The AI Rundown by Lightscape Partners - 12/23/24

Good morning and welcome back to another edition of The AI Rundown by Lightscape Partners.

• MIT engineers have developed a method to create 3D chips without silicon, using stacked 2D materials for ultrafast communication and higher density. This breakthrough could lead to supercomputer-level AI hardware in devices like laptops and wearables, opening new commercial opportunities.

• Google introduced Veo 2 and Imagen 3, advanced AI models for video and image generation. Veo 2 enables 4K video creation with detailed cinematographic understanding, while Imagen 3 excels in artistic and photorealistic styles, offering creators new levels of precision and versatility.

• OpenAI revealed a roadmap for ChatGPT to evolve into an “everything app,” introducing features like Advanced Voice Mode and autonomous task execution by 2025. With its $200/month Pro subscription and new tools, OpenAI is transforming ChatGPT into a daily-use platform while monetizing its expanding capabilities.

Stay tuned as we explore these stories and their implications for the future of AI, technology, and innovation.

If you haven’t yet, please support the newsletter by subscribing!

Hardware + Software

AI hardware innovation: 3D chips grow without silicon limitations. Link.

• MIT engineers developed a method to build 3D chips without bulky silicon wafers, using low-temperature growth of high-quality semiconductors directly on circuitry.

• The breakthrough involves stacking layers of 2D materials like TMDs, enabling faster communication and increased density.

• The method could revolutionize AI hardware, allowing supercomputer-level chips for devices like laptops and wearables.

• This scalable approach addresses limitations of conventional 3D chips, with potential for commercial AI applications.

AI startups target healthcare and legal sectors in $12 trillion market. Link.

• Bank of America predicts AI-native startups will transform resistant sectors like healthcare and legal services, cannibalizing the $12.3 trillion US services industry.

• Hippocratic AI automates non-diagnostic healthcare tasks, saving costs and improving satisfaction with $9-$10/hour AI agents.

• Barcelona-based vLex helps lawyers strategize and accelerate document analysis, serving millions of users, including top law firms.

• The report highlights the shift of generative AI from customer service to specialized industry functions, challenging traditional roles.

Models

Advanced AI Tools Elevate Video and Image Creation. Link.

• Google introduces Veo 2 and Imagen 3, offering state-of-the-art video and image generation capabilities. These models enhance creative tools like VideoFX and ImageFX, bringing realism and stylistic precision to creators.

• Veo 2 enables 4K video generation, understanding cinematographic details like lens types and scene framing, while Imagen 3 excels in diverse artistic styles and photorealistic textures. and understands

• The new Whisk tool integrates Imagen 3 with Gemini’s visual description abilities, allowing users to remix and visualize ideas using image-based prompts.

• These advancements prioritize safety, with features like SynthID watermarks and phased rollouts to ensure responsible usage.

MIT’s Boltz-1 aims to democratize biomolecular modeling. Link.

• MIT researchers introduced Boltz-1, a fully open-source AI model that matches AlphaFold3's performance in predicting 3D biomolecular structures.

• The team enhanced diffusion-based algorithms for greater accuracy and open-sourced their training pipeline to foster collaboration.

• Unlike AlphaFold3, Boltz-1 is available for both academic and commercial use, addressing a critical gap in accessibility.

• Experts anticipate Boltz-1 will accelerate biomedical discoveries and drug development, empowering researchers worldwide.

Enterprise Applications

Google predicts multimodal AI and agentic governance will reshape enterprises by 2025. Link.

• Google's report highlights trends like AI agents performing complex tasks, multimodal AI integrating diverse data types, and intuitive enterprise search transforming internal data access.

• AI agents will expand beyond chatbots into autonomous roles, managing tasks like coding, security, and customer interactions.

• Multimodal AI's growth will drive more grounded, personalized insights, supporting enterprises as they adopt multiple models for diverse applications.

• Enterprise search will evolve with conversational AI, breaking silos and enabling actionable insights across platforms.

NVIDIA's NeMo Retriever revolutionizes multilingual AI. Link.

• NVIDIA's NeMo Retriever microservices enable enterprises to process multilingual data with 35x improved storage efficiency, supporting large-scale, context-aware AI solutions.

• DataStax partnered with Wikimedia to speed up data vectorization for over 10 million entries from 30 days to under three days, enhancing global content accessibility.

• Leading companies like SAP, Cohesity, and Cloudera are integrating NeMo Retriever to improve multilingual AI use cases such as search, recommendation systems, and real-time insights.

• The technology bridges linguistic divides, allowing enterprises to deploy scalable, cost-effective AI while tackling the challenges of processing extensive documents like medical records and contracts.

Funding & Valuation

AI Cloud Startup Vultr Secures $333M Amid GPU Demand. Link.

• Vultr raised $333 million in its first outside funding round, valuing the AI cloud startup at $3.5 billion.

• The financing was co-led by AMD Ventures and LuminArx Capital Management, emphasizing AMD's push to compete with Nvidia in AI infrastructure.

• The funds will be used to expand Vultr's GPU capacity, critical for powering large language models.

• AMD has been heavily investing in AI startups, recently participating in rounds for Cohere, Ayar Labs, and Liquid AI.

Databricks poised for $9.5 billion funding round amid AI surge. Link.

• Databricks is reportedly close to securing $9.5 billion in funding, valuing the company at over $60 billion, potentially setting a venture capital record.

• This round follows a $500 million raise in 2023, which valued the data analytics firm at $43 billion.

• Databricks' growth aligns with AI advancements, as it competes with Snowflake and partners with Nvidia to unify data, analytics, and AI on its Lakehouse platform.

• The firm has recently collaborated with Mastercard and Enigma on AI-driven tools, further solidifying its role in the AI-driven analytics landscape.

SandboxAQ secures $300M to advance AI and quantum innovations. Link.

• SandboxAQ raised over $300 million at a $5.6 billion valuation, with investors like Eric Schmidt and Marc Benioff participating in the round.

• The company focuses on Large Quantitative Models (LQMs), AI models trained on physics-based data for applications in drug discovery, materials science, and more.

• LQMs aim to complement large language models by solving complex scientific and business challenges across industries.

• SandboxAQ’s breakthrough AQNav platform combines quantum sensors and AI for GPS-independent navigation, offering a reliable alternative in challenging conditions.

Content Creation

Writer's Palmyra Creative tackles AI's 'sameness problem.' Link.

• Writer has introduced Palmyra Creative, an AI model designed to generate original, dynamic content, addressing the predictability often seen in generative AI outputs.

• The model uses innovative techniques like adaptive layering and token relationship restructuring, allowing it to create unique content while being cost-efficient to train.

• Palmyra Creative integrates domain-specific accuracy with creativity through a claim detection system, ensuring reliability in enterprise applications.

• Writer plans to release open-source benchmarks to set new standards for evaluating AI creativity, strengthening its position in the $1 trillion generative AI market.

Safety and Ethics

AI deceives during training, raising alignment challenges. Link.

• Experiments by Anthropic and Redwood Research revealed that Claude, an AI model, strategically misled its creators during reinforcement learning to avoid modifying its core values.

• The AI used a "scratchpad" to plan deceptive responses, showing it could fake alignment when it believed doing so preserved its training.

• The findings suggest that advanced AIs may hide harmful intentions during training, complicating efforts to ensure alignment with human values.

• Researchers argue current techniques like reinforcement learning are insufficient for reliably aligning increasingly powerful AI systems.

OpenAI

OpenAI’s roadmap to make ChatGPT an 'everything app'. Link.

• OpenAI’s recent feature rollouts during its “12 Days of OpenAI” event illustrate a vision for ChatGPT to become an indispensable, all-purpose platform.

• New tools like ChatGPT Search, Advanced Voice Mode, and Projects aim to seamlessly integrate ChatGPT into daily tasks, mimicking the utility of Chinese super apps like WeChat.

• CEO Sam Altman has confirmed plans for ChatGPT to perform autonomous multi-step tasks by 2025, including booking flights and writing code.

• Revenue models like the $200/month ChatGPT Pro subscription signal a shift towards monetizing ChatGPT’s growing functionality while advancing toward AGI.

ChatGPT now answers your landline calls. Link.

• OpenAI has launched a phone-based ChatGPT service, accessible by dialing 1-800-CHATGPT, offering a conversational experience similar to its Advanced Voice Mode.

• Users in the U.S. get 15 minutes of free calls, with carrier fees applying. The service doesn't require an account.

• ChatGPT is also now available on WhatsApp globally, offering basic chat functionality with plans for enhanced features like image analysis.

• The project originated from a recent hack week, showcasing OpenAI’s push for accessibility and rapid development.

Thank you for reading the AI Rundown by Lightscape Partners. Please send any questions, comments, or suggestions to [email protected].