Top News
OpenAI insiders are demanding a “right to warn†the publicÂ
Thirteen current and former employees of OpenAI and Google DeepMind have published a proposal demanding the right to warn the public about the potential dangers of advanced artificial intelligence (AI). The signatories, including AI pioneers Geoffrey Hinton, Yoshua Bengio, and Stuart Russell, are calling for AI companies to allow them to voice their concerns about the technology without fear of retaliation. They are not seeking to divulge intellectual property or trade secrets, but rather to raise concerns about potential risks. The proposal also calls for the abolition of nondisparagement agreements that prevent insiders from voicing risk-related concerns. The signatories argue that the current whistleblower protections are insufficient as they focus on illegal activity, whereas many of the risks they are concerned about are not yet regulated.
OpenAI Insiders Warn of a ‘Reckless’ Race for Dominance
OpenAI Offers a Peek Inside the Guts of ChatGPT
A new research paper from OpenAI, authored by the disbanded “superalignment” team including ex-employees Ilya Sutskever and Jan Leike, introduces a method to identify how the AI stores concepts that might cause misbehavior. Similar to Anthropic’s recent work, this technique uses an additional model to make the internal workings of GPT-4 more transparent, identifying patterns representing specific concepts like profanity. The paper, accompanied by released code and visualization tools, aims to mitigate unwanted AI behaviors and improve model control.
Extracting Concepts from GPT-4
Chinese short video app Kuaishou unveils Sora-style product amid AI race
Kuaishou, a Chinese short-video app, has launched a text-to-video service similar to OpenAI’s Sora, as part of the race among Chinese Big Tech firms to catch up with US counterparts in AI applications.
Other News
Tools
Google’s updated AI-powered NotebookLM expands to India, UK and over 200 other countries – Google expands AI-powered NotebookLM to over 200 countries, adding new features and language support, allowing users to generate summaries and ask questions based on their documents, with plans for future integrations and a focus on privacy.
‘Apple Intelligence’ will automatically choose between on-device and cloud-powered AI – Apple is set to unveil a new AI system called Apple Intelligence, which will offer AI features across its platforms and apps, including a chatbot powered by OpenAI, and will use a unique algorithm to determine whether tasks can be processed on-device or in the cloud.
ChatGPT voice just got a big privacy upgrade for background conversations – ChatGPT has introduced a new Background Conversations setting for voice chats on iPhone and Android, allowing users to continue conversations while using other apps or when the screen is off, with the option to enable or disable the privacy feature.
Asana introduces ‘AI teammates’ designed to work alongside human employees – Asana introduces ‘AI teammates’ to work alongside human employees, aiming to create transparency and structure around AI to help move work inside an organization, while emphasizing the importance of human oversight.
Hugging Face and Pollen Robotics show off first project: an open source robot that does chores – Hugging Face and Pollen Robotics have collaborated to create an open-source humanoid robot, Reachy2, which can autonomously perform household chores and safely interact with humans and animals, using a novel training process and open-source dataset.
Stability AI debuts new Stable Audio Open for sound design – Stability AI has released Stable Audio Open 1.0, a specialized model for creating short audio samples such as drum beats and instrument riffs, allowing users to fine-tune the model on their own custom audio data.
Business
Zoox to test self-driving cars in Austin and Miami – Zoox, Amazon’s self-driving unit, is expanding its autonomous vehicle testing to Austin and Miami, focusing on fine-tuning its technology in different urban environments.
Microsoft to invest $3.2 bn in AI in Sweden – Microsoft plans to invest $3.2 billion in AI infrastructure in Sweden, including training 250,000 people and increasing capacity at its data centers, while also addressing concerns about AI-generated deepfakes.
Intel announces new AI chips to challenge Nvidia’s dominance – Intel unveils new AI chips to challenge Nvidia’s dominance, with a focus on better performance, power efficiency, and lower pricing.
Amazon’s AI ‘Private Investigator’ flags damaged products, improving the customer experience and preventing waste – AI ‘Private Investigator’ flags damaged products, improving customer experience and preventing waste by helping Amazon associates decide the fate of flagged items.
AI video start-ups race ahead as Big Tech competition looms – AI video-generation start-ups like Pika are racing ahead with significant venture capital investment, aiming to compete with industry giants like OpenAI and Google in the rapidly evolving AI video creation space.
Spark Capital, Jared Leto Back AI Video Startup Pika – AI video startup Pika, backed by Spark Capital and Jared Leto, raises $80 million in funding, signaling investor confidence in generative AI.
Nvidia’s Latest Investment is an AI Startup Focusing on Video Search – Nvidia invests in a Korean AI startup focused on video search and analysis, co-leading a $50 million investment in Twelve Labs Inc.
Salesforce to open new AI center in London as part of $4 billion UK investment – Salesforce is opening an AI center in London as part of a $4 billion investment in the UK, aiming to foster industry collaboration, facilitate AI training, and create 500,000 AI-related jobs, while also investing in UK startups.
Shutterstock Made $104 Million Licensing Assets to AI Devs Last Year – Shutterstock has made $104 million in revenue through licensing its data to AI companies, and its partnership with Reka is projected to earn the company $138 million in 2024.
ChatGPT is back online after an outage – OpenAI’s ChatGPT experienced multiple outages, including a major one during the daytime in the US, but the issues were eventually resolved.
Elon Musk ordered Nvidia to ship thousands of AI chips reserved for Tesla to X and xAI – Elon Musk diverted thousands of AI chips reserved for Tesla to his other companies, X and xAI, causing delays and raising concerns about his conflicting interests and priorities.
Research
Improving Alignment and Robustness with Short Circuiting – Preventing harmful outputs and adversarial attacks in AI systems through a new approach called “short-circuiting” that controls representations responsible for harmful outputs without sacrificing utility.
Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality – Transformers and state-space models (SSMs) are closely related, with SSMs like Mamba matching or outperforming Transformers at small to medium scale, leading to the development of a new architecture, Mamba-2, that is faster and competitive with Transformers in language modeling.
LLMs achieve adult human performance on higher-order theory of mind tasks – Large language models (LLMs) have achieved adult-level performance on higher-order theory of mind tasks, with GPT-4 and Flan-PaLM reaching or exceeding human performance, indicating significant implications for user-facing LLM applications.
Show, Don’t Tell: Aligning Language Models with Demonstrated Feedback – Aligning language models with demonstrated feedback through a method called DITTO allows for fine-grained style and task alignment, outperforming other methods by 19% points.
Scalable MatMul-free Language Modeling – Eliminating matrix multiplication from large language models while maintaining strong performance at billion-parameter scales, the proposed MatMul-free models achieve on-par performance with state-of-the-art Transformers, reducing memory usage and offering a GPU-efficient implementation, as well as a custom hardware solution on an FPGA for increased efficiency.
σ-GPTs: A New Approach to Autoregressive Models – A new approach to autoregressive models challenges the fixed order assumption by adding positional encoding for on-the-fly modulation, offering advantageous properties such as sampling and conditioning on arbitrary subsets of tokens.
SpatialRGPT: Grounded Spatial Reasoning in Vision Language Model – Grounded spatial reasoning in vision language models is the focus of the SpatialRGPT, aiming to enhance understanding and application of AI in various fields.
To Believe or Not to Believe Your LLM – Identifying and quantifying uncertainty in large language models, particularly distinguishing between epistemic and aleatoric uncertainties, can help detect unreliable outputs and hallucinations, with potential applications in improving model reliability and accuracy.
Guiding a Diffusion Model with a Bad Version of Itself – Using a smaller, less-trained version of a model to guide image generation leads to significant improvements in image quality and variation without compromising prompt alignment.
Perplexed by Perplexity: Perplexity-Based Data Pruning With Small Reference Models – Small language models can effectively determine high-quality subsets of large-scale text datasets, improving the performance of larger language models through perplexity-based data pruning.
Alice in Wonderland: Simple Tasks Showing Complete Reasoning Breakdown in State-Of-the-Art Large Language Models – State-of-the-art large language models show a dramatic breakdown in reasoning capabilities when faced with simple, common sense problems, leading to overconfidence in wrong solutions and nonsensical explanations.
Transformers need glasses! Information over-squashing in language tasks – AI struggles with “information over-squashing” in language tasks, leading to the need for improvements in processing and understanding textual data.
OmniH2O: Universal and Dexterous Human-to-Humanoid Whole-Body Teleoperation and Learning – OmniH2O is a learning-based system that enables human control of a full-sized humanoid through various methods, including teleoperation and autonomy, showcasing versatility and dexterity in real-world tasks.
Are We Done with MMLU? – Analysis reveals errors in the popular MMLU benchmark, leading to the creation of MMLU-Redux, a subset of manually re-annotated questions, and advocating for revisions to enhance its future utility and reliability.
Concerns
AI chatbots got questions about the 2024 election wrong 27% of the time, study finds – AI chatbots, including Google’s Gemini 1.0 Pro and OpenAI’s GPT-3, provided incorrect information 27% of the time when asked about voting and the 2024 election, raising concerns about the potential impact on voter decisions.
FTC and DOJ reportedly opening antitrust investigations into Microsoft, OpenAI, and Nvidia – FTC and DOJ open antitrust investigations into Microsoft, OpenAI, and Nvidia, with the FTC looking into potential antitrust issues related to investments made by technology companies into smaller AI companies and the DOJ leading inquiries into Nvidia.
Amazon’s AI Warehouses Isolate Workers and Hinder Union Organizing, New Report Finds – Amazon’s use of AI and robotics in its warehouses isolates workers and hinders union organizing, according to a new report by Oxford University researchers.
India’s election wasn’t the deepfake doomsday many feared – AI-generated content in India’s 2024 election, while present, has been more of a parody than a malicious threat, with platforms struggling to effectively moderate and label it.
Microsoft Will Switch Off Recall by Default After Security Backlash – Microsoft is making Recall an opt-in feature in Copilot+ compatible versions of Windows, introducing new security measures, and requiring authentication to access Recall’s stored data after facing criticism for silently storing screenshots of user activity every five seconds.
Adobe scolded for hosting ‘Ansel Adams-style’ images generated by AI – Adobe faces backlash for hosting AI-generated images labeled as “Ansel Adams-style,” prompting the Adams estate to publicly scold the company for unauthorized use of the late photographer’s work.
Policy
Israel reportedly used fake social media accounts to influence US lawmakers – Israel used fake social media accounts to influence US lawmakers in support of its war on Gaza, with the Ministry of Diaspora Affairs funding a $2 million campaign using hundreds of impersonating accounts to target members of congress.
Google Play cracks down on AI apps after circulation of apps for making deepfake nudes – Google Play is cracking down on AI apps, issuing new guidelines to prevent the circulation of inappropriate and harmful content, such as deepfake nudes, and requiring developers to rigorously test their AI tools and models.
AI-generated X-rated content is officially allowed on Musk’s X – AI-generated X-rated content is officially allowed on Musk’s X, with the stipulation that it must be consensual and properly labeled.
Analysis
What does the public in six countries think of generative AI in news? – Public opinion on generative AI in news across six countries is presented, showing widespread awareness and use of AI tools, with younger people more likely to use them; public expectations on the impact of generative AI in various sectors are documented, with most expecting a large impact on society and news media; public opinion on the use of generative AI in journalism is explored, revealing skepticism about its use and concerns about trust and transparency; the report concludes that public debate and experience with generative AI will shape its
Expert Opinions
Former OpenAI researcher foresees AGI reality in 2027 – Former OpenAI researcher predicts the arrival of AGI by 2027, foreseeing AI machines surpassing human intelligence and national security implications.
A realistic path to robotic foundation models – Progress in the field of robotics is often overshadowed by advancements in language models, but new companies are emerging to build general robotics models that work on many robots, for many tasks, with a general interface, with the goal of making robotics more accessible and cost-effective.
Source: Read MoreÂ
