Anthony

@GergelyOrosz AI accelerates the first 90% of development, but not so much the last 90%.

One thing that endlessly frustrates with Anthropic, a $300B+ dollar company, where most code is written with AI: Their landing page for paying customers, Claude .ai has been broken for weeks UX-wise, and no one notices or cares or fixes: It "loses" stuff I type while it loads:

@mattcassinelli Containers on iOS would be great. github.com/apple/containe…

Can you imagine if Apple gave us Terminal on iPad. God that'd be sick.

@awnihannun Thanks for everything you’ve done, and especially for creating a thriving community around MLX!

Today is my last day at Apple. Building MLX with our amazing team and community has been an absolute pleasure. It's still early days for AI on Apple silicon. Apple makes the best consumer hardware on the planet. There's so much potential for it to be the leading platform for AI. And I'm confident MLX will continue to have a big role in that. To the future: MLX remains in the exceptionally capable hands of our team including @angeloskath, @zcbenz, @DiganiJagrit, @NasFilippova, @trebolloc (and others not on X). Follow them or @shshnkp for future updates.

@awnihannun I've been thinking the same thing. The compute power of all currently active Apple devices is equivalent to a non-trivial percentage of all data center capacity. This is probably the scenario that @exolabs is aiming for.

Inference compute scarcity seems plausible. Hyper growth in demand with limited hardware or data center energy supply means token shortage. This could be a reason to run more AI locally.

@anemll The main limiting factor for this now is the agentic capabilities of smaller on-device models.

strong reasoning and built-in tools, running locally on the iPhone. autonomous, low-latency ops

Siri should be an on-device version of OpenClaw

@bcherny @peakcooper Some more feedback: Please implement more MCP features like reacting to "tools changed" notifications and automatically reconnect to MCP servers after disconnections.

@peakcooper Actively working on improving this. Keep the feedback coming

Claude Desktop app takes a full 10 seconds to switch from the 'Code' tab to the 'Chat' tab. It then takes another 10 seconds for the input to become usable. It then takes another 10 seconds for your prompt to be submitted after you clicked Enter. I am convinced, absolutely convinced that nobody at anthropic actually uses this. It's impossible. I refuse to believe

@anemll It could be interesting to expose this on an MCP server in Swift. github.com/DePasqualeOrg/…

My little remote harness works with iPhone mirroring too, which is useful for on-device ANE workflow. I give Claude a task, it implements and tests it with actual UI/device. Some UI inconsistencies are only visible on the real device—like iPhone 17 Pro's notoriously bad UI scaling (e.g., still not fixed in the X app after 6 months).

Here's my recent presentation at Swift Barcelona about building agentic apps with the new Swift AI and Swift MCP packages. youtube.com/watch?v=ekOzNd…

@nickoates_ It's even easier now on Apple devices with my Swift version. github.com/DePasqualeOrg/…

idk how this is taking Apple so long. you literally just type `npm i ai` and you’re good to go.

@rauchg I implemented almost all the missing functionality in the Swift SDK for MCP. Human feedback is still necessary for evaluating engineering tradeoffs and finding the right API design. github.com/DePasqualeOrg/…

Skeptics are saying that agents / opus are not actually yielding new or improved software. That it's just hype. Reply with the most interesting things you've shipped or solved with the latest generation of ai coding tools. I'd love to hear the anecdotes and see the links.

@jordansblog Thanks for your feedback. I'm working on this open-source package, which could enable this feature in the future. github.com/DePasqualeOrg/…

@DePasqualeOrg I'm trying local chat for the first time on MacOS. is there a way to enable spoken (TTS) responses? I'm a voiceover user, so this would be helpful. Thanks

A new version of Local Chat is now available with support for several much-requested models, including: - Gemma 3n (text only) - SmolLM 3 - DeepSeek V3 Thanks to everyone in the growing MLX Swift community who helped port these models to Swift!

The talk will include demos using the MCP package for Swift that I've been working on. github.com/DePasqualeOrg/…

I'll be giving a talk about MCP in Swift at the SwiftBarcelona meetup on January 22. Come by if you're in town! meetup.com/swiftbarcelona…

@bcherny @bcherny, Claude Code keeps asking me for permission to run commands that I've explicitly allowed in ~/.claude/settings.json. Is this a bug, or am I missing something?

I'm Boris and I created Claude Code. Lots of people have asked how I use Claude Code, so I wanted to show off my setup a bit. My setup might be surprisingly vanilla! Claude Code works great out of the box, so I personally don't customize it much. There is no one correct way to use Claude Code: we intentionally build it in a way that you can use it, customize it, and hack it however you like. Each person on the Claude Code team uses it very differently. So, here goes.

@ldenoue In principle, any model can be implemented in MLX. It looks like this one hasn't implemented voice cloning, but it's on their roadmap.

@DePasqualeOrg Thanks for your work. Do you know is this Soprano TTS model can run in MLX? x.com/wildmindai/sta…

Loading MLX models is about to get a lot faster in Swift, thanks to the performance work I’ve done on swift-transformers. In my testing, I'm seeing load times go from ~2300 ms down to ~500 ms. This makes it possible to interact with the model immediately after app launch.

The results of this test are even better: loading time goes from ~3900–4600 ms to ~300–360 ms after all my optimizations. github.com/ml-explore/mlx…

@ivanleomk Most recently a small base model for experimenting with typing completions. Performance during inference in Swift is generally pretty close to Python, and I’m on a mission to close the remaining gaps.

@DePasqualeOrg What models are you currently running on MLX? Been tempted to use it for local apps but not sure how far it can go

@guitaripod @awnihannun M3 MacBook Pro. The same model on an iPhone 16 Pro loads in ~600 ms after my optimizations.

@EricCO2Removal That's a great tip. Keep in mind that this is a very new project, with many conceivable paths for future development. My current focus and point of differentiation at this early stage is implementing a wide range of models in MLX in Swift.

@DePasqualeOrg How many models does one need once we hit a certain level ;) Need new features that make whichever models are there, better... If doing model, maybe some realtime ones (this has vad) github.com/collabora/Whis… Then it can be used widely for captioning/translation/live meeting notes.

In the past week, I've added two new text-to-speech models (Chatterbox Turbo and CosyVoice 3) and one new speech-to-text model (Fun-ASR) to mlx-swift-audio. I've also improved performance across all models. You can try them out in the example apps in the repo (link below).