MindStudio Desktop App is Out + Llama 3 is an Open Source GPT-4

This week, we released our official Desktop app for MacOS and Windows.

In parallel, the AI industry kept moving at breakneck speed, with Meta releasing Llama 3 8b and 40b, two open source models that surpass Mistral in most benchmarks. We also got a new GPT-4 level model, Reka Core, and a new Batch API from OpenAI with a 50% discount on token cost.

Keep reading to learn more!

🗞️ Industry news

1. Llama 3 8b and 40b generally available, 400b in training & set to surpass GPT-4
   

   

   
   Llama 3 is here and it’s as groundbreaking as open source enthusiasts expected. Other than the very limited 8k context window, Llama 3 excels across the board, and the 40b variant outperforms Gemini 1.5 Pro and Claude 3 Sonnet.
   
   This is the first time we have an open source model that is this capable. And it’s just the beginning.
   

   

   
   Meta also announced Llama 3 400b is in training. From early benchmarks, Llama 3 400b gets MMLU 84.8 in pre-trained and 86.1 in instruct. For context, Claude 3 Opus is 86.8, meaning that Llama 3 has the potential to basically match SOTA (state-of-the-art) performance for 1/10th or even 1/20th of the cost when self-hosted.

   
   MindStudio is looking to add support for all Llama 3 models, and will release updates soon.
   

2. Reka releases “Core”, multimodal LLM that competes with GPT-4
   
   

   

   
   Reka is a fairly new startup on the AI scene, and just surprised the community with a multimodal LLM named “Core”.
   
   Reka Core performs better than Gemini 1.5 Pro and Gemini Ultra in most benchmarks, and includes multimodal capabilities such as Vision to understand images. The context window is 128k, the same as GPT-4 Turbo, though.
   
   What’s perhaps most interesting here is the cost of training and developing a foundation model. What previously required big-tech capitals (in the hundreds of millions) seems now attainable by smaller companies like Reka.
   
   It’s good to see competition driving innovation and lowering cost for the new layer of intelligence for all of humanity.
   

3. Grok 1.5 Goes Multimodal: Vision Announced Over the Weekend
   
   

   

   
   X AI, owned by Elon Musk, announced vision for their newest Grok 1.5 last week on Sunday.

   
   Similarly to Core and Llama 3, Grok-1.5V seems to nearly match GPT-4 performance. OpenAI is deemed to reply with a new GPT in the summer, given how many competitors are starting to pop up with alternatives to their leading model.
   
   Grok 1.5 doesn’t stand out in many ways, but is quite unique in how it’s trained. Being the only model with open and unrestricted access to X (previously Twitter), Grok 1.5 has the potential to become the best model when looking for real-world news and events.
   

4. OpenAI Launched the “Batch API” - Get a Response in 24h and Save 50% on Token Cost
   OpenAI has a new API endpoint, and it’s a weird one.
   
   From now on, users will be able to “batch” requests and receive a response within 24h. As a thank you for the long waiting time, OpenAI will grant a 50% discount on all token cost. This includes models like GPT-4 and GPT-4 Turbo.
   
   This can be a great option for large workflows that don’t require immediate responses and want to save up on cost.

🔥 Product Updates

This week, we released the MindStudio desktop app. You can download it from here for Windows and MacOS. We also revamped the website to accommodate the new menu item.

The desktop app lets you:

• Access all your favorite MindStudio features right from your desktop

• Enjoy a faster and more responsive interface

• Benefit from improved performance and stability

• & get that 🤯 “wow” feeling when seeing the MindStudio logo next to Slack on your device!

The team is also working on:

• Adding Image Generation Capabilities: you can already use DALL-E 3 in MindStudio with your own API key, but we want to make the process easier. Soon, you’ll be able to generate images from the editor just like you “send a message” block;

• New models: there are a plethora of new, advanced models that were released in the past 10d. MindStudio will work hard to include the most relevant ones like Gemini 1.5 Pro, GPT-4 Turbo (new), and Llama 3;

• Certifications: our product team is working to get SOC2 compliance in the coming weeks. After that, we’ll look into HIPAA and GDPR to build up our Enterprise offering. We know that certifications are crucial for enterprise team looking to handle PII or sensitive data, and we want to offer the best platform for your needs.

💡 Tip of The Week

You can use send message blocks to extract portions of a larger prompt and output the response.

For example, I have a workflow where a custom function scrapes a lot of data. We’re talking 50k+ tokens or more.

To save cost, I’m using Haiku to go through the scrape only one time and performing all the operations I need:

• Summarize key data

• Outline the structure of the content

• and more

Then, a series of send message blocks cuts and slices the refined output, now in a separate variable. Here’s how it looks like in the editor:

By doing so, I’m saving tons of tokens. Instead of resending the 50k output every time, I’m only processing the refined output (more like 1-2k tokens) and getting an even smaller output every subsequent interaction. They’re then all saved in variables I can re-use in my final prompts to the LLM.

If you want to learn more about techniques to save up on token cost, take a look at this masterclass video on our YouTube channel.

🤝 Community Events

You can now register for all our upcoming workshops on our redesigned event home page. You can find it here.

Here are the three closest workshops from today:

We heard your feedback and decided to focus on live builds and features currently tricky to learn and/or with a steep learning curve like RAG.

As a general reminder, our Discord group is an invaluable source of information, news, and more. The entire MindStudio team is active on the platform.

We host these hangouts every Friday, approximately at the same time unless something happens :)

🌯 That’s a wrap!

AI is moving faster than usual, and we’re happy to be part of this groundbreaking revolution. 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