TLDR AI 2025-12-24

NotebookLM is testing a new 'Lecture' format for Audio Overviews. The feature will generate a comprehensive AI lecture of roughly 30 minutes. Lectures can be produced in different languages, depending on the user settings. A sample lecture is available in the article.

NotebookLM has introduced Data Tables, a new feature that helps users organize and analyze information from sources in a structured format. The feature synthesizes sources into clean, structured tables ready to export to Google Sheets. It is now rolling out to all users. Screenshots of the feature are available in the article.

Agents can follow complex instructions, use tools, and work autonomously for hours. However, ask them the same question twice, and they have to start from scratch. We've made agents capable, but haven't yet figured out how to make them learn. This article looks at different types of memory and how they could be implemented into agents.

There's no wrong choice when it comes to AI. The tool you choose should match how you work. Try out both Claude and Codex and see which one fits. Every AI tool has its strengths and weaknesses, and the only way to discover what they are is by using them.

Stirrup is a framework for building agents that lets models choose their own approach to completing tasks. It has best practices and tools built in and is fully customizable. Stirrup features a skills system that extends agent capabilities, flexible tool execution, context management tools, flexible provider support, and multimodal support.

SpecBundle Phase 1 is a set of production-ready EAGLE-3 checkpoints trained with industry partners to improve real-world speculative decoding. The release focused on instruct-tuned models and shipped alongside SpecForge v0.2, which added major system refactors and multi-backend support.

ExecuTorch is a solution for deploying AI models on-device. Built by PyTorch for privacy, performance, and portability, ExecuTorch powers KPWA meta's on-device AI across Instagram, WhatsApp, Quest 3, Ray-Ban Meta Smart Glasses, and more. It allows developers to deploy LLMs, vision, speech, and other multimodal models with familiar PyTorch APIs. The tool can accelerate research to production with seamless model export, optimization, and deployment.

OpenAI's use of Weekly Active Users (WAU) rather than Monthly Active Users (MAU) renders its user base scale incomparable to other large consumer technology products. ChatGPT's user retention is likely quite low, meaning many users cycle in and out every month, inflating MAU relative to WAU. Dividing MAU would reveal relatively weaker unit economics that are directly comparable to other consumer technology products.

AI is making formal verification go mainstream. Random testing will play an important role in the future of software engineering. As autoformalization tools get better, we will have many more formal specifications. Random testing benefits from these formal specifications in different ways than formal verification, but both have their places.

Poetiq saw results as high as 75% at under $8/problem using GPT-5.2 X-High on the full PUBLIC-EVAL dataset.

Gemini 3 Flash was Arnaud Autef's first release as distillation TL.

Mother Nature can host human intelligence using only 20 watts of power in a space equivalent to the space inside your skull.