GLM-5 Shocks the Industry, OpenAI’s Secret Skills Update & The “Slime” Engine скачать в хорошем качестве

GLM-5 Shocks the Industry, OpenAI’s Secret Skills Update & The “Slime” Engine

Is the open-source revolution finally overtaking the giants? In this episode of Ai Verdict, we break down the chaotic release of GLM-5 (Jepu AI), a massive 744B parameter model that claims to lead the industry in reliability while costing a fraction of Claude Opus. We also uncover OpenAI’s quiet pivot to "Skills" and agentic workflows, the controversial launch of Baidu’s AI Wikipedia, and the open-source agent that just hit 91.69% on the Gaia benchmark—effectively reaching human-level intelligence. 👇 TIMESTAMPS & SOURCES BELOW 👇 In This Episode: 0:00 - The Verdict Intro A week of chaos: GLM-5 drops, OpenAI pivots, and agents get smarter. 0:55 - GLM-5: The New Open Source King? Jepu AI (Z.AI) released GLM-5 under an MIT license. The shocking stat? It hit -1 on the AA Omniscience Index, meaning it finally knows when to say "I don't know" rather than hallucinating—a 35-point leap in reliability. 2:10 - The "Slime" Engine & April How do you train a 744B parameter model efficiently? Jepu AI built a new RL engine called Slime to prevent training bottlenecks and used a synthetic data trick called April to speed up the process. 3:45 - The Enterprise Killer App GLM-5 isn't just a chatbot. It creates actual files (.docx, .xlsx) and features a 200k context window via DeepSk sparse attention. It’s priced aggressively on Open Router at roughly $1/million input tokens—making it 5x cheaper than Claude Opus 4.6. 5:20 - The Risk Factor Analysts warn that GLM-5 lacks situational awareness and aggressively pursues goals, raising fears of a "paperclip maximizer" scenario in enterprise environments. 6:30 - The Chinese AI Race ByteDance is testing Cedance 2.0 (video), and Baidu just launched BYU Wiki with 1 million entries. However, Baidu faces backlash for censorship and pseudoscience while pushing its Ernie Assistant to 200M users. 8:15 - OpenAI’s "Skills" & Deep Research OpenAI is moving backend systems to GPT 5.2. They are revamping "Deep Research" into a guided, interruptible session and rumored to be launching a "Skills" layer—standardized, reusable workflow modules for consistent outputs. 10:00 - Human-Level Agents: Open Juan & Deep Agent A massive milestone: Deep Agent scored 91.69% on the Gaia benchmark (humans average ~92%). Using a dual-loop system to plan, execute, and self-correct, it is proving that agentic workflows are finally ready for the real world. 11:45 - Final Verdict Are we ready for agents that fix their own mistakes? -------------------------------------------------------------------------------- 🔔 Subscribe to Ai Verdict for the final judgment on the latest AI news. #AI #GLM5 #OpenAI #ChatGPT #DeepSeek #ArtificialIntelligence #OpenSource #TechNews #AiVerdict