NerfGun

I loaded the research papers OpenIntelligence's search is built on into the app and asked it to explain its own search. It's three steps: BM25 for the words, Sentence-BERT for the meaning, and Reciprocal Rank Fusion to merge the two rankings. For each one the video shows the paper's own lines and the Swift that does it. github.com/Gunnarguy/OpenInt…

In September, 5 people downloaded OpenIntelligence from a link inside the ChatGPT app. Neat.

People in 90 different countries have downloaded my iOS/macOS apps. 739 of the 1,666 downloads are from the US, and the rest are from everywhere else, which is wild.

I build pretty much everything with AI agents now, and when a model hands you something, it's hardly gold. The whole job is knowing when plausible isn't true. That's the unglamorous part, and it's the part that separates vibecoded slop from a working product.

Five apps shipped and four on the App Store now, built by MEEEE OpenAssistant, my first, 727 downloads. OpenIntelligence, 418 downloads and 35 purchases in almost 9 months. OpenResponses, 315. OpenCone, 193. OpenManual, 50 in six weeks. About 18 out of every 100 people who open OpenIntelligence's App Store page download it, and OpenResponses is closer to 38. Figuring out the marketing side is never ending

Builders: what's the one part of your codebase you'd never let an AI agent touch on its own?

OpenIntelligence 5.7 is out on iOS and macOS, and this one's about getting things in and out of the app. Siri and Shortcuts can ask your libraries now, Spotlight and Visual Intelligence find your documents, JSON and JSONL files come in as records, and you can share a page or a file to it from any app. apps.apple.com/app/apple-sto… #iOS27 #RAG

OpenIntelligence (my iOS RAG app) has three answer modes, and this is every step each one takes, racing side by side. For one question in a worked example with five phrasings, Standard runs 78 steps, Deep Think about 231 and Maximum about 430, because Deep Think and Maximum search fully for each phrasing. The race counts steps, not seconds, so it isn't how long each one takes. apps.apple.com/app/apple-sto…

Over half the internet is now AI generated content. Will we ever go back?

I started actually posting on X about a month ago and I'm at 57 followers. I build iOS apps solo, primarily around Apple Intelligence and OpenAI's API, four on the App Store so far. If you like to build stuff, wanna collaborate or you're just into techy stuff, let's connect 🤙🏻

OpenIntelligence is almost 9 months old. 418 downloads, 35 purchases, and about 18 out of every 100 people who open its App Store page download it. I still have no idea what I'm doing.

I asked OpenResponses what time the sun sets in Chicago on Saturday, November 7, and with Computer Use on it went to timeanddate.com in a real browser to find out instead of answering from memory. The answer comes from the page it read, and every step it took is listed under it, with the address it was on. apps.apple.com/app/apple-sto…

OpenCone used to search one Pinecone index at a time, so a question that needed two of them meant asking twice. In the video a sample trip itinerary is in a travel index and a school handbook is in a school index, and I ask once: “Does my Lisbon trip clash with anything at school?” It searches only those two, GPT-6.1 Sol writes the answer, and each source says which index it came from. apps.apple.com/app/apple-sto…

OpenIntelligence 5.7 is actively being worked on, not released yet. Addresses, phone numbers, dates and amounts are found anywhere in a document now, not only in its tables. In the video I ask a sample trip itinerary for its travel insurance's 24-hour number, and Image makes a card of the answer to post. apps.apple.com/app/apple-sto…

OpenIntelligence 5.7 is actively being worked on, not released yet. Addresses, phone numbers, dates and amounts are found anywhere in a document now, not only in its tables. In the video I ask a sample trip itinerary for its travel insurance's 24-hour number, and Image makes a card of the answer to post. apps.apple.com/app/apple-sto…

OpenIntelligence 5.7 is actively being worked on, not released yet. Addresses, phone numbers, dates and amounts are found anywhere in a document now, not only in its tables. In the video I ask a sample trip itinerary for its travel insurance's 24-hour number, and Image makes a card of the answer to post. apps.apple.com/app/apple-sto…

I asked OpenManual how to clean the steam wand on a Breville espresso machine. It had found the 76-page manual on breville.com, so the answer came back as quoted lines from that manual with the page each one is on. Nothing gets reworded, and if the manual doesn't say, OpenManual doesn't either. apps.apple.com/app/apple-sto…

I picked GPT-6.1 Sol in OpenResponses, set its reasoning effort to High and asked it for a Swift function. In the video it reasons first and then writes the code, and under the answer are the tokens it used and every event the API streamed back. apps.apple.com/app/apple-sto…

OpenIntelligence once filed a paper about dopamine and serotonin under "API Reference." It's my on-device RAG app, and its Atlas, which groups your documents by topic, was matching its labels on pieces of words, so it found "api" inside "therapies" and "rest" inside "interest." The same check lived in three copies of the same code, and the Atlas was reading a different copy from the one I found first, so fixing just that one would've changed nothing on screen. Now all three use one shared matcher that looks for whole words.

Post into the void, see what happens Post into the void, see what happens Social media is weird