🚨 Catch up with the AI industry, July 18, 2025

🗞️ Today's Top AI Stories:

MIT AI Model Predicts Nuclear Waste Long-Term Stability

A new AI-powered model developed by MIT researchers can accurately predict the long-term effects of nuclear waste on underground disposal systems. This breakthrough uses high-performance computing software called CrunchODiTi to simulate complex interactions over millennia, matching real-world experimental results from Switzerland. This capability is crucial for validating the safety of geological repositories and building public trust in nuclear waste management solutions.

MIT's "Smart Coach" Helps LLMs Master Code and Text Switching

Researchers at MIT have created "CodeSteer," a "smart coach" LLM that guides larger language models to seamlessly switch between textual reasoning and code generation. This innovation significantly boosts accuracy on computational tasks like multiplication and Sudoku, even enabling less sophisticated LLMs to outperform more advanced ones. Inspired by human trainers, CodeSteer helps LLMs select the most effective method for problem-solving.

OpenAI Reflections: A First-Hand Look at Its Rapid Growth

A former OpenAI insider shares reflections on the company's culture and rapid expansion, growing from 1,000 to over 3,000 people in a year. The post highlights a bottoms-up, meritocratic research environment where good ideas often emerge from individual initiatives. It also notes the challenges of scaling, from communication breakdowns to varied team dynamics, and reveals the company's reliance on Slack for all internal communication.

Human Brain's Astonishing Data Compression Beats AI

New research reveals the human brain's remarkable efficiency in data compression, far surpassing current AI capabilities. Our brains can discard irrelevant information while retaining only essential components, allowing for incredibly fast and adaptive learning. This fundamental difference in processing highlights a key area where biological intelligence still holds a significant advantage over even the most advanced artificial neural networks.

NVIDIA Navigates US Politics Amidst Trump's AI Chip Ambitions

NVIDIA is increasingly finding itself at the intersection of geopolitics and AI, as the company navigates the complexities of US policy regarding AI chip exports to China, particularly under potential new administrations. The article explores how the US government's stance, including the possibility of a Trump presidency, could impact NVIDIA's strategies and the broader global competition in AI hardware.

OpenAI and Google's Cloud Battle Heats Up for AI Dominance

The competition between OpenAI and Google Cloud is intensifying as both tech giants vie for supremacy in the enterprise AI market. OpenAI's strong push for broader adoption of its models is directly challenging Google's cloud services, which are heavily investing in AI infrastructure and offering their own powerful models. This rivalry is driving innovation and choice for businesses seeking to integrate advanced AI into their operation

---

🔔 Tooling and model updates:

OpenAI ChatGPT Agent: Introduced a new AI agent mode for Pro, Plus, and Team subscribers that autonomously performs complex tasks like scheduling, competitive analysis, and presentation creation in a virtual environment, significantly boosting productivity.

Google Deep Search & AI-Powered Business Calling: Rolls out Gemini 2.5 Pro and Deep Search for AI Pro and AI Ultra subscribers in Search, and AI-powered business calling for all US Search users, offering advanced reasoning and automating information gathering from local businesses.

Adobe Firefly: Adds new video capabilities, including Generate Sound Effects (beta) and Text to Avatar (beta), allowing users to easily add custom sound effects via text/voice or create avatar-led videos from scripts, enhancing video storytelling.

AWS Amazon Bedrock Knowledge Bases & Amazon S3 Vectors (Preview): Announces Amazon S3 Vectors (preview), the first cloud object storage with built-in support to store and query vectors at a low cost, significantly reducing costs for storing and querying large vector datasets for RAG applications.

---

🎥Video of the day:

François Chollet: How We Get To AGI

Key Takeaway 1: The Shift from Pre-training Scaling to Test-Time Adaptation (TTA) [03:03] Explanation: The AI field is moving beyond just scaling models. True progress in fluid intelligence now lies in test-time adaptation, where models dynamically learn and adapt to novel situations during inference, rather than relying solely on pre-trained knowledge.

Key Takeaway 2: Intelligence as a Process of Invention, Not Just Skill Automation [06:19] Explanation: AGI redefines intelligence as the ability to handle entirely new problems and invent solutions. This crucial shift means focusing on enabling autonomous invention and scientific discovery, not just automating existing tasks or skills.

Key Takeaway 3: The Need for Combining Different Types of Abstraction [25:39] Explanation: Achieving true AGI requires merging value-centric (intuition, pattern recognition) and program-centric (reasoning, planning) abstractions. This integrated approach, guiding discrete program search with deep learning, is key to unlocking invention beyond automation.

---

🛠️ Tool of the day:

(Introduction by Dan the creator)

Lovart AI is introduced as "The World's First Design Agent," a platform that enables seamless collaboration between human designers and AI on a single canvas to create and edit designs using natural language.