TLDR AI 2025-09-23

How do you price agent outcomes?

Does AI break usage-based pricing?

How do you iterate and experiment on pricing models?

Expect to find answers to these questions—and some interesting takeaways—in Metronome's new webinar series.

Coming up:

• October 2nd: Ryan Campbell, Snowflake's Director of Product Finance, reveals how Snowflake built their Monetization Operating Model to connect product strategy with execution across finance, product, GTM, and engineering.
• November 4th: Martin Casado from Andreessen Horowitz shares what he's seeing across the market as companies experiment with new monetization strategies for the AI era.

Register here for all sessions

Nvidia and OpenAI signed a letter of intent to build data centers powered by 10 gigawatts of Nvidia systems, with investments potentially reaching $100 billion to support future AI model development.

Oracle has promoted Clay Magouyrk and Mike Sicilia to co-CEO roles, signaling a leadership transition as it focuses on scaling its AI infrastructure capabilities. Former CEO Safra Catz has moved into the role of executive vice chair of the board.

The mysterious Jony Ive hardware project at OpenAI is taking shape. OpenAI is working on multiple AI devices, including a smart speaker without a display, glasses, a voice recorder, and a wearable pin, targeting late 2026 or early 2027 releases. The company contracts with Apple suppliers and is actively poaching Apple hardware employees.

OpenAI argues its new Responses API is part of the inevitable evolution from turn-based chats to persistent agentic reasoning that maintains state across conversation turns. The company claims GPT-5 achieves 5% better performance on TAUBench through preserved reasoning state and reports 40-80% improved cache utilization. It also adds hosted tools so developers don't have to build their own retrieval pipelines from scratch.

AI coding requires a mindset shift, emphasizing ownership of AI-generated code and exploiting opportunities for maximum efficiency. Success lies in treating AI coding like management, focusing on creating and leveraging productive gradients where minimal effort yields significant rewards. Junior engineers might have an advantage over seniors in adapting to this new role, as embracing AI demands stepping out of comfort zones and fostering a strong sense of responsibility.

Grok 4 is the largest (known) training run to date. Researchers estimate it cost $490 million, required enough electricity to support a 4,000-person town for a year, and had a carbon footprint roughly equivalent to annual emissions from 3 airplanes.

Hear directly from the Salesforce engineering and developer teams on how they're using Slack and AI-powered agents to speed up code deployment and get time back by automating routine requests. See their workflow in action.

Tool calling is many orders of magnitude more costly than calling a plain old function from code. People should design their agentic systems according to the limit on how many tool calls their agents can effectively make. Using a tool call to add two numbers once probably doesn't matter, but scaling the problem up to 1,000 numbers will require a long wait and may exceed context window limits. Calling a function many times in a loop is one of the most common ways to solve a problem with code.

Qwen recently made several announcements on X, including official FP8 quantized versions of its Qwen3-Next models, Qwen3-TTS-Flash, Qwen3-Omni, and Qwen-Image-Edit-2509. Qwen3-Omni is a 30-billion-parameter model that supports text, audio, and video input and text and audio output. More details about the announcement are available in the article.

Seal Showdown is a new benchmarking tool that allows users to test various AI models head-to-head and vote on which one performs better. Unlike LMArena, it captures real preferences to more closely reflect how everyday users feel about various models. Rankings are derived from conversations on Scale's Outlier platform, so each user's country, education level, profession, language, and age can be verified. This allows the platform to show which models are most popular according to specific regions, languages, ages, and use cases.

DeepMind recently released the third version of its Frontier Safety Framework, which explores how generative AI systems can become threats. The safety framework details critical capability levels for assessing AI models that define the points at which behavior becomes dangerous. It also discusses ways for developers to address these risks. Misaligned AIs may ignore human instructions, produce fraudulent outputs, or refuse to stop operating when requested. For now, developers can prevent this by monitoring AI's chain of thought outputs, but future models may hide their reasoning, making it impossible to determine with certainty that a model is working against the interests of its human operator.

Join us live on September 25, 9 a.m. PT as we highlight the power of AI. CTA: Register now.

PyTorch's TorchAO team released pre-quantized versions of popular models using int4 and float8 precision optimized for both mobile and server deployments.

Some features will initially only be available to Pro subscribers, and some new products will have additional fees because of the associated costs.

Facebook Dating has introduced an AI chatbot to help users refine profiles and request tailored match suggestions, aiming to combat swipe fatigue and improve user experience.