TOGETHER WITH HYPERMODE.

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

Here’s what’s on the menu today:

• OpenAI's new AI stops and thinks 🤖 🧠

  OpenAI released o1 and o1-mini.

• Google’s new AI models 👾 🎮

  Google introduced DataGemma.

  New here? Subscribe!😎

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

MAIN COURSE

OpenAI’s new model has reasoning skills 🧠 🤖

Yesterday, OpenAI released o1 and o1-mini. They're the first in a series of "reasoning" models designed to handle complex queries better than previous models.

What is o1?

It's an AI model trained to solve problems step-by-step, similar to human reasoning.

Is it any good?

Yupp. It scored among the top 500 U.S. students in math, outperformed PhDs on science tests, and ranked in the 89th percentile of competitive programmers.

How is it different from GPT-4o?

It stops and “thinks“ before spitting out answers. It uses reinforcement learning and "chain of thought" processing, which makes it better at coding, math, and more.

Any cool features?

It shows its reasoning process in real-time, using human-like phrases like "I'm curious about" and "Let me see."

Who can use it?

ChatGPT Plus and Team users get access now. Enterprise and Edu users next week. Free users will get o1-mini later.

Is it expensive?

Very. API access costs $15 per million input tokens and $60 per million output tokens – much pricier than GPT-4o.

Any downsides?

It's slower than GPT-4o for some tasks and still has issues with hallucinations.

SPONSORED BY HYPERMODE.

Integrate AI into your Application, effortlessly

Don’t let complex tools slow you down.

Hypermode helps developers integrate AI-powered search and recommendations with just a few lines of code.

Ship faster without getting bogged down by backend complexity.

Built by the team who brought you Vercel, Astronomer, Sentry, and VS Code.

SIDE SALAD

Google’s new AI models 👾 🎮

Yesterday, Google introduced DataGemma. These are new versions of its Gemma models that use trusted statistical data to reduce AI hallucinations.

What is DataGemma?

It's an AI model family that integrates Google's Data Commons, a repository of 240 billion data points from trusted sources.

How does it work?

It uses two techniques:

• RIG: Queries trusted sources before responding.

• RAG: Retrieves relevant data before generating answers.

How accurate is it?

Using RIG, factual accuracy jumped from 5-17% to about 58%. With RAG, it achieved 98-99% accuracy for specific numerical values.

Any limitations?

It struggles with data coverage and drawing inferences from statistics. Also, it's an early release for research purposes only.

Why is this important?

It's a significant step towards more trustworthy AI, especially for fields requiring precise data.

What about previous versions?

Last month, we covered Gemma 2. It’s a small, 2-billion-parameter model that outperforms larger models, in conversational tasks.

YOUR DAILY MUNCH

Think Pieces 🧠

Why is Canada developing AI military technology? The goal is to assist with tasks by 2030, ensuring humans remain responsible for important decisions.

Startup News 💰

Oracle introduced over 50 AI Agents in its Cloud Applications Suite. The aim is to automate business tasks like scheduling and order management.

Research 👨‍🔬

INTRA — a method that helps models understand how objects can be used by learning from pictures of people interacting with them, without needing pairs of specific images.

OneGen — a method that enables LLMs to both generate text and retrieve information in a single step without needing separate models.

FURRY FRIENDS

Say “hi” to Freya. 🐶

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?

Character Mashups… 🧥

Who is your favorite? 👇

Ideas? Comments? Complaints?

Respond to this email or hit me up on 𝕏.

Until next time 🤖😋🧠