Historical Moment for AI: Open Source Beats Closed Source + New Models to Fine Tune

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.

This week, MindStudio released copy/paste + 14 new models from Meta, Mistral, Google, Stable Diffusion and Eleven Labs. Copy/paste has been on our feature request board for a long time, and we’re sure power users will love it.

This week has been one of the most intense so far this year. In the same week, we got new models from Meta and Mistral, a new major update to ChatGPT called “SearchGPT” rivaling Perplexity AI, Udio v1.5 with the best song quality ever, an ETA for Grok 2 by xAI, and more Generative AI features for Illustrator by Adobe.

Continue reading to learn more. There’s lots to find out this week!

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. Meta releases Llama 3.1 8b, 70b, and 405b. The largest model supposedly beats GPT-4o
   
   

   

   Copyright Meta

   
   
   Meta has announced the release of the Llama 3.1, family their latest and most capable open-source AI model collection. The family consists of Llama 3.1 8b, 70b, and 405b.
   
   The flagship 405B parameter model is seemingly as powerful, if not more powerful, than state-of-the-art (SOTA) models like GPT-4o and Claude 3.5 Sonnet.
   
   

   Key points:
   

   • Llama 3.1 405B excels in general knowledge, steerability, math, tool use, and multilingual translation;

   • The new models feature a 128K context length, finally unlocking the real potential of the model family. The previous 8k context length was one of the most impactful limits of the previous generation;

   • These models are and will remain open source and ready to fine tune;

   • Over 25 partners offer day-one support, including major cloud providers and tech companies. MindStudio will soon join them and offer the entire family in the coming days;

   • New safety tools like Llama Guard 3 and Prompt Guard are introduced to promote responsible AI development.
     

   
   Unlike other popular models, Llama 3.1 is not multimodal. Meta AI offers image generation, but it’s a separate model dedicated to the task.
   
   Overall, Meta impressed the entire community with an incredibly capable 405b model, but most importantly, a much needed refresh to the context length of Llama 8 and 70b.
   
   The smaller models in the family are where it’s at. Llama 3.1 8b runs at 750 tokens per second on Groq, beating any other model of a similar size in cost, speed, and performance. Llama 3.1 70b is probably the best model to fine tune to reach GPT-4o level performance in your niche.
   
   Llama 3.1 405b is the best model, but it’s also the most expensive and will never be able to run on your local machine. It’s also pricey to train. While interesting, it doesn’t provide much as a differentiator other than being open source. The other two definitely do!
   
   We can’t wait to see what you build with Llama 3.1 in MindStudio.
   
   

2. Mistral releases Mistral Large v2, reaching state-of-the-art performance at a great price/performance ratio
   
   

   

   Copyright Mistral - Large v2 performance data

   
   
   Mistral AI has announced the release of Mistral Large 2, their latest flagship AI model.
   
   This new generation offers significant improvements in various capabilities, including code generation, mathematics, reasoning, and multilingual support.
   
   

   Key points:
   

   • Mistral Large 2 has 123 billion parameters, a 128k context window, and supports dozens of languages and 80+ coding languages;

   • The model sets a new frontier in performance/cost ratio, achieving 84.0% accuracy on MMLU benchmark;

   • It vastly outperforms its predecessor and competes with leading models like GPT-4o, Claude 3 Opus, and Llama 3 405B in code generation and problem-solving;

   • Improved capabilities in following precise instructions and handling long multi-turn conversations, as demonstrated by various benchmarks.

   • Excels in multiple languages, including English, French, German, Spanish, and many others, outperforming several other models on multilingual benchmarks;

   • Accessible via Mistral's "la Plateforme" and through partnerships with major cloud service providers like Google Cloud, Azure, Amazon, and IBM. It will soon be available in MindStudio as well

   
   To be honest, there’s nothing that really sets the model apart other than the multilingual capabilities and a slightly better cost per performance.
   
   Mistral Large is NOT open source like its smaller siblings, and can’t be fine tuned - at least for now.
   
   

3. OpenAI is testing a competitor to Perplexity AI and challenges Google’s Search dominance with SearchGPT
   
   

   

   The new SearchGPT - copyright OpenAI

   
   OpenAI is testing SearchGPT, a prototype of new AI search features designed to provide fast, timely answers with clear sources by combining AI models with web information.

   
   OpenAI is collaborating with publishers to ensure proper attribution and protect journalism, with clear in-line citations and links to sources. Additionally, a system is being launched for publishers to manage how they appear in SearchGPT, separate from AI training data opt-outs.

   
   As in current OpenAI style, the feature is being released to a small group of users and publishers for feedback, with general availability coming “in the future”.
   
   It’s incredibly frustrating to see OpenAI keep hyping up features that are half-baked and not available to the public. It seems like the current focus for the company is publishing blog posts, and not AI models.
   
   SearchGPT was long expected when the company acquired search.chatgpt.com and it was kind of an obvious move. This is not a real innovation, given many companies like Arc and Perplexity already offer AI search, but was a given as a next step for web search in ChatGPT.
   
   Let’s see when we’ll actually get our hands on it.
   
   

4. Runway AI leak shows the company might have used thousands of YouTube videos to train their Gen 3 video model
   
   A leaked internal spreadsheet reveals that Runway, a multi-billion dollar AI company, allegedly trained its Gen-3 AI video generator on YouTube videos.

   
   The spreadsheet shows links to thousands of YouTube channels and videos from major brands, media outlets, and popular content creators.

   
   According to a former employee, Runway used open-source software and proxy servers to scrape videos, avoiding detection by YouTube.

   
   The document contains multiple spreadsheets categorizing videos by themes, camera movements, and specific actions for targeted training.

   
   The training data reportedly includes links to piracy sites for animated movies and other copyrighted content.

   
   This is hardly surprising. Contrary to image and text, there aren’t many openly available videos. For models to get this good, companies will need to find ways to train them on high-quality data avoiding controversies like these.
   
   

5. Adobe launches new Generative AI features in Illustrator and Photoshop after T&C controversy
   

   
   Adobe is rolling out new generative AI features to Illustrator and Photoshop, powered by their latest Firefly Vector AI model, aiming to speed up creative workflows.
   
   

   New features include Generative Shape Fill for adding detailed vectors to shapes, improved Text to Pattern for creating scalable vector patterns, and Style Reference for generating outputs mirroring existing styles.

   
   The “Generate Image” feature is now generally available on desktop and web apps, with improved Enhance Detail for Generative Fill and a new Selection Brush tool.

   
   Following the latest controversy on their updated T&C, Adobe insists its Firefly models are ethically trained on licensed content.

This is not all, but I don’t think reading a 9,000 words newsletter is how you wanted to spend your Friday. Let’s run through other cool updates quickly:

• You can now fine-tune GPT-4o mini from the OpenAI playground. OpenAI is offering 2m free training tokens per day until September 2024. You can call fine-tuned models through a custom function in MindStudio;

• GPT-4o mini is seemingly outranking all models minus its big brother on the lmys arena, raising doubts about the validity of the rankings. While GPT-4o mini is a great model, it definitely can’t compete or outrank Claude 3.5 Sonnet. This reiterates the need for testing the models for your actual workflow rather than relying on generic benchmarks;

• Udio released v1.5, one of their biggest updates yet. The update includes the ability to generate 2 minutes songs (previously 30s), much clearer sounds and vocals, a new design, and much more. I tested it out and the results were insane, give it a go for free;

• In an interview with Jordan Peterson, Elon Musk announced Grok 2 finished training and should be released in August. Grok 3 is in training;

• Google released Gemini 1.5 Flash with a context window of 32k in the free Gemini app. It’s also coming to Google Messages, including for consumers in the EU. This is still a capped limit, as the real context length for 1.5 Flash is a whopping 1M tokens;

• Microsoft confirmed Bing isn’t allowed to scrape Reddit anymore and can’t serve results from the platform. Following Google’s licensing deal with the social platform, only Googlebot seems to be allowed to serve Reddit posts;

• The Verge published a controversial piece about the popular AI VC fund a16z about their endorsement for Trump. We keep politics out of this newsletter, but if you want to join the discussion it’s here.

Wow, this was intense. I really hope you enjoy AI as much as I do, or this might look like torture. We usually stick to 3 big news stories per week, don’t worry, it’s only for this release!

Let’s take a break. What do you think about how AI companies are training their models?

🔥 Product Updates

You can now display multiple user inputs on the same page within an application so that users can see everything that’s required from them. Group questions on different pages to organize inputs and create a better user experience.

If your current workflows consolidate inputs into a single input block, these will now be displayed together on the same page for the end user.

This was actually one of my first feature requests, as it’s the best way to build mini tools you can embed on your website for SEO and marketing purposes.

While this might sound like a smaller release than usual, it’s just a small step towards our next big update.

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.

Copy-pastable blocks will be JSON code. This means you can create your prompt library to resell it on your website, provide templates for your clients as an agency, streamline your builds with standardized prompts, and so much more.

This is going to turn the automation builder into a powerhouse and simplify complex builds while making you more productive and expand the reach of what you can do to promote MindStudio and your services.

💡 Tip of The Week

The image above is the reason why I asked for form-like inputs even before joining MindStudio, and I can’t be happier this is now a reality.

What you see above is the result of Ahrefs adding free AI writing tools on their website. Free tools have always been an effective SEO strategy, and AI makes it all more effective.

The SEO company is driving millions of clicks to tools you can build in a couple of minutes in MindStudio and embed on your website. This is much more effective than a blog post, and usually works faster due to lower competition.

I did the same thing for a previous client of mine, and the free tools overtook the traffic to all other pages combined in less than 3 months, with one gathering over 90k visits per month.

This is the before/after adding free AI tools to the website:

AI tools work for SEO - and they can work amazingly well.

Will it work all the time? No.

Most of the tools I published were a failure. But it’s SO EASY to build more to meet your leads when they’re looking for your products that’s not a big deal. It literally takes less time than writing a blog post.

Will it work forever? Also no.

Traffic will probably start to decline in the future. No SEO strategy works forever other than generic best practices like “write good content”. Take advantage of marketing opportunities when they arise. For the past year, AI tools worked pretty well and you can research your competitors on Ahrefs to see if it might work in your niche as well.

With the new input format, you can now build marketable assets to embed on your website that:

• Preview what you can do for your clients: words are cheap. SHOW your potential customers that you know what you’re talking about and can deliver results that blow ChatGPT out of the water;

• Limit how much they can spend: MindStudio offers advance rate limiters to avoid spending too much per lead. For example, you might set the max per user to $1 - cheaper than one click on Google Ads;

• Improve time on site: the more people stay on your site, the better they remember your name and product. It’s also beneficial on a technical standpoint;

• Links, links, links: people love to add links to useful tools and research-backed blog articles. MindStudio can now help you build both in a jiffy.

While this might seem like a tiny redesign, it’s a major change in how clients see your workflows. Some things just work better in a form.

🤝 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