Daily AI Newsletter: Sam Altman is considering turning OpenAI into a regular company

In today’s email:

🔥 1-bit LLMs Could Solve AI’s Energy Demands
☠️ Hacker Releases Jailbroken "Godmode" Version of ChatGPT
👀 Paul Graham claims Sam Altman wasn’t fired from Y Combinator
🧰 10 new AI-powered tools and resources. Make sure to check the online version for the full list of tools.

Sam Altman is considering turning OpenAI into a regular company

Sam Altman, the CEO of OpenAI, is considering restructuring the organization into a traditional for-profit company, according to a report by The Information. Currently, OpenAI operates as a "capped-profit" company, where its for-profit arm is governed by a nonprofit. This structure was adopted in 2019 to better align with OpenAI's mission of developing safe artificial general intelligence (AGI) while maintaining the ability to raise capital effectively.

Altman's tenure at OpenAI has not been without controversy. He was briefly ousted in November after the board accused him of not being transparent in his communications. Helen Toner, a former board member, alleged on "The TED AI Show" podcast that Altman had withheld information, misrepresented company activities, and even lied to the board. She also claimed that the board was unaware of ChatGPT's release in November 2022 until it was publicly announced on Twitter. Toner further accused Altman of failing to disclose his financial interests in OpenAI, specifically his ownership of the OpenAI Startup Fund.

OpenAI has since stated that Altman no longer owns or controls the venture capital fund, describing its initial ownership structure as a temporary arrangement. Despite the controversies, OpenAI chairman Bret Taylor emphasized the company's commitment to its mission of ensuring AGI benefits all of humanity. Representatives for OpenAI did not immediately respond to requests for comment regarding these issues.

1-bit LLMs Could Solve AI’s Energy Demands

Large language models (LLMs), which power AI systems like ChatGPT, are becoming increasingly powerful but also more energy-intensive and computationally demanding. To address this, researchers are exploring ways to significantly reduce the size and energy consumption of these models by using 1-bit quantization. This method drastically reduces the precision of the mathematical parameters that store the models' memory, representing them with just 1 or -1. This reduction not only makes the models smaller and faster but also more environmentally friendly, enabling them to potentially run on smaller devices like cellphones.

There are two main approaches to achieving 1-bit LLMs: post-training quantization (PTQ) and quantization-aware training (QAT). PTQ quantizes a full-precision network's parameters after training, while QAT trains the network from the beginning with low-precision parameters. Recently, researchers at ETH Zurich, Beihang University, and the University of Hong Kong introduced a PTQ method called BiLLM, which uses 1-bit for most parameters and 2-bits for critical weights, significantly reducing memory usage with minimal performance loss. Conversely, Microsoft Research Asia's BitNet, a QAT method, showed even greater energy efficiency and speed improvements.

The advancement of 1-bit LLMs has demonstrated their potential to rival full-precision models in performance while drastically cutting energy and memory requirements. However, current hardware, like GPUs made by Nvidia, isn't fully optimized for these low-precision models. Future developments in hardware that can natively support 1-bit operations could further enhance the efficiency and performance of these models. Researchers believe that as these 1-bit models and specialized hardware evolve together, they will pave the way for more sustainable and accessible AI technologies.

Google explains AI Overviews’ viral mistakes and updates, defends accuracy

Google recently addressed criticisms regarding AI Overviews, explaining their distinct operation compared to chatbots and other LLM products. Unlike generating output solely based on training data, AI Overviews integrate a customized language model with Google's core web ranking systems to identify high-quality, relevant results. This integration aims to reduce the hallucination problem seen in broader LLM applications, providing users with text outputs and relevant links to explore further.

Google acknowledged some high-profile mistakes, such as misinterpreting queries or nuances and the lack of quality information. Viral instances like the “How many rocks should I eat” query, which was mistakenly linked to a satirical article, and the “using glue to get cheese to stick to pizza” query, which relied on Reddit forums, were highlighted. These errors stemmed from the system's difficulty handling satirical content and over-reliance on user-generated content as reliable sources.

To improve, Google has limited the inclusion of satirical and humorous content and updated detection mechanisms for nonsensical queries. They’ve also restricted user-generated content for potentially misleading advice and refined systems for specific topics like news and health to ensure quality and accuracy. Despite these challenges, Google claims that AI Overviews provide high-quality information, with accuracy rates comparable to their quote-based Featured Snippets, though they did not provide specific data to support this assertion.

Perplexity AI’s new feature will turn your searches into shareable pages

Perplexity AI has introduced a new feature called Perplexity Pages, which enables users to create detailed and visually appealing web pages from their search queries. This feature is available to both free and paid users in the library section, where they can enter prompts like “Information about Sahara Desert” to start generating a page. Users can tailor the tone of the generated text by selecting an audience type and can further customize the content by rewriting, reformatting, or adding new sections. The generated pages can include relevant media items such as images and videos and are searchable through Google, making them easily shareable.

Henry Modisett, Perplexity’s head of design, highlighted that the company aims to leverage its core technology to enhance Perplexity as a research tool in a more shareable format. The feature allows users to curate information rather than generate content solely through AI, emphasizing the human aspect in making decisions about the page’s organization and content. Although the AI engine begins answering questions and forming the page instantly, it takes a couple of minutes to complete, during which Modisett focused on creating a pleasant user experience through animations.

The rollout of Perplexity Pages is initially limited to a select number of users, with plans to expand access to all users eventually. Currently, pages can be created on the web but are viewable on both web and mobile apps. As the feature integrates different sources prominently, it aims to balance AI-generated content with human curation, impacting areas like blogging, journalism, and the broader web ecosystem.