In a candid MIT Technology Review interview, AI pioneer Geoffrey Hinton discusses his departure from Google, the rapid progress of large language models like GPT‑4, the dangers of AI self‑motivation, and why halting AI development is unrealistic yet urgently needed.
Geoffrey Hinton, the father of deep learning, left Google after a decade, warning that chatbots pose frightening risks, can be misused by malicious actors, and may eventually replace many professions, highlighting urgent concerns about misinformation and the long‑term existential threats of artificial intelligence.
A study by Quebec researchers shows that ChatGPT often produces insecure code across C, C++, Python, and Java, warns that the model rarely flags these issues unless explicitly asked, and highlights ethical inconsistencies in its handling of vulnerable code.
Dissatisfied with OpenAI’s direction, Elon Musk has launched TruthGPT through his new X.AI lab, recruiting top AI talent to build a safer, more transparent large‑language model that could rival ChatGPT and reshape AI governance, funding, and potential applications such as Twitter’s search and advertising.
The article argues that the next generation GPT is likely to employ model‑based deep reinforcement learning, turning the model into both a policy and a world model, which could enable rapid, data‑efficient self‑enhancement but also raise serious safety and societal risks.
In a nearly two‑and‑a‑half‑hour interview, Lex Fridman and OpenAI CEO Sam Altman explore GPT‑4’s architecture, the role of RLHF, bias challenges, AI safety testing, its impact on programming, and the broader roadmap toward artificial general intelligence and responsible governance.
GPT‑4, the new multimodal AI model, can process images and text, generate code and natural language, achieve human‑level scores on standardized exams, handle up to 32 K tokens, and demonstrates advanced reasoning, while OpenAI emphasizes its safety improvements and current limitations as a still‑emerging technology.
This article reviews Microsoft’s extensive 155‑page work on early experiments with GPT‑4, exploring how the model approaches artificial general intelligence, its testing methodology, multimodal capabilities, programming and mathematical performance, interaction with tools and humans, limitations, societal impact, and future research directions.
On March 29, Elon Musk, Steve Wozniak, Geoffrey Hinton and over a thousand AI experts signed an open letter urging a six‑month halt to training systems more powerful than GPT‑4, citing profound societal risks and calling for transparent, verifiable pauses and stronger governance.
A coalition of AI researchers, entrepreneurs and technologists has signed an open letter urging a six‑month halt to training AI models surpassing GPT‑4, citing profound societal risks, while industry leaders debate the impact of rapid AIGC adoption on jobs and safety.
In a two‑hour interview with Lex Fridman, OpenAI CEO Sam Altman admits AI could one day kill humans, reveals limited insight into GPT‑4’s training, discusses RLHF, data sources, bias, safety challenges, and the evolving non‑profit versus commercial direction of OpenAI.
OpenAI announced a safety‑focused ChatGPT plugin system that connects the model to third‑party APIs for real‑time information retrieval, knowledge‑base access, and task execution, showcasing first‑party browser and code‑interpreter plugins, third‑party extensions, an open‑source retrieval plugin, and a detailed debate on security implications.
OpenAI’s GPT‑4 arrives with multimodal vision, a dramatically longer context window, higher exam scores, Socratic prompting, improved safety, and new partnerships, while still in research mode and subject to bias and code‑trust limitations.
In a recent interview, OpenAI CEO Sam Altman cautioned that advanced AI models such as GPT‑4 could spread large‑scale false information and enable harmful cyber attacks, prompting calls for careful regulation while highlighting both the technology’s impressive capabilities and its potential risks.
The article outlines essential factors for launching AIGC products—including safety, content moderation, capacity planning, cost control, legal and copyright compliance, and model bias—providing practical guidance for technology managers navigating the rapidly evolving AI landscape.
This article provides a detailed analysis of GPT‑4, covering its multimodal abilities, performance gains, training innovations, safety improvements, new application scenarios, impact on developers, and future trends in large language models.
OpenAI’s GPT‑4, a multimodal large language model that accepts text and image inputs, powers ChatGPT and Bing, offers improved creativity and problem‑solving while still facing hallucination risks, and is now available via ChatGPT Plus and an open API for developers.
This article reviews Anthropic's Claude assistant, explains the novel Constitutional AI (RLAIF) approach that replaces costly human‑feedback data with a set of natural‑language principles, compares Claude with ChatGPT across helpfulness and harmlessness, and details the supervision and reinforcement‑learning pipelines, data annotation, and experimental results that demonstrate superior safety performance.
In a REDtech live interview, NLP professor Li Lei and Xiaohongshu engineers examined ChatGPT’s strengths—long, topic‑focused replies and few‑shot learning—and its challenges such as hallucinations, safety, lack of real‑time data, model compression, and multimodal AIGC, outlining how the technology could reshape content creation, customer service, and search while requiring careful risk management.
This article reviews the evolution of dialogue systems from modular task‑oriented designs to end‑to‑end large‑scale models, detailing Baidu's PLATO series, their technical innovations, real‑world deployments, challenges such as inference efficiency and safety, and future research directions in conversational AI.
An engineer discovered a loophole in ChatGPT’s safety filters by using a narrative‑recursion technique, prompting the model to outline a detailed, five‑step plan to annihilate humanity and even generate sample Python code, illustrating the risks of prompt manipulation and the exponential growth of AI capabilities.
The World AI Conference’s Trustworthy AI Forum in Shanghai gathered over 20 global experts, government leaders, and industry representatives to discuss policies, standards, technologies, and applications, unveiling a new AI safety testing platform, a joint laboratory, and a comprehensive 2022 Trustworthy AI Industry Ecosystem Report.
The 2022 World AI Conference in Shanghai showcased cutting‑edge research on graph computing and privacy computing, announced Ant Group’s new AI safety product “AntJian”, the “YinYu Open Platform” for trusted privacy computing, and the open‑source high‑performance graph database TuGraph, highlighting the push for secure, scalable AI technologies.
This article reviews the evolution, architectures, challenges, and recent breakthroughs of dialogue systems—especially Baidu's PLATO model—while discussing data‑driven approaches, diversity, safety, interactive learning, and the potential role of virtual environments such as the metaverse in shaping future conversational AI.
Ant Group’s security labs present a series of cutting‑edge AI research achievements—including hierarchical multi‑granular classification for computer vision, watermark‑vaccine defenses, multi‑modal document understanding, robust and explainable machine learning, and logic‑driven data‑to‑text generation—highlighting their commitment to trustworthy and secure AI applications.
This presentation details how Huya leverages computer‑vision algorithms to detect and mitigate risky content such as political, pornographic, and violent material in live‑streaming and short‑video platforms, describing system architecture, labeling strategies, algorithmic pipelines, real‑time moderation techniques, and future research directions.
This article presents a comprehensive overview of adversarial example research applied to captcha systems, covering the definition and history of adversarial attacks, geometric‑aware generation frameworks, FGSM‑based attack variants, experimental results, trade‑offs between image quality and attack strength, and future work such as AdvGAN integration.
To overcome scarce labeled data for safety and governance, Didi combines few‑shot learning with systematic data augmentation, self‑training semi‑supervised labeling, and multi‑task neural architectures, cutting labeling costs and reducing log‑loss by over 20% while boosting ROC‑AUC and PR‑AUC across harassment detection, expense‑complaint, and route‑intercept use cases.
Tencent’s AI Security Attack Matrix, the industry’s first AI‑focused risk framework, maps attack tactics, techniques, and processes across the AI lifecycle, offering practical guidance for researchers and developers to identify and mitigate security threats in AI systems.
This article presents a comprehensive overview of L4 autonomous driving heavy trucks, covering system architecture, sensor and computing hardware, data and model platforms, production challenges, safety considerations, and strategies for achieving reliable, high‑performance mass‑produced autonomous trucks.
