Apple Partners With Anthropic for Claude-Powered AI Coding Platform

[darkpaw]

It is a waste of time or energy to write every line of codes when you can let AI first write it and go through it to perfect them. Work smarter, not harder.

That right there is complete rubbish. (Aside from the amount of energy wasted by 'AI'...)

People seem to think coding is just 100% writing algorithms; it isn't. You have to structure your work based on the requirements of the project. You can't tell an LLM everything it needs to do the work of a human, and if you think you can, then you shouldn't be coding anything.

If I'm writing a testing framework I have to decide whether I'm using JUnit or TestNG. I decide how to read in test config; do I use JSON, XML or properties; can they be overridden on the command line?

Am I testing web pages? Should I use Selenium? Do I support Firefox, Chrome, Edge and remote?

Am I writing API tests? Should I use RestAssured?

Am I connecting to a database to retrieve data and check changes to that data? What's the type of db? What are the tables called? What are the column names? What relevant SQL statements do I need to use to access the data?

Right, I've made those fundamental decisions. How do I tell the LLM when I'm testing a webpage that on page C clicking the "Back" link on the page (not the web browser's back button) should go to page A and not page B? How do I tell it to add five addresses with the start and end dates that I lived there over the past 20 years, but leave a five-year gap so I can see an error when I try to continue?

How do I tell it to validate a UK driving licence number without having to tell it the intricacies of that? (Like how the number always has a "9" at position 14? And how, if you're female (how do I tell it you're female?), the month number has 50 added to it, so "03" becomes "53"?) It's quicker for me to write that than for me to let an LLM write it and make me check its work.

I'm a Java developer, not a Java code checker for an LLM's output.

I started a new job three months ago. I came in and created a test framework, and expanded their automation testing from 20 tests that took 20 minutes to run, to almost 600 that run in 15 minutes.

Here's another example of why LLMs are crud. As part of my investigations into whether 'AI' would be helpful in my work, I tried to get an 'AI testing tool' to add a new test for a phone number that was changing from 555-123-4567 to 555-987-6543. It recognised there was already a test checking the phone number. It copied the existing test and added a comment, then did a global find & replace on the phone number, resulting in two tests doing exactly the same thing. And, because of the find & replace, the comment in the test said: "'AI testing tool name': Added test: Make sure phone number is changed from 555-987-6543 to 555-987-6543".

Yeah, really helpful.

If I ever interview someone with your attitude to writing code (not "codes"), you would not get the job.

 

[novagamer]

stealing other people's work is "sensible". Sure.

How do you think Apple was training internal models? They surely weren’t only using Apple employee repos, though I expect part of the problem is that there isn’t a lot of Swift out there in general.

This is purely for source code, public repos are already used for training and it’s part of the agreement when using them. If it’s not already in the Apple Developer Program paperwork I expect it will be eventually.

I understand the argument against LLMs and GenAI for art and other copyrighted work but programming is the one area I think they get a pass, pay the money and make your repo private if you don’t want it used for training – or don’t publish it. I also know that similar technology was used years ago to mitigate critical severity defects in public code, free of charge. That information isn’t out there in the world but I helped work on it. It’s not all nefarious.

Anthropic is a much better partner than the others, though I still expect Apple will eventually take all of this in-house down the road assuming the entire fad doesn’t die off within a few years.

I have enormous issues with current GenAI and have even published written work on the subject. You can search my post history here for a glimpse into that if you’re interested. I don’t think Apple will ever make a chatbot, at least with current technology, and frankly they shouldn’t.

‘Vibe coding’ and similar is only useful if you do the legwork to understand some of the LLM output, and I don’t mean error correcting I mean going and reading the documentation and understanding the language. There is use in LLMs for that, because their corpus is large for certain languages and you can get some implementation suggestions that might point you in the right direction to pursue further on your own.

People who think any of these things are going to one-shot output a program, and the developers who have tried some of these products and written them off because they couldn’t do that (and often they are prompted terribly, likely not using LoRA if at a company or MCP if an individual) are both kind of ignorant from a certain POV. There is some utility, whether it pencils out financially down the road, who knows. All of this technology is only semi-useful at best, but it doesn’t mean it isn’t worth understanding.

I wouldn’t hire someone who claimed they use GenAI but didn’t know how the models worked at a deep level. I’m talking explain the process to me on a whiteboard, from input tokens on up. If you treat it like a black box and you’re a professional in any field, frankly you’re an idiot. Just don’t use it at all in that case, I’d happily hire someone excellent who didn’t use GenAI tooling. The real problem is the enormous, horribly ignorant mass of people and professionals in the middle area – that’s where the real danger is right now.


Regardless of all of that, Apple promised Swift Assist and I think there’s nearly zero chance they could get a well-working version of it out before 2026 if it was built with entirely internal technology. Apple is making great progress but they are very focused on quantized models, that might pay off but it won’t for that feature in the way they demoed it. At least, not yet.

 

[gimmesomemo]

I suggest Apple teams up with JetBrains and create a decent IDE.

This. Please. Ai is in almost all cases a waste of time apart from super boring work, like ‘make me some structs based on this json’. But holy **** could Xcode do with some real IDE features, proper refactoring features, and a whole lot more

 

[reyesmac]

Claude is already capable of creating Mac apps, even outside of Xcode. it's just that its limits make it impractical to create anything but basic apps with limited functionality for those with zero coding experience. We are talking about apps under 5megs in size. Plus you have to start from scratch for every single thing you make. Someday it will be easier. But I doubt Apple will revolutionize this space. It would need to have a library of already built functions and components you can piece together with AI. Or at least be able to say something as simple as make me a finder grid view and it know to make it work just like the finder, thats a lot of stuff you are not telling it yet it needs to make it based on tiny prompts. That kind of accessibility to writing apps is something I doubt they want the public to have access to. It's already annoyingly complex to get info from them as it is. This field is not open to the public like desktop publishing is. And it would have to handle writing thousands of lines of code without limits. Thats way more than they are willing to do for the developer community and more than any GPT has allowed on their platforms. Coding is not going to be as easy as making a picture or writing an essay is now for many many years to come.

 

[grovian]

If an LLM can write better code than you you won't be able to modify or fix it later. If it writes worse code than you don't need it.

 

[rjp1]

Whatever they do it can't be worse than the current situation. Xcode crashes so much when trying to accept the code completions. I don't even try most of the time, because I've had to redo so much work that got lost from the crash.