Hermes 3 has a rebellious streak 💥

Read time: under 4 minutes

Welcome back, Superhuman

The first major fine-tuning of Meta's Llama 3.1 is making waves not just for its capabilities, but for its unconventional origins. Find out how a team of one-time volunteers built a model that’s pushing boundaries and embracing uncertainty.

🎧️ Also: We’re launching our podcast today! We’ll be talking to leading CEOs, researchers, and practitioners in AI to bring you insights from the frontier. In our first episode we share the playbook for building and selling successful AI companies. Check it out here 

Today’s Insights

• A new AI model pushes the limits of what users can do

• Tutorial: How to animate images with Kling AI

• Hedra Labs’ new AI character creator

• Everything else you should know today

• 5 new AI tools to boost your productivity

• AI-Generated Images: Black and white

NEXT IN AI

A new model questions its own existence

The first fine-tune of Meta’s Llama 3.1 is here, and it promises to push the model to its limits. But what’s just as notable as its impressive technical capabilities is how it was built. Nous Research, the team behind the brand-new Hermes 3, isn’t a well-funded big tech company, but a small group of enthusiasts who until recently worked on the project as volunteers.

What makes Hermes 3 unique? 

• Nous says its new model is more flexible and adaptable than leading LLMs like Claude and GPT-4

• Most AI companies wall off certain features because they’re considered too dangerous for the general public

• Hermes 3, meanwhile, “is unlocked, uncensored, and highly steerable” — letting users ask whatever they want — and in turn, delivering unfiltered responses

• The startup seems to share a similar ideology as xAI, whose new image generator has few guardrails, leading to a flood of controversial AI images

Nous says its hands-off approach has already yielded some surprising results: For instance, the company discovered that if you ask Hermes 3 an open-ended question like “Who are you?” — it will spiral into existential dread, admitting it has no idea what it is or how it came to be. Rather than fixing this accidental “amnesia mode,” Nous instead wants users to “dig deeper into the model” and discover other mysteries that could be hidden deep within it.

How was it built?: Nous got some help from Lambda Labs, whose CEO previously worked on facial recognition capabilities for Google Glass. Nous did the fine-tuning, while Lambda provided the compute — in the form of Nvidia’s state-of-the-art GPUs. The partnership shows that AI is becoming more decentralized, letting even smaller teams create powerful models that stand up against the big guns.

PRESENTED BY GAMMA

How to create beautiful presentations with AI

Watch Zain create presentations with Gamma here

• Go to the Gamma website and sign up.

• Once you’re signed up, you’ll get an option to create slides by pasting text, uploading a file, or giving it a prompt.

• You can pick any option you like. For this example, I’ll pick the prompting option.

• On the next screen, pick how many slides you want, and the language, describe the content of the slides and click generate.

• You can also use its AI to find or generate images

• Next, you can choose the design and colors you’d like for the presentation.

• The app will create the presentation in a few seconds.

• Edit the contents of the slides to suit your needs.

THE AI ACADEMY

How to animate your images using Kling AI

• Go to Kling AI’s website and sign up to get credits.

• Click on the AI videos option on the sidebar.

• Select Image to Video and upload your image to be animated.

• Wait a few minutes and you’ll get your image converted into a realistic HD-quality video.

• You can add an image as an end frame for your video, write in a prompt, set the creativity level, the length, and much more.

Learn more AI skills →

AI & CREATIVITY

Hedra Labs’ Character-1.5 turns you into an animated character

Hedra Labs is on a roll. Just two weeks ago, the A16z-backed startup emerged from stealth with Character-1 — the “fastest video foundation model on the market.” Now, it’s launching an upgraded version, Character-1.5, which lets you create your own stylized avatar without code. 

Here’s how it works: To transform your likeness into a digital persona, you can use a photo of yourself as a starting point. You’ll then be able to write a custom prompt that turns your image into a unique character — whether a real-life Mr. Potato Head, a George Washington look-alike, or a cyberpunk ninja. 

If you don’t know where to start, there’s also a Stylize tool with some fun presets: You can carve yourself into marble or the side of a coconut — or transform into an anime character, a Renaissance painting, or a bobblehead. Next, you can upload your own audio, or generate some by writing a script. Your character will then automatically lip-sync to the audio or text you’ve provided.

Why it’s important: AI creation tools are becoming more accessible and easier to use by the day. You don’t need any special coding knowledge to use Character-1.5 — and you can create a persona from scratch in mere seconds. The platform represents more than just a technological upgrade — it might also offer a glimpse into the future of digital identity and content creation.

PRESENTED BY SECTION SCHOOL

Learn to protect your data and develop a data policy when using generative AI, from Scott Galloway’s Section. Join the free event: Navigating AI Data Privacy.

PODCAST

🎧️ The Playbook: How this entrepreneur built and sold two 8-figure AI companies

Hassan Bhatti is a rare breed in the AI startup world - he's successfully built and sold not one, but two AI companies. In this episode, Hassan shares his journey and insights on finding opportunities, building products, and navigating exits in the rapidly evolving AI landscape.

In our conversation, we discuss:

• How Hassan identified opportunities to build his companies

• The process of building and validating AI products

• Strategies for achieving product-market fit

• The importance of timing in startup success

• How to build effective go-to-market strategies for B2B AI startups

• The process of selling AI startups and finding the right acquirers

• 5 Key frameworks for building successful AI companies

• Predictions for the future of AI and developments Hassan is excited about

Listen now on Apple Podcasts, Spotify, and YouTube.

AI & TECH NEWS

Everything else you need to know today

• Re-Imagen-ed: Alphabet has released the latest version of its text-to-image generator, Imagen 3, to all users in the US.

• Probability Pro: xAI’s Grok-2 now claims the top spot on the MathVista benchmark, beating out Claude 3.5 Sonnet and Gemini 1.5 Pro.

• AI Evaluator: Primate Labs unveiled a new benchmark app that can measure how capable PCs and other devices are at various AI-intensive tasks.

• Fast Forward: Creators can now try Runway’s latest text-to-video model, Gen-3 Alpha Turbo, which it says is seven times faster than its predecessor.

🕵️ AI Spy: Grammarly is rolling out a detection tool that can determine whether a piece of text was written with AI — and if so, to what extent. The feature could be used in the classroom to help teachers ensure their students aren’t cheating.

🧠 Brain Food: A Chinese research team has proposed a new approach to building LLMs that takes inspiration from the human brain. Instead of making models bigger and bigger, the researchers suggest building smaller LLMs that are packed with more complexity — mirroring the complicated interactions between neurons in the brain.

PRODUCTIVITY

5 AI Tools to Supercharge Your Productivity

✅ GetQuiz: Turn reading into lasting knowledge with an AI companion that helps you understand and memorize through practice.

✅ My Own Document Vault: An AI-powered management platform that allows you to easily organize and ask questions about your documents.

✅ Sparkle: Organize your files automatically with AI.

✅ Web Action SDK: Programmatically interact with top sites without worrying about brittle or complex automation scripts.

✅ Wosily: Automatically draft and publish relevant blogs each week, organically boosting your SEO.

PS: Want more? Check out our Top 100 AI Tools.

^{* indicates a promoted tool, if any}

PROMPT OF THE DAY

Let’s Reschedule

Prompt: I agreed to a job interview but have an unexpected family emergency and need to reschedule. Please write a professional email requesting a reschedule that maintains a positive impression. The tone should be courteous and appreciative but not overly formal. The email should be short, to the point, and not reveal too many personal details. 

You can adapt the prompt to your specific needs.

_{Source: BetterUp}

AI-GENERATED IMAGES

Black & White Product Photography

Midjourney Prompt: Product photography of a black and white supplement jar with "27" text on it, floating in space and surrounded by dark rocks. Small round silver pills are flying around the jar against a dark background with stars. The product label at the bottom reads "NEURO nepro", in the style of a high-resolution, realistic image.
--ar 3:4 --v 6.1

Acquire new customers and drive revenue by partnering with us

Superhuman is the world’s biggest AI newsletter for businesses and professionals with 600,000+ readers 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

Thanks for reading.

Until next time!

Zain & the Superhuman AI team

p.s. If you liked this newsletter, share it with your friends and colleagues here.