Sunday Rundown #77: Mistral's Moment & Unsliceable Pizza

Sunday Bonus #37: Making NotebookLM podcast hosts debate each other.

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): a goodie for my paid subscribers.

Let’s get to it.

🗞️ AI news

Here are this week’s AI developments.

---

👩‍💻 AI releases

New stuff you can try right now:

1. Allen Institute open-sourced Ai2 OpenScholar—an LLM for scientific research with more accurate citations and fewer hallucinations than GPT-4o. (Try it here.)

2. Black Forest Labs released a suite of models under the “FLUX.1 Tools” umbrella:

   1. FLUX.1 Fill: Image inpainting and outpainting (a la Midjourney)

   2. FLUX.1 Depth & FLUX.1 Canny: Restyle images while keeping the structure/layout the same (a la ControlNet or Adobe Structure Reference.)

   3. FLUX.1 Redux: Restyle images while keeping the character or scene consistent (a la --cref in Midjourney)

3. DeepSeek has a new reasoning model called R1 Lite Preview that performs on par with OpenAI’s o1-preview. (Try it for free at chat.deepseek.com.)

4. ElevenLabs has two launches out:

   1. A nifty auto-regenerate feature that automatically identifies and fixes any issues in generated audio clips.

   2. Conversational AI that lets you build and customize your own realistic voice agents. (Try it here.)

5. Google added a memory feature called “Saved Info” to Gemini Advanced, so you can ask it to remember important details about you and your interests.

6. Mistral stepped up big time this week:

   1. The company overhauled its free chat interface, LeChat, which now includes:

      1. Canvas (a la ChatGPT) - a side panel to the left of your chat that outputs code, etc.

      2. Image generation powered by FLUX models from Black Forest Labs.

      3. Ability to search the web (but are we ready for this?)

   2. Mistral’s new multimodal model, Pixtral Large, gives frontier models like Claude 3.5 Sonnet. GPT-4o, and Gemini 1.5 Pro a run for their tokens.

   3. Try all of the above for free at chat.mistral.ai.

7. OpenAI released an update to GPT-4o, making it better at creative writing and working with uploaded files.

8. Perplexity now lets you buy products directly from its search results (as long as you’re a Pro user based in the US).

9. Pickle is a new tool that lets you create an avatar of yourself and steer it using just your voice during online meetings.

10. PlayAI (formerly PlayHT) released a beta version of PlayDialog, its voice model that can create emotive, human-like dialogue from written scripts. (Try it here.)

11. Runway introduced “Expand Video” that lets you change the aspect ratio of an existing video and automatically fills in in the blanks.

12. Suno’s brand-new V4 is out for Pro and Premier users, with better quality audio, a lyrics assistant, the ability to remaster older tracks, better cover art, and more.

13. TikTok rolled out its Symphony Creative Studio to all advertisers, featuring AI-powered video generation, digital avatars, and more.

---

🔬 AI research

Cool stuff you might get to try one day:

1. Apple is reportedly planning to launch a conversational version of Siri—called LLM Siri—to compete with the current chatbot offerings…in 2026.

2. Microsoft Teams users will soon be able to clone their voice and dub their speech into nine different languages during calls.

3. OpenAI is rumored to be looking into making a web browser and might be about to release the “Live Camera” feature for ChatGPT.

4. YouTube will soon roll out automatic dubbing, which dubs any videos you upload into almost a dozen languages

---

📖 AI resources

Helpful AI tools and stuff that teaches you about AI:

1. “2024: The State of Generative AI in the Enterprise” [report] - figures and insights from Menlo Ventures.

2. “ChatGPT Foundations for K–12 Educators” [course] - a collaboration between Common Sense Media and OpenAI.

3. “Growing Up: Navigating Generative AI’s Early Years” [PDF] - a big report on AI adoption by Wharton School.

Alternatively, listen to my NotebookLM podcast summarizing the findings from the two reports above (+a third article):

---

🔀 AI random

Other notable AI stories of the week:

1. Microsoft has partnered with HarperCollins to train future AI models on its content. (Not without controversy.)

🤦‍♂️ AI fail of the week

I’m so sorry, DeepSeek. My nonsensical riddle broke you.

---

💰 Sunday Bonus #37: How to make NotebookLM podcast hosts debate each other

If you know me, you know I love NotebookLM.

I was a fan before it was cool—before “Audio Overviews” made NotebookLM go viral:

NotebookLM: Beginner Guide

Since then, I shared lots of tips for NotebookLM:

• Compare sources & extract consensus with NotebookLM

• Steering “Audio Overviews” in NotebookLM

• Creating an “Audio Overview” without any underlying sources

But as much as I enjoy the lifelike chit-chat of NotebookLM’s podcast hosts, I must admit I got quite tired of how agreeable they always are.

They tend to “Yes, and…” a lot and build on each other’s arguments.

It’s all very echo-chambery.

If you want to truly explore the pros and cons of an issue, this constant agreement isn’t very helpful.

So I wondered: Could I get the hosts to take two opposing stances and stick to their guns?

Turns out, this wasn’t as easy as I’d first assumed.

The hosts are apparently hardcoded to be amicable, so it wasn’t as simple as just asking them to “debate the pros and cons”—they would inevitably slide back into finishing each other’s sentences again.

Also, you’re limited to just 500 characters of instructions, so there isn’t much room to steer the conversation.

But I’m proud to announce that, after dozens of tweaks, I seem to have cracked the code.

Listen to how I forced the hosts into a meta-discussion about the value of discussions:

As you can hear, they’re still friendly and respectful, but I finally got each of them to pick a point of view and stick to it.

How did I do it?

Let me show you: