This week's latest generative AI updates - September 10, 2024

In this week's edition of the Eureka generative AI newsletter, learn about new model developments, the latest research, and the state of generative AI adoption in enterprises.

SymphonyAI news and resources 

• News: SymphonyAI Wins the 2024 Oracle Partner Awards – EMEA Business Impact Category 

• Video—Financial Services: How SensaAI for Sanctions reduces false positives – AI for sanctions screening software 

• Blog—Financial Services: 5 reasons to move financial crime compliance to the cloud 

• Webinar recording—Enterprise AI: Cutting through the hype to unleash enterprise AI value 

Latest AI model developments 

• Strawberry Model Boosts Reasoning: OpenAI will launch its new AI model, Strawberry, this fall. Strawberry aims to overcome current model limitations, improving performance in handling symbolic or ambiguous problems. 

• DeepMind & UC Berkeley Optimize LLM Inference-Time Compute: Researchers found that optimizing inference-time compute can boost large language model (LLM) performance without needing larger models or extensive retraining. By refining answer proposals and verifiers, smaller LLMs matched larger ones, achieving superior results with just 25% of the usual computation. 

• Cerebras Unveils Fastest Inference: Cerebras Systems launched an AI inference service—powered by its WSE-3 chip, promising 100x better price performance with a pay-per-query model, enhancing real-time, interactive AI applications. 

• LTM-2 Mini Code Handling: Magic AI's LTM-2-mini model boasts a 100 million-token context window, significantly enhancing AI operations. Its efficiency surpasses Llama 3.1 by 1000x, and the new HashHop benchmark improves contextual assessments.  

• GameNGen AI Redefines Development: Google's GameNGen AI can simulate classic games like Doom without relying on a traditional game engine, generating game environments solely from text inputs. This breakthrough offers a new way to create immersive experiences.  

AI-driven innovation: empowering developers and enterprises 

• OpenAI's File Search Tool: OpenAI's File Search Tool automates document parsing, chunking, and embedding creation, enabling efficient retrieval through vector and keyword searches. It significantly streamlines content discovery in large document sets, enhancing productivity.  

• NVIDIA Launches NIM Agent Blueprints: To accelerate and streamline AI workflows and generative AI app development for enterprises. Key use cases include digital avatars, drug discovery, and PDF data extraction. 

• Progress Software's Semaphore 5.10: Introduces AI-assisted knowledge modeling to enhance semantic model creation, improving productivity and decision-making.  

• Elastic Integrates Claude Models: The Elasticsearch Open Inference API now integrates with Anthropic's Claude models, enabling real-time analysis of proprietary data for enhanced decision-making and operational efficiency while reducing latency and costs. 

• Nous Research DisTrO Optimizer: Nous Research's DisTrO optimizer boosts AI model training efficiency by up to 10,000 times and reduces inter-GPU communication, enabling training over consumer-grade internet.  

• Vectara Portal Democratizes AI: Vectara Portal is a no-code platform allowing non-developers to create AI tools for chat, search, and summarization.  

 AI in financial services  

• MSATP Partners with Blue J: The Maryland Society of Accounting and Tax Professionals (MSATP) has partnered with Blue J to provide AI-driven tax research solutions for solo and mid-sized firms. 

• Coinbase Executes AI Transaction: Coinbase announced the first autonomous crypto transaction by AI agents. 

• AI Risk and Fynancial Partnership: AI Risk, Inc. and Fynancial have partnered to integrate AIR-GPT into wealth management with features like pre-meeting readouts, audio transcriptions, predictive responses, and sentiment analysis. 

AI regulation, responsibility, and public trust 

• Health AI Antidiscrimination Requirements: The US Department of Health and Human Services mandates health organizations to manage discrimination risks in AI tools by May 2025, emphasizing accountability and enhancing health equity and outcomes.  

• X's AI Misinformation Fixes: X (formerly Twitter) modified its AI chatbot Grok, following concerns from five state secretaries about election misinformation. Grok now directs users to CanIVote.org and Vote.gov.  

• California Passes AI Safety Bill: California's legislature passed SB 1047, an AI safety bill requiring companies to implement safety measures and conduct rigorous testing for models costing over $100 million. Governor Gavin Newsom has until September 30 to decide on the bill. 

• OpenAI & Anthropic’s AI Safety: OpenAI and Anthropic's agreement with the U.S. AI Safety Institute to send AI models for safety testing aims to establish responsible AI development standards.  

• AI Chatbots in Police Reports: U.S. police departments are testing AI chatbots like Axon's "Draft One" to expedite incident reporting. While advocates highlight efficiency gains, critics worry about errors, biases, and legal implications, stressing the need for oversight. Axon CEO Rick Smith insists officers remain accountable for report accuracy. 

The ongoing battle for AI training data 

• Websites Block Apple's AI: Several prominent websites, including Facebook and The New York Times, are blocking Apple's AI crawler, Applebot-Extended, to prevent free content access for AI training. In contrast, Google forces compliance while Apple seeks fair negotiations. 

• Google Restricts AI Search: Google, Microsoft, and ChatGPT are restricting AI-generated content for the 2024 U.S. election due to misinformation risks, while Perplexity maintains unrestricted outputs but encourages user verification.  

• Baidu Blocks Google Scraping: Baidu has updated its robots.txt file to block Google and Bing from scraping its Baike service, which contains nearly 30 million entries.  

The road to AI maturity: overcoming challenges and risks 

• Enterprise AI Satisfaction Declines: The ISG report shows increasing outsourcing in AI and automation, yet reveals enterprise dissatisfaction with these services, as reflected in lower customer experience scores. Generative AI received the lowest score at 68.46.  

• Generative AI Adoption Challenges: A Deloitte report highlights that two-thirds of companies are increasing generative AI investments for efficiency, but only 38% track productivity changes and 68% have scaled less than 30% of projects.  

• Generative AI Faces Scaling Issues: Deloitte’s report reveals that despite increased investment in generative AI, challenges in data management and risk hinder scaling, with only 30% of AI projects reaching production. Critical issues include data quality, privacy, and security. 

• AI Firms Shift Focus: AI firms are heavily investing in hardware and data centers, but struggle to translate generative AI into reliable products. Key challenges include cost, reliability, and privacy.  

Struggling to get past POCs and get lasting value from AI? This webinar, Cutting through the hype to unleash enterprise AI value, provides practical strategies from experts from Microsoft, MIT, Constellation Research, and SymphonyAI.

 Other generative AI models 

• Google Releases New AI Models: Google introduced three experimental AI models: Gemini 1.5 Flash-8B, Gemini 1.5 Pro, and updated Gemini 1.5 Flash, focusing on developer feedback and advancements.  

• Dracarys Open Source Models: Abacus.ai unveils Dracarys, a new open-source family of large language models optimized for coding, enhancing models like Qwen-2 and Llama-3.1. 

• Aleph Alpha Unveils New EU-Compliant AI: Aleph Alpha launched two open-source language models, each with 7 billion parameters, designed for EU regulation compliance and performance parity with top models.  

• LG & Google Clouds Collaboration: LG AI Research is partnering with Google Cloud to enhance EXAONE 3.0 and ChatEXAONE AI, achieving 56% faster processing, 35% less memory usage, and 72% reduced costs.  

• SambaNova Improves Speed: SambaNova Systems set a milestone by processing 114 tokens/second with Meta’s LLaMA 3.1 405B model, verified by Artificial Analysis. Their SN40L chip optimizes data movement, enabling rapid real-time applications and expanded use cases.  

• Alibaba Unveils Qwen2-VL: Alibaba's Qwen2-VL model analyzes videos over 20 minutes, supporting real-time analysis, summarization, and multilingual capabilities.  

• Progress Enhances AI Modeling: Progress Software's Semaphore 5.10 introduces AI-assisted knowledge modeling to enhance semantic model creation, improving productivity and decision-making. 

• Cohere Upgrades Enterprise Models: Cohere announced significant upgrades to its Command R AI models, improving coding, math, reasoning, and latency for enterprise clients.

Notable research  

• Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling: Generating synthetic data from weaker, cheaper models is more compute-optimal for training LLM reasoners, offering better coverage, diversity, and performance gains than data from stronger, expensive models. 

• Text2SQL is Not Enough: Unifying AI and Databases with TAG: The paper introduces the TAG framework, which unifies language models and databases to answer complex natural language queries with high accuracy. 

• Generative Verifiers: Reward Modeling as Next-Token Prediction: The paper proposes GenRM, a next-token prediction-based verifier that unifies solution generation and verification, outperforming traditional methods in reasoning tasks and enhancing LLM performance. 

• Agentic RAG for Time Series Analysis: The paper proposes an agentic RAG framework for time series analysis using a multi-agent architecture.  

• Diffusion Models Are Real-Time Game Engines: GameNGen is a neural diffusion model that powers real-time game simulation by predicting frames interactively, achieving high visual fidelity.  

• Eagle: Exploring The Design Space for Multimodal LLMs with Mixture of Encoders: The paper presents "Eagle," a multimodal LLM design that uses multiple vision encoders with simple fusion strategies, achieving superior performance in visual tasks and setting new benchmarks in multimodal perception and understanding. 

 Top funding and M&A announcements 

• Magic Secures $320M Investment: Magic, a generative AI coding startup, raised $320 million, totaling $465 million in funding, to develop supercomputers with Google Cloud for 160 exaflops performance.  

• Cribl Reaches $3.5B Valuation: Cribl has raised $319 million in a Series E round, boosting its valuation to $3.5 billion, a 40% increase since 2022.  

• Codeium Valued at $1.25B: Codeium has raised $150 million in Series C funding, reaching a $1.25 billion valuation.  

• Yale Commits $150M to AI : Yale University is investing $150 million over five years to boost its AI infrastructure, enhancing access to tools, research support, and collaboration. This initiative includes increasing GPUs, hiring 20 AI faculty members, and integrating AI across all fields. 

• Story Protocol Secures $80M: Story Protocol raised $80 million in Series B funding, valuing the company at $2.25 billion.  

• Slingshot AI Mental Health Funding: Slingshot AI, a mental health startup from New York and London, secured $30 million in seed funding. 

• Supio Raises $25M in Series A Funding: Supio, an AI platform for personal injury law firms, secured $25M in Series A funding, raising its total to $33M.  

• Viggle AI Secures $19 Million: Viggle AI, a Canadian startup, raised $19 million in Series A funding led by Andreessen Horowitz to enhance its generative AI technology for character animation. Their JST-1 technology simplifies animation using text prompts.  

• Bland AI Secures $16M: Bland AI has secured $16 million in Series A funding, totaling $22 million. The platform automates enterprise phone calls with realistic AI agents, featuring voice cloning, multi-language support, and analytics.  

• Nvidia Considers OpenAI Investment: Nvidia is considering a $100 million investment in OpenAI, valuing the startup at over $100 billion. Thrive Capital, Apple, and Microsoft are also interested. The funding will boost OpenAI’s computing power and operations. 

• Capgemini Acquires Syniti Expertise: Capgemini has agreed to acquire Syniti, bolstering its data management capabilities for large-scale SAP transformations like migrations to SAP S/4HANA.