marshalla

I have two $100/mo Anthropic accounts (watch out the single $200 doesn't raise weekly limits) and one $100/mo Gemini account. I have Opus generate review prompts for Argon and paste them into agy.

I tried setting up a script so Claude could automatically call agy, but it kept hitting timeouts, and letting Claude control that tends to burn my usage too quickly anyway (and if I'm not copy-pasting into different terminals, where's my value add in all this?).

Argon hallucinates, so I rarely use it for impl, but I've actually had better results with it than Opus for SwiftUI and AppKit work.

I try not to run more than three sessions at once, or I get anxiety about having to keep all the session caches fresh and end up typing silly things into some of them every 30-45 min so they don't expire or auto-compact.

"I've even had it transcribe words before I've finished saying them! " lol idk if this part is actually a good thing. It's great when you always speak in predictable sandwiches.

I do think Parakeet is pretty amazing, though, so some of these product announcements have limited interest to me, especially since it's cloud and costs per hour, even if it's a low cost.

Parakeet is trained on text that has had ITN done on it, so it prints digits, which messes up window merges. I'm training that out of it: https://github.com/MimicScribe/parakeet-spoken-form

Neat, can I get mine (yes another one) listed on there? https://mimicscribe.app/

It's just about done. I've got a custom version of Parakeet Ultra I'm training now that does digits better.

I set the battery to low power mode on the website and it didn't do anything and now I'm profoundly disappointed.

On a more serious note, I think on-device object/face recognition search is a great idea. Having the cta button on the site go to Polar is a little unexpected. It'd be nice to be able to try the app without going through the hoops of running it in a Docker container, if that's the same app or system?

This guy cloned Photoshop and apparently it works. First commit Sep 29, 2026. https://github.com/storytold/photocraft https://github.com/storytold/photocraft/commits/main/?after=...

Applications built using deterministic code to work might just not be a thing people need to pay for anymore.

Maybe work relating to training ML can still have value? The compute cost and time involved in it add to the moat.

I'm kind of proud of how popular this thread became, even though some parts collected some downvotes.

Key takeaways: People don't like subscriptions for on-device apps, and people really don't like telemetry.

Oh neat, I hadn't even seen or tested SpeexDSP. Pretty happy with localVQE 1.2, it's a little extra on the CPU, but the cancellation is pretty complete.

This one looks super nice. Trying it now. I've got one I'm working on full-time and trying to get people to pay for, and in some ways this beats it. I'm using localVQE for echo cancellation. I really just need to give up and get some corporate job

Ouch. A subscriber cancelled with a P.S. saying the real-time transcription left something to be desired and that I should use GPU models with more accuracy.

Long story short: I got a batching model to behave more like a streaming model, ~2s updates, and it runs on the ANE so it’s efficient annnd it’s more accurate than all feasible alternatives annnd it only took me like a month. Annnd this is a meeting transcription / notetaker app for macOS.

I used a bot to help me generate precise language around the numbers for this post. Forgive me.

Transcription can happen from audio on the system channel without a bot in the call, so if the remote user can hear it, a model can convert it to text.

The noises confuse the ASR models, I presume.

Someone could still record audio and manually transcribe it or take notes with a pen and pencil... just sayin'. Some of this just comes down to having to trust the other party.

I do take some issue with my voice almost certainly being used to train models when using certain services.

My prediction is that soon every mid-size+ company will have their own hardware running an open-weight model and all employees will be required to connect to it and it alone for inference.

Nice, how well does WhisperKit work for diarization?

lol fair that's very dev of you. The next gen of models will prob one shot my app anyway.

I like the idea of having a completely headless auto recording system. I was against it initially over privacy concerns but with completely on-device its a better idea. I might snag the concept

I'm building one that finally seems to be reasonably stable https://mimicscribe.app/.

Speaker identification is a huge trick to do on device. I've masked by sentence and capture embeddings within that mask to get things more accurate.

If you try it that would be super cool and it's basically free rn while I figure out how to market.

I'm testing obsidian integration rn. It's currently available through mcp but that kind of defeats the offline transcription thing

venice.ai? Ok nice I'll check it out. thanks. ya I added the openai compat endpoint, no brainer.

Right? I feel like local model support is something that has strong ideological appeal and might influence someones feeling towards the app, but when it comes down to it most people will probably just use a cloud model for larger tasks unless they have beastly hardware. It's like how I have an Android in part because I might one day flash the ROM.

Ollama just exposes models via an OpenAI compatible endpoint though (I'm pretty sure), so adding that standard is probably a good idea. The prompts are a bit tuned for Gemini. I'd have to test how much that matters.

lol that'd be a trick. I'd have it purposely misidentify to cheaper brands to mess with people.

Hey thanks! Good call. I had limited success getting Qwen 3.5 9B working for some of the longer prompts that require lots of json output. I feel like completely on-device is so close to being usable for this stuff though. I should revisit this, actually.

No joke about the indexing feeling slow. Could that be run as a lower priority task? handy app!

Matt, Thanks!

I just tried Hedy, same concept, also a great tool. It's todos are nice.

MimicScribe works well with any Apple silicon Mac so it'll feel snappy on an M1 with 8GB of RAM even. It uses Apple's on-device ML accelerator, the ANE. https://mimicscribe.app/docs/performance