Sam Altman's Insights on AI

PLUS: Databricks Boosts Data Quality with Lilac, Nvidia's GTC 2024 Keynote Highlights and more.

Author

Wes Roth
March 20, 2024

Today:

  • Sam Altman's Insights on AI

  • DeepMind Co-founder Heads Microsoft's AI

  • Stability AI introduces Stable Video 3D

  • DeepMind's TacticAI Transforms Soccer Strategy

  • Databricks Boosts Data Quality with Lilac

  • Nvidia's GTC 2024 Keynote Highlights

Sam Altman "SHOCKING UPDATES" to Lex | AGI, GPT-5, Elon Musk, Ilya, Living in a Simulation and Sora

Sam Altman's appearance on The Lex Freedman podcast delving into OpenAI's future plans, includes GPT-5 and the controversy with Elon Musk. Altman expresses admiration for Ilia, emphasizing their collaboration. Musk's desire for total control and conflicts over OpenAI's direction are highlighted. Altman advocates for open-source AI tools, stressing the importance of cooperation.

Altman advocates for both open and closed-source models, discouraging non-profits from transitioning to for-profits. The conversation touches on Sora, suggesting it's trained on synthetic data, possibly from Unreal Engine 5, demonstrating advanced world simulation. Overall, Altman calls for cooperation, viewing conflict as detrimental to progress in AI development.

Google DeepMind co-founder joins Microsoft as CEO of its new AI division

Google DeepMind co-founder, Mustafa Suleyman, joins Microsoft to lead its new AI division, overseeing consumer products like Copilot, Bing, and Edge. Suleyman, now CEO of Microsoft's AI team, brings expertise from his tenure at DeepMind, acquired by Google in 2014. 

Despite past controversies leading to his leave from DeepMind, Suleyman's vision aligns with Microsoft's AI mission. Microsoft reaffirms commitments to partnerships with OpenAI and Mistral, while pushing forward with AI integration in Windows and Surface devices. Suleyman's appointment signals Microsoft's ambition to pioneer AI advancements accessible to all users.

Stability AI brings a new dimension to video with Stable Video 3D

Stability AI introduces Stable Video 3D (SV3D), enhancing its generative AI model lineup. SV3D allows for 3D video rendering, expanding on Stable Video technology's capabilities. It generates multi-view 3D meshes from a single image input, catering to gaming and e-commerce applications. With a monthly fee for commercial use, SV3D targets creators and developers, while non-commercial users can access model weights for free. 

SV3D represents a leap forward in 3D generation, providing coherent views from any angle and optimizing 3D meshes. Its variants, SV3D_u and SV3D_p, offer flexibility in generating orbital videos and accommodating camera path specifications.

Google DeepMind’s New AI Model Can Help Soccer Teams Take the Perfect Corner

Google DeepMind's TacticAI, an AI model, aids soccer teams by predicting corner kick outcomes and suggesting strategic adjustments. Developed with Liverpool FC, it utilizes player-tracking data to analyze player positions and movements during corners. TacticAI predicts ball contacts and offers recommendations to maximize scoring chances for attackers or minimize them for defenders. It streamlines video analysis, highlighting critical players and weaknesses in defensive setups. 

The model enhances coaches' efficiency, potentially refining corner routines during matches or aiding pre-match preparation. TacticAI aims to establish a common language for describing corner patterns and could evolve into a natural language interface for coaches.

Databricks acquires Lilac to supercharge data quality efforts for gen AI apps

Databricks acquires Lilac, a Boston-based startup specializing in data understanding and manipulation tools. Lilac's technology will enhance Databricks' data intelligence platform, aiding users in improving dataset quality for large language model (LLM) applications. By integrating Lilac's team and solutions, Databricks aims to streamline data exploration and refinement processes. 

This acquisition underscores Databricks' commitment to providing end-to-end tooling for developing high-quality generative AI applications using proprietary data. The move aligns with industry trends toward comprehensive AI solutions and enhances Databricks' competitive position in the market.

Nvidia's GTC 2024 Keynote Highlights

Nvidia's GTC 2024 event has grown into a massive AI expo, drawing 16,000 attendees and 300,000 online viewers. Analysts highlight Nvidia's prowess in accelerated computing, essential for AI. Here are some of NVIDIA’s biggest announcements.

Nvidia aims for AI accessibility via cloud partnerships, challenging industry norms. Despite AI's transformative potential, its monetization remains uncertain. Nvidia leads with its AI Foundry ecosystem, though others struggle to grasp its economic impact. Expect more partner announcements focusing on sustainability and GPU-powered efficiency. Follow for updates on GTC developments.

🧠RESEARCH

Latent Adversarial Diffusion Distillation (LADD), speeds up image synthesis while overcoming limitations of prior approaches like ADD. LADD leverages pretrained latent diffusion models for simpler training and better performance. Applied to Stable Diffusion 3 (8B), it yields SD3-Turbo, matching top text-to-image generators with just four sampling steps.

PERL, or Parameter Efficient Reinforcement Learning from Human Feedback, enhances alignment of Large Language Models (LLMs) with human preferences. Leveraging Low-Rank Adaptation (LoRA), it offers comparable performance to conventional Reinforcement Learning from Human Feedback (RLHF) but with faster training and reduced memory usage. PERL introduces two new preference datasets for further research in RLHF.

Larimar introduces a brain-inspired architecture for Large Language Models (LLMs) featuring distributed episodic memory. This memory enables efficient one-shot updates of knowledge without costly re-training. Results demonstrate Larimar's accuracy matching competitive baselines while offering 4-10x speed-ups and flexibility. Mechanisms for selective forgetting and input context length generalization are also proposed.

LLaVA-UHD introduces a versatile large multimodal model capable of efficiently perceiving high-resolution images in any aspect ratio. Overcoming limitations of previous models, it employs innovative strategies including image modularization, compression, and spatial schema. Outperforming established models on multiple benchmarks, LLaVA-UHD achieves superior accuracy with significantly less data and computational resources.

MVEdit presents a novel approach for open-domain 3D object synthesis by adapting multi-view diffusion with controlled editing. Leveraging ancestral sampling and a training-free 3D Adapter, it ensures 3D consistency and high-quality textured meshes. Achieving superior trade-off between speed and quality, MVEdit demonstrates state-of-the-art performance in various 3D generation tasks.

🛠️TOP TOOLS

DataStax - simplify the development of RAG (Retrieval-Augmented Generation) applications for GenAI developers

Adsby - AI co-pilot for Google search ads

AdGen AI - AI-powered creative generator that crafts irresistible ad copy and visuals, perfectly tailored to multiple ad channels

Powder - Compelling short-form videos from your streams 

Hyperplace - Create dedicated places for your interests, collect relevant ideas and discover how they come together in meaningful new ways

📲SOCIAL MEDIA

🗞️MORE NEWS

Hippocratic hits $500mn valuation as tech investors seek new bets in AI

Healthcare start-up Hippocratic AI has reached a valuation of $500 million following a $53 million fundraising round led by General Catalyst and Premji Invest. The company is developing AI 'agents' to assist hospitals and clinics, with a focus on low-risk, non-diagnostic, patient-facing healthcare. Hippocratic has partnered with Nvidia to reduce latency in AI agent responses, creating a real-time interaction experience for patients. Investors are increasingly looking to invest in AI start-ups with narrower focuses, as Big Tech companies have dominated the AI investment landscape. FINANCIAL TIMES

The new Inflection: An important change to how we’ll work

Inflection, based in Palo Alto, unveils plans to focus on AI studio business, offering its personal AI model, Pi, to developers and enterprises. Co-founders Mustafa and Karén depart to launch Microsoft AI division. Sean White steps in as CEO. Pi remains accessible, emphasizing privacy. Inflection-2.5 to be hosted on Microsoft Azure, expanding reach. INFLECTION

Saudi Arabia Plans $40 Billion Push Into Artificial Intelligence

Saudi Arabia plans to establish a $40 billion fund for investing in artificial intelligence (A.I.), aiming to become a dominant force in the tech market. Partnering with top VC firm Andreessen Horowitz, the initiative signals the nation's economic diversification and global influence ambitions, backed by its vast wealth. THE NEW YORK TIMES

SAP and Nvidia Announce Deal to 'Fine Tune' AI Models

SAP and Nvidia join forces to enhance AI models for enterprise applications, leveraging SAP's extensive access to corporate data and Nvidia's technology. SAP aims to fine-tune large language models for specific industries, utilizing Nvidia's NIM microservices platform for improved connectivity and automation. The collaboration anticipates transformative impacts on business processes. BARRONS

Our progress on generative AI in health

Google Health showcases progress in generative AI for healthcare, unveiling MedLM for Chest X-ray and Gemini model advancements. Research demonstrates state-of-the-art performance in medical reasoning tasks. Collaborations with Fitbit aim to personalize health coaching. AI systems like AMIE assist clinicians, potentially revolutionizing healthcare. Efforts focus on overcoming complex challenges for improved patient care. GOOGLE

Citizen scientists and AI take a cosmic cruise to discover 430,000 new galaxies

Citizen scientists and AI collaboration uncovers 430,000 new galaxies, including 30,000 rare ring-shaped ones, using Subaru Telescope data. The "GALAXY CRUISE" project enlisted 10,000 volunteers to analyze vast data sets, supported by AI classification. This joint effort expands our understanding of galaxy morphology and the power of AI in astronomical research. SPACE.COM

How AI taught Cassie the two-legged robot to run and jump

Researchers utilize reinforcement learning to teach the two-legged robot Cassie dynamic movements without explicit training for each task. AI enables Cassie to run 400 meters, perform standing long jumps, and adapt to new scenarios through simulation training and task randomization. This approach accelerates learning and enhances robot agility, paving the way for humanoid robots to interact more effectively with the physical world. MIT

What'd you think of today's edition?

Login or Subscribe to participate in polls.

What are MOST interested in learning about AI?

What stories or resources will be most interesting for you to hear about?

Login or Subscribe to participate in polls.

Reply

or to participate.