🤖 AI HOT Daily · Aug 20, 2026
Liquid AI ships LFM2.5 QAD quantized checkpoints recovering 97% precision loss; GLM-5.3 launches, scoring 60 on the AA Intelligence Index to tie the top open-source model at lower cost; FastMetal generates a video locally on Mac in 30 seconds; Google Search adds 5 AI study tools; Replit launches Free Mode powered by GPT-5.6 Luna; OpenRouter announces it is joining Stripe; LMSYS approaches B300 performance serving DeepSeek-V4-Pro on H20; Apple analyzes LLM human-like behaviors; GitHub Copilot app manages work with the My work panel.
1. Liquid AI Releases LFM2.5 QAD Quantized Checkpoints
Four Q4_0 GGUF checkpoints trained with quantization-aware distillation recover 97% of BF16 average precision loss at native memory and speed.
2. GLM-5.3 Debuts: 60 on AA Index, Top Open-Source Tie
Strong at complex coding, defensive security, and long-horizon tasks, it scores 60 on the AA index — on par with closed flagships, tied with Kimi K3 for best open model at the lowest per-task cost.
3. FastMetal Generates Video on Mac in 30 Seconds
A 5-second 480P clip is generated entirely on a Mac — no CUDA, no cloud, just 3.9 GiB RAM, with DiT, DMD sampler, and decoder running on Metal via MLX.
4. Google Search Adds 5 AI Study Tools
AI Mode's generative UI goes live globally in English with interactive visuals and custom simulations, and practice quizzes are free in AI Overviews and AI Mode, covering ACT, SAT, and more.
5. Replit Launches Free Mode Powered by GPT-5.6 Luna
Lets users turn ideas into working software without token worries, offering quick answers, suggestions, and project analysis, switching to GPT-5.6 tier for deeper reasoning.
6. Claude Code v2.1.236: New Default Model Env Var
The new ANTHROPIC_DEFAULT_MODEL env var sets the default model for new sessions, while /model selections still override and persist across restarts.
7. OpenRouter Announces It Is Joining Stripe
Processing 10+ trillion tokens daily across 400+ models for 10M+ developers and companies, OpenRouter will keep operating independently under its original name and mission.
8. Pushing DeepSeek-V4-Pro Limits: H20 Optimization
LMSYS reaches 271 output tokens/s serving the 1.6T-param MoE model on a single H20 node, narrowing the gap to B300's 383.7 tokens/s to ~1.4x.
9. Apple Research: Multidimensional Look at LLM Human-Like Behaviors
Examining prevalence, impact, and controllability of behaviors like expressing feelings, building rapport, and setting boundaries — using LLM-as-a-judge plus human eval over 21,000 samples.
10. GitHub Copilot app: Manage Work with My Work Panel
The My work panel centralizes pull requests and issues with All, Active, Review requests, and Done views, custom views, and batch-spawned agent sessions.
11. How Slack Builds Human-Agent Teams
Slack's CPO advocates public-by-default channels so agents learn from visible conversation, connecting meetings, email, and calendar to cut duplication, with Claude agents handling drafting, summarizing, and monitoring.
12. Databricks: Designing Effective Genie Agents from a Single Prompt
Where generic agents often grab only the first relevant table, Genie Agents use more precise prompting to locate and answer user questions more accurately.
Sources and verification
This legacy briefing did not retain its original source URLs. Verify safety, financial, policy, and breaking-news claims with authoritative primary sources before acting on them.