GPT-4o mini, Big Tech Secretly Train AI, and New Startup by Tesla's Ex-Head of AI
PLUS HOT AI Tools & Tutorials
Hello and welcome to the weekly roundup by The Creators’ AI!
After a very brief lull (a week? maybe two?), we're back to the “new week, new AI model” rule. Plus, we've got another investigation proving that corporations aren't being completely honest with us (no way) and the launch of an EdTech startup by Andrej Karpathy.
Let's go discuss it.
You know who really benefits from AI development and makes money from it? Accenture. If you want to find out why this is so, follow this newsletter:
This Creators’ AI Edition:
Featured Materials 🎟️
News of the week 🌍
Useful tools ⚒️
Weekly Guides 📕
AI Meme of the Week 🤡
AI Tweet of the Week 🐦
(Bonus) Materials 🎁
Featured Material 🎟️
OpenAI Unveils GPT-4 mini
OpenAI has unveiled the GPT-4o mini, its latest small AI model. The company claims this platform is cheaper and faster than OpenAI's current advanced AI models. GPT-4o mini is already available to developers, as well as through the ChatGPT web app and a mobile app for consumers. What's unusual is that enterprise users will be the last to get access to the new AI. Let's see what GPT-4o mini can do.
Technical Сapabilities
The GPT-4o mini is OpenAI's smallest model, replacing the GPT-3.5 Turbo. And here's a set of boring numbers for you. The Artificial Analysis tests show that the model scores 82% on MMLU, a benchmark test for measuring reasoning, compared to 79% for Gemini 1.5 Flash and 75% for Claude 3 Haiku.
On the MGSM, which measures mathematical reasoning, the GPT-4o mini scored 87%, compared to 78% for Flash and 72% for Haiku.
The image below compares it with other models. Price here is a combination of input and output tokens.
In addition, OpenAI says that the GPT-4o mini is significantly more affordable to operate than all of its previous models. Specifically, this AI is 60% cheaper than GPT-3.5 Turbo. Right now, GPT-4o mini supports text and vision in APIs; in the future, the model will start working with video and audio.
Keep your mailbox updated with key knowledge & news from the AI industry
For developers working on the OpenAI API, GPT4o mini is priced at 15 cents per million input tokens and 60 cents per million output tokens. The model has a context window of 128k tokens, supports up to 16k output tokens per request, and has knowledge through October 2023.
With an improved tokenizer common to GPT-4o, processing non-English text promises to be more efficient.
The exact size of the GPT-4o mini is not known. The company says the model is about on par with direct competitors: Llama 3 8b, Claude Haiku, and Gemini 1.5 Flash. That said, the AI from OpenAI is very fast - with an average output speed of 202 tokens per second. This is more than 2x faster than GPT-4o and GPT-3.5 Turbo.
In ChatGPT, Free, Plus, and Team users can access GPT-4o mini starting today instead of GPT-3.5. So go for it!
As AI evolves, small models are becoming increasingly popular among users. This makes sense, as they are faster, more cost-effective, and are already a valuable option for simple tasks where the developer repeatedly calls on the model.
News Of The Week 🌍
Investigation: Big Tech Firms Secretly Train AI Using YouTube
This is a serious investigation that, unfortunately, doesn't surprise much (but that doesn't make it any less important). According to Proof News, Apple, Nvidia, Anthropic, and other companies used materials from thousands of YouTube videos to train their models. We're talking about subtitles from 173,536 videos posted on 48,000 channels. Creators claim that the videos were used without their knowledge.
Proof News has compiled a video version of their investigation if you're too lazy to read the article. You can watch it here:
Well, such situations are becoming more and more numerous (remember the Perplexity case), and the relationship between copyright and AI training is becoming more and more acute. The question is what such news will lead to in the future and how it will affect the industry.
Andrej Karpathy’s Startup Aims to Apply AI Assistants to Education
Tesla's former head of AI and OpenAI researcher has launched a new company. It's called Eureka Labs and focuses on education with AI. The idea is that the startup aims to use the latest advances in generative AI to create digital teachers who can guide students through different learning materials. According to Karpathy, this platform will allow people to approach “an ideal experience for learning something new.”
The founder discussed his plans in this X post. He also announced that Eureka Labs' first course will be “the world's obviously best AI course, LLM101n.”
A startup developing AI for learning is not exactly a new story. You know what I am talking about if you are familiar with Praktika AI. Nevertheless, Karpathy's extensive experience may well guarantee a decent realization of this idea.
Sharing is caring! Refer someone who recently started a learning journey in AI. Make them more productive and earn rewards!
The Information: Meta Will Release Largest Llama Model on July 23
The new model, with 405 billion parameters, will be multimodal and able to understand and generate images and text. The report follows the April release of the smaller Llama 3 models with 8 billion and 70 billion parameters, which have already gained significant traction among developers. The upcoming release marks an important milestone for Meta, coming about a year after the launch of Llama 2.
Save the date. July 23.
Google, Microsoft, Nvidia и OpenAI Teamed Up to Promote AI Security
Several major tech companies are joining the Coalition for Secure AI (CoSAI) to share information on making AI 'secure by design.' The initiative aims to address a “fragmented landscape of AI security” by providing access to open-source methodologies, frameworks, and tools.
The current lineup includes all those names you know well: Google, OpenAI, Microsoft, Amazon, Nvidia, Intel, IBM, PayPal, Cisco, and a few other AI-related companies.
The founders of CoSAI promise to help organizations large and small integrate AI safely and responsibly, helping them harness its benefits while mitigating its risks. How exactly this will work, we'll find out (or not) in the near future.
YouTube Music is Testing an AI-Generated Radio Feature
YouTube Music's new feature is reminiscent of existing AI-based playlist generation tools. You may have seen similar features from Spotify, Amazon Music, and Deezer. Users who access the new feature have a new “Ask for music any way you like” card in their feed. After tapping on the card, the app will open a dialog UI where you can enter a custom request or select a suggested one.
According to TechCrunch, while the feature is currently only available for select users, YouTube plans to make it available to more users in the future.
Useful Tools ⚒️
fastn.ai – The no-code AI orchestration platform developers love
Builco – Build MVPs with Next.js using AI in minutes
RewriteBar – Improve your writing in any macOS app with AI help
VirtualReception.AI - Your AI Answering Service. Never Miss a Call
Flow Studio – Turn text description into the movie
Flow Studio is a text-to-movie app that allows you to turn your written stories into short films. It generates fully produced, film-quality 3-minute videos with plots, consistent characters, and automatically matched background music.
It's a great platform for lazy people with good ideas.
Share this post with friends, especially those interested in AI stories!
Weekly Guides 📕
How To Use AI To Find Top Leads & Generate Custom Sales Pitches
Easily Create Your First GPT Today | Custom GPTs Tutorial
How to make videos like Ray William Johnson (using Leonardo AI)
How To PROPERLY USE Generative AI in Lightroom
AI Meme Of The Week 🤡
Great, now AI is taking care of nature!
AI Tweet Of The Week
You know, like… That's weird, even for Twitter. The video is here.
(Bonus) Materials 🏆
TTT models might be the next frontier in generative AI
How Ai Is About To Transform The World’s Economy
The AI-Powered Future of Coding Is Near
OpenAI Touts New AI Safety Research. Critics Say It’s a Good Step, but Not Enough
These AI Models Are Pretty Mid. That’s Why Companies Love Them.
How was this week in AI? Share your content and ideas in the comments to this post so we can discuss or include them in the next edition!