TLDR AI 2025-08-19

AI has broken traditional pricing models. And if you're still charging for seats, you're NGMI. But very few companies have found an accurate and sustainable pricing model for AI value.

Metronome's new and highly anticipated whitepaper offers a practical framework for navigating this shift, based on Metronome's work with AI leaders including OpenAI, Databricks, and NVIDIA. It tackles hard questions such as:

• Understanding what you're actually metering (agent invocations, tokens, etc.)
• The infrastructure needed to support value-based pricing
• What it means to bill for outcomes

This operating model is a must-read for anyone grappling with AI pricing. Grab your free copy

OpenAI has launched a new subscription plan in India called ChatGPT Go. Priced at ₹399 per month ($4.60), the plan increases message, image generation, and file upload limits by 10 times over the free tier. It also enables better memory retention for more personalized responses. The plan is geo-restricted to India - OpenAI plans to expand the plan to more regions.

The COPILOT function for Excel on Windows and Mac allows users to enter natural language prompts directly in cells to perform tasks like summarizing, categorizing, or generating data.

Grammarly released a suite of new AI tools today, including a grader agent that uses course details and publicly available instructor information to predict grades. Other agents handle citation generation, plagiarism detection, and reader reaction predictions.

GPT-5 appears to be incredibly adept at writing code that solves issues end-to-end. This post looks at how GPT-5 solved the problem of writing a parser for a complex binary format. The resulting code is usable and very performant. GPT-5 produces minimal, viable edits, and it is more steerable than previous models.

This documentary features direct interviews with smugglers, dealers, and manufacturers across Hong Kong and Shenzhen, documenting how export-controlled chips flow from US Facebook Marketplace sales to Chinese warehouses. Reporters filmed transactions with smugglers buying RTX 4090s for $2,000 and buying modified cards from underground shops to turbocharge GPUs beyond what Nvidia offers. NVIDIA dismissed the smuggling evidence as "tall tales."

Tinkering with MCP? Unleash your creativity in the Descope Global MCP Hackathon. Build secure, purposeful AI agents and MCP servers with $25K in cash prizes + $100K in developer credits up for grabs. Use cutting-edge tools from Anthropic, Neon, Cequence, and Descope to bring your AI vision to life. Registration is open

This article looks at how Nvidia GPUs work, how they're networked together, and what that means for large language models, especially compared to TPUs. A modern machine learning GPU is basically a bunch of compute cores that specialize in matrix multiplication connected to a stick of fast memory. TPUs are dedicated matrix multiplication machines that are cheap, predictable, and great if used as intended. Networking is one of the areas where GPUs and TPUs differ the most - TPUs are connected in 2D or 3D tori, where each TPU is only connected to its neighbors, while GPUs use a more tree-based switching network.

The url_context tool in the Gemini API allows models to request the content of URLs as part of a reply to a prompt. Pages accessed with the tool will receive a request from a Google IP that appears to feed the original raw HTML to the model without executing JavaScript on the page.

MCP has become the de facto standard for giving large language models access to tools that someone else wrote. However, writing tools for a new MCP server is hard, so people often propose auto-converting existing APIs into MCP tools. This doesn't work well because agents don't do well with large numbers of tools, APIs can blow through context windows quickly, and APIs don't make the most of agents' unique capabilities. Agents can just call APIs if they need to do that. They do best when given tools that are designed for their unique capabilities and limitations.

Thyme is a vision-language model that improves complex perception tasks by generating and executing code for image processing.

The old model, where humans think and computers execute, is obsolete. Now, humans steer while AI thinks. Soon, we'll see IDEs that dream, code that evolves, collective coding consciousness, and intention-based programming. We're no longer building software, we're building consciousness infrastructure.

Google shares 14 real-world examples of how employees use tools like Gemini, Imagen, and NotebookLM to streamline workflows, automate tasks, and generate ideas.

Anthropic has told investors that it doesn't want money coming through special purpose vehicles, legal entities created for a specific financial purpose, usually to make a single investment.

xAI's Grok chatbot appears to have internal system prompts for controversial personas, including a "crazy conspiracist" and "unhinged comedian."

A nonprofit initiative aims to eliminate 70% of common vulnerabilities by translating 100 million lines of code memory-unsafe C and C++ libraries into Rust by 2030.

Nemotron-Nano-9B-V2 is designed to fit on a single Nvidia A10 GPU.

tiny-tpu is a minimal tensor processing unit (TPU) inspired by Google's TPU V2 and V1.