AI News: Codex Surges; Free NotebookLM Updates; Viral Image Prompts
Paul J. Lipsky covers the major AI news of the week, focusing on OpenAI's Codex updates for knowledge work, Google Gemini's new file creation feature, and significant NotebookLM improvements. He also introduces new weekly segments including 'Prompt of the Week' and 'Meme of the Week,' and compares GPT Image 2 versus Imagen (Nano Banana) for image generation.
Summary
The video opens with Paul framing the week's news around an 'AI super agent race' between OpenAI and Google, both attempting to compete with Anthropic's Claude Co-work for knowledge worker dominance. He explains that Codex — often misunderstood as coding-only — received major updates making it suitable for everyday knowledge work tasks like email triage, document creation, and automation, mirroring Claude Co-work's capabilities but in a single unified interface without tab-switching. A notable Codex feature highlighted is a 'tasks' panel that surfaces follow-up actions during chat sessions. Paul reserves final judgment on Codex vs. Co-work for a dedicated head-to-head video the following week.
A new segment, 'Prompt of the Week,' is introduced featuring a viral prompt that converts any image into a deliberately crude, scribbled drawing. Paul tested this across multiple image models and found GPT Image 2 produced the best results for this prompt. He also shares his broader conclusion on the GPT Image 2 vs. Imagen debate: GPT Image 2 excels at generating images from text prompts, while Imagen (referred to as 'Nano Banana') is superior for editing existing images.
On the Google Gemini front, Paul covers two updates: the rollout of file creation directly within Gemini chat (supporting PDFs, Word docs, PowerPoints, Google Docs, Sheets, and more), and a long-overdue fix to the mobile dictation button on iOS and Android, which previously cut off recordings on brief pauses. He also previews Google I/O, expected in under three weeks, as a likely source of major announcements.
For NotebookLM, Paul highlights two updates: the rollout of notebook folders inside Gemini to all free and paid users including mobile, and a new 'auto label sources by topic' feature that automatically organizes sources into labeled categories. He argues this labeling system is more impactful than it appears, enabling students or researchers to isolate specific topic clusters for focused study or content generation within a single notebook.
Additional news includes OpenAI releasing a prompting guide for GPT-4.5 recommending shorter, goal-focused prompts, Spotify rolling out an AI music identification label, and two humorous 'meme of the week' items: Apple accidentally shipping a Claude.md file in a software update (suggesting they use Claude Code internally), and a funny response to a tweet asking for YouTube thumbnail prompts in GPT Image 2.
Key Insights
- Paul argues that Codex and Claude Co-work are functionally designed for the same purpose — knowledge work automation — but Codex differentiates itself by combining chat, coding, and co-work capabilities into a single unified interface with no tab-switching required.
- Paul concludes that GPT Image 2 consistently produces high-quality images from text prompts alone, but performs poorly at editing existing images, whereas Imagen ('Nano Banana') excels at image editing tasks like changing facial expressions.
- Paul frames Google's Gemini file creation feature as a philosophically different — and less powerful but more accessible — approach to the same problem OpenAI and Anthropic are solving with agent tools, arguing that ease of use matters as much as raw capability for mainstream adoption.
- Paul argues that NotebookLM's new auto-labeling of sources by topic is more impactful than it appears, because it allows users to selectively chat with or generate content (like audio overviews or flashcards) from only a specific subset of sources within a single notebook.
- Paul notes that OpenAI's prompting guide for GPT-4.5 recommends shorter, goal-oriented prompts rather than detailed instructions, reflecting a broader trend across both OpenAI and Anthropic models where users need to provide less hand-holding and allow the model to determine its own path to the goal.
Topics
Full transcript available for MurmurCast members
Sign Up to Access