Apple Will 'Watch Everything Burn' When AI Bubble Bursts - Ed Zitron

[loopy123]

I question his judgement. Has he actually written any code? Claude code is , in my experience, a 10-50x accelerator in software development especially in the hands of an experienced developer. The great thing is a senior developer, who maybe has limited ,say, python experience, becomes an A+ python dev when using CC. This often “gets under the skin” of people with existing python skills, but the goal of a company is to ship products not stroke egos.

The code it writes is not “slop”, it’s usually at a high technical standard following best practices. Now, Claude might go down the wrong track when solving a problem, but 99% of the time it’s very impressive. If an organization is not getting value ( better, faster shipping products) from AI, it’s their fault, not the tools.

 

[loopy123]

It’s really not. I just asked it a question and it completely whiffed the answer. ChatGPT and Grok both nailed it.
On Device requests have also been slower and less accurate than the older version of Siri, especially when responding to text messages by voice while I’m wearing headphones.

It’s a fairly early beta , so your experience is quite pointless to be reporting here.

 

[loopy123]

Haven’t paid a dime for AI usage except for utility costs. Everything is run locally on device or on desktop and no plans to change. Apple Studio Max hardware is amazing. 😎

I have found local LLMs on my max are pretty crap for real coding work. They are like ChatGPT from 2 years ago

 

[expensiveram]

I have found local LLMs on my max are pretty crap for real coding work. They are like ChatGPT from 2 years ago

It will take several years for a local model to catch up the top cloud based models. The huge amount of bandwidth on server boards isn’t available yet on any local hardware even if you spend 100K.

The Rubin server boards have an insane amount of bandwidth that no desktop will have for a decade.

 

[throAU]

The big thing that will pop the AI market is this:

A pair of apple studios can run a 1T model locally in under 200 watts.

Yes, currently expensive, but this sort of capability is already insane and will get even more so as the smaller models continue to become more powerful. 30b models are doing things 500b models struggled with 12 months ago.

What is currently only possible on huge cloud servers today will be running in your pocket inside of 5 years.

 

[throAU]

I have found local LLMs on my max are pretty crap for real coding work. They are like ChatGPT from 2 years ago

The harness makes a huge difference.

Qwen3-coder in ~30b size is doing things for me when running under llama-cpp in open code that gpt3.5 totally failed at 18 months ago.

 

[Morlock]

There’s a perfectly reasonable argument here that AI spending is in a bubble. Massive capex, questionable ROI, subsidized usage, and hundreds of startups chasing the same market absolutely deserve scrutiny. But that is a very different claim from “LLMs aren’t useful,” “they’re all basically the same,” or “customers won’t pay for them.” Several people in this thread are already making that distinction: the dot-com bubble bursting didn’t make the internet disappear, and an AI infrastructure correction wouldn’t make AI disappear either.

The part I really don’t buy is the Apple victory lap. Apple spent years telling us Apple Intelligence and a radically improved Siri were important. If Apple had actually decided AI was a bad economic bet, that would be interesting. But struggling to execute and then spending less than competitors is not automatically evidence that Apple foresaw the bubble. Sometimes being behind turns out to save you money; that doesn’t mean being behind was the strategy.

And the “AI services aren’t differentiated” argument is especially hard to take seriously when people in this very thread are describing questions that ChatGPT and Grok answer correctly while Siri misses them. That is differentiation. Whether that differentiation ultimately justifies present valuations is a separate question.

The subscription-cost chart has the same problem. “A user could theoretically consume $700 of metered tokens on a $20 plan” does not tell us that the average $20 subscriber costs $700 to serve. You need actual average usage, marginal inference cost, and the distribution of heavy versus light users. One commenter correctly identified oversubscription as the key missing variable.

So I think the interesting debate is not “AI is fake versus AI is the future.” It’s: How much of the current investment survives once the market discovers what people are genuinely willing to pay for? Maybe a lot of AI companies die. Maybe data-center spending gets crushed. Maybe only a few major model providers survive. None of that requires pretending the technology itself has little value.

And if that happens, Apple may indeed benefit from having spent less.

But “Apple benefited from the crash” and “Apple was right about AI” are not the same sentence.