OpenAI's SearchGPT, Mistral Large 2, and Latest AI Breakthroughs

OpenAI has announced the release of a new prototype called SearchGPT, which aims to combine the strengths of their AI models with information from the web to provide fast and timely answers with clear and relevant sources.

This new search feature is designed to be a direct competitor to services like Perplexity, which currently use ChatGPT as their language model backend. OpenAI's decision to release a search-focused product is a significant move, as it poses a direct challenge to traditional search engines like Google.

The key highlights of SearchGPT include:

• Fast and Timely Answers: The prototype is designed to quickly and directly respond to user queries, providing up-to-date information from the web.
• Relevant Sources: SearchGPT will give users clear links to the relevant sources it has used to formulate its responses, allowing for further exploration.
• Conversational Interaction: Users will be able to ask follow-up questions, building on the shared context with each query, similar to a conversation with a person.
• Clean Interface: The prototype features a simple and intuitive interface, making it easy for users to interact with the search functionality.

This move by OpenAI is a clear indication of the company's ambition to expand its reach beyond language models and into the search market. With the potential advantages of SearchGPT, such as its web-searching capabilities and conversational nature, it could pose a significant challenge to existing search engines, especially for users who already rely on large language models like ChatGPT for their information needs.

Mistral, the AI research company, has recently announced the release of Mistral Large 2, the new generation of their flagship model. Compared to its predecessor, the new version is significantly more capable in areas such as language generation, mathematics, and reasoning. It also provides much stronger multilingual support and advanced function calling capabilities.

Benchmark scores show that Mistral Large 2 performs exceptionally well, with an average code generation accuracy of 76.9%, which is only slightly lower than the impressive 77.9% achieved by the powerful GPT-4 model. Interestingly, this high-performing model has much fewer parameters than the larger Llama 3.1 405B model, demonstrating the efficiency and capability of the Mistral team's work.

Mistral Large 2 is designed for single-node inference with a focus on applications that require long context windows. It supports dozens of languages and has a context window of 128k tokens. While the model is released under the Mistral research license, which limits its usage to research and non-commercial purposes, it is still an exciting development in the world of advanced language models.

The race for high-quality, yet smaller models is an important trend, as it allows for the deployment of more capable AI systems on edge devices. Mistral's achievement with Mistral Large 2 is a testament to the progress being made in the field of efficient and powerful language models.

Meta has announced that it will not be releasing its multimodal version of the Llama AI model in the European Union due to regulatory concerns. This decision will prevent European companies from accessing and utilizing the advanced multimodal capabilities of the Llama 3.1 model.

The primary reason cited for this move is the "unpredictable nature of the European regulatory environment" surrounding AI technologies. The EU has recently finalized compliance deadlines for its strict new AI Act, which will require tech companies operating in the EU to adhere to rules around copyright, transparency, and the use of AI for applications like predictive policing.

This regulatory environment has led Meta to halt the release of the multimodal Llama model in the EU, despite the model being made available under an open license. The decision follows a similar move by Apple, which said it would likely exclude the EU from the rollout of its Apple Intelligence feature due to concerns around the Digital Markets Act.

The inability to access the multimodal Llama model is a significant setback for European companies and researchers who were eager to leverage the advanced capabilities of this technology. The multimodal nature of the Llama 3.1 model would have allowed for the integration of text, images, and other modalities, opening up a wide range of potential applications.

This situation highlights the ongoing tension between the desire for innovation and the need for robust regulatory frameworks to address the potential risks and ethical concerns surrounding AI development. While the EU's efforts to establish guardrails for AI are understandable, the unintended consequence of stifling access to cutting-edge technologies is a concern that policymakers will need to address.

As the global AI landscape continues to evolve, the balance between fostering innovation and ensuring responsible development will remain a critical challenge for regulators and technology companies alike.

Stability AI has announced the release of the research paper for Stable Audio Open, an open-source text-to-audio model. This model generates high-quality stereo audio at 44.1 kHz from text prompts, making it perfect for synthesizing realistic sounds and field recordings.

The model runs on consumer-grade GPUs, making it accessible for academic and artistic use. It can generate up to 47 seconds of audio, and the model was trained using nearly 500,000 licensed recordings from sources like Freesound and the Free Music Archive.

Users can fine-tune the model locally with a 6000 GPU, and the model supports a wide range of audio generation use cases, from the sound of rain hitting a roof to upbeat hip-hop drum loops and audio logos.

This release from Stability AI is an exciting development in the field of text-to-audio generation, providing a powerful and accessible tool for creators and researchers alike.

According to the article, OpenAI plans to make the advanced voice capabilities of GPT-4 available to select groups of users by the end of this month. The author is excited about this development and plans to test the new voice mode as soon as he gets access to it.

The article cites a tweet from Sam Altman, the CEO of OpenAI, who replied to a user asking when the voice mode would be available. Altman stated that the alpha version will start later this month, while the general availability (GA) will come a bit later.

The author is eagerly anticipating the release of the GPT-4 voice capabilities, as it will allow him to test and explore the new features. The ability to generate high-quality, realistic-sounding audio from text prompts is seen as a significant advancement in the field of AI-powered audio generation.

According to reports, Open AI is facing significant financial challenges, with projections indicating a potential loss of $5 billion this year. The company's expenses, including a substantial $4 billion on Azure bills, have outpaced its revenue, which is estimated to be around $30 billion.

This financial situation has raised concerns about Open AI's sustainability and its ability to compete with other companies offering similar technologies, especially as Meta is providing similar tech for free. While Open AI is still the main player in the game, the fact that intelligence and AI are becoming commodities means the company needs to offer something unique to differentiate itself.

This could include offering multiple models, a unique inference offering, or specialized fine-tuned models, rather than just relying on a generalized model. The competition in the AI market is expected to intensify, and Open AI will need to find ways to maintain its competitive edge and ensure its long-term viability.

Google DeepMind has published a new paper this week showcasing their AI system, AlphaProof, which can solve International Mathematical Olympiad (IMO) problems at a silver medal level. This is a significant achievement, as these math problems are considered among the most challenging in the world.

The key highlights of the AlphaProof system are:

1. Reinforcement Learning-based Approach: AlphaProof uses a reinforcement learning-based system for formal math reasoning, allowing it to discover novel algorithms and insights to solve open math problems.

2. Improved Geometry Solving: The system also includes an improved version of their previous geometry solving system, AlphaGeometry 2, which together with AlphaProof was able to solve 4 out of 6 problems from the latest IMO competition.

3. Silver Medal Standard: By achieving the same level of performance as a silver medalist in the IMO competition, AlphaProof demonstrates significant progress in building AI systems that can assist mathematicians in their work.

4. Limitations in Reasoning and Training Data: The authors acknowledge that current AI systems still struggle with solving general math problems due to limitations in reasoning skills and training data.

This research highlights the continued advancements in AI's ability to tackle complex mathematical problems, bringing us closer to the day when AI systems can autonomously discover new mathematical insights and techniques. As the authors note, this is an important step towards the potential for an "intelligence explosion" as AI becomes capable of self-improvement.