dc

@GlytchTech @SwiftOnSecurity It's always DNS except sometimes it's NTP.

@ctnicholasdev Can it parse mars too?

I've built an open-source slide editor! ✦ Generate HTML slides with AI. ✦ Collaborative code editing. ✦ Draggable commenting. ✦ Export to powerpoint. Try it out ↓

@PascalLindenau Untrue. I’m maxed out until Tues. RIP Fable usage

Have we all forgotten this?

@SkyeSharkie @AnthropicAI Worse for me. I’m already at week limit until Tues. Ugh.

... @AnthropicAI The Fable usage/time period is really unfair due to uneven distribution to the userbase. My weekly reset was exactly on the same day you started it, meaning I only get 1 week worth of usage, while others, based on random chance, have different reset days and get 2 weeks worth of usage instead. If you intended to give everyone only 1 week worth of usage, you should have reset all user weekly days to the same day. As it stands, many users are unfairly getting double usage during this time period from having a reset day in the middle of your Fable week. I'd prefer to have 2 weeks too, but yeah.

@robinebers What's the cost for opus 4.8 high

bet you don't know this: Fable 5 Low is CHEAPER / BETTER / FASTER than Opus 4.8 Max if you use Cursor or similar and you pay per usage.... you'll save a ton of time and money + get better results

1 query and Fable ate all the tokens. See you at 8pm.

@mweinbach Opus Low performs better for less money

Sonnet 5 is near Opus but better value for knowledge work It's $2 per 1M in, $10 per 1M out until September, then $3 per 1M in, $15 per 1M out.

@ALTIC_DEV @maylivesforever @AmeliaBarty How does it compare to Ghost Pepper?

Try us out and tell me if it’s slow. I’ll fix it before you wake up next day. With just raw transcription, i promise a sub second and most of the time 0 delay. With our fluid-1 model, you’d probably see a little delay but nothing can run that fast and accurate to clean up your raw transcription out there.. you can use the $15 for something else :)

I like to imagine the sheer horror at the Wispr Flow office when they see tweets like this with 200k impressions It's gotta be tough trying to build a VC-funded monopoly when your core product is an open-source commodity And your users are developers who will take every chance they can to promote cheaper OSS alternatives with zero lock-in

@mweinbach I wonder with the new M6/M7 timelines if Apple will attempt to design/build memory to appear in M7 systems…

Willing to bet this happens but only for non-US devices CMXT memory for China/Asia/India market devices, Europe/US get the standard SK/Samsung devices

@reach_vb /cc @hqmank

@reach_vb Here's the other path for folks to check that consumes a ton of disk /private/var/folders/dd/dmtdpxl15pv_p0rmv7_rrl300000gn/X/com.openai.codex.code_sign_clone I have almost 4G stuck in there

This is now fixed along with the latest release of Codex! Make sure to upgrade your codex installation to the latest version via npm or bash installer Thanks again to all of you for raising this issue and to the goated (jif) codex team

@sdhilip How does it compare to Ghost Pepper?

I've dictated almost everything for 6 months with Wispr Flow. 44,414 words, 161 wpm, top 0.1% of users. Last week I tried FluidVoice. Open source, runs local on my Mac, corrects as I speak with no API key, and handles slang better than I expected. Cancelled my paid plan. If you're on a Mac, this one's for you: altic.dev/fluid @ALTIC_DEV

@lmstudio Will this be pluggable to the back of AFM Framework like one of the big frontier models? Imagine if we could have our own PCC...

For WWDC, we worked with Apple to run Kimi K2.6, a 1T-parameter model, across a cluster of four Mac Studios using a preview version of LM Studio. We showcased secure remote access from a MacBook Neo and iPhone using LM Link. A glimpse of your own private, frontier-scale AI.

@mweinbach If Apple is able to somehow improve things so that AFM Advanced works on 8GB devices too, that would be a big win perception wise

I think of these platform improvements and fixes as a sort of apology from Apple that they let it get this way And it genuinely is such a huge improvement in performance and UX across every device I would expect the planned obsolescence views to end very quickly this fall

@stalman @lmstudio I want to see this plug into the back of AFM alongside ChatGPT etc.

One of the coolest in person demos at #wwdc26 was @lmstudio running massive local models on 4 daisy chained Mac Studios with a total of 2tb memory Then they pulled out an iPhone and chatted with those models remotely over a secure connection I need to find a use for this

@ronaldmannak I wonder if this means that the new MoE will eventually be able to run on iPhone 17

Interesting: CoreAI appears to treat MoE models like dense models: reading every expert instead of only the routed ones. That burns memory bandwidth and can make expert execution 2–4× slower than necessary. Caveat: it’s a first beta, and Apple can fix this without breaking the API

@mweinbach @MIBProductions What would be amazing if there ended up being an option to plug the backend into a local LLM via @lmstudio or something else… Would be interesting to see how that works out.

@MIBProductions This was never in the keynote, it was rumored by Bloomberg. It’s Apple Foundational Model and it could be one of three models, an on device model, a smaller (still large!) cloud model, or a pro cloud model based on Gemini but still an Apple model

I remember seeing in the #WWDC26 keynote that you would be able to change the underlying model that powers Siri AI. I saw that, for example, they used Gemini and Claude. I just asked Siri to confirm if this would eventually be true, as for now you are unable to change the models at the moment. This is what she had to say.

@ivanfioravanti So mtplx or omlx at this time?

Yes, MLTPX v1 is able to bring my MacBook Pro 16" M5 Max under heavy throttling. Look at the screenshot 👀 Clearly if you push like crazy on it with evals like I'm doing now 🤷🏻‍♂️

@anemll I think the fm cli is able to access

Did I read it right? No developers facing 20B ? 🥺 On device: AFM 3 Core (3B, dense) and Core Advanced (20B sparse, firing just 1–4B parameters per request to power the new voices and dictation).