AI News — Friday, September 18, 2026

10
Crusoe raises $3.9B to build massive data centers and small modular 'AI factories'

Crusoe Energy Systems has secured $3.9 billion to significantly expand its infrastructure for AI, focusing on building large data centers and smaller, modular 'AI factories' to meet growing computational demands.

TechCrunchindustry
9
Google DeepMind launches institute to widen the AGI debate

Google DeepMind has established a new institute dedicated to fostering a broader public and academic discussion around the development and implications of Artificial General Intelligence (AGI).

TechCrunchindustry
8
PrismML hopes its tiny LLM will change how we all use AI

PrismML is developing a compact Large Language Model (LLM) with the ambition of revolutionizing everyday AI interactions through its efficiency and accessibility.

TechCrunchproduct
8
LimiX-2: A Contextual Mechanism Network Towards General Structured-Data Intelligence

Researchers introduce LimiX-2, a novel contextual mechanism network designed to advance general intelligence in processing and understanding structured data.

Hugging Faceresearch
8
Introducing Astra for Law

OpenAI announces Astra for Law, a new specialized AI product designed to assist legal professionals with various tasks, enhancing efficiency and accuracy in the legal sector.

OpenAI Blogproduct
7
Reimagining advertising with AI

OpenAI explores how AI can fundamentally transform the advertising industry, offering new tools and strategies for more effective and personalized campaigns.

OpenAI Blogindustry
7
How Cooley is accelerating IPO work with ChatGPT

Law firm Cooley is leveraging ChatGPT to streamline and accelerate the complex processes involved in initial public offerings (IPOs), demonstrating practical AI application in legal services.

OpenAI Blogindustry
7
Rethinking Critic Learning in PPO: Understanding and Mitigating Value Flattening

This research paper delves into the challenges of critic learning in Proximal Policy Optimization (PPO) and proposes methods to understand and alleviate the issue of value flattening.

Hugging Faceresearch
6
Confidence Comes from Experience: Experiential Confidence Estimation from Reasoning to Agents

A new paper explores how AI agents can develop and estimate their confidence based on past experiences, improving their reasoning and decision-making capabilities.

Hugging Faceresearch
6
ProgramDistill: From Interactive Web Apps to Verifiable Reference-Guided SWE Tasks

ProgramDistill introduces a method to convert interactive web applications into verifiable, reference-guided software engineering tasks, enhancing automated code generation and verification.

Hugging Faceresearch
6
Making global data easier to explore

Google is enhancing its AI capabilities to make global datasets more accessible and easier to explore, facilitating insights and research across various domains.

Google AI Blogproduct
6
Building AI to accelerate science and improve lives

Google highlights its ongoing efforts to develop AI applications specifically designed to accelerate scientific discovery and improve human well-being across diverse fields.

Google AI Blogindustry
6
AI for everyone in every language

Google is committed to making AI universally accessible by expanding its capabilities to support every language, aiming for inclusive AI experiences worldwide.

Google AI Blogproduct
5
Agora: Git as Shared Memory for Collective AutoResearch

Agora proposes using Git as a shared memory system to facilitate collective autonomous research, enabling multiple AI agents to collaborate more effectively on scientific discovery.

Hugging Faceresearch
5
Show a model your old code and it writes your old bugs: 32 runs, 0% reuse

An experiment reveals that AI coding models, when exposed to old code, tend to replicate existing bugs rather than learning from them, highlighting challenges in code reuse and error propagation.

Dev.toresearch