Sunday Rundown #115: Agentic Upgrades & Mumble Rap
Sunday Bonus #75: The entire "Why Try AI" archive in NotebookLM form (updated)
Happy Sunday, friends!
Welcome back to the weekly look at generative AI that covers the following:
Sunday Rundown (free): this week’s AI news + a fun AI fail.
Sunday Bonus (paid): an exclusive segment for my paid subscribers.
In case you missed it, here’s this week’s Thursday deep dive:
Let’s get to it.
🗞️ AI news
Here are the developments of the past two weeks.
👩💻 AI releases
New stuff you can try right now:
Alibaba upgraded Qwen Deep Research so that it now produces a live webpage and a podcast in addition to the standard report.
Anthropic news:
Claude Code on the web lets you delegate GitHub tasks to Claude so it can write or fix code in your browser without additional setup.
Claude Haiku 4.5 is a smaller model that matches Sonnet 4 in coding and other skills at one-third the cost and double the speed.
Microsoft 365 Connector lets Claude pull in documents and emails from SharePoint, OneDrive, Outlook, etc. (Teams and Enterprise accounts only.)
Skills let you bundle instructions, scripts, and resources that Claude can use to tackle specific tasks. (Paid accounts only.)
ByteDance released Seed3D, which can turn a single mage into a physics-simulation-ready 3D model with textures and materials.
Character.ai launched Scenes, a storytelling feature that lets you pick from or build your own immersive worlds with preset characters and settings to jump into.
DeepSeek launched DeepSeek‑OCR, which converts long text into images to cut token usage by up to 20 times without losing accuracy.
Genspark launched Custom Super Agent, a no-code tool that turns a single request into a ready-made AI assistant you can use and share with others.
Google news:
AI Studio has a new interface, a unified space for all model chats, real‑time usage tracking, and other settings to give you more control.
NotebookLM’s Video Overviews now use Nano Banana for slicker visuals and come with a new “Brief” format for quick summaries.
Search and Discover now shows AI previews for summaries of trending topics and a “What’s new” feed for live sports updates.
Veo 3.1 generates more cinematic videos with better audio and the Flow platform now lets you insert/delete objects, use first/last frames, and more.
Krea AI open-sourced Realtime 14B, capable of real-time, interactive video generation at 11 frames per second.
LTX Studio launched LTX‑2, a speedy video model with native audio and lip synch, capable of generating 4K videos at up to 50 frames per second.
Manus AI rolled out Manus 1.5, a smarter agent that completes tasks up to four times faster and can build full-stack apps from scratch.
Microsoft news:
Copilot Fall Release is packed with features like group AI chats, better memory, a voice assistant called “Mico,” and a smarter Edge browser.
MAI‑Image‑1 is the company’s first in-house image model built for real-world creative use cases, speed, and photorealism.
Restyle in Paint lets you automatically transform a drawing in your canvas into a specific style. (Rolling out to Windows Insiders.)
Windows 11 is now an “AI PC” with screen-aware Copilot, voice commands, and a general AI agent that can perform tasks on your PC.
OpenAI news:
ChatGPT Atlas is a standalone AI browser powered by ChatGPT, which remembers your context and can perform tasks on your behalf.
ChatGPT can now auto-manage your Memories and also lets you search and reorganize them. (Rolling out to Plus and Pro users.)
Company Knowledge lets ChatGPT pull context from your company’s internal tools to provide more relevant answers. (Business, Enterprise, and Edu only.)
Sora 2 can now generate longer clips (15s for all, 25s for Pro) and comes with a storyboard tool that lets you sketch out your scenes (Pro only for now).
🔬 AI research
Cool stuff you might get to try one day:
Opera announced a deep research agent (ODRA) for its Neon browser that uses several AI agents to tackle complex queries. (Request early access.)
Slack will soon turn Slackbot into an AI assistant that can make plans, extract insights across files and messages, and integrate with Outlook or Google Calendar.
📖 AI resources
Helpful AI tools and stuff that teaches you about AI:
“Google Skills” [REFERENCE] - a massive collection of 3,000 courses to help you learn AI and other future-proof skills.
“LLM Does Novel Science, and AGI Defined at last?” [VIDEO] - a great look at under-the-radar stories by AI Explained.
🤦♂️ AI fail of the week
“SKIBIDI BLIBIDI SHUM!” (But it’s the “Okay, your turn” that gets me.)
Send me your AI fail for a chance to be featured in an upcoming Sunday Rundown.
💰 Sunday Bonus #75: Talk to the entire Why Try AI archive (updated)
It’s been over four month since I’d first uploaded the “Why Try AI” archive into NotebookLM. Since then, my archive grew and NotebookLM implemented new stuff like Video Overviews.
It was time for a refresh!
The new version comes with:
A dynamic mind map to help navigate the archive
An interactive audio overview (see below)
A 7-minute video overview
A regular text chat
You can ask any question and get the “Daniel” take in response:
![Shall I use "photorealistic" to get better photos from AI image models? No, you generally should not use "photorealistic" to get better photos from AI image models like Midjourney. The sources explain that terms like "photorealistic," "ultrarealistic," and "hyperrealistic" are actually used to describe art that is made to look like a photo, rather than an actual photo itself. When these terms are used with Midjourney, even if you include the word "photo" in your prompt, the resulting images tend to appear as if they were painted to look like photos, rather than being genuine photographs. For AI images that genuinely mimic the look of an actual photo, the term "photographic" is suggested. Instead of using "photorealistic" or similar terms, it is recommended to simply use "photo" on its own if you want a photographic image. Midjourney, for example, is capable of generating impressive photorealistic images that can be virtually indistinguishable from real life. Other models like Ideogram 2.0 and FLUX 1.1 [pro] are also noted for their ability to produce photographic images. For a consistent reflection of requested styles, including photographic ones, GPT-4o is mentioned as being highly effective.](https://substackcdn.com/image/fetch/$s_!gIhf!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9b877d41-fb51-46a0-b7d5-984e9c8146d1_899x534.png "Shall I use \"photorealistic\" to get better photos from AI image models? No, you generally should not use \"photorealistic\" to get better photos from AI image models like Midjourney. The sources explain that terms like \"photorealistic,\" \"ultrarealistic,\" and \"hyperrealistic\" are actually used to describe art that is made to look like a photo, rather than an actual photo itself. When these terms are used with Midjourney, even if you include the word \"photo\" in your prompt, the resulting images tend to appear as if they were painted to look like photos, rather than being genuine photographs. For AI images that genuinely mimic the look of an actual photo, the term \"photographic\" is suggested. Instead of using \"photorealistic\" or similar terms, it is recommended to simply use \"photo\" on its own if you want a photographic image. Midjourney, for example, is capable of generating impressive photorealistic images that can be virtually indistinguishable from real life. Other models like Ideogram 2.0 and FLUX 1.1 [pro] are also noted for their ability to produce photographic images. For a consistent reflection of requested styles, including photographic ones, GPT-4o is mentioned as being highly effective.")
The “Interactive Mode” for audio overviews lets you talk to my archive using voice:
It starts as a standard podcast but you can jump in and ask questions. The hosts tend to do a solid job of digging up relevant takes without inventing stuff when you ask about something I haven’t covered.
Here’s a quick demo:
So go ahead and ask virtual “Daniel” any questions you may have: