GPT-4o mini is here, New Dynamic Features in MindStudio & Firecrawl integration

You're receiving this email because you registered to one of our workshops. You can unsubscribe at the bottom of each email at any time.

MindStudio released a first-party integration with Firecrawl for scraping web URLs, new voices for the text-to-speech models, and a few quality of life improvements like a better audio player and LLM extraction for scraped URLs.

The industry didn’t sleep either. OpenAI broke its long-lasting silence and released a new model, GPT-4o mini, replacing the previous OG GPT-3.5 Turbo. GPT-4o mini scores better than Claude 3 Haiku and Gemini 1.5 Flash while being substantially cheaper.

In less enticing news, Meta confirmed they don’t plan on releasing Llama 4 to EU customers and a third-party audit confirmed Apple, Anthropic, and other companies used YouTube videos to train their AI. MKBHD replied claiming this is a double-breach given he pays for every transcription.

Continue reading to learn more!

As a reminder, we’re now welcoming partners that want to build AIs for their clients. Sign up for extra support, training resources, and more here.

🗞️ Industry news

1. GPT-4o mini is here. It’s cheaper than GPT-3.5 Turbo, multimodal, and smarter than similarly priced competitors
   
   

   

   Originally published by OpenAI in its intro article

   
   
   OpenAI just dropped GPT-4o mini, their most budget-friendly small model yet. It's a game-changer for AI apps, making intelligence way more affordable for everyone.
   
   GPT-4o mini scores 82% on MMLU and beats other small models like Gemini 1.5 Flash and Claude 3 Haiku across various benchmarks.
   
   Key insights:
   

   • We added GPT-4o mini minutes after release, and it’s currently priced at $0.20/1M tokens in input and $0.78/1M tokens in output in MindStudio;

   • It supports text and vision. You can already use GPT-4o vision in your MindStudio workflows. It’s unclear whether the vision capabilities are also significantly cheaper than the big brother GPT-4o;

   • OpenAI built in the same safety features as GPT-4o, including filtering out problematic content and using techniques like RLHF;

   • The model is now live in the Assistants API, Chat Completions API, and Batch API. ChatGPT users (Free, Plus, Team) can access it too.

   
   I haven’t tested GPT-4o mini enough to form a nuanced opinion, but it does seem to perform quite well. Claude 3 Haiku has been my go-to for basic workflows until now, but GPT-4o mini might replace it.
   
   Just to contextualize how crazy cheap the model is, Sam Altman mentioned the state-of-the-art model in 2022 cost 100x more.
   
   The top user by usage in MindStudio used a total of 1.25 billion tokens. Assuming 80% is input and 20% is output using GPT-4o mini, that’s $395.
   
   Yes, you read that right. Less than $400 to write approximately 900 million words. I remember when writing an AI article cost $29 each…
   

2. Meta won’t release Llama 4 in the EU due to confusing regulations and slow responses from the bloc
   
   

   

   
   
   Meta's skipping the EU for their new AI that handles video, audio, images, and text. Why? They're spooked by the EU's regulatory maze.
   
   Even though it's an open license model, European companies are left out in the cold.
   
   The EU just locked in compliance deadlines for its AI Act. Companies have until August 2026 to play nice with rules on copyright, transparency, and AI uses like predictive policing.
   
   Apple's pulling similar moves, likely keeping its Apple Intelligence out of the EU, at least for now.
   
   A text-only version of Llama 3 400b will likely still be in the roster for EU consumers, but nothing after that.
   
   The EU has always been faster at regulating tech companies, but this time around they might be pushing their hand too hard. The AI regulations are indeed quite blurry and hard to follow, and the bloc is notorious for heavy red tape and bureaucracy.
   
   As a European citizen, it’s sad to see companies like Apple and Meta openly skipping the EU market without thinking too much about it. It highlights the EU is not embracing innovation, yet again, and might be left behind in the economy of the future.
   

3. Apple Intelligence and Claude are trained on YouTube videos from more than 48,000 channels (to no one’s surprise)
   

   

   
   It looks like AI companies have been sneaking YouTube content into their training datasets.
   
   Over 170,000 YouTube videos' subtitles were snagged without permission for AI training. We're talking about content from 48,000+ channels.
   
   Apple, Anthropic, Nvidia, and Salesforce are all caught up in this. The training datasets they used included videos from all sorts of YouTube channel: from MrBeast and MKBHD to major news outlets like ABC News and BBC. Salesforce and Apple replied for comments, and you can read their response here.
   
   MKBHD, one of the largest tech channels on YouTube, isn’t happy, calling it an "evolving problem." Creators are just starting to realize their content has been used without consent. He claims it’s even worse given he has humans transcribing all his videos for accessibility purposes.
   
   This YouTube dataset is part of a bigger open-source collection called The Pile, put together by EleutherAI. The tech giants didn’t directly scrape the videos, but they purchased data from a company that did.
   
   Both YouTube's CEO and Google's big boss say using YouTube content (even transcripts) for AI training violates their terms.
   
   To be honest, this shouldn’t really surprise anyone. It was clear that all of the models included data from YouTube transcripts. The real question is not whether it’s ok to scrape YouTube videos, but whether it’s ok to scrape UGC without explicit user consent as long as it’s not replicated 1:1 by the models.
   
   There’s an ongoing copyright battle between the NYT and OpenAI that might give us more details into how the courts interpret copyright laws in relation to LLMs.

🔥 Product Updates

You can now use GPT-4o mini in MindStudio, text & vision included!

Please note the model isn’t on par with the state-of-the-art options like Claude 3.5 Sonnet, GPT-4o (the big brother), and Gemini 1.5 Pro. BUT, it should outperform Claude 3 Haiku and in my initial testing it’s quite good at outputting valid JSON - very useful for workflows that include sending data externally.

MindStudio also shipped a new “Scrape URL” block, replacing the old community function. The new block lets you choose between two providers:

• Default: our previous scraper. This will scrape the content of the page and save it as plain text in a variable;

• Firecrawl: our newest provider. Firecrawl will scrape the website and return the content in markdown in a variable. In addition to the base scraping functionality, Firecrawl has tons of useful extras like:

  • Taking screenshots;

  • Using the internal LLM function to find specific components in the web page and extract them as JSON;

  • Pass in specific headers.

We also shipped new voices for our text-to-speech models from OpenAI. You can now choose between:

• Alloy

• Echo

• Fable

• Onyx

• Nova

• Shimmer

All with their own unique accent and tone. The update comes together with a new audio player that lets you control the volume, open the host URL, and looks more visually pleasing:

Finally, you will find new options to direct what kind of images you want DALL-E 3 to generate, including options for size and details. All combinations come with a different price tag, and you can refer to our pricing page to check the current price.

These are just our first steps into enhancing your multimodal workflows. We want to give you as many options as possible to automate entire portions of your role, not just a small text-based workflow!

What’s coming next:

• The new scrape block with Firecrawl is just the beginning. Now, you can scrape a page for data. Next, we'll upgrade Firecrawl to do full website crawls. Finally, you'll be able to save the crawl results as a data source;

• MindStudio has a brand new learn page with a new design, a plethora of new tutorials, and a more organized view of upcoming webinars;

• We are going to release copy-pastable blocks very soon. You will be able to copy automation blocks within the same workflow or bring them from one app to another. It will work with all automation blocks and will include prompts within Generate Text blocks. This will help speed up complex applications significantly (I’m testing it out and it’s awesome).

💡 Tip of The Week

I’m an AI nerd. Many of you reading this are too.

Too often, nerds are so focused on what cool new feature is coming next they get stuck in the shiny object syndrome and forget just how far we’ve come.

On top, you can see the difference between an AI-generated image in 2021 (displayed by OpenAI to show how cool it was) and one in 2024.

I can show similar examples for text-only AI, going from autocomplete GPT-2 to full articles for less than a cent with GPT-4o mini or Claude 3 Haiku.

AI Intelligence is one of the only technologies that went from incredibly expensive to affordable for most people and companies in the span of a couple of years.

Take advantage of it.

You don’t need yet another DALL-E to start generating images.

You don’t need GPT-5 to completely revolutionize how you create content.

You don’t need your voice cloned in a professional studio to start editing with AI.

You have everything you need to completely transform how you work. It might not get you 100% there yet, but it will save you HOURS and make previously impossible workflows for small individuals possible.

Let me give you a personal example.

I built my first blog at 13y old and never stopped working in digital marketing. I wrote hundreds of articles in multiple languages, managed websites from small one-pagers to e-commerce shops, and every single time creating content was a very, very complicated process.

It usually involved hiring multiple junior writers, a couple of managers, establishing whole pipelines to fact-check, and more. In the hyper-competitive tech space, it’s not uncommon to pay over $500 for an article.

With AI, a quality article can cost as little as $0.10 or so. Then, you can get a human reviewer to ensure factuality and add the final touches. You save money, they save the tedious part of the job.

As a solo content marketer w/ AI in a previous role, I was able to publish nearly an article per day, add dozens of AI-first tools to generate more traffic, and even outreach to blogs and journalists to get featured on their blogs. This would have been impossible to handle as a one-person team in the past.

Again, will it work all the time?

No, it won’t. Most of this newsletter is 100% written by me, AI only proofreads it in the end and helps with news summaries.

But it will work for the vast majority of content you’ll publish: glossary pages, programmatic SEO pages, e-commerce page descriptions, SEO metadata, open graph values, alt tag for images (thousands in one go), accessibility tags, and more.

All of this is now not only possible, but very easy to achieve. And this is just ONE example in ONE niche.

Just wanted to take a moment to remember how far we’ve come, and what’s already possible, today, with AI.

🤝 Community Events

If you want to hangout with our team, we usually host a Discord event every Friday @ 3PM Eastern. Join our Discord channel to keep up to date with the hangouts - our entire team is active there.

You can register for upcoming events on our brand new events page here.

Our new webinar series is up on there as well, with the following on-demand webinars:

• MindStudio for Partners: Level Up Your Agency With AI Workflows

• Customize AI Applications With Your Data

Plus, we have new weekly and bi-weekly events:

• How To Build Your First AI Assistant (the best webinar for beginners!)

• Which Models to Choose for Your AI

• First Steps to Bring AI Into Your Workflow

Thank you for being an invaluable member of our community, it’s always great to see many of you join multiple workshops 🔥

If you’re interested in any topic in particular, feel free to reply and I’ll do my best to include it in the next releases. We’re going to update all of these soon.

🌯 That’s a wrap!

Stay tuned to learn more about what’s next and get tips & tricks for your MindStudio build.

You saw it here first,

Giorgio Barilla
MindStudio Developer & Project Manager @ MindStudio