SOCIETY

📈 AI May Replace "Almost All" Forms of Work, Says Larry Summers

Larry Summers, former U.S. Treasury Secretary and now a board member at OpenAI, believes artificial intelligence has the potential to replace “almost all” forms of work. In a recent interview, Summers compared the potential impact of AI to the Industrial Revolution, suggesting that AI could eventually replace most forms of human work, especially cognitive tasks. Summers predicts that AI will be capable of everything from building houses to medical diagnostics, with a particular focus on the "cognitive work" of white-collar workers. He also highlighted that emotional intelligence (EQ) will become more important than intelligence quotient (IQ) as AI advances.

Although Summers acknowledges that AI has the potential to revolutionize the economy, he also warns that its impact will take time to materialize. He argues that realizing productivity gains from a new idea takes years of investment, research, and development. Summers' view on AI and work is a reminder of the importance of preparing for the changes that technology can bring to the job market and the global economy.

OPENAI

🤝 Jony Ive and Sam Altman Plan AI Device

Jony Ive, former chief designer at Apple, and Sam Altman, CEO of OpenAI, have joined forces to develop a personal device powered by artificial intelligence. The duo is seeking funding for the project, which is still in its early stages and does not resemble a smartphone. Altman, a major investor in Humane AI pin, a screenless wearable device, may be creating a similar product with Ive.

Discussions about the partnership started last year but have only now become official, with the still-unnamed startup seeking funds from major investors. Ive aims to raise up to $1 billion in funding, and OpenAI may own a part of the business. Possible investors include Thrive Capital, an investor in OpenAI, and Emerson Collective, founded by Laurene Powell Jobs.

OPEN SOURCE

🛠️ SWE-Agent: New Open Source Programming Agent

SWE-Agent is an open-source software engineering tool that transforms language models, like GPT-4, into powerful agents for fixing bugs and issues in real GitHub repositories. SWE-Agent offers a simple interface for language models to navigate through repositories, view, edit, and execute code files. A key feature of the tool is its linter, which checks code syntax before allowing edits to be made, helping to prevent errors and ensuring that any changes made by the agent are syntactically correct. Additionally, SWE-Agent includes a specialized file viewer and a directory search tool, making it easier for language models to navigate and understand code repositories.

The agent's efficacy is demonstrated by its metrics. In the full SWE-bench test set, SWE-Agent solves 12.29% of problems, achieving cutting-edge performance. This shows the power of using language models as software engineering agents and the importance of a well-designed interface, like the ACI (Agent-Computer Interface), to maximize their capabilities.

APPLE

🤖 ReALM: Reference Resolution as Language Modeling

Researchers at Apple presented "ReALM: Reference Resolution As Language Modeling," which proposes a new strategy for resolving ambiguous references in dialogues, such as pronouns and demonstratives. While easily understood by humans within a specific context, these represent a significant challenge for computational systems. The paper presents a method by which language models are applied to address and simplify the problem of identifying exactly what these references align with within a conversation, essentially treating it as a task of language modeling.

The work demonstrates notable advances over previous approaches, evidenced by an improvement of over 5% in accuracy rates on reference resolution tasks, even with smaller-scale models. Moreover, it compares its performance not only with previous versions of language models, like GPT-3.5, but also with the more advanced GPT-4, showing that the smallest ReALM model already matches GPT-4 in efficacy, while larger models significantly exceed this benchmark. This innovation aims to improve interaction with voice assistants and other conversational systems, allowing a more natural and intuitive understanding of user references, facilitating a more seamless use of such technologies.

TSMC

🌐 Earthquake in Taiwan and the Risk to the Chip Industry

The recent earthquake in Taiwan brought significant concerns to the global semiconductor industry. The quake, considered the largest in 25 years, caused damages and disrupted operations of chip manufacturing giants, such as the Taiwan Semiconductor Manufacturing Co. (TSMC). TSMC, which produces about 80% to 90% of the world's most advanced chips, is crucial for technologies like artificial intelligence, smartphones, and electric vehicles.

Despite the evacuation of employees and the halt of operations, TSMC expects its most advanced factory, located in the city of Tainan, to resume full recovery soon. The company reported that more than 80% of the tools have been recovered and there was no damage to the most critical equipment, including extreme ultraviolet lithography systems. However, some production lines will require more time to return to normal due to the earthquake's greater impact.

The situation in Taiwan highlights the vulnerability of the global supply chain and the strategic importance of the island in the semiconductor market. As the world watches Taiwan's recovery, the need for diversification and resilience in chip production becomes increasingly evident.

STABILITY AI

🎵 Introducing Stable Audio 2.0

Stable Audio 2.0 is the latest innovation from Stability AI, setting a new standard in AI-generated audio. This model enables the creation of complete, high-quality tracks with coherent musical structure and up to three minutes in duration in stereo 44.1kHz, from a single natural language prompt. Besides the text-to-audio capability, Stable Audio 2.0 introduces audio-to-audio generation, allowing users to upload audio samples and transform them into a wide range of sounds using natural prompts. This update also expands the generation of sound effects and style transfer, offering artists and musicians more flexibility, control, and an elevated creative process.

Built on Stable Audio 1.0, which debuted in September 2023 as the first commercially viable AI music generation tool capable of producing high-quality music in 44.1kHz, the model is now available for free use on the Stable Audio website and will soon be accessible through the Stable Audio API.

APPLE

🤖 Apple Studies Domestic Robots as "The Next Big Thing"

Apple is considering venturing into the field of personal robotics, which has the potential to become one of the "next big things" for the company. According to sources familiar with the situation, Apple has teams exploring two different types of domestic robots in search of new revenue sources beyond the iPhone. Apple's hardware engineering division and the artificial intelligence and machine learning group are overseeing work on personal robotics. The domestic robot project is still in the early research and development phase.

This initiative comes after Apple abandoned its efforts to develop an electric car, along with an effort to create next-generation smartwatch displays. Apple's search for a bold new product category includes this foray into robots, in addition to preparing new iPads and corresponding accessories for May. Apple's entry into the domestic robotics market could represent a significant advance for the company, known for its hardware innovation. With iPhone sales stagnating, representing more than half of its revenue, Apple seeks to remain at the forefront of hardware innovation.

GOOGLE

🌐 Google Explores Paywall Option for AI

The tech giant Google is considering implementing a paywall for premium AI-generated content, according to a recent report. The company, used by over a billion people worldwide, may be planning a significant shift in its business model by adding certain AI-powered search features to its premium subscription services. These services already offer access to the new AI assistant called Gemini, Google's version of the popular ChatGPT service. While Google's traditional search engine remains free and continues to display ads, the company is exploring the possibility of placing some of its main offerings behind a paywall.

Google has faced challenges in adapting to the AI revolution, especially after the controversy generated by Gemini, which created incorrect images in response to text prompts. Despite these setbacks, Google remains the go-to for most internet users when it comes to searching for information.

OPENAI

🖼️ Editing Images with DALL·E

OpenAI has launched the DALL·E editing interface, which allows users to intuitively edit AI-generated images. With this tool, it's possible to select specific areas of the image to add, remove, or update elements. For example, by highlighting a field in the image and using the command "add cherry blossoms," a result incorporating the desired flowers into the image is achieved.

Additionally, the interface offers options to undo and redo selections, ensuring greater control over the editing process. Users can also edit images conversationally, providing a command in the conversation panel, without the need to use the selection tool.

RESEARCH

🛡️ "Many-Shot Jailbreaking": Challenging AI Security

The "Many-Shot Jailbreaking" study by Anthropic explores a technique of "jailbreaking"—a method that can be used to circumvent the security measures implemented by developers of large language models (LLMs). The technique, called "many-shot jailbreaking," is effective in both Anthropic's own models and those produced by other AI companies. The research reveals that by including large amounts of text in a specific setup, this technique can force LLMs to produce potentially harmful responses, despite having been trained not to do so. "Many-shot jailbreaking" takes advantage of the dramatic increase in the size of LLMs' context window, which has gone from a long article (~4,000 tokens) to the size of several long novels (1,000,000 tokens or more).

Anthropic has confidentially shared the details of "many-shot jailbreaking" with other AI researchers and has already implemented mitigation measures in its systems. The goal is to accelerate progress towards a mitigation strategy and foster a culture where exploits like this are openly shared among LLM providers and researchers.

🤖 ChatGPT can now be used instantly, without the need for sign-up. OpenAI is making AI more accessible to everyone, aiming to allow anyone curious about its capabilities to experience ChatGPT. Learn more.

🔬 SyntheMol, a new artificial intelligence tool developed by researchers at Stanford Medicine, is creating recipes for chemists to synthesize new drugs to combat antibiotic-resistant bacteria. The tool generated structures and chemical recipes for six new drugs intended to kill resistant strains of Acinetobacter baumannii, one of the main pathogens responsible for deaths related to antibiotic resistance. Learn more.

📱 Samsung is considering updating its voice assistant Bixby with generative artificial intelligence technology. The company seeks to redefine the role of Bixby, making it smarter and capable of conducting more natural conversations and supporting Samsung ecosystem products. Learn more.

👨‍💻 Replit announced "Code Repair," a low-latency AI program repair agent. Code Repair automatically fixes code in the background as developers work. Additionally, Replit introduced "Replit Teams," a new functionality that allows collaboration, development, and deployment of software in teams. Learn more.

🎮 Microsoft is testing an AI chatbot for Xbox support, promising to be a useful tool for users. The virtual agent, powered by AI, aims to help users with issues efficiently, using natural language to facilitate and speed up the support process. Learn more.

🌐 The United Kingdom and the United States announced a partnership to test the security of the most powerful artificial intelligence models. The agreement, signed by the UK's Secretary of State for Science, Innovation, and Technology, Michelle Donelan, and the U.S. Secretary of Commerce, Gina Raimondo, establishes a plan for collaboration between the two governments. The AI security institutions of both countries will develop a common approach to AI security testing, involving the use of the same methods and underlying infrastructure. Learn more.

📸 A recent study showed that feeding an AI image generator with just 1,000 additional images was able to reduce stereotypes generated by AI. Researchers managed to make the image generator produce less offensive images simply by adding a small amount of training data. Learn more.

📅 Billionaire investor Steve Cohen predicts that artificial intelligence (AI) technology will help implement a four-day work week. Cohen, owner of the New York Mets, expressed his belief that a shorter workweek is inevitable, citing the advent of AI and generally lower productivity on Fridays as contributing factors. He mentioned that although he does not foresee his traders taking Fridays off while the markets are open, most people will, at some point, enjoy a three-day weekend. Learn more.

🌐 Opera announced that it will now allow users to download and use large language models (LLMs) locally on their computers. This functionality, being initially released to Opera One users who receive developer updates, will allow users to choose from more than 150 models from over 50 families. The company is using the open-source Ollama framework in the browser to run these models on the user's computer. Learn more.

🚗 Waymo and Uber Eats launched a driverless food delivery service in the Phoenix metropolitan area. Uber Eats customers in Tempe, Mesa, and Chandler can have their orders delivered by an autonomous vehicle. The partnership is working with selected merchants, including local favorites like Princess Pita, Filiberto’s, and BoSa Donuts. Learn more.

🛒 Amazon's “Just Walk Out” technology, which promotes a cashierless shopping experience, apparently relies on hundreds of workers in India to ensure its accuracy. Although Amazon has touted the technology as being fully powered by AI, reports suggest that many “Just Walk Out” transactions are reviewed by real people. Learn more.

🖥️ Microsoft 365 announced major updates to Copilot, including priority access to the GPT-4 Turbo model and improvements in image generation with Microsoft Designer. Commercial users of Copilot now have unlimited access to the number of conversations per day and the amount of file uploads. Additionally, starting next month, the number of image generations per day will increase from 15 to 100, utilizing the DALL-E 3 model. Learn more.

This week's video brings an update from Humane on its AI Pin device. The company said that the first deliveries start on April 11, and demonstrated the device's functionalities. Click here or on the image above to watch.

📊Elicit: a tool that accelerates the analysis of research articles, automating time-consuming research tasks like summarizing documents, extracting data, and synthesizing findings. The platform searches over 200 million academic articles from the Semantic Scholar corpus, covering all academic disciplines. Link.

💞 Maia: an artificial intelligence assistant designed specifically for couples, offering a private chat space to improve communication and deepen connections. Maia helps plan personalized dates, navigate disagreements using relationship frameworks, and initiate meaningful conversations. Link.

🗣️ Jessica Que: a virtual assistant specialized in creating content for LinkedIn, offering a fast and efficient service for experts wishing to highlight their expertise on the platform. With just a 5-minute interview, Jessica can generate up to two weeks of relevant content, helping professionals position themselves as authorities in their niches. Link.

🚀 Astra: an artificial intelligence copilot designed for agencies, automating the administrative side of managing a service business. The tool offers automated project planning, calendar integration for work sessions, and efficient communication, including progress updates and scope of work. Link.

⚙️Run Open Source Models Locally

Several smaller open-source models can be run offline on your laptop — keeping data private and avoiding the censorship that accompanies other models. Step by step:

1. Download LM Studio - a free platform for local AI models.

   

2. Within LM Studio, search for a model that can be run on your computer. In this example, we searched for Mistral and selected dolphin-2.2.1 from the left listing, which is an uncensored version of the Mistral model.

3. On the right side, select and download a version that your GPU supports (the program indicates if the GPU supports it). Generally, the smaller the size, the faster. In this example, we downloaded the Q4_K_S version.

4. Open the Chat tab, make sure dolphin is selected, and start sending commands!