Jan 23, 2025
Welcome to Edition 2 of Fine-Tuned by Genloop – your guide to the evolving world of LLM customization. The past two weeks have been eventful. Donald Trump has returned to the presidency, new policies on US citizenship are emerging, the LLM leaderboard is shifting, OpenAI's massive "Project Stargate" has been announced, and NVIDIA has unveiled its latest GPUs. In this edition, we break down these developments, highlight top research breakthroughs, and share our thoughts.
AI Industry Highlights
1. OpenAI’s Stargate: The $500B Bet on AI Infrastructure
Sam Altman and President Trump have announced Project Stargate, a historic $500 billion initiative to build AI infrastructure across the United States over the next four years. This marks the largest private AI investment to date.
Key Players and Contributions:
OpenAI: Responsible for AI model development.
SoftBank: Led by Masayoshi Son, handling financial structuring.
NVIDIA: Providing the chips.
Oracle: Assisting in system construction and operations.
Microsoft: Ensuring OpenAI continues to leverage Azure.
Arm: Participating as a technology partner.
2. DeepSeek R1: Advancing AI Reasoning with Reinforcement Learning
DeepSeek AI has introduced its DeepSeek-R1-Zero and DeepSeek-R1 models, pioneering reinforcement learning (RL) without supervised fine-tuning (SFT). The models exhibit remarkable reasoning capabilities, although early versions faced readability challenges and language mixing issues.
Key points:
DeepSeek-R1-Zero: Trained purely via RL, demonstrating emergent reasoning behaviors.
DeepSeek-R1: Incorporates multi-stage training and cold-start data to enhance reasoning performance.
Performance: Matches OpenAI-o1-1217 on reasoning tasks.
Open-Sourced: Includes six distilled models (1.5B to 70B parameters) based on Qwen and Llama.
3. Apple Faces AI Challenges in Notifications
Apple has temporarily paused AI-generated notification summaries for news and entertainment apps due to backlash over inaccuracies. The move comes after an incident where the BBC reported misleading information due to AI summarization errors.
Apple will refine its summarization model before reintroducing the feature in a future iOS update.
The controversy highlights the difficulties of deploying small language models (SLMs) for real-world applications.
Well, this difficulty in getting small language models to work is precisely why companies like ours exist.
4. NVIDIA Unveils New GPUs and AI Innovations
At CES 2025, NVIDIA announced groundbreaking hardware and AI platforms:
RTX 5090, 5080, 5070 Ti, and 5070 GPUs: Built on the Blackwell architecture, priced between $549 and $1,999.
GB10 AI Superchip: Designed for AI applications and humanoid robotics.
NVIDIA Cosmos™: A new platform enabling real-world simulation for AI systems, promoting advancements in robotics and autonomous vehicles.
Featured Blog Posts
We've got two fascinating reads that showcase how the AI landscape is evolving
1. eBay's e-Llama: Bringing AI to e-Commerce
eBay has taken an innovative hybrid approach to implementing LLMs in e-commerce, leveraging its vast marketplace data spanning 190 global markets. They're developing both their own LiLiuM family of models and adapting existing ones like Meta's Llama. Their latest creation, "e-Llama," comes in 8-billion and 70-billion parameter versions, specifically tuned for e-commerce applications. What's particularly interesting is their strategic decision to maintain both in-house and adapted models, allowing them to balance control and performance.
2. Domain Memory Agents Rise Amid US AI Export Controls
Read the complete article here
The United States' former administration proposed to expand AI export restrictions on closed general models and high-performance computing chips. These regulations create a three-tier system:
Tier 1 (Including Australia, Japan, Taiwan, the UK, and most of Europe): Full access with allied computing requirements
Tier 2 (Israel, Saudi Arabia, Singapore): Gradual access increases over time
Tier 3 (China, Russia, Iran, North Korea): Complete restriction from advanced AI technology
This is particularly significant as it affects models exceeding 1 trillion parameters and introduces computational constraints that could reshape the global AI landscape. Interestingly, open-weight models remain unrestricted, potentially boosting open-source AI development.
Research Corner
Our team has been diving deep into groundbreaking research papers, and two particularly caught our attention:
1. Mind Evolution: DeepMind's Leap in LLM Thinking
Google DeepMind has introduced a game-changing approach to how LLMs process information. Their "Mind Evolution" concept combines:
Divergent thinking: Exploring multiple solutions in parallel
Convergent thinking: Carefully evaluating and refining ideas through another LLM
The results are impressive. Using this approach, Gemini 1.5 Flash achieved:
95.6% success on TravelPlanner tasks
85% success in Meeting Planning challenges
Gemini 1.5 Pro pushed these numbers even higher, reaching 100% and 98.4% respectively
What makes this special? Unlike previous approaches requiring formal problem specifications, Mind Evolution works directly with natural language, making it immediately applicable to real-world challenges.
2. NVIDIA's Cosmos: Revolutionizing Physical AI Systems
NVIDIA has unveiled its Cosmos World Foundation Model Platform, and it's generating serious buzz for good reason. This platform could fundamentally change how we develop AI systems that interact with the physical world.
Key highlights that caught our attention:
Safe Learning Environment: Cosmos creates digital twins of the physical world where AI systems can learn without risk. Imagine robots practicing complex tasks thousands of times without real-world consequences or costs.
Smart Development Pipeline: They've implemented a clever two-step approach - first training a general model on vast video datasets, then fine-tuning it for specific tasks. What previously took months or years of real-world training can now be accomplished much faster.
Democratic Access: Perhaps most exciting is NVIDIA's decision to make this platform available through open-source licensing. This means everyone from startup innovators to established research labs can build on this technology.
Our take? While there's still work to be done on physics simulation accuracy, Cosmos marks a significant milestone in making physical AI systems more practical and accessible. The enthusiastic response at CES suggests we're looking at a potential game-changer for robotics and autonomous systems development.
Looking Forward
The intersection of political changes, technological advancements, and research breakthroughs is creating an exciting landscape for AI development. Thank you for reading! Share your thoughts with us on LinkedIn or Twitter, and don't forget to subscribe to stay updated on the latest in LLM customization.
About Genloop
Genloop delivers customized LLMs that provide unmatched cost, control, simplicity, and performance for production enterprise applications. Please visit genloop.ai or email founder@genloop.ai for more details.