Johnny (@johnny_xx) header
Johnny (@johnny_xx) profile picture

Software Engineer Mac, iOS, AI. Usually building something open source

Joined February 2009
Photo by Johnny (@johnny_xx)Photo by Johnny (@johnny_xx)Photo by Johnny (@johnny_xx)Photo by Johnny (@johnny_xx)Photo by Johnny (@johnny_xx)Photo by Johnny (@johnny_xx)Photo by Johnny (@johnny_xx)Photo by Johnny (@johnny_xx)Photo by Johnny (@johnny_xx)Photo by Johnny (@johnny_xx)
Made an app for all the stuff i save and never find again links, notes, screenshots, pdfs. it searches the text inside everything, even voice notes. Tons of features, too long to list free on mac, iPhone and iPad. Built it for myself, maybe some of you can find it useful as well.
4
1
116
Johnny retweeted
WHY DID I JUST GET 3 BANKED RESETS ON CLAUDE "new model reset" FABLE 5.2 ANY MOMENT NOW
2
1
30
3,041
Adding a way to clean cache on my IOS app. Why don't all apps do this? Apple should enforce this.
2
91
I am am amazed by the feedback. The app will be open source soon and hopefully build a community around it. A lot of people interested in contributing.
1
2
7
Bug in CLaude?. Been using opus high for 2 hours straight and doesn't go below 45% For sure a bug never seen anything like this.
1
2
79
one hour later it finally went to 46%
19
I was very surprised, in 2 days this app got to more than 800 users, 48k views on reddit, more than 150 upvotes and almost 100 comments with praise but above all suggestions for the app. real engagement with users. It is so great to build software that people actually use it and contribute with ideas . So great.
5
apple bought ascend.io, an agentic ai engineering startup xcode agents or siri, which one is it for?
1
2
109
coding is basically solved. so i spent a day helping my wife's international logistics job using only AI.. it did not go how i expected
1
1
66
she handles import/export, manifests, tons of paperwork forms and manifests? models filled them fine and saved a lot of time but i gave it a complete 1:1 example file, same customer, same destination, every step from the first quote to the last operation. it still couldn't run the process end to end by itself. no judgment, constant questions. and not just safety ones) , slips on details when the docs are full of numbers and a large body of data. On an activity where a wrong number can mean thousands in fines and extra costs, the trust is not there yet. great assistant.. saves a lot of time, not a replacement. no public model is AGI yet. So all that fear of being replaced even for admin work.., as of today we are not there yet.. unless companies want to loose so much money in fines and costs.
20
Fable 5.1 orchestrating 151 agents on ultracode burned my 5 hour limit in 15 minutes. I know not smart.. but i really needed to work on a critical fix very quickly. Couldn't complete the verification but gave me enough clues
1
51
Are Claude models, specially Opus and Fable, more dependable than a person now?
44
HERMES AGENT CAN NOW SNEEZE LIKE A GREAT DANE. 🔊 I'm dead serious. 😂 New official ElevenLabs plugin. Tell Hermes what you wanna hear, and it'll design the prompt, pick the duration, and GENERATE the sound. Cat vs. mug. Sad trombone. Monday alarm. SOUND ON. 🪽
1
3
13
1,389
Building a fun goofy project, just because…. a security drone with a Parrot, a Pi 5 and Claude orchestrating. What works at the moment: YOLO on a Pi 5 + AI HAT for detection, 30fps ByteTrack to follow the detected movement. PID loop to keep them centered Simple rule: sensor trips, fly, search, track, return A sensor fires, the drone flies to it and sends snapshots to Claude. It works, sort of. The hard part isn't control, it's latency. The flight training and detection in the sim took roughly 4 days running 24/7. Real flights are surprisingly consistent with the sim
6
1
5
331
Back the simulator training, today is very windy and that completely messed up the route. Didn't account for that initially. Also the camera is so bad on this cheap drone, that is introducing miss categorization. Today it crashed during a real flight because it miss detected a target. The latency is still there, as you can see on the sim run, it flew past a target and the detection triggered seconds later and then yes it followed the target.
36
I have to spend my usage now that i have downgraded to just a Chatgpt plus account and i still can't understand how such a big company can reach this point
41
what's the one file you still won't let agents touch besides env's?
33
For the nerds out there: microcontroller makers... what about a Mac console with flashing, serial plotting and library management, plus on device explanations of what a sensor stream is doing. Interest?
32
I have been building a scorer for claude code. yes, the harness already has one which it uses to make decisions about what subagent it uses but it could be better or at least more personalized for your particular tasks. it scores a lot with the intent of being the cheapest possible which is good, but not always what is intended. The script inspects the code, scores complexity and uncertainty and the consequences of getting it wrong. This part is very important and not baked on the herness. from there i choses the subagent model not only based o the reasoning level but against a data base of examples of how each model performed in the past in similar task. I know what you are about to say, but this will consume a lot of tokens, defies the purpose. Not exactly. The only extra tokens consumed are comparing the codebase ( which would have to read anyway ) with the examples on success and failures on the database but because each subagent will only address on function of the code, the overhead in tokens could not be significant if the final result is faster and of better quality. The goad here is not to have it it being the cheapest but having the best result.
1
1
79