bilater

yes

nice! looks cool.

glad you liked it!

haha

I did something similar using ffmpeg WASM in the browser. But it does have limits sadly. You can't really do heavy tasks. https://vididoo.vercel.app/

Cool. I think low hanging fruit to add drag n drop. Some options are libraries like dnd-kit or hello-pangea/dnd.

Many people will analyze this from a tech perspective. I urge you not to. This isn't about whether the technology is good or bad. The real reason they succeeded is their mastery of marketing, particularly TikTok and influencer marketing. Understanding distribution is far more important than knowing how to build something, and this is truer now than ever before. I know it stings but its the truth.

What is this dumb article? Gemini pro 2.5 is getting talked a lot about in AI circles. AI Studio team is cracked and has been able to avoid the committee virus. Google has problems but this piece is written by someone who is not knowledgeable about this company and the space at all.

I'm sick of em dashes cause somehow that's become the tell its AI generated text.

yeah i guess it depends on your app...if your whole paid tier relies on access to protected pages where the check happened in middleware then its a big issue, but if have additional checks there such as checking userid and subscription inside routes then its not as big of a deal as the user in theory wont be able to do anything.

BTW ppl are talking about why middleware should be used for auth and, while I don't like this pattern, it is the adopted pattern for app router in nextjs and services like clerk and supabase use it heavily.

Will hopefully write a blog about it soon but its basically Mistral OCR + Pinecone (not ideal but I wanted something up and running).

yup keep it simple. hybrid search on supabase with keyword + pg vector. use a good embedding model with inner product (not cosine so its faster). there was a good article about learnings in this space recently here https://news.ycombinator.com/item?id=43299659

na you're right this would be a dumb move with a huge blow back

I think the more critical thing is an open UI protocol.

Let me explain - imagine generative AI is good enough we can just generate a UI on the fly. Take to its extreme every user can have a personal block of UI components catered to their preferences (dark mode, blue color scheme, bigger fonts etc). Then instead of every business designing their own UI for their website they just send over the information and the UI is compiled for every user baed on their own personal set of blocks.

We would very quickly need to have some sort of standard protocol to make this work. I think that would be a way more efficient world because companies can focus on the content and not on tweaking design. And every user has a lot more control over their user experience.

Of course a lot of companies may want to control the experience themselves so maybe its not one way or the other but a good chunk of websites cna use this pattern and in time it may actually become an advantage as users expect you to follow this UI protocol.

I had my own spin on deep research which you might find this easier to navigate: https://github.com/btahir/open-deep-research

thanks. yeah there is an annoying limit on vercel's end for body size of 4.5 mb. there are solutions to get around it by uploading directly to a storage solution but i wanted to keep the app simple for folks who want to dive into the code.

haha yes I was very happy to get the domain

thanks!

Added!

thanks!

thanks!

its not something you build with just a prompt. at least not yet.

thanks!

Its open source.

Yes - unfortunately it seems they don't read images in the pdf.

Agreed - I am surprised they did are not using Pixtral to read images as well.

Thanks! Cursor definitely helps me speed things up haha.

OK I've been critical of Mistral AI but credit where credit is due. Mistral OCR seems cool.

So cool in fact, I got distracted and ended up building an open source PDF parser and chat app!

Presenting Auntie PDF - your all-knowing guide that unpacks every PDF into clear, actionable insights.

You can upload a pdf or point to a public link, parse it, and then ask questions. All open source and free.

The problem with having such a specific, prescriptive formula for success is that it never actually works out that way. Sure, there are high-level principles, the PostHog team executes brilliantly, and I love the product, but I think we're really bad at connecting the dots on what actually made something successful. Instead, we assign credit to random things just to make it all make sense. A lot of times, it's the equivalent of saying, "Bill Gates eats this cereal for breakfast, so if I do that, I should become a billionaire too."

I wet through this journey myself with Deep Search / Research https://github.com/btahir/open-deep-research

I think it really comes down to your own workflow. You sometimes want to be more imperative (select the sources yourself to generate a report) and sometimes more declarative (let a DFS/BFS algo go and split a query into subqueries and go down rabbit holes until some depth and then aggregate).

Been trying different ways of optimizing the former but I am fascinated by the more end to end flows systems like STORM do.