Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Apple Partners With Anthropic for Claude-Powered AI Coding Platform
- Thread starter MacRumors
- Start date
- Sort by reaction score
>As it turns out, Apple engineers testing Swift Assist found that it was making up information and could slow down app development in some situations. Apple's work with Anthropic could help with Swift Assist and future tools.
Claude/Gemini/ChatGPT is 100x better than any LLM that Apple can do. I hope they replace Swift Assist and genmojis etc with built-in support to Claude/Gemini/ChatGPT
Claude/Gemini/ChatGPT is 100x better than any LLM that Apple can do. I hope they replace Swift Assist and genmojis etc with built-in support to Claude/Gemini/ChatGPT
I would be happy with some well focused guidance planning the app and sample code for specific functions that I may need to be create.
I'm not all in favor of having AI writing my whole app either. Then when it's time to troubleshoot or tweak, I will have to spend time trying to understand the code I didn't write.
Maybe, but then it's not my code. We'll see when it comes out. As with everything else, there will be a transition period to adjust to the new tech, then it will adopted mainstream.
It will have its pros and cons. If anyone can 'write' code using AI, there will be no need for programmers, hardly anyone will be buying software.
We have tried a wide variety of tools and models. Org even hired contractors to find us tools. None of them are able to actually learn from our projects enough to produce working code. People have had success with non-coding tasks where success is more subjectively defined. I would love to see the code that proponents of these tools are actually having written.
As it happens we just released an episode of our podcast today interviewing an engineer from Anthropic (she had hinted there should be some news by the end of the week, but hadn't said what it was). It was a very interesting interview, actually. It's a C++ podcast, and there is a bit of C++ in there, but most of it is pretty general (or at least generalises to Swift), so if you skip the preamble, community news etc, to this timestamped linked, I think it's worth a listen:
Well, I hope it's better than what's currently in Xcode. For example:
I typed:
let myDate = Date.now.advanced(by:
Xcode suggested this:
let myDate = Date.now.advanced(by: .day, value: 1)
Xcode then pointed out that this code causes these two errors:
Extra argument 'value' in call
Type 'Timelnterval' (aka 'Double') has no member 'day'
In other words, utter and total nonsense. It hallucinated a load of useless crud.
That being said, I loathe all this AI shizzle. I'm being forced to investigate some AI tools at work, which just takes time away from my actual work, and provides little to no benefit.
I typed:
let myDate = Date.now.advanced(by:
Xcode suggested this:
let myDate = Date.now.advanced(by: .day, value: 1)
Xcode then pointed out that this code causes these two errors:
Extra argument 'value' in call
Type 'Timelnterval' (aka 'Double') has no member 'day'
In other words, utter and total nonsense. It hallucinated a load of useless crud.
That being said, I loathe all this AI shizzle. I'm being forced to investigate some AI tools at work, which just takes time away from my actual work, and provides little to no benefit.
This is good. Apple needs to accept that they won't be a near term player in AI. ChatGPT, Gemini and, Claude are evolving too quickly, so let's leverage their work for Apple customers and products.
I also think the premise of on-device AI is too far from the reality of today's hardware (except may be a maxed out $14k spec. Mac Studio). Again - I don't think it unreasonable to let users make a choice between absolute privacy (i.e. stuff that can be done 100% on-device) vs. compromised privacy (i.e. Cloud AI options).
I also think the premise of on-device AI is too far from the reality of today's hardware (except may be a maxed out $14k spec. Mac Studio). Again - I don't think it unreasonable to let users make a choice between absolute privacy (i.e. stuff that can be done 100% on-device) vs. compromised privacy (i.e. Cloud AI options).
Let's hope we get Claude as alternative for ChatGPT in Apple Intelligence, I much prefer Claude.
So apple failed again with their in house xcode coding LLM which they announced last WWDC
Last edited:
Too bad that wasted years on Apple Car and Vision NOT SO PROs. I mean Vision Pro has a great future maybe ten years from now when it’s down in weight. The price isn’t the problem, it’s a solution that isn’t worth the hassle of wearing it. Apple Car alone was over $10B. They just thought AI wasn’t trustworthy? Build your own and adapt or die. That’s the rule of business. This is the problem with bean counters and COOs running innovation companies. Tim is clueless on anything but how to make money. If he treated his employees, customers, developers, or other stakeholders half as good as his shareholders, Apple would be a force.
Did Apple use this tool to develop Apple Intelligence? That would explain the current mess.
You might want to try cline. It learns from the tasks you give and from the existing code and its knowledge grows in a so called memory bank which becomes part of the code base. It is not perfect, but so far the best of all ai tools our company has tried.
^^^ This is the kind of mealy-milquetoast gruel you end up with when you use LLMs.
Xcode's existing AI code writing would be better if it generated code that compiled.
While it is sometimes good at prompting new variable names that old code completion simply can't do, it fails at doing what I needed code completion for: reminding me of the proper spelling of existing vars and functions, and reminding me what member properties or functions are available on a given type given my spotty memory.
While it is sometimes good at prompting new variable names that old code completion simply can't do, it fails at doing what I needed code completion for: reminding me of the proper spelling of existing vars and functions, and reminding me what member properties or functions are available on a given type given my spotty memory.
"As it turns out, Apple engineers testing Swift Assist found that it was making up information and could slow down app development in some situations. Apple's work with Anthropic could help with Swift Assist and future tools."
This definitely holds water as I've found the same thing with Swift code from the major LLMs. I think there just isn't enough Swift data for them to be reliable. Hopefully they can get this sorted out with Anthropic.
This definitely holds water as I've found the same thing with Swift code from the major LLMs. I think there just isn't enough Swift data for them to be reliable. Hopefully they can get this sorted out with Anthropic.
Will Claude (or any other AI) just focus on coding ? or will it also manage the whole project, from idea to release ? like a multitude of specialized agents.
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.