What Top Execs Read Before the Market Opens

The Daily Upside was founded by investment professionals to arm decision-makers with market intelligence that goes deeper than headlines. No filler. Just concise, trusted insights on business trends, deal flow, and economic shifts—read by leaders at top firms across finance, tech, and beyond.

Join 1M+ professionals who stay informed before the rest of the market catches up.

Today:

• Sakana AI Redefines Reinforcement Learning

• Court Backs Anthropic On Fair Use

• OpenAI Builds AI Office Suite

• DeepMind Unveils On-Device Robotics AI

• Thinking Machines Builds Custom AI

Sakana AI New Model Sparks a RL Revolution

Sakana AI’s new open-source study flips reinforcement learning on its head. Instead of rewarding a student model for right answers, it rewards a small “teacher” model for explanations that help a larger student solve problems. 

A 7-billion-parameter teacher beat much bigger models on math and science tests, cut training time from months to one day and slashed costs. The approach could let even cheap hardware build sharper, more general reasoning systems.

WATCH THE VIDEO ON YOUTUBE

Court Backs Anthropic On Fair Use

A US judge ruled that Anthropic could legally train its Claude AI on copyrighted books because the training counts as “fair use” — a legal rule that lets limited copying if it helps create new work — yet storing seven million pirated books still broke copyright. The court will decide damages in December. This is the first big judgment on fair use for generative AI and shapes how tech firms gather data going forward.

Why this matters

1. Sets a clear early legal guide for how AI models may learn from books, steering future court fights.
   

2. Shows that copying sources must be lawful; keeping illegal files can still cost millions.
   

3. Influences data access costs and speed, which could tilt competitive balance among AI labs.

OpenAI Builds AI Office Suite

OpenAI is building new ChatGPT features for making and editing documents together, similar to Google Workspace and Microsoft Office. The plan adds shared editing, built-in chat, a browser, hardware device, and social feed, turning ChatGPT into a full work hub. This move may strain its partnership with Microsoft, which owns 49%, as both firms renegotiate terms. No launch dates were given, but the suite could challenge today’s top software leaders.

Why it matters

• Competitive jolt – Adding full office tools to ChatGPT challenges Google and Microsoft’s grip on daily work software, pushing everyone to speed up AI upgrades.
  

• Ecosystem control – Owning writing, sharing, and browsing in one place (a connected app family, or “ecosystem”) lets OpenAI collect better data to train models and open fresh ways to earn money.
  

• Partner tension – A direct Office rival tests the unusual Microsoft-OpenAI alliance, hinting that future big AI partnerships may face more overlap and conflict.

DeepMind Unveils On-Device Robotics AI

Google DeepMind unveiled Gemini Robotics On-Device, a vision-language-action model that runs entirely on a robot’s hardware, eliminating network delays or outages.It follows plain-language commands and lets two-arm robots unzip bags, fold clothes, zip lunchboxes, draw cards, pour dressing and assemble items.Developers can customize the model with just 50-100 demonstrations, enabling quick learning, private data handling and broader everyday use. The launch shows Google’s push toward edge-AI autonomy.

Why it matters

• Local brains – Proves powerful AI can live inside the machine, not the cloud, cutting delays and keeping data private.
  

• Fast learners – Needs only 50-100 human demos to master new jobs, showing progress toward data-efficient training.
  

• Toward useful robots – Combines natural-language understanding with skilled two-arm control, inching AI closer to helpful home and factory robots.

🧠RESEARCH

Drag-and-Drop LLMs: Zero-Shot Prompt-to-Weights

"Drag-and-Drop LLMs" create task-specific model updates instantly from just a prompt—no training needed. This method skips the usual tuning process and delivers results much faster and more efficiently. It works across various tasks like coding and reasoning, even without seeing labeled data, showing strong accuracy and flexibility.

Light of Normals: Unified Feature Representation for Universal Photometric Stereo

This paper introduces a new method for photometric stereo—reconstructing 3D surface shapes from light and shadow. It tackles two major problems: separating lighting effects from surface shape and preserving fine surface details. The approach works well under any lighting, capturing complex textures more accurately than previous techniques.

OmniGen2: Exploration to Advanced Multimodal Generation

OmniGen2 is an open-source AI model that can generate both text and images using two separate decoding systems. It handles tasks like image editing and subject-specific generation without losing its ability to produce clear text. Despite its smaller size, it rivals larger models in performance and supports future research with public tools.

Phantom-Data : Towards a General Subject-Consistent Video Generation Dataset

Phantom-Data is a new dataset that improves how AI generates videos of specific subjects by keeping identity consistent across different scenes. It avoids common issues in current methods by using cross-scene image pairs. This leads to better video quality and more accurate results when following text prompts.

Vision-Guided Chunking Is All You Need: Enhancing RAG with Multimodal Document Understanding

This paper introduces a smarter way to break down complex PDF documents using vision-enabled AI models. Unlike standard text methods, it handles tables, visuals, and page-spanning content more accurately. The result is better document understanding and improved performance in AI systems that answer questions based on retrieved information.

🛠️TOP TOOLS

LogoAI - Creating professional logos, cohesive visual identities, and automating brand promotion.

BoredHumans - Offers over 100 free AI tools, designed to cater to a wide range of creative, entertainment, and productivity needs.

Duplichecker - Designed to ensure the originality and integrity of written content.

Mnml AI - AI-powered design assistant specifically designed for architects, interior designers, and related professionals.

Getimg - AI-driven image generation and editing tool designed to create high-quality visuals from text descriptions.

📲SOCIAL MEDIA

🗞️MORE NEWS

• Former OpenAI CTO Mira Murati’s startup, Thinking Machines Lab, will build custom AI tools for businesses using reinforcement learning and open-source model components, aiming for faster, cheaper development of company-specific solutions.

• Microsoft introduced Mu, a tiny, fast AI model built for on-device use in Windows Settings. Running on Copilot+ PCs’ NPUs, Mu translates natural language into system actions with high speed, low memory use, and strong accuracy.

• Court documents reveal OpenAI and Jony Ive’s team are exploring a pocket-sized AI device—not a wearable—still a year from launch. The project sparked a trademark lawsuit over similarities with another in-ear tech startup, iyO.

• Scale AI exposed private client and contractor data in unsecured Google Docs, including confidential projects for Meta, Google, and xAI. Contractors' personal details, pay info, and flagged behavior were also publicly accessible, raising major cybersecurity concerns.

• Amazon is building a 1,200-acre AI data center in Indiana to support Anthropic, using 2.2 gigawatts of power and millions of gallons of water annually. It’s part of a growing trend of supersized infrastructure for AI.