TLDR AI 2025-07-29

The new wave of AI tools leads with PLG. If you're not offering an easy way into your product, you're not gonna make it. But being overly-generous with expensive inference and API calls can jeopardize your business's viability.

Metronome's new blog offers a framework for choosing between 5 different PLG billing models — PayGo, threshold billing, prepaid credits, and more — based on company safeguards vs customer adoption barriers.

The blog also looks at an example of a hybrid billing model from Snowflake and dives into the importance of using real-time signals to prevent revenue leakage and customer churn.

Read the blog →

Microsoft introduced Copilot Mode in Edge, an experimental AI-driven feature designed to transform traditional web browsing. It assists users by predicting next steps, summarizing content, and reducing tab overload to streamline interaction with the web.

Startup Z.ai (formerly Zhipu) launched GLM-4.5, claiming it costs less to operate than DeepSeek due to its smaller size and ability to decompose tasks into sub-goals. The startup, backed by the Chinese government, previously gained OpenAI's attention for securing significant contracts in China despite competition from larger firms.

Tesla secured a $16.5 billion agreement with Samsung to manufacture its AI6 chips at a new Texas facility.

xAI is preparing to introduce several major updates to its Grok app. One substantial addition is the Imagine feature, which unlocks Grok's new generative AI models for images and videos. The relatively few restrictions on Grok's image and video generation models could spark viral use once rollout widens. There is no firm launch date, but it is likely xAI will time the rollout to coincide with the release of GPT-5.

Researchers tested 19 LLMs across 76,977 participants and found that specialized post-training techniques boosted AI persuasiveness by up to 51%, far exceeding gains from simply using larger models. The most effective strategy was prompting models to deploy dense factual information, but this came with a troubling tradeoff: the most persuasive AI conversations were the most inaccurate.

Writer's Action Agent is a researcher, a strategist, a data scientist, and a programmer. Give it a goal and it puts together a multi-phase plan, then executes to deliver exactly what you need: reports, slide decks, spreadsheets, and more. Give Action Agent a project to tackle today -- try the beta for free.

mcp-context-protector is a security wrapper for LLM apps using the Model Context Protocol (MCP). It defends against line jumping attacks with features designed to help expose malicious server behavior. The wrapper proxies tool calls and other operations to the downstream server, making it compatible with any MCP server and any MCP-compliant host app. Developers are able to add any sort of control suited to their organization's use case.

Efficiency Leverage (EL) is a metric that quantifies the computational advantage of Mixture-of-Experts models. Based on experiments with over 300 models, these authors present scaling laws that predict EL using MoE configuration parameters.

System design and proper software engineering are required for building effective AI agents. Developers need to focus on clear instructions, lean context management, robust tool interfaces, and automated validation loops instead of finding a silver bullet of a great prompt or an advanced framework. Error analysis is paramount - models should be used to help developers understand where agents fail so that those failure modes can be systematically addressed. The goal is to create reliable and recoverable agents that fail gracefully and can be improved iteratively.

Ads gave us an Internet that was mostly free and open. AI has introduced a new problem for content creators: a new wave of bots crawling sites for content to train AI, often taking value and offering nothing in return. This is causing content creators to put up walled gardens or bot protections. The prevalence of AI-generated content is causing people to seek more authentic spaces, resulting in a cozy web of niche communities.

Harmonic's AI model, Aristotle, offers 'hallucination-free' answers for questions involving mathematical reasoning. The startup is focused on creating mathematical superintelligence and eventually wants to help users in all fields that rely on math. It will eventually release an API to let enterprises access Aristotle, as well as a web app for consumers. Aristotle achieved gold medal performance on the 2025 International Math Olympiad through a formal test, meaning the problems were translated into a machine-readable format.

Google Chrome is rolling out store reviews in the U.S., using AI to summarize feedback from Google Shopping and popular review sites.

Anthropic will impose weekly rate limits starting August 28 for Claude Pro and Max subscribers to prevent misuse and manage intensive usage patterns.

Microsoft's Copilot Appearance is a real-time animated character that responds with facial expressions during conversations.