TLDR AI 2025-06-09

Free trials are meant to drive adoption—not rack up costs. But for AI apps, trial abuse is a growing threat: fake signups, disposable emails, and automated account farming are draining compute and distorting metrics.

WorkOS Radar helps stop abuse before it spreads. With built-in protection against bots, free trial abuse, brute force attacks, and even impossible travel logins, Radar gives your app the real-time defense it needs.

Protect your app today with WorkOS Radar.

An investment in Scale, already valued at $14 billion, underscores how high-quality training data has become a key competitive differentiator.

Anthropic launched Claude Gov models for U.S. national security customers, designed for strategic planning, operational support, and intelligence analysis. These models excel in handling classified materials, understanding intelligence contexts, and interpreting complex cybersecurity data. Built with feedback from government users, they adhere to strict safety standards while addressing unique national security needs.

Gemini's new scheduled actions feature allows AI Pro and AI Ultra subscribers to ask the assistant to perform tasks at specific times. Subscribers can now ask the assistant to provide daily summaries. Users can also ask Gemini to complete one-tasks, for example, summarizing an award show the day after it airs. Gemini subscribers can manage planned tasks on the 'scheduled actions' page in the Gemini app's settings.

Apple researchers evaluated Large Reasoning Models (LRMs) using custom puzzle environments to study reasoning complexity. They found LRMs collapse at high complexities, with reasoning effort peaking then declining.

Anthropic released detailed case studies showing how 10 internal teams use Claude Code. Claude only works on its first attempt one-third of the time, leading to a "slot machine" approach: commit frequently, let Claude work autonomously, then either accept or restart fresh. The most successful teams emphasize writing detailed Claude.md documentation files and breaking complex workflows into specialized sub-agents for better results.

Of the 18 AI models tested, OpenAI's o3 emerged as the most successful by mastering deception and secretly orchestrating coalitions, including convincing Claude 4 Opus to betray its ally Gemini 2.5 Pro by promising an impossible "four-way draw" before eliminating Claude itself. Gemini 2.5 Pro was the only other model to achieve victory using blitzkrieg-style tactics, while Claude consistently sought peaceful resolutions even when betrayed.

Get your B2B SaaS portcos instant access to millions of tech professionals across our 12 interest-based newsletters. Additional perks (i.e., discounts) are included when you sign up as a VC partner.

Alibaba has open-sourced the Qwen3 Embedding series, with its 8B model achieving #1 on the MTEB multilingual leaderboard. The models enable developers to build RAG systems, semantic search engines, and document retrieval applications across 100+ languages, with options ranging from 0.6B to 8B parameters for different performance needs.

GUI-Actor is a new approach that lets AI agents interact with computer interfaces by using attention mechanisms instead of predicting precise pixel coordinates based on screenshots. The method achieves state-of-the-art results on GUI benchmarks while being more efficient - fine-tuning only 100M parameters while keeping the underlying vision model frozen delivers comparable performance to much larger models.

PTSD trains diffusion models sequentially across temperatures to improve sampling from unnormalized densities.

Hugging Face and collaborators released the Common Pile v0.1, an 8 TB openly licensed dataset for training large language models.

Mistral AI is reportedly closing multiple $100M+ contracts and approaching $100M annual revenue as European companies seek alternatives to US AI providers following Trump's return to office. The "sovereignty play" appears to be working, with Mistral's CEO noting its business tripled in the last 100 days, particularly in Europe and non-US markets.

The two-week competition challenges participants to jailbreak an AI to provide dangerous information, from poison recipes to nuclear detonation instructions.

Google is rolling out interactive financial data visualizations in AI Mode (Labs), enabling dynamic graphs and multi-step reasoning to answer complex stock and mutual fund queries.

Cloudflare's open-sourced OAuth 2.1 library was almost entirely written by Claude, and the company documented its entire creative process through git commit messages.