JPMorgan steps into the LLM game

Greetings!

Welcome to The Menu Magic - Finance & AI weekly newsletter

In today’s email:

• JPMorgan steps into the LLM game

• A very stupid gadget to replace the iPhone

JPMorgan steps into the LLM game

Dear friend,

So, JPMorgan just dropped this new thing called DocLLM, and it's all about making sense of those tricky documents we deal with every day. But let me break it down without the tech jargon.

DocLLM is a new tool for understanding documents. It's not just about reading words but also understanding how the page looks. Imagine e-books, online articles, and other documents with fancy layouts – DocLLM can handle them without needing a ton of prep work. It's like giving the model eyes to see the structure of the document itself.

Now, here's where it gets cool. DocLLM doesn't get confused by long documents. It knows where one-page ends and the next begins. This is a big deal because most other models struggle with super long stuff or focus on single pages. Plus, in training, it learns from examples of what the output should look like – either text or images – making it super handy.

But wait, there's a catch. DocLLM's secret sauce is making sure it fills in missing parts during training, so it doesn't give you random or disconnected info. The tricky part? Figuring out the best tools to get this info. They mention using Microsoft Azure API for better results for OCR, but there's still room to explore other options.

So, what does this mean? It's like DocLLM is upgrading how language models learn, making them understand not just words but the whole document vibe. It's a new way for anyone to deal with diverse and complex documents. And hey, if researchers find even better tools, it could get even more powerful.

Now, the real talk. The dataset they used – CDIP – is a bit biased. I mean, 16 million pages about the tobacco industry? Talk about niche. And, guess what, finely-tuned models still outshine this whiz kid in certain scenarios. So, if you're dealing with ever-changing docs, maybe stick to the finely-tuned crew.

Now, the big question: Is this DocLLM the next big thing, or is it going to be that cool gadget no one notices? Think of it like BloombergGPT's sibling, stepping into the scene. Will other financial institutions jump on the bandwagon?

What's your take on this? Can you see other banks and financial folks giving DocLLM a high five, or is it destined to be the unsung hero in the AI saga?

You can find more about it here

Catch up soon, and let me know what you think!

Rabbit: A very stupid gadget to replace the Iphone

During CES, a company called Rabbit launched their first AI-based device. They promote it as the evolution of smartphones because they don't use apps. Instead, with a couple of buttons, you interact directly with the AI.

The 25-minute presentation in Las Vegas only shows one interesting thing: more people are trying to create AI-native devices, but their learning functionality could be something intriguing. It's like showing your AI companion how to do certain specific tasks.

I think the project will fail, but those learning capabilities (borrowed from AI agents) could be very interesting for other usages and apps...

You can find more about the R1 Rabbit device here

I'd love to hear your feedback on today's newsletter! Is there a specific type of content you'd like to see more of in the future? Since I'll be releasing a new edition each week, I welcome any suggestions or requests you may have. Looking forward to hearing your thoughts!

The Menu Magic is written by Francisco Cordoba Otalora, an entrepreneur living in London.

Share this newsletter with a friend