> Then, when it was time to review or merge something, we spent even more time digging through the code just to understand what the commit or PR was supposed to do.
> Now I delegate that work to AI.
> […]
> If one day passes and nobody on the team reviews the PR, we merge it
It’s not clear to me whether or not the author is reviewing the code their agent is committing on their behalf, so maybe I shouldn’t be surprised that they don’t seem to care whether or not their team is reviewing it either.
But I am surprised. The author says they used to spend quite a lot of time writing developer guides, hopefully by building consensus among their peers. It seems to me like a big shift if it’s now a workplace where they might not even read each other’s code changes.
Or maybe the dev guides were top-down directions on the way they want the team to do things, and now AI is super compliant and they’re living the dream.
It’s certainly not a new phenomenon. I appreciate this XKCD [1], with a chart of “How long can you work on making a routine task more efficient before you’re spending more time than you save”
It’s not the final word, since automation has other benefits: documenting the procedure’s steps, reducing human errors, increasing consistency, etc.
> How often does a restaurant update a menu? Those that need to update frequently already solved that by having a chalk board.
One advantage of a digital ordering system: it can track inventory and mark a dish as unavailable after N have been ordered.
Some restaurants might have menu updates in the middle of service quite frequently, if they have a daily special with a fixed quantity. Most restaurants probably have an unavailable dish occasionally. Or if a beverage choice becomes unavailable, or a new beer is put on tap.
I don’t think ease of menu updates would be the deciding factor for any restaurant, I think it’s more likely to be based on the experience they want their customers to have
> I’m sure there’s a technical Unicode word for this process of matching codepoints for e.g. search and uniqueness of rendering
That’d be Unicode Normalization. I don’t have an opinion on the best source for more details, so here’s a link from unicode.org https://www.unicode.org/reports/tr15/
I don’t know enough to know whether or not there are still Arabic-specific issues, either in the spec or the implementations.
The example in the article of copy/paste/search is interesting. I think it’s equally likely to be a RtL issue as a normalization bug, but I haven’t done anything significant with either topic.
I don’t watch baseball (maybe a couple games a year), but I think it could use some sound / audio, so you can have it running while doing something else.
IDK if there’s an easy way for the average person to get a live audio broadcast feed from games, so maybe your target market would be listening to that instead.
I’m thinking it could use some sound effects, for balls, strikes, hits, etc. I only tuned in for a couple pitches and then it was between innings, so maybe the more significant events already have something, and I just wasn’t patient enough to experience them.
I was looking away when the last out of the inning happened (or maybe changing views?). Is there a display of what caused the out, and maybe an animation of the fielders coming into the dugout, or does it flash up the “between innings” screen pretty quickly?
It might be nice to have a significant event summary available somewhere. It feels hard to believe that this would catch someone’s attention well enough that they’re watching the whole thing, and without audio cues / replays, I know I wouldn’t be interested in watching it for any length of time.
From this article, it does sound like it’s a newer development:
> Goals for Hacking for Defense
> A decade ago, our goal for the class was to teach students Lean Innovation methods while they engaged in national public service. We wanted to familiarize students with the military as a profession and help them better understand its expertise, and its role in society. We also hoped the class would show our sponsors a methodology that builds problem understanding before writing requirements.
> The class still does all this, but now that the DoW is buying from startups and defense venture capital is abundant, the class has turned into a national security incubator. Most of our teams form defense companies.
I think the difference is in the scale. I think Apple’s asking you to index / share all of the data from your app. Virtually no user is going to copy & paste every single message into ChatGPT.
And so I think there’s a meaningful difference between the app donating all the data automatically, vs making the data available to the user who is free to share what they want.
I think the server-side stuff will be a mix of users & developers paying. I have seen this info in several places:
> PCC delivers a powerful server model without compromising privacy: data is never stored, used only for the request, and independently verified. It's integrated with the OS and iCloud, so there's no authentication or API keys, no token cost to developers, a daily per-user limit (higher with iCloud+), and eligibility for apps under 2M downloads.
> Maybe you missed the 'or sent to private cloud' part of the announcement, it's not just local-llm only.
I saw that. Maybe you’re unfamiliar with Apple’s Private Compute Cloud? It’s intended to allow cloud computation on data without making the data available to anyone, which I think backs up my interpretation that apple’s stance is “no one should have this data, not even us”
We designed Private Cloud Compute with core requirements that go beyond traditional models of cloud AI security:
* Stateless computation on personal user data: PCC must use the personal user data that it receives exclusively for the purpose of fulfilling the user’s request. User data must not be accessible after the response is returned to the user.
* Enforceable guarantees: It must be possible to constrain and analyze all the components that critically contribute to the guarantees of the overall PCC system.
* No privileged runtime access: PCC must not contain privileged interfaces that might enable Apple site reliability staff to bypass PCC privacy guarantees.
* Non-targetability: An attacker should not be able to attempt to compromise personal data that belongs to specific, targeted PCC users without attempting a broad compromise of the entire PCC system.
* Verifiable transparency: Security researchers need to be able to verify, with a high degree of confidence, that our privacy and security guarantees for PCC match our public promises.
- - - -
Second, according to their press release ([1] and a sibling comment elsewhere in this chain), they’ve been trying to find a way to allow interoperability without giving full access to everything. Unsuccessfully, so far. So it’ll be interesting to see where it goes, but I’m sympathetic to their current stance.
I don’t have the EU perspective, which might be changed by things like GDPR, but I prefer Apple’s stance that “no one should have this data, not even us”.
One reason is that the data on a user’s phone isn’t solely owned by them. Some of it is shared with other people, or “belongs” to someone else: chat, email, shared documents, photos of people, contact information, etc.
In a corporate environment, this is more explicit: you have access to company information, so the IT department controls what apps you can install / run, because individual EEs won’t always make the best choices.
Second, I think app developers are more likely to share more data, if they know that the shared data doesn’t leave the user’s control. And that (presumably) makes the feature work better. If I’m developing an app, I’ll think twice about indexing any sensitive data, if I don’t know where it was going to end up.
I’ve got a 2023 Mac Studio M2, and was dismayed by the M3 & later. So I’ve been trying to track down more details. That specific device list is only for:
> Apple’s most powerful on-device model and the features it enables, like expressive voices and more advanced dictation, […]
On other devices, I think there’s still on device support (just not with the “most powerful model”), for these devices:
> Apple Intelligence and Siri AI in iOS 27, iPadOS 27, macOS 27, watchOS 27, and visionOS 27 are available on iPhone 16 models or later, iPhone 15 Pro, iPhone 15 Pro Max, iPad mini (A17 Pro), MacBook Neo (A18 Pro), iPad models with M1 or later, Mac with M1 or later, Apple Vision Pro, Apple Watch Series 9 or later, Apple Watch Ultra 2 or later, and Apple Watch SE 3 when paired with an Apple Intelligence-enabled iPhone nearby.
They’re actively asking developers to index all the content in their apps, to provide Personal Context that Siri can use for user requests. And to create/index the actions available in the app.
So, where developers comply, all of that content is now accessible to those alternative implementations.
It’s not full read/write of the phone, and it’d exclude obvious secrets like passwords, but it is quite far reaching access.
I don’t know what sort of restrictions they can put on the alternative implementations. Can I vibe code one and have it live in a week? or is there a minimum bar?
Sounds like Apple offered to implement that, but it didn’t fit the EU’s requirements.
I suspect if you paid apple enough money, and were willing to prove that your personal Private Cloud Compute did meet their requirements, it wouldn’t be impossible.
I think there’s a case that Apple’s commitment to privacy here will increase participation by 3rd party developers.
For example, if I’m maintaining a secure chat app, I think I’d be more likely to adopt the APIs to share the chat messages with the system AI due to Apple’s promises that the data will either be processed On Device, or in their Private Compute Cloud.
If I instead believe that sharing the chat messages with the system AI would cause those messages to be sent to unknown-to-me other entities, I think I’d be less likely to participate in the new API.
This user might be okay with their data going to this other provider, but what about the people they’re messaging? I have a responsibility and a commitment to _all_ of my users to protect their data.
I might not be able to control what any specific user does with the data, but proactively writing the code that sends the chat messages to this other system is something that I have control over.
Clicking through to the query for the first chart, I see the peak of ~300k in May of 2020, and it was ~3k in April of 2026 (the last complete month). I’m flabbergasted.