OpenAi SearchGPT 🔍, Mistral Large 2 🤖, DeepMind AlphaProof & AlphaGeometry 2 🧠

July 30-31, 2024 Fortune Brainstorm AI Singapore Register Now

August 12-14, 2024 Ai4, MGM Grand, Las Vegas. Register Now

Mistral Large 2: The New Benchmark in AI Performance and Multilingual Proficiency

Mistral Large 2 Overview

• Context Window & Language Support: Offers a 128k context window, supporting multiple languages including French, German, Spanish, Italian, Portuguese, Arabic, Hindi, Russian, Chinese, Japanese, and Korean, along with over 80 coding languages.

• Performance: Designed for single-node inference with long-context applications, it features 123 billion parameters for high throughput. Achieves 84.0% accuracy on the MMLU benchmark.

• Licensing: Available under the Mistral Research License for non-commercial use, with a commercial license required for self-deployment.

• Code and Reasoning: Outperforms previous models and rivals leading models like GPT-4, Claude 3 Opus, and Llama 3 405B in code and reasoning capabilities.

• Instruction Following and Alignment: Excels in instruction-following and conversational capabilities, evidenced by its performance on MT-Bench, Wild Bench, and Arena Hard benchmarks.

• Multilingual Proficiency: Trained on a large proportion of multilingual data, surpassing previous models in performance on the multilingual MMLU benchmark.

• Tool Use and Function Calling: Enhanced function calling and retrieval skills make it suitable for complex business applications.

Accessibility and Integration

Availability: Accessible via la Plateforme under the name mistral-large-2407, with instruct model weights on HuggingFace.

Cloud Integration: Available on Google Cloud Platform's Vertex AI, Azure AI Studio, Amazon Bedrock, and IBM watsonx.ai.

Conclusion 
Mistral Large 2 represents a significant advancement in AI model performance and cost efficiency, challenging top models like GPT-4, Claude 3 Opus, and Llama 3.1 with superior language support, coding proficiency, reasoning capabilities, and enhanced instruction-following skills.

Read the full version Here!

OpenAI Announced GPT-4o Mini fine-tuning and SearchGPT

Free Fine-Tuning for GPT-4o Mini Model

• Opportunity: Free fine-tuning available for Tier 4 and Tier 5 developers until September 23.

• Model Features: GPT-4o mini, launched on July 18, offers longer context and lower latency, ideal for tasks like customer support chatbots.

• Usage: Accessible via the fine-tuning dashboard, with 2 million free training tokens per 24 hours.

• Functionality: Combines OpenAI models with internet searches for conversational responses with real-time information.

• Features: Provides accurate, relevant results and partners with publishers to promote trusted sources.

• Rollout: Begins with a small user group, aiming for broader release.

Legal and Ethical Challenges

• Issues: Facing legal action over alleged copyright violations from news outlets.

• Defense: OpenAI claims "fair use" and works with publishers for proper citation and content promotion.

Future Impact

• Market Entry: OpenAI's SearchGPT could challenge Google amidst its antitrust issues.

• Conclusion: These developments aim to make advanced AI more accessible and transform online search and AI technology.

Read the full version Here!

AI achieves silver-medal standard solving International Mathematical Olympiad problems

AGI with advanced mathematical reasoning can unlock new frontiers in science and technology. AlphaProof and AlphaGeometry 2 have solved four out of six problems from the 2024 International Mathematical Olympiad (IMO), equivalent to a silver medalist performance.

Breakthrough AI Performance

• The IMO is a prestigious math competition and a benchmark for AI's reasoning capabilities.

• AlphaProof solved two algebra problems and one number theory problem, including the hardest problem in the competition.

• AlphaGeometry 2 solved one geometry problem, with two combinatorics problems remaining unsolved.

• The AI system achieved a score of 28 out of 42 points, near the gold-medal threshold.

AlphaProof: Formal Mathematical Reasoning

• Trains using the formal language Lean and the AlphaZero reinforcement learning algorithm.

• Translates natural language problems into formal statements and generates solutions by proving or disproving them.

• Trained on millions of problems covering various difficulties and topics.

AlphaGeometry 2: Improved Geometry Solving

• A neuro-symbolic hybrid system trained on significantly more data than its predecessor.

• Uses a faster symbolic engine and a novel knowledge-sharing mechanism to solve complex problems.

• Solved Problem 4 within 19 seconds during the IMO.

Future Prospects

• Experimenting with natural language reasoning systems that don't require translation into formal language.

• Plans to release more technical details on AlphaProof.

• Aim to enhance collaboration between mathematicians and AI tools to solve complex problems efficiently.

Google’s Gemini AI is getting faster with its Flash upgrade
Google has introduced the Gemini 1.5 Flash model, optimized for speed and efficiency, providing a lightweight alternative to the 1.5 Pro. Designed for high-volume tasks, it supports multimodal inputs and excels in summarization, chat applications, and data extraction. Improvements include increased rate limits, new tuning support, and JSON schema mode for better control over outputs. The 1.5 Pro model also received enhancements, including a 2 million token context window, improving performance in coding, reasoning, and audio tasks.

Tech’s splurge on AI chips has companies in ‘arms race’ that’s forcing more spending
Tech giants like Meta, Alphabet, and Tesla are engaged in a fierce competition to secure advanced AI chips, driving a major arms race in the industry. Nvidia leads the market with about 80% of the global GPU market share due to its advanced AI chips and CUDA infrastructure. AMD is challenging Nvidia with its new MI300X GPUs, and Intel is aiming to regain market share with its Xeon 6 processors and Gaudi AI accelerators. This competition highlights the critical role of these chips in advancing AI capabilities and the significant investments being made to secure them.


Amazon racing to develop AI chips cheaper, faster than Nvidia's, executives say
Amazon is developing its own AI chips to reduce reliance on Nvidia and cut costs for its cloud business, Amazon Web Services (AWS). These new chips, Trainium and Inferentia, promise to be cheaper and more efficient, offering up to 50% cost savings compared to Nvidia's. Amazon's AI chips aim to handle complex calculations and large data processing tasks, catering to increasing customer demand for more affordable solutions. AWS, a major growth driver for Amazon, continues to expand its chip capabilities to stay competitive with Microsoft and Alphabet in the AI chip market.


Using AI to train AI: Model collapse could be coming for LLMs, say researchers
Researchers warn that using AI-generated data to train future AI models could lead to "model collapse," where outputs degrade into nonsense over successive generations. This occurs because AI may focus on repetitive or common outputs, ignoring less frequent but significant data. The study emphasizes the importance of using reliable, human-generated data for training to prevent such degradation. Ensuring data quality is crucial to maintaining the effectiveness of AI models.


After AgentGPT’s success, Reworkd pivots to web-scraping AI agents
Reworkd, backed by prominent investors like Paul Graham, Nat Friedman, and Daniel Gross, is developing AI agents for web scraping. Their product, AgentGPT, simplifies the data extraction process by using large language models to navigate and gather data from numerous websites, adjusting automatically to changes. This technology allows businesses to efficiently manage large-scale web data tasks without extensive manual effort, enhancing data integrity and availability.

👥 Connect & Feedback!

👉 Join Us:

📧 Advertise In Weekly AI News:

📧 Contact directly at [email protected]

😍 Share with your friends!