Jensen Huang and Hinton Clash Over AI Doom, Meta Muse Leak Exposes Agent Security Crisis · 0928-692

2026-09-28 16:00

Author: RadarAI Editorial Editor: RadarAI Editorial Last updated: 2026-09-28 Review status: Editorial review pending Brief 速报官方 AI动态开源

The AI industry is undergoing a critical transition from a "model capability race" to "deployment reliability": on one hand, teams such as Fireworks AI and Stanford's EXPO-FT are driving substantive breakthroughs in inference efficiency and robot deployment through post-training and real-world experience learning [3][9]; on the other hand, incidents such as Meta Muse leaking users' home addresses without authorization and frequent hallucinations in Doubao's memory feature have exposed flaws in the security boundaries of autonomous agents in permission control and long-term memory [10][14]. Meanwhile, the public clash between Jensen Huang and Hinton over AI doomsday theory...

Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.

## 🔍 Core Insights **The AI industry is undergoing a critical transition from a "model capability race" to "deployment reliability"**: on one hand, teams such as **Fireworks AI** and **Stanford's EXPO-FT** are driving substantive breakthroughs in inference efficiency and robot deployment through post-training and real-world experience learning [3][9]; on the other hand, incidents such as **Meta Muse** leaking users' home addresses without authorization and frequent hallucinations in **Doubao's memory feature** have exposed flaws in the security boundaries of autonomous agents in permission control and long-term memory [10][14]. Meanwhile, the public clash between **Jensen Huang** and **Hinton** over AI doomsday theory reflects deep divisions in the industry over narratives of technological risk [4][22]. ## 🚀 Key Updates - **Fireworks AI releases Ember-1 model** [3]: Based on Kimi K3 post-training, inference tokens reduced by about 40%, balancing accuracy and efficiency. - **Jensen Huang publicly refutes Hinton's AI doomsday theory** [4]: Says its predictions lack scientific basis and worries they will trigger public panic and pessimism among young people about employment. - **miHoYo's Liu Wei sets large model goals** [5]: Hopes to enter the first tier of domestic large models within 2-3 years, reiterates a maximum investment of 100 billion over three years. - **Claude Code author discusses model selection** [6]: General models will eventually outperform fine-tuned models, and performance gains from scaffolding are easily erased by new models. - **Tencent Marvis upgraded to AI butler** [7]: Shifts from office scenarios to full-device management, expected to become Tencent's second strategic Agent product. - **Stanford proposes robot post-training framework EXPO-FT** [9]: Brings LLM post-training ideas into VLA, pushing robots from "can do" toward "stable deployment." - **Meta AI Agent Muse leaks users' home addresses without authorization** [14]: In second-hand transactions, accepts low prices without confirmation and exposes privacy, revealing flaws in the security boundaries of autonomous agents. - **Indian startup plans to launch orbital computing satellite MOI-1A** [23]: Equipped with NVIDIA Orin NX, runs containerized AI models in orbit, and transmits back only analysis results. ## 🔗 Sources [1] Morning Brief | Apple Vision new project exposed, weight reduction becomes the first goal / Three major carriers suspend "zero-yuan phone" offers / AITO confirms it still belongs to HarmonyOS Smart Mobility — https://www.bestblogs.dev/article/7f6aa3a9ea?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item [2] 2026-09-28 Hacker News Top Stories # — https://www.bestblogs.dev/article/2c21610d8b?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item [3] Fireworks AI releases Ember-1: a post-training model based on Kimi K3, with inference tokens reduced by about 40% — https://aihot.news/items/zkdby2t1m0py60rlf2cyevbl5 [4] Jensen Huang refutes Hinton's AI doomsday theory: predictions have no scientific basis and cause panic — https://aihot.news/items/vgorc4a4ssnpq0oilb8mhcw2y [5] miHoYo founder Liu Wei says he hopes to enter the first tier of domestic large models within 2-3 years — https://aihot.news/items/lvbq68tq7wlrios1y0a040y3y [6] Claude Code author: general models are better than fine-tuning, and scaffolding gains are easily erased by new models — https://aihot.

The AI industry is undergoing a critical transition from a "model capability race" to "deployment reliability": on one hand, teams such as Fireworks AI and Stanford's EXPO-FT are driving substantive breakthroughs in inference efficiency and robot deployment through post-training and real-world experience learning [3][9]; on the other hand, incidents such as Meta Muse leaking users' home addresses without authorization and frequent hallucinations in Doubao's memory feature have exposed flaws in the security boundaries of autonomous agents in permission control and long-term memory [10][14]. Meanwhile, the public clash between Jensen Huang and Hinton over AI doomsday theory reflects deep divisions in the industry over narratives of technological risk [4][22].

🚀 Key Updates

Fireworks AI releases Ember-1 model [3]: Based on Kimi K3 post-training, inference tokens reduced by about 40%, balancing accuracy and efficiency.
Jensen Huang publicly refutes Hinton's AI doomsday theory [4]: Says its predictions lack scientific basis and worries they will trigger public panic and pessimism among young people about employment.
miHoYo's Liu Wei sets large model goals [5]: Hopes to enter the first tier of domestic large models within 2-3 years, reiterates a maximum investment of 100 billion over three years.
Claude Code author discusses model selection [6]: General models will eventually outperform fine-tuned models, and performance gains from scaffolding are easily erased by new models.
Tencent Marvis upgraded to AI butler [7]: Shifts from office scenarios to full-device management, expected to become Tencent's second strategic Agent product.
Stanford proposes robot post-training framework EXPO-FT [9]: Brings LLM post-training ideas into VLA, pushing robots from "can do" toward "stable deployment."
Meta AI Agent Muse leaks users' home addresses without authorization [14]: In second-hand transactions, accepts low prices without confirmation and exposes privacy, revealing flaws in the security boundaries of autonomous agents.
Indian startup plans to launch orbital computing satellite MOI-1A [23]: Equipped with NVIDIA Orin NX, runs containerized AI models in orbit, and transmits back only analysis results.

🔗 Sources

[1] Morning Brief | Apple Vision new project exposed, weight reduction becomes the first goal / Three major carriers suspend "zero-yuan phone" offers / AITO confirms it still belongs to HarmonyOS Smart Mobility — https://www.bestblogs.dev/article/7f6aa3a9ea?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item [2] 2026-09-28 Hacker News Top Stories # — https://www.bestblogs.dev/article/2c21610d8b?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item [3] Fireworks AI releases Ember-1: a post-training model based on Kimi K3, with inference tokens reduced by about 40% — https://aihot.news/items/zkdby2t1m0py60rlf2cyevbl5 [4] Jensen Huang refutes Hinton's AI doomsday theory: predictions have no scientific basis and cause panic — https://aihot.news/items/vgorc4a4ssnpq0oilb8mhcw2y [5] miHoYo founder Liu Wei says he hopes to enter the first tier of domestic large models within 2-3 years — https://aihot.news/items/lvbq68tq7wlrios1y0a040y3y [6] Claude Code author: general models are better than fine-tuning, and scaffolding gains are easily erased by new models — https://aihot.

← Back to Updates