Meta-focused post argues Meta should have followed the DeepSeek recipe, scaling to 5T parameters, training with vast compute and data, then distilling into smaller models. It describes RL training to exhaustion and postulates profit through distillation into 30B, 100B, and 500B variants, while criticizing Meta’s Avocado model as delayed and subpar. Source-x

2. Massive AI breakthrough expected early 2026, says Morgan Stanley

Morgan Stanley warns that a massive AI capability jump, driven by unprecedented compute scaling at U.S. labs, could arrive in early 2026. The development could deliver rapid productivity gains while causing job disruption and power shortages as intelligence becomes a key economic resource. Source-x

3. Anthropic frees 1M-token context in Claude 4.6

Anthropic announced that longer context windows no longer incur extra charges, enabling models to handle up to a 1,000,000-token context. The 1M-context feature is now generally available in Claude Opus 4.6 and Claude Sonnet 4.6, expanding long-form processing without additional cost. Source-x

📰 Featured

Open Source & Code

RTX-5090 Hits ~2000 TPS with QWEN 3.5-27B — A benchmark reports around 2000 transactions per second on an RTX-5090 with QWEN 3.5-27B, noting performance is highly workload-dependent and non-cached tests can vary. Source-reddit
14B Model Beats Claude Opus 4.6 on Ada Code Gen — Fine-tuning Qwen2.5-Coder-14B-Instruct via QLoRA yields Ada-code-generation results that reportedly outperform Claude Opus 4.6 on Ada tasks, marking progress in open-model code tasks. Source-reddit

Industry & Hardware

Joining SpaceX and xAI to build superintelligence with Elon Musk — A developer announces joining SpaceX and xAI to work closely with Elon Musk on frontier-scale AI, emphasizing hardware-digital intelligence synergy and a high-agency culture. Source-x
AI Compute Bottlenecks: Logic, Memory, and Power — A deep-dive identifies the triad of bottlenecks to scaling AI compute and discusses economic and supply-chain factors across labs, hyperscalers, foundries, and fab equipment. Source-x
Lemonade v10 Adds Linux NPU Support and Multimodal Capabilities — Lemonade v10 expands to Linux platforms, adds NPU support, and bundles multimodal tools under a unified base URL with cross-platform portability and a local AI app ecosystem. Source-reddit
Perplexity Computer Goes Mobile with Cross-Device Sync — Perplexity now supports mobile use with cross-device synchronization; iOS is live with Android coming soon, enabling seamless task management across devices. Source-x

Tools & Platforms

Launch Claude Code sessions on laptop from your phone — Claude Code now allows spawning a local laptop session from the mobile app via Remote Control, with Team/Enterprise (version >= 2.1.74) support and aiming to speed up startups. Source-x
Claude Code: Local laptop sessions, mobile-first workflow enhancements — The update also speeds session startups and hints at upcoming GitHub integration for mobile orchestration. Source-x
Anthropic frees 1M-token context in Claude 4.6 — Context-window expansion without extra cost enhances long-form tasks on Claude Opus 4.6 and Sonnet 4.6. Source-x
Grok Imagine Turns 7 Images Into a Video — Grok Imagine adds capabilities to convert a set of images into video, broadening multimodal storytelling options. Source-x

⚡ Quick Bites

First cloud to deploy NVIDIA Vera Rubin NVL72 for validation — Cloud deployment initiates validation of the Vera Rubin NVL72 for AI workloads. Source-x
Spatial-TTT Advances Streaming Visual Spatial Intelligence with Test-Time Training — Explores test-time adaptation for streaming visual-spatial AI tasks. Source-huggingface
MADQA Benchmark Tests Strategic Reasoning in Multimodal Document Agents — MADQA evaluates strategic reasoning in multimodal doc agents. Source-huggingface
IndexCache accelerates sparse attention via cross-layer index reuse — IndexCache optimizes sparse attention workloads. Source-huggingface
Video-Based Reward Modeling for Computer-Use Agents — Uses video signals for reward modeling in agent training. Source-huggingface
DreamVideo-Omni Enables Omni-Motion Multi-Subject Video Customization — Multi-subject video customization via DreamVideo-Omni. Source-huggingface
Context Gateway Compresses Agent Context Before LLM Inference — Reduces context size to speed up LLM inference. Source-github
Spine Swarm: AI Agents Collaborate on a Visual Canvas — AI agents collaborate on a shared visual canvas. Source-rss
Can I Run AI Locally? Guide to Local Inference — Practical guide to local AI inference. Source-rss
OpenRAG Launches AI-Powered RAG Platform for Documents — OpenRAG delivers document-focused retrieval-augmented generation. Source-github
Microsoft BitNet: 1-bit LLM Inference Framework Boosts CPU Speed — BitNet enables faster CPU-based LLM inference. Source-github
Tennessee grandmother jailed after AI face recognition error links her to fraud — Report on misidentification by AI facial recognition. Source-rss
Innocent Woman Jailed After AI Facial Recognition Misidentification — Misidentification case details. Source-rss
CLI Is the Essential Interface for AI Agents (Part 2) — Debates CLI as the primary interface for AI agents. Source-reddit
Claude’s Interactive Chart UI Praised for Usability — Claude gains a user-friendly interactive chart UI. Source-x
Prompt caching auto-injects Anthropic breakpoints, saves 90% of tokens — Prompt caching reduces token usage substantially. Source-rss
Amazon Employees Say AI Increases Workload, Study Confirms — A study corroborates increased workload due to AI adoption. Source-rss
OneCLI: Vault for AI Agents in Rust — Rust-based tool for AI agent vault management. Source-github
Claude now creates interactive charts, diagrams and visualizations — Claude adds visual generation features. Source-rss
Atlassian CEO: AI won’t replace people, yet layoffs continue — CEO commentary on AI adoption and ongoing layoffs. Source-rss
Blind user seeks local LLMs to rival Claude Code and Codex — Accessibility-focused demand for local LLMs. Source-reddit
Why can’t we have small SOTA-like coding models? — Community discussion on small SOTA coding models. Source-reddit
Non-Chinese LLMs Currently Relevant, Reddit Discussion — Reddit discussion on non-Chinese LLM relevance. Source-reddit
Turn 10,000 API Endpoints Into One CLI Tool via OpenAPI — OpenAPI-based approach to unify 10k endpoints into a single CLI. Source-reddit
Chipotle’s Free Support Bot; Claude Code Costs — Costs and performance considerations for Claude-based support bots. Source-x
Grok Imagine Turns 7 Images Into a Video — Grok Imagine converts seven images into video content. Source-x

Generated by AI News Agent | 2026-03-13