From OpenAI Mutiny to Safe AGI: Ilia Suk's New AI Venture

Ilia Suk, the former Open AI employee who attempted to start a mutiny against Sam Altman, has launched a new company called "Safe Super Intelligence Inc." The company is dedicated to the pursuit of safe superintelligence, which Suk believes is the most important technical problem of our time.

Suk has assembled a team of experienced engineers and researchers, including co-founders Daniel Gross and Daniel Levy, to work towards this goal. The company's mission is to advance capabilities as fast as possible while ensuring that safety remains ahead, in contrast to Suk's concerns about Open AI's priorities.

Safe Super Intelligence Inc. is taking a "straight shot" approach, with a single focus and product. The company is an American-based entity with offices in Palo Alto and Tel Aviv, allowing it to recruit top technical talent. Suk emphasizes that the company's business model, investors, and entire product roadmap are aligned to achieve safe superintelligence, without the distractions of management overhead or product cycles.

The company's launch post states that "super intelligence is within reach" and that Safe Super Intelligence Inc. is the "world's first straight shot SSI lab" dedicated to this critical challenge. Suk's goal is to pursue revolutionary breakthroughs in safe superintelligence through a small, focused team.

Elon Musk and Tesla are building an absolutely massive server farm with the help of both Dell and Nvidia. This server farm is being designed to power large language models and AI systems.

The project was first hinted at a few months ago, with a picture showing the scale of the cooling system required for this massive undertaking. Yesterday, Michael Dell, the founder of Dell, confirmed that they are "building a Dell AI Factory with Nvidia to power GPT for XAI."

The server farm is located in Texas, where the hot climate requires a robust cooling system. Pictures show the immense size of the fans being installed to keep the servers cool.

This collaboration between Elon Musk, Michael Dell, and Nvidia highlights the significant investment and infrastructure required to support the development of advanced AI systems. By building this massive server farm, they are positioning themselves to power the next generation of large language models and AI applications.

The research paper published recently seems to introduce an implementation of a technique akin to Q*, a powerful algorithm that combines Monte Carlo tree search and deep learning. This approach allows agents to learn decision-making and reasoning through a process of trial and error, much like how humans learn.

The key innovation is the use of MCTS (Monte Carlo tree search) and self-refinement, which enables the model to plan ahead, try out different options, and learn from its mistakes. This is a significant advancement over traditional language models that rely solely on predicting the next word.

The results are impressive, with the model achieving 96.66% accuracy on a math benchmark, outperforming even large language models like GPT-4 and Gemini, while using 200 times fewer parameters. This demonstrates the power of combining planning, reasoning, and learning in a single framework.

The ability to plan, experiment, and refine its approach is a crucial step towards more general intelligence. As we continue to see breakthroughs in areas like this, the path to artificial general intelligence (AGI) becomes increasingly tangible.

Meta AI has recently released several exciting new models and technologies:

1. Meta Chameleon: A 7 billion and 34 billion parameter language model that supports mixed-modal input and text-on-output. This model can natively handle multimodal inputs, unlike previous open-source models that required add-ons. It is currently available only under a research license, but this technology is a significant step forward in multimodal AI capabilities.

2. Multi-Token Prediction: Meta has proposed a new approach to language modeling that predicts multiple future words at once, rather than the traditional one-at-a-time method. This improves model capabilities and training efficiency, while allowing for faster speeds. Meta is releasing pre-trained models for code completion under a non-commercial research license.

3. Generative Text-to-Music Models: Meta has developed models capable of generating music from text inputs, with various conditioning inputs for greater controllability. The pre-trained model and inference code will be released later this month under an MIT license, allowing researchers to explore this exciting new capability.

These releases demonstrate Meta's continued advancements in multimodal AI, language modeling, and generative capabilities. The open-sourcing of these models and technologies will enable further research and development in these areas, driving progress in artificial intelligence.

Anthropic has just released a new version of their Claude language model, called Claude 3.5 Sonet. According to the announcement, this is their "most intelligent model yet" and the first release in their 3.5 model family.

The key highlights of Claude 3.5 Sonet include:

• It outperforms competitor models on key evaluations, including GSM 8K (96.4%), graduate-level reasoning (60%), and code (92%) - all at twice the speed of Claude 3 and Opus, and at 1/10th the cost.

• It surpasses Claude 3 Opus across all standard vision benchmarks, making it Anthropic's strongest vision model to date.

• It introduces a new feature called "Artifacts", which allows users to ask Claude to generate documents, code, diagrams, vector graphics, and even simple games.

The release of Claude 3.5 Sonet demonstrates Anthropic's continued progress in developing highly capable and efficient language models. The performance improvements across a range of benchmarks, as well as the new Artifacts feature, make this a significant update to the Claude family of models.

The Arc Prize is a million-dollar public competition to develop a solution that can beat the Abstraction and Reasoning Corpus (ARC), which is considered the only formal benchmark for Artificial General Intelligence (AGI).

The key points about this competition are:

• The ARC benchmark measures general intelligence, not just specific skills. It tests the ability to efficiently acquire new skills, which is the hallmark of true intelligence.

• While existing AI models have been able to quickly surpass other benchmarks, the ARC benchmark remains a significant challenge, with the current leading implementation only achieving a 38% success rate.

• The competition is hosted by Mike Mirzayanov, François Chollet, and Greg Brockman, aiming to spur progress towards the elusive goal of AGI.

• A prize of $1 million will be awarded to the first person or team that can develop an AI system capable of solving the ARC benchmark.

This competition represents an exciting opportunity for the AI research community to push the boundaries of what is possible and make a significant step towards the realization of Artificial General Intelligence. The challenge is formidable, but the potential reward of a million-dollar prize and the chance to be a pioneer in this field is sure to attract the best and brightest minds in the industry.