Simone Civetta @viteinfinite.bsky.social · 24/06/2025Installed Beta 2 on my iPhone. The Liquid Glass effect is nice, but too present in such a small screen estate. It feels somehow “baroque”. As screens are larger, I find it much more pleasant on iPadOS and macOS. 010
Simone Civetta @viteinfinite.bsky.social · 19/06/2025Interesting, it's kind of counterintuitive. But will try. 100
Simone Civetta @viteinfinite.bsky.social · 17/06/2025And that’s all I know! For more resources: - WWDC videos - machinelearning.apple.com/research/app...machinelearning.apple.comUpdates to Apple's On-Device and Server Foundation Language ModelsWith Apple Intelligence, we're integrating powerful generative AI right into the apps and experiences people use every day, all while… 000
Simone Civetta @viteinfinite.bsky.social · 17/06/2025📊 Telemetry No data is collected by Apple via telemetry. 100
Simone Civetta @viteinfinite.bsky.social · 17/06/2025🧪 Adaptor Training Toolkit Allows model adaptation via Fine-Tuning (presumably LoRa) but, as mentioned in the videos, you should re-run the FT each time the underlying model is updated (which means at every iOS release or so) 100
Simone Civetta @viteinfinite.bsky.social · 17/06/2025🔐 Safety model A built-in guardrail AI model is automatically added and enforced by Apple. It cannot be modified or disabled. No API for building your own as of now. 100
Simone Civetta @viteinfinite.bsky.social · 17/06/2025🔄 Background inference Foundation Models framework supports background inference! But subject to system-level constraints: – Rate-limited – Cancelled during camera or AAA games usage – Throttled or terminated in case of resource pressure 100
Simone Civetta @viteinfinite.bsky.social · 17/06/2025⚙️ Hardware execution Runs primarily on ANE (Apple Neural Engine). This means you can use GPU while running inference — but prolonged concurrent usage can cause system stress, which may lead the system to terminate your app, even if in foreground. 100
Simone Civetta @viteinfinite.bsky.social · 17/06/2025🧬 Model versioning Same model version across all supported devices — no per-device variation. Whether you’re running an old iPad, the latest iPhone or even macOS, it’s exactly the same model. Perfect for testing and consistency across your multiplatform apps! 100
Simone Civetta @viteinfinite.bsky.social · 17/06/2025🔢 Context size Not officially documented as it might change anytime. But today the limit stands at 4096 tokens. 100
Simone Civetta @viteinfinite.bsky.social · 17/06/2025Foundation Models framework is beautiful. I had the opportunity of grabbing a WWDC lab, and here’s a bunch of info you might find useful. 🧵 111
Simone Civetta @viteinfinite.bsky.social · 16/06/2025Sure. That said, it’s not about models but about features: Cursor has agent modes, custom modes, MCP integration and frequent updates. I don’t see how Xcode can keep up with this, but I hope to be wrong. 110
Simone Civetta @viteinfinite.bsky.social · 14/06/2025Oops, forgot mentioning. I’m on the Beta, it wouldn’t work otherwise. 010
Simone Civetta @viteinfinite.bsky.social · 14/06/2025On macOS, System Language and Siri must be set to the same language (and the setting is irritantly reflected on your iPhone). Then make sure Apple Intelligence is turned on. At that point it should work. (Or at least it does for me. I’m using English (UK)) 100
Simone Civetta @viteinfinite.bsky.social · 14/06/2025Ha! Try again or double check the settings on the simulator. (Maybe locale / language?) It should work. (Unless I hallucinated last night) 200
Simone Civetta @viteinfinite.bsky.social · 14/06/2025By far the Foundation Models framework. Wonderful API IMO, and so feature rich. 020
Simone Civetta @viteinfinite.bsky.social · 14/06/2025Can’t agree more. For what it’s worth I submitted it as a feedback in the satisfaction form. The level of “self-referential” of the last two videos is cring-ish. 010
Simone Civetta @viteinfinite.bsky.social · 14/06/2025Thinking about Dia (and @perplexity_ai’s Comet - which I haven’t tried yet). But in an AI-first world, apps who have the most potential are IMO those like @raycastapp, offering a deep integration with the system that goes way beyond browsing. And they’re far from being sherlocked 000
Simone Civetta @viteinfinite.bsky.social · 12/06/2025macOS Beta is usable as a daily driver on my MBPro M1. And by usable, I mean "usable". It's not perfect, but it doesn't crash. Mean issues so far: far from buttery smooth (esp. Safari, which is sluggish to say the least) + quick battery drain. The UI is not radically different, but it's pleasant. 000
Simone Civetta @viteinfinite.bsky.social · 11/06/2025As much as I loved “Letter to Arc members” by @browsercompany I feel that for my own personal use, Dia is a step in the wrong direction. I find that features like command bar (with actions) or the swipe to change the profile were true productivity enhancers. 000
Simone Civetta @viteinfinite.bsky.social · 11/06/2025At this point the most disappointing Apple product is the Apple Watch. Sure they sold millions of it but it now sits in a purgatory between being too power hungry to have a long lasting battery life and but too constrained to run a SLM or proper background tasks. 000
Simone Civetta @viteinfinite.bsky.social · 11/06/2025I’m glad to see such a feature-rich Foundation Models framework and its API so pleasantly “swifty”. A simple inference API would have already been good, but FMf goes way beyond that. 010
Simone Civetta @viteinfinite.bsky.social · 10/06/2025Why do I feel such a strong urge to try your alarm app? 010
Simone Civetta @viteinfinite.bsky.social · 10/06/2025The ChatGPT integration in Xcode will certainly useful at some point. But at the moment is still ages behind Cursor (via the xcode-build-server integration, cf @dimillian.app's tutorial). 130
Simone Civetta @viteinfinite.bsky.social · 09/06/2025Not sure Background Task APIs are the right answer but it's certainly a step in the right direction. #WWDC25 000
Simone Civetta @viteinfinite.bsky.social · 09/06/2025Ok, I lied. I was also hoping for these new iPad features! #WWDC25 000
Simone Civetta @viteinfinite.bsky.social · 09/06/2025Shortcuts looks neat, even as a Raycast user. #WWDC25 010
Simone Civetta @viteinfinite.bsky.social · 09/06/2025To me the only meaningful update to watchOS would be a truly minimal version that would allow the battery to last at least 4 days. 000
Simone Civetta @viteinfinite.bsky.social · 09/06/2025Screen Visual Intelligence hopefully available (in Europe) before Winter 2029. #WWDC25 000
Simone Civetta @viteinfinite.bsky.social · 09/06/2025Ok, translations. But just give me multilingual Siri for putain sake. #WWDC25 000
Simone Civetta @viteinfinite.bsky.social · 09/06/2025Not entirely sure that widgets in CarPlay are the safest idea. Looking forward to seeing how it works in real life. That said, Siri is still, by FAR, the biggest irritant while using CarPlay. #WWDC25 000
Simone Civetta @viteinfinite.bsky.social · 09/06/2025That Liquid Glass shader is giving me shivers. #WWDC25 000
Simone Civetta @viteinfinite.bsky.social · 09/06/2025Spoke too soon. The *only* API I was actually waiting for, "Foundation Model framework" is official! 000
Simone Civetta @viteinfinite.bsky.social · 09/06/2025\#WWDC is starting now, and the fact that it’s kicking off with an #F1TheMovie promo brings back painful memories of that Apple Music-focused WWDC. 000
Simone Civetta @viteinfinite.bsky.social · 23/05/2025Most notably you get to refine the plan before any code is written. So when you finally move to the “Code” phase, you already understand the why and the how, which makes the AI’s output easier to read, review and trust. 000
Simone Civetta @viteinfinite.bsky.social · 23/05/2025Roocode (and the tool it's forked from, Cline) supports this flow beautifully. You can switch into Planning Mode or Architecture Mode with read-only access, different models (e.g., ones that reason better), and prompts geared toward high-level thinking. 100
Simone Civetta @viteinfinite.bsky.social · 23/05/2025Tools like Cursor often jump into code editing immediately, and you only realize things are off after 2–3 files have been touched. At that point, you’re rolling back changes, losing time, even with checkpoints. That’s why planning first is more than handy. It’s foundational. 100
Simone Civetta @viteinfinite.bsky.social · 23/05/2025I ask the AI to generate a high-level plan or architecture first. Then I review, refine, and align it with what I actually want to build. This planning step helps keep things coherent across files—and prevents one of the biggest AI coding pains: misaligned edits. 100
Simone Civetta @viteinfinite.bsky.social · 23/05/2025Lately, I’ve been using Roocode more and more in my AI coding workflow—so much so that I miss it when working in Cursor. My preferred workflow now starts with a planning phase before letting the AI touch a single line of code. 🧵 101
Simone Civetta @viteinfinite.bsky.social · 23/05/2025You also have a bookmarking feature in Notion. The app is bloated IMO, but on the plus side, you can store basically anything there, and let it become a "second brain". Same goes for Obsidian. 100
Simone Civetta @viteinfinite.bsky.social · 23/05/2025If you're on iOS, maybe just the Reading List? It's basic, but gets the job done. 110
Simone Civetta @viteinfinite.bsky.social · 16/05/2025Sure, AI can re-summarize the file. But let’s be honest: it often leaves you with a weird feeling— Like you don’t fully get the code, And you’re not truly in control of your own software. 000
Simone Civetta @viteinfinite.bsky.social · 16/05/2025Without making the micro-decisions yourself—turns, signs, stoplights—you don’t build memory. Same with code. Open the file one day later, and it feels unfamiliar. That slows you down. 100
Simone Civetta @viteinfinite.bsky.social · 16/05/2025Second: memory. Even if you understand what the AI did—you didn’t write it. And that matters. It’s like sitting in the passenger seat: you arrive, but couldn’t retrace the way back. 100
Simone Civetta @viteinfinite.bsky.social · 16/05/2025First, reading. AI generates tons of code—fast. We’re not used to reading hundreds of lines in one go. But to trust what it suggests, you must understand it. And that takes time and focus. 100
Simone Civetta @viteinfinite.bsky.social · 16/05/2025Coding with AI brings lots of challenges. Two I keep running into: – Reading – Memorisation Let’s talk about them. 100
Simone Civetta @viteinfinite.bsky.social · 02/04/2025 🤯 @piaskowyk.bsky.social showing JS code running faster than C++! 031
Reposted by Simone CivettaSoftware Mansion @swmansion.com · 13/02/2025It’s time for the Llama generation speed benchmark! 🚀 We’ve compared quantized Llama 3.2 1B QLoRA and the full precision model. Results: ⚡Quantized model: 9.56s ⌛Full precision model: 19.14s It’s 2x faster with quantization! 📈 Try it yourself with our example app ⭐ github.com/software-man... 0134
Simone Civetta @viteinfinite.bsky.social · 11/02/2025Choosing the right engine is key in local LLM/LMM inference; it can significantly impact speed, quality, docs, compatibility and portability. Here's a comparison among the most starred ones as of Geb 2025: llama.cpp, MediaPipe, MLC-LLM, MLX, MNN - PyTorch ExecuTorch. 001
