NO/FOMO

Independent AI signal, once a day

The AI briefing worth opening.

ISSUE DATE2026-01-18DEFAULT EDITION
This issue
—
All time
—

Hacker News

6 stories
01

Gaussian Splatting – A$AP Rocky "Helicopter" music video

The recent music video for A$AP Rocky's track "Helicopter" features a prominent integration of Gaussian Splatting technology, signaling a significant adoption of advanced 3D reconstruction and rendering techniques in mainstream entertainment. Gaussian Splatting, known for its efficiency and quality in synthesizing novel views from sparse input images, showcases its capabilities by creating highly detailed and photorealistic scene representations throughout the video. This application moves the cutting-edge computer vision technique beyond academic research into high-profile creative productions. The successful deployment in a music video highlights the growing potential for generative AI and advanced graphics tools to revolutionize visual effects and immersive content creation, offering a compelling alternative to traditional 3D modeling and rendering pipelines. It marks a pivotal moment for how digital content can be crafted, enabling more dynamic and visually striking experiences across various media forms.

Gaussian Splatting3D ReconstructionComputer VisionGenerative AIVisual EffectsComputer VisionGenerative AIMachine Learning
02

Predicting OpenAI's ad strategy

The article 'Predicting OpenAI's ad strategy' explores potential future monetization avenues for OpenAI beyond its current API services and premium subscriptions. As OpenAI continues to expand its user base and develop more sophisticated AI models, the introduction of advertising could represent a significant shift in its business model. The analysis delves into various hypothetical advertising approaches OpenAI might consider, such as context-aware ads integrated within AI responses, sponsored content, or a freemium model supported by targeted advertising. It considers the ethical implications, user experience impact, and technical challenges associated with implementing an ad-based revenue stream for AI-powered services. The discussion also touches upon how such a strategy could leverage OpenAI's vast data and AI capabilities to create highly personalized and effective advertising solutions, while balancing the need for revenue growth with user privacy and trust. The overall aim is to forecast how a leading AI research and deployment company might navigate the complex landscape of digital advertising.

OpenAIAI MonetizationAdvertising StrategyBusiness ModelGenerative AIArtificial IntelligenceLarge Language ModelGenerative AI
03

Starting from scratch: Training a 30M Topological Transformer

This report details the intricate process of training a 30-million parameter "Topological Transformer," dubbed TauFormer, entirely from scratch. The project investigates the feasibility and challenges inherent in developing novel transformer architectures without leveraging pre-trained weights or extensive foundational models. A core aspect is the integration of specific topological properties into the transformer's design, hypothesizing an improvement in its capacity to process complex data structures and spatial relationships. This endeavor offers critical insights into the computational resources, unique data preparation strategies, and methodological innovations required for bootstrapping advanced deep learning models, especially those exploring non-traditional inductive biases. The findings are expected to significantly contribute to the broader understanding of transformer scalability, architectural experimentation, and the efficacy of specialized modifications within the field of machine learning. The study closely monitors the model's performance and training stability throughout its development lifecycle.

Topological TransformerTransformer ArchitectureDeep LearningMachine LearningModel TrainingMachine LearningDeep LearningArtificial Intelligence
04

Show HN: Figma-use – CLI to control Figma for AI agents

Dan has developed "Figma-use," a command-line interface (CLI) designed to empower AI agents with direct control over Figma for design tasks. This innovative tool offers approximately 100 commands, enabling AI to create and manipulate shapes, text, frames, components, modify styles, and export assets within Figma. A key feature is its JSX importing capability, which is significantly faster—around 100 times—than existing Figma plugin APIs, and it seamlessly integrates with any Large Language Model (LLM) coding assistant. The project addresses a critical limitation where the official Figma API is primarily read-only, preventing AI from actively designing. Figma-use allows AI to genuinely create design elements like buttons, build layouts, and generate entire component systems, overcoming the verbose JSON schema issue of prior solutions that consumed excessive tokens. Built with a tech stack including Bun, Citty, an Elysia WebSocket proxy, and a custom Figma plugin, it leverages Chrome DevTools for enhanced performance when handling numerous objects.

FigmaCLIAI agentsDesign automationLLM integrationArtificial IntelligenceAI AgentLarge Language Model
05

Erdos 281 solved with ChatGPT 5.2 Pro

A recent social media post by Neel Somani has ignited discussions within the artificial intelligence community, asserting that a speculative 'ChatGPT 5.2 Pro' has successfully resolved Erdos Problem 281. While the brevity of the original tweet offers minimal supporting details, the declaration, if validated, would signify an extraordinary milestone in AI's capacity for advanced mathematical reasoning and problem-solving. Erdos Problem 281, an unproven conjecture in number theory, presents a formidable challenge, and its solution by an AI system would showcase an unprecedented level of sophisticated computational and theoretical understanding. The claim has prompted considerable debate regarding the current and future potential of large language models to address complex, long-standing open mathematical problems, emphasizing the crucial need for rigorous methodology and independent verification to substantiate such a profound scientific assertion. The AI research landscape is now poised for further insights and peer-reviewed confirmation of this groundbreaking development.

ChatGPTLarge Language ModelMathematical Problem SolvingErdos ProblemsArtificial IntelligenceArtificial IntelligenceLarge Language ModelAI Agent
06

Show HN: GibRAM an in-memory ephemeral GraphRAG runtime for retrieval

GibRAM, or Graph in-buffer Retrieval and Associative Memory, is an experimental, in-memory GraphRAG runtime designed to enhance information retrieval for complex, regulation-heavy documents. It addresses a critical limitation of traditional flat RAG pipelines, which often struggle to identify and retrieve interconnected content such as references, definitions, or clauses. The developer, finding existing GraphRAG implementations cumbersome due to their reliance on separate systems for graph storage and vector indexing, created GibRAM to integrate these components. This innovative solution unifies entities, relationships, text units, and embeddings within a single process, making it particularly suitable for short-lived analysis tasks that demand efficiency and minimal overhead. GibRAM's ephemeral nature streamlines the handling of interconnected data, providing a more effective and contextually aware framework for data analysis in domains like regulatory compliance.

GraphRAGRetrieval Augmented GenerationIn-memory computingEphemeral runtimeVector indexingArtificial IntelligenceNatural Language ProcessingMachine Learning