Split-Brain Discovery Shatters AI Chain-of-Thought Safety, Step-Star Drops Step-5-Preview Model | AI Daily Brief
Audio in Mandarin Chinese · English transcript below
⚡ Chain-of-thought safety roots shattered by split-brain discovery; Step-5-Preview's 600B params storm in as top chatbots fail accuracy tests and Nigeria AI terror attack forces industry reckoning between capability surge and safety collapse.
Foundation models are achieving simultaneous breakthroughs in parameter scale, context length, and modality coverage, with technological competition now plunging into the deep waters of ultra-long contexts and full-modality interaction. Yet it is worth noting that while StepFun and TypeSafe AI are pushing capability boundaries with relentless urgency, the emergence of AI-enhanced terrorism in Nigeria and the collective inaccuracy of five mainstream chatbots are thrusting the sharp contradiction between technological sprint and safety governance into the foreground.
Today's Top 3 Headlines
- AI Industry News
🤖 AI Safety "Chain-of-Thought" Defense Breached: "Split-Brain" Exposes Inference Vulnerability
New research reveals Chain-of-Thought reasoning in AI safety suffers from "split-brain" vulnerabilities, letting model safeguards be bypassed. For AI product teams, CoT explanations alone are now insufficient—more robust alignment and monitoring layers are urgently needed.
Source ↗ - AI Industry News
🤖 Step-5-Preview: 600B/27B Params, 1M Context from Jieyue Xingchen
Step-5-Preview from Jieyue Xingchen sparks Reddit buzz with 600B/27B params and 1M context window. Devs get long-text & vision capabilities, but non-commercial license blocks enterprise deployment.
Source ↗ - AI Industry News
🤖 TypeSafe AI ships Jev 1.14: images, Box types, 262k context
TypeSafe AI shipped Jev 1.14 with image processing, Box types, and a 262k context window. Developers can now chunk long docs less and handle mixed text-image tasks directly, expanding the cost-performance tradeoff in model selection.
Source ↗
+4 more headlines
- 🤖 5 Major AI Chatbots Flunk Accuracy Test: DeepSeek, ChatGPT, Gemini, Claude, Grok All Fail
- 🤖 AI-Enhanced Terrorism Emerges: Boko Haram Uses ChatGPT, Claude, DeepSeek, Gemini
- 🤖 Apple Market Cap Hits $5T for First Time; OpenAI Accuses Ex-Employee of Leaking Secrets to Join Apple
- 🤖 Figma launches AI Agent workflows, opens code layer public beta
Podcast highlight
The main course was that sucker punch, and the sides—rankings and spec sheets—both gotta be taken with a grain of salt.