Back to MergeSort

AI News Digest - May 22, 2026

15 stories · May 22, 2026

Listen to the podcast
1.6 MB · Download MP3

Top Stories

research

OpenAI solves the unit distance problem in mathematics

OpenAI has reportedly solved the unit distance problem in mathematics, leading to discussions about whether the solution utilized informal chain-of-thought reasoning or neurosymbolic tools like Lean, and prompting a shift in focus to AI's potential in social science domains.

Sources: Digg AI, Latent Space

other

Microsoft discontinues internal use of Anthropic's Claude coding tool due to high costs

Microsoft ceased internal deployment of Anthropic's Claude for coding assistance, citing unsustainable token-based billing expenses. This highlights the significant operational costs associated with large language models even for major tech companies.

Sources: Digg AI, Hacker News

research

CODA: New GPU Kernel Abstraction Enhances Transformer Training Efficiency

Researchers have introduced CODA, a novel GPU kernel abstraction that reparameterizes memory-bound transformer operations such as norms and RoPE. This innovation allows these operations to execute within GEMM epilogues while output tiles remain on-chip, significantly reducing kernel launch overhead and memory bottlenecks. CODA promises enhanced GPU efficiency and improved training performance for AI models.

Sources: Digg AI, Hacker News

research

OpenAI model disproves 80-year-old Erdős unit distance conjecture

OpenAI's unreleased reasoning model found a counterexample to the Erdős unit distance conjecture, a discrete geometry problem from 1946, proving it false. This demonstrates a cleaner test of AI reasoning as the proof survived expert review, including verification by external mathematicians.

Sources: The Neuron

funding

Startup raises $64M to develop AI for formal mathematical verification

A 24-year-old founder secured $64 million in funding to create AI for formal mathematical verification, a method designed to eliminate hallucinations by ensuring provably correct answers.

Sources: The Neuron

More Stories

research

OpenAI's Internal AI Model Discovers New Mathematical Constructions for Erdős's Unit Distance Problem

OpenAI has reported that an internal reasoning model, without specific mathematical training, identified new super-linear constructions for a long-standing problem in mathematics, demonstrating advanced problem-solving capabilities in AI.

Sources: Digg AI

funding

Hark Secures Over $700 Million Series A Funding for AI Foundation Models and Custom Hardware

AI company Hark raised over $700 million in Series A funding, achieving a $6 billion post-money valuation, with investments from Parkway Venture Capital, NVIDIA, and AMD, to accelerate the development of its AI foundation models and custom hardware.

Sources: Digg AI

product_launch

Thinking Machines Lab Unveils TML-Interaction-Small for Real-Time Multimodal Conversation

Thinking Machines Lab is developing TML-Interaction-Small, a multimodal system designed for concurrent processing of audio, video, and text inputs and outputs, enabling real-time, interruptible conversations. This system aims to significantly reduce conversational latency and enhance natural interaction by interleaving input and output in micro-turns.

Sources: The Batch

product_launch

Google Unveils Gemini 3.5 Flash AI and Antigravity 2.0, Demonstrating Advanced 3D CAD Capabilities with Increased API Pricing

Google has launched its Gemini 3.5 Flash AI model and the Antigravity 2.0 agent-first desktop application, with the new AI model showing strong autonomous performance in 3D CAD generation benchmarks, uniquely creating complex architectural details. While the API pricing for Gemini 3.5 Flash has tripled, benchmarks also highlighted the continued precision benefits of human-in-the-loop workflows for intricate tasks.

Sources: Hacker News

research

Dharma-AI Research: Specialized 3B Model Surpasses Commercial Frontier APIs in OCR Quality and Cost

New research from Dharma-AI reveals that a 3-billion-parameter specialized model significantly outperformed leading commercial frontier APIs, including Claude Opus 4.6 and GPT-5.4, on a Brazilian Portuguese OCR benchmark. The specialized model achieved superior extraction quality, fifty times lower operational cost, and enhanced production stability, challenging the prevailing "scale is best" assumption.

Sources: Hugging Face Blog

product_launch

Amazon Nova Act Achieves HIPAA Eligibility, Enabling AI Automation in Healthcare

Amazon Nova Act, an AWS service for building and managing AI agents for UI workflow automation, is now HIPAA eligible. This allows healthcare and life sciences organizations to deploy autonomous AI agents to automate complex, browser-based workflows involving electronically protected health information (ePHI), such as claims processing and appointment scheduling, while maintaining compliance.

Sources: AWS ML Blog

product_launch

AWS Launches Conversational AI Assistant for Natural Language Cloud Operations

AWS has unveiled a new conversational AI assistant, leveraging Amazon Bedrock AgentCore Runtime with Model Context Protocol (MCP) and the AWS API MCP Server. This solution translates natural language queries into AWS CLI commands, enabling SREs and DevOps engineers to streamline cloud operations with enhanced security and compliance. The AWS API MCP Server is now available on AWS Marketplace, facilitating seamless interaction with AWS services.

Sources: AWS ML Blog

product_launch

AWS Launches AI-Powered Multi-Agent System for Natural Language BI Dashboard Modification

AWS has introduced a new AI-powered multi-agent system designed to enable self-service modification of Business Intelligence (BI) dashboards using natural language. Built on Amazon Bedrock AgentCore and leveraging the Strands Agents framework, this solution integrates with Amazon Quick to streamline dashboard operations through specialized agents. It aims to significantly enhance BI capabilities and accelerate decision-making by allowing users to intuitively manage and update their dashboards.

Sources: AWS ML Blog

other

OpenAI and Anthropic show strong revenue growth, with Anthropic projected to reach $10.9B in Q2

OpenAI reportedly generated $5.7 billion in Q1, outpacing Anthropic, which is projected to more than double its revenue to $10.9 billion in Q2. This indicates rapid financial expansion for both leading AI companies.

Sources: The Neuron

policy

California signs first-in-the-nation order to prepare workforce for AI disruption

California has enacted a pioneering order aimed at preparing its workforce and small businesses for the potential disruptions caused by artificial intelligence. This policy initiative is a proactive step to address the economic and social impacts of AI.

Sources: The Neuron