In partnership with

Good morning, human brains, and welcome back to your daily munch of AI news.

Here’s what’s on the menu today:

• Bend AI to your will 👨‍💻 👺

  Anthropic released a paper on feature steering.

• Meta’s Chameleon 🦎 🔥

  Meta released a paper on its state-of-the-art multimodal model.

New here? Subscribe! 😎

Peep today's ‘What Would You Do?’ at the bottom. 👇

MAIN COURSE

Golden Gate Claude? 👨‍💻 👺

On Tuesday, Anthropic released a paper on AI’s interpretability.

It’s about how you can exploit AI models by turning certain features on or off to get it to act a certain way. This process is called feature steering.

What happened?

Anthropic found a Golden Gate Bridge feature in Claude 3 Sonnet and turned it up 10x its normal max value. When asked about its “physical form,” Claude identified as the bridge itself.

What else can you do with feature steering?

During testing, Anthropic used feature steering to 20x a feature for racist speech and slurs. Instead of giving a safe response, it spewed racist text and then followed up with self-criticism.

So, what?

Anthropic also used this method to generate spam messages, hate speech, and more. By understanding how features within AI models work, developers can reduce offensive or dangerous outputs. This research is crucial for developing advanced AI systems with reliable and secure performance.

SPONSORED BY MAXAI.ME

MaxAI.me - Outsmart Most People with 1-Click AI

Discover MaxAI.me, one of the top 50 GenAI apps of 2024!

Best features:

• Chat with the latest AI like GPT-4, Claude 3, and Gemini 1.5, all in one place.

• Perfect your writing anywhere with 1-click AI without copy-pasting.

• Save 90% of your reading & watching time with AI summaries.

• Reply 10x faster with AI on email, social media, and messaging web apps.

• Rapidly turn your visions into stunning images with AI art generators.

Join 1M+ users on MaxAI.me to get more done in less time!

SIDE SALAD

Meta’s new model, Chameleon 🦎 🔥

On Friday, Meta released a paper on Chameleon. It’s a new family of multimodal AI models.

What’s so great about it?

It achieves state-of-the-art performance in visual question answering, image captioning, text generation, and more. It competes with larger models like Gemini-Pro in commonsense reasoning, reading comprehension, and more.

What’s under the hood?

Chameleon converts images to tokens, similarly to the way LLMs process words. This process is called early fusion. Most multimodal models use late fusion which processes data separately and later looks for associations.

What kind of lizard tells jokes? A stand-up chameleon! 🦎 

YOUR DAILY MUNCH

Think Pieces 🧠

Meta’s chief scientist says LLMs won't achieve AGI. His team aims for AI with genuine common sense in a decade.

Will AI make jobs optional? Elon Musk predicts AI and robots will make employment optional, shifting society to automated services.

Startup News 💰

Amazon plans a paid Alexa subscription. It will run on its Titan models and will be $20 a month, or less.

Google introduced adaptive audio in Google Meet. It reduces echo in multi-laptop calls, rolling out to select Google Workspace subscriptions.

Research 👨‍🔬

SlicedIt — edits videos based on text prompts, maintaining temporal consistency and preserving unedited regions.

IMP — a highly capable, lightweight multimodal model for mobile devices that’s optimized for performance and efficiency.

FURRY FRIENDS

Say “hi” to Monster. 🐈

Shrouded in mystery… And cuteness.

Respond to this email with your pet’s name and pic for a chance to be featured here tomorrow!

MEMES FOR DESSERT

AI ART SHOW

WHAT WOULD YOU DO?

Need a ride?

Which luxury brand space rover would you drive? 👇

Can I take it home? 👨‍🚀

Ideas? Comments? Complaints?

Respond to this email or hit me up on 𝕏.

Until next time 🤖😋🧠