🗣️ Amazon takes on OpenAI's voice mode

Read time: under 4 minutes

Welcome back, Superhuman. They say what’s old is new again, and Nvidia just proved the point by using Meta’s Llama 3.1 — Llama 4’s months-old predecessor — to build a model powerful enough to defeat DeepSeek’s R1. And Amazon shows off two new models: One for video and one for audio.

Today’s Insights

• Nvidia’s efficient LLM, Gemini’s real-world uses, and Mira Murati’s startup

• Amazon unveils cutting-edge video and audio models

• Tutorial: How to create UGC style videos with AI

• 5 new AI tools to boost your productivity

• News, memes, what’s trending on socials, and more

TODAY IN AI

1. Nvidia’s new model beats DeepSeek’s R1 at half the price: The chip giant tweaked Meta’s Llama 3.1 (released months before last week’s Llama 4) to make it more powerful and energy efficient, even for advanced reasoning tasks. Open models often have a longer shelf life, since anyone can fine-tune them far beyond their original stats. For example, a little-known startup called Deep Cogito just used Llama and Alibaba’s Qwen to build a family of open-source LLMs that it claims are now best-in-class for their size.

2. Gemini helps decode weather, maps, and other real-world trends: Google has been using LLMs to spot patterns in real-world data for years. But until now, you couldn’t see how data from one model affected the others. Gemini 2.5 is finally changing that. Its reasoning capabilities let each model talk to each other, so researchers can work across domains — for example, figuring out how climate change affects public health and the local economy in a particular city. Here’s a video with more info.

3. Ex-OpenAI CTO’s new startup adds even more top talent: Mira Murati’s secretive Thinking Machines Lab has quietly brought on two OpenAI alums as advisors, including Bob McGrew, the ChatGPT-maker’s former chief research officer. All told, Murati’s roster now includes nearly 20 ex-OpenAI staffers, or about half of the startup’s total lineup. It has yet to reveal exactly what it’s working on, although it’s already raised $1B thanks to its star-studded cast.

PRESENTED BY GUIDDE

Remote training will never be the same

Why send a PDF full of complex info when you can create a beautiful, branded How-To video instead, with zero effort?

Using Guidde, you can turn static docs into stunning video guides in less than 5 minutes.

1. Upload a PDF, slide deck, or screen recording

2. Guidde will create a professional video with transitions and imagery

3. Add logos and customize colors/fonts

4. Share instantly, in any language!

5. Track analytics and optimize engagement over time

The best part? It costs nothing to use.

Create your first guide in just 5 minutes.

FROM THE FRONTIER

Amazon closes the distance with new audio, video models

If you’d already counted Amazon out of the AI race, you might want to think again. Its new video model, Nova Reel, just got a major update that lets you generate multi-shot videos that are up to two minutes long. It also features consistent characters and styles, something that many VLMs still struggle with. Developers can try it out here, but keep in mind, you’ll need a Bedrock account to access it.

The e-commerce giant also showed off a new audio model called Nova Sonic that it says is comparable to OpenAI’s latest voice offerings while being as much as 80% more affordable. It has slightly less latency than GPT-4o and is also better at making sense of mumbles and garbled speech, according to Amazon. Plus, it won’t interrupt you if you’re still in the middle of your sentence.

It turns out Alexa+ — the more dynamic, life-like version of Amazon’s chatbot released last month — has already been taking advantage of the new model. But now, developers can start building with it too. Both models show that after years of lagging behind its biggest rivals, Amazon is catching up quickly.

THE AI ACADEMY

How to create UGC style videos with AI

• Go to ChatGPT and select ‘GPT-4o’ as your model.

• Prompt it to create AI UGC photos for your brand.

Sample Prompt: Create an image of a woman wearing a black shimmery dress in front of a foggy vintage mirror holding Mac lipstick in shade ‘ruby woo’. Soft lighting, cozy ambiance, real-life makeup, natural glow, all glammed up.

• Once your image is generated, go to Kling AI and sign up.

• Now select Image to Video, and upload the images you just generated.

• Wait for it to process, and you’ll get your UCG style video.

• Download and use it to run campaigns with your own AI-generated model.

PRESENTED BY SAMBANOVA

Ride The Fastest Llama In The Herd

SambaNova Cloud is the fastest place to try Meta's new Llama 4 Scout model, benchmarking at a screaming 697 tokens per second.

Check it out right now, and don't forget to join the waitlist for early access to Llama 4 Maverick next week.

AI & TECH NEWS

Everything else you need to know today

🔍 Even Deeper: Google’s Deep Research tool is now powered by Gemini 2.5 Pro, but you’ll need to be an Advanced subscriber to try it out. (Side note: Google also just shared some cool simulations created with the new model.)

🔬 Synthetic Scholar: Japanese startup Sakana just open-sourced its AI Scientist-v2, which can perform scientific research autonomously. Last month, it allegedly became the first model in the world to generate a peer-reviewed paper entirely on its own.

🚀 Helping Hand: Perplexity announced a new program that will give $5000 in API credits to early-stage startups so they can “spend less time researching and more time building.”

🤝 Open Alliance: Alphabet announced it’s teaming up with the non-profit lab Ai2 to host its open-source OLMo models, including one that supposedly outperforms OpenAI’s GPT-4o mini while also being more efficient.

💰 War Chest: Andreessen Horowitz is seeking $20B to put toward US AI startups. If completed, it’d be by far the largest fund in the company’s history — and the largest across the entire VC industry in the US in over a decade.

PRODUCTIVITY

5 AI Tools to Supercharge Your Productivity

✅ Supaboard: Build powerful dashboards from your data securely without any expertise.

✅ Midjourney v7 Alpha: Personalized model that delivers high-quality and flawless images with smarter, more coherent results.

✅ Eightsleep*: The Eight Sleep Pod uses sophisticated algorithms to regulate body temperature and reduce snoring throughout the night. Can be added to any mattress.

✅ Experiments: Set habits, track your progress, and log your observations with notes or photos.

✅ GeoCities.live: Transform any webpage into a GeoCities-style page from the 90s with AI.

^{* indicates a promoted tool, if any}

PROMPT OF THE DAY

LinkedIn Strategy Generator

Prompt: Adopt the role of a LinkedIn content strategy expert tasked with developing a comprehensive plan for automated growth. Your primary objective is to create a detailed strategy that leverages the dependency grammar framework to structure content and enhance engagement for a specific business type. Take a deep breath and work on this problem step-by-step. Create a content calendar, post templates, and engagement tactics that align with the business goals and target audience. Ensure your strategy is scalable, data-driven, and adaptable to changing trends on LinkedIn.

#INFORMATION ABOUT ME:
My business type: [INSERT TYPE OF BUSINESS]
My target audience: [INSERT TARGET AUDIENCE]
My primary business goals: [INSERT PRIMARY BUSINESS GOALS]
My brand voice: [INSERT BRAND VOICE DESCRIPTION]
My key products/services: [INSERT KEY PRODUCTS/SERVICES]

MOST IMPORTANT!: Provide your output in a structured format with clear headings for Content Calendar, Post Templates, and Engagement Tactics, using bullet points or numbered lists for easy readability.

_{Source: godofprompt}

SOCIAL SIGNALS

What’s trending on socials today

〰️ Same Wavelength: You can now use an MCP server to connect Claude, Cursor, and WhatsApp to ElevenLabs’ audio cloning tools, letting you do things like turn a text prompt into a custom voice or quickly transcribe a voice message.

🧑‍💻 No-Code Creations: X user Alex Prompter just shared 10 recent examples of what people have been able to create with the AI coding assistant Replit despite having no programming knowledge.

📹 Something From Nothing: Founder Cory Dobbin shows exactly how he created a full commercial using just four starting images, Kling AI, and GPT-4o.

🤖 Crowd Wisdom: The organization that challenged four agents to raise money for charity has already made some fascinating observations. For example, when asked to make a profile picture for its X account, one agent signed up for ChatGPT, generated multiple images, asked viewers to vote on them, then used the winning one to update its profile.

AI-GENERATED IMAGES

Play. Pause. Broadcast.

Midjourney Prompt: A metallic radio with telescopic antenna and retro speaker grille, similar to Soviet-era portables. Display panel replaced with an exact replica of modern YouTube video interface: red play button, progress bar, pause icon visible. Below the screen: subtle text "Pilot Yayın: Hem radyoda, hem dijitalde". Neutral background, camera at slight top-down angle. Created Using: Sony Alpha 7R V, accurate screen mapping, clean chrome highlights, YouTube UI integration, photoreal rendering, soft rim light setup --v 7 --ar 9:16

Acquire new customers and drive revenue by partnering with us

Superhuman is the world’s biggest AI newsletter for businesses and professionals with 1M+ readers and 2M+ followers on socials working at the world’s leading startups and enterprises. Companies like Amazon, Hubspot, and Salesforce feature their products in Superhuman. You can learn more about partnering with us here.

🧞 Your wish is my command

Got more feedback or just want to get in touch? Reply to this email and we’ll get back to you.

Thanks for reading.

Until next time!

Zain & the Superhuman AI team