MindStudio AI to build MindStudio Workflows + OpenAI working on Project "Orion"

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 finished refactoring the API, built a workflow-builder to generate MindStudio apps in a minute or two for any use case, and added a UI for the renaming feature.

The industry was also busy cooking. OpenAI hinted at “Project Orion”, with the Information reporting they’re using Strawberry to train Onion. Magic reached a 100M context window, OpenAI is raising at a $100b valuation with interest from Apple and Nvidia, Meta announced Llama is approaching 350M downloads on Hugging Face, Google launched their GPTs clone and a few new experimental models.

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

OpenAI wakes up and signs agreement with the Feds while hinting at a new model and raising @ $100b valuation

After an incredibly disappointing summer, this week The Information reported OpenAI is actively working on a new model using synthetic data generated by Project Strawberry.

The article is somewhat unclear, but here’s what we know:

• We thought Project Strawberry would be the new model. Now, it looks like Project Strawberry is being used to train Orion, and that it might be too slow to be released publicly;

• It’s now much easier to understand why we got so many trailers and no new features. OpenAI is bleeding money left and right, and they’re looking to raise. The more outstanding promises, the more likely they are to get a significant influx of cash to keep developing;

• OpenAI and Anthropic signed an agreement to release their new powerful models with the Feds before releasing them publicly. This contrasts with the company's previous ethos, where Sam Altman claimed that the best way to gauge a model's capability was to release it as soon as possible.

Google launches 3 new experimental models, a clone of custom GPTs, and in-browser Gemini chat

Google is on a rampage. In a week, Google announced:

• A new, smaller variant of gemini 1.5 Flash-8b, one of the most cost-effective models with solid capabilities;

• A stronger Gemini 1.5 Pro model;

• An improved Gemini 1.5 Flash model;

• A new Imagen model, Imagen 3, with the ability to draw people.

All of these are available immediately in AI Studio and/or Gemini Chat, with the industry-leading 2 million context window.

In addition, Google is working on adding value to Gemini Advanced, arguably one of the best alternatives to ChatGPT Premium / Teams.

Gemini Advanced will now come with Custom Gems, a concept very similar to ChatGPT custom GPTs and Claude’s Projects. It’s Google’s idea of personalized AI solutions.

As with most other custom chatbots, Gems are quite underwhelming and nowhere near the power and flexibility of an end-to-end AI builder like MindStudio, but they can be a good starting point to test maxi-prompts you can then transfer as system prompts to MindStudio!

Magic reaches 100M context window - what’s next for RAG and Open Source AI?

Magic is advancing ultra-long context AI models that can process up to 100 million tokens during inference, revolutionizing areas like software development by allowing AI to reason with extensive codebases and documentation.

Current evaluation methods for long-context models have flaws, so Magic developed "HashHop," a more robust benchmark to measure a model's ability to retrieve and connect large amounts of information.

The news comes together with Meta announcing Llama reached 350M downloads on HuggingFace, further strengthening the case for Open Source AI.

Food for thought:

• The world is now deciding if AI is a dangerous technology, amid the ranks of nuclear weapons, or a positive technological revolution. It’s interesting to see leading tech companies position themselves on the spectrum between “we need to keep everything closed up or the world will explode” and “let’s release weights, code, and everything in between”;

• We’re expanding context to extreme levels, which might make RAG feel outdated in a couple of years. While it’s still very useful for specific workflows, RAG falls short when you need the AI to know a lot about a huge corpus of data, for example a codebase, a whole book, or more. Context window went from a few thousands words to 2 million in a little over 2 years.

In other news, NVIDIA reported higher-than-expected quarterly earnings but investors were still unhappy, leading to a significant 6% drop, and it’s thinking to invest in OpenAI to continue innovating.

With all players caught up to GPT-4o and OpenAI losing tons of money every day, they’re forced to release something incredibly cool soon. With interest from Apple, NVIDIA, and likely another round from Microsoft, OpenAI could get cash it desperately needs to move forward.

🔥 Product Updates

You can now rename blocks in the editor, and we deployed a UI that goes along with the feature.

To rename a block right click on the block and select “Rename”. You can rename all functional blocks, but not the “start” or “terminator” blocks.

Next week, we will also release version 2 of our API, which comes with huge improvements to the stability, performance, and dev resources:

Here’s what’s new:

• Workspace-Level API Management: Manage your API settings directly from your workspace, with access to high-level statistics on successful and failed requests;

• API Keys Management: View, create, and delete API keys easily. Track when each key was created and last used;

• API Logs: Access detailed logs of API requests, including request and response data, HTTP status codes, and source information;

• Debugger Improvements: API requests are now visible from the debugger and tagged as “API” for easy identification during development;

• Help & Documentation: Access sample requests and documentation directly from the API settings in your workspace.

If you’re currently using the API, keep an eye on your inbox to receive further updates.

Lastly, copy-paste was the last piece of the puzzle to make it possible to generate automated workflows in MindStudio. So, I’m excited to tease the upcoming workflow builder.

The workflow builder is one of the biggest AIs I’ve built in MindStudio. The prompts alone are around 30 pages, and it includes a lot of MindStudio trainer’s learnings + new insights.

The workflow builder will simply ask a few questions and then:

• Outline all the required inputs;

• Choose the best models for you;

• Generate the prompts for all generate text, generate image, and generate audio blocks;

• Assign variables;

• Structure the WHOLE workflow, creating JSON blocks from start to finish.

This means that you can get a fully functional workflow in one click in around 2 to 3 minutes. The workflow isn’t a simple one block AI app, but usually includes multimodality, factors in variables like the desired complexity, prompt length, and token expenditure.

The builder is currently in alpha stage and the team is testing it internally. But don’t worry, we’re not OpenAI and will release it very soon. We’re targeting the end of next week for a first beta.

Here you can see a very brief demo.

After a few weeks in beta, I’m looking to include the workflow builder in MindStudio Trainer together with a new version, MindStudio Trainer v3. This should come around the end of September or beginning of October.

💡 Tip of The Week

Don’t obsess over tokens, and use the context window of models like Gemini 1.5 Flash to test out massive prompts versus chain prompting.

There are two core strategies to prompt an LLM:

• Chain prompting: multiple prompts, one after the other, that focus on one task. For example, if you’re writing a blog post, you might want to generate the title first, then the outline, then the actual sections;

• Mega prompts: these where the only option at the beginning. They’re usually massive, multi-pages prompts with a huge set of instructions for the LLM to follow.

Realistically speaking, you’ll likely need both in your workflows.

For example, the new workflow builder includes chain prompting to first outline the required inputs, then generate the JSON for them, then outline the best models, generate the prompts, and finally create the workflow.

Each of these steps is a massive prompt, especially the last one, which is over 10 pages of a standard document.

Nowadays, models can ingest a huge number of tokens, and while it can sound scary, cost fell dramatically over the last year and is not that big of an issue anymore. If you want to save money, try all workflows with something like Llama 3.1 8b or Gemini 1.5 Flash. They’re so cheap you’d struggle to spend over two cents even for significantly large prompts.

Remember, AI is not a chatbot, and MindStudio is your personal IDE that helps craft your way to professional custom AI solutions. Don’t limit yourself and keep building!

🤝 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 (new video!)

• 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