Let them Cook in the State of Cupertino

A response to “Something is Rotten in the State of Cupertino”, addressing Apple’s AI trajectory and what the future likely holds.

from Amit Sharma, founder and CEO of Evitar.ai 

—————————————

Introduction: From Rotten to Simmering

John Gruber’s piece “Something is Rotten in the State of Cupertino” pulls no punches. He argues that Apple overhyped its “Apple Intelligence” initiative – unveiling flashy Personalized Siri features that it couldn’t actually deliver on schedule . In his words, Apple “announced something [it] couldn’t do” and wound up looking “out of their depth… in over their heads,” with key features still missing in action. 

Ouch. Gruber’s frustration is palpable (and frankly, understandable): Apple promised Siri smarts like never before, from reading your screen to automating tasks between apps, but months later those promises remain vaporware. The whole saga, complete with concept videos instead of real demos, looked bad – “bullshit,” even, to use Gruber’s colorful term .

So, is something truly rotten in Cupertino’s state? Or is the kitchen of innovation just taking a bit longer to serve the meal? In this article, we’ll stir in some broader industry context, dash in examples of how even Google and others stumble in the AI race, and ultimately argue that Apple’s ambitious bets might simply need more time to cook. Yes, hold Apple accountable – but also, let them cook. 🍲

A Big Leap vs. the Slow Giants

OpenAI’s 2022 leap with ChatGPT stunned the tech world. An upstart blew past the incumbents, delivering a conversational AI that felt almost sci-fi. Caught on their heels, the big players scrambled.

Google – the supreme AI research titan – suddenly looked flat-footed, even “fallen asleep” at the wheel . Early 2023 saw Google rush out its Bard chatbot to regain momentum, only to fumble the debut. (Bard famously erred in its first demo, knocking $100 billion off Alphabet’s market value in one day – talk about a wake-up call!). Microsoft hustled to bake AI into Bing Search and Office. The AI spring had sprung, and the giants had to mobilize fast.

Fast forward 18 months: some of the giants have indeed mobilized. It just took them some time – which shouldn’t be surprising. Behemoths like Google and Microsoft have massive user bases and reputations to protect, so they move deliberately, not break-things fast. But once they get going, they really get going. Google is a prime example: after the rocky start, it began shipping impressive AI-powered features at scale. A few highlights:

• Latest Gemini models especially with their image generation and multimodal capabilities 

• Smarter Workspace & Gmail: Integrated generative AI across its productivity apps..

• AI in Search: Google’s Search Generative Experience (SGE).

• Pixel Magic.

OpenAI lit the path, but the likes of Google have the resources (and incentive) to catch up. It took a year of internal retooling and some public stumbles, yet Google is now delivering. Apple may have been surprised by that leap too – and it’s still earlier in its sprint, arguably – but history suggests that these giants can cover a lot of ground once they hit full stride.

Doable Unknowns vs. Uncertain Unknowns in AI

We need to distinguish two categories of innovation challenges. Think of them as “doable unknowns” versus “uncertain unknowns.” The doable unknowns are hard problems whose solutions are mostly engineering and scale – you know what you want to achieve, and it’s presumably possible, but it’s a heavy lift. The uncertain unknowns are the moonshots – fundamental breakthroughs that may or may not be attainable soon, where you’re not even sure if your approach will pan out.

Most of Apple’s much-vaunted Apple Intelligence features are in that first category: doable unknowns. Getting Siri to understand what’s on your screen, or sift your photo library for “the picture from the BBQ last Saturday” – these are difficult engineering problems, sure, but not mysterious new science. Apple largely knows what needs to happen (e.g. create a local index of your content for Siri to query; integrate an “Intents” API so Siri can act in other apps). It’s mostly a matter of software integration, UI design, and processing power. Difficult, but doable. In fact, none of the individual features Apple announced last year (apart from maybe the on-device image generation Playground) felt conceptually impossible – they felt like natural evolutions of Siri, albeit ones that should have been around by now. That’s partly why the delays sting: these capabilities seemed within reach with enough focus.

The uncertain unknown – the real X factor – has been something deeper: foundational AI intelligence. By this I mean the large-scale language and multimodal models (LLMs) that exhibit remarkably human-like understanding and generative abilities. These are the “foundation models” that can be adapted for countless tasks. Until recently, it was uncertain who (if anyone) could crack the next leap in AI performance. As it turned out, only a handful of players pulled it off:

• OpenAI – with GPT-4, 4-o and now GPT-4.5, arguably the first AI models to really blow the lid off conversational and creative tasks.

• Google – with Gemini, aiming at GPT-4-level and beyond, and already powering features in Search and Workspace.

• Meta (Facebook) – with LLaMA and LLaMA 2/3, open(ish)-source models. 

• Anthropic – with Claude 3.5 and now 3.7.

• Elon’s xAI – with Grok-3 which achieves SOTA like capabilities. 

• DeepSeek – the newcomer from China, with its DeepSeek-R1 model.

Apple is not on that list. The “uncertain unknown” of creating a cutting-edge foundation model has (so far) been cracked by OpenAI, Google, Meta, Anthropic, xAI, DeepSeek… but not by Apple. And this is a challenge that makes all the difference. With a state-of-the-art LLM under the hood, those “doable” Siri features become far easier to actually do. Without one, Apple’s stuck trying to teach Siri new tricks with an aging brain, or worse- an outsourced brain.

Why has Apple lagged here? A big factor is Apple’s own philosophy. With notable exceptions, the company has a strict approach to data and privacy that, ironically, works against the needs of training giant AI models. Modern AI training gobbles up massive amounts of data – essentially reading the entire internet (and then some) to glean patterns. Apple, however, is extremely cautious about data collection. Siri’s design famously prioritizes privacy: it does a ton on-device and sends minimal data to the cloud. Apple doesn’t hoard your every query to improve its AI; in fact, Apple barely retains Siri requests at all in identifiable form.

Apple Intelligence as a platform is built so that it’s “aware of your personal information without collecting your personal information,” using on-device processing and only sending anonymized bits to Apple’s servers when absolutely necessary . Good for privacy? Absolutely. But it means Apple hasn’t been amassing a trove of real-world Siri interactions or personal data to fuel an AI brain. Nor has Apple been willing to scrape random internet content without clear permission – a lot of the web training data that other models used might violate Apple’s strict user privacy stance or its legal caution.

Apple also historically favored on-device AI powered by specialized hardware (the Neural Engine in iPhones) for things like image recognition, rather than huge cloud AI models. That kept them focused on compact models, not those sprawling 100-billion-parameter beasts. The result: when the AI revolution hit, Apple didn’t have a GPT-4 equivalent ready to go.

The good news for Apple is that the landscape has changed dramatically in just the last year. The emergence of strong open-source models – like DeepSeek-R1 and Meta’s LLaMA series– provides Apple a very real path forward. Apple doesn’t necessarily have to build a giant LLM from absolute scratch behind closed doors; they can leverage the progress made by the open AI community (and possibly even partner with or adopt some of those models under the hood). 

This is huge: it means Apple can stand on the shoulders of giants. They can use Apple’s computing muscle (they’re reportedly deploying Apple Silicon in servers for Private Cloud Compute ), use Apple’s expertly curated data (maybe all those on-device models and knowledge graphs), and incorporate the open model weights as a baseline. Essentially, now that a handful of groups have proven it’s possible to build these AI brains, and some are even giving them away, Apple’s job switches back to a “doable unknown.” It’s no longer whether Apple can get a great LLM – it’s how and when they’ll integrate one. And that is a much more comfortable problem to solve.

Hype Happens to Everyone

Let’s address another criticism from the Daring Fireball article: that Apple engaged in a bit of hype or even dishonesty by showcasing Siri features long before they were ready. Fair enough – Apple did jazz everyone up about personalized AI features that, as of today, still haven’t shipped. That’s frustrating. But framing this as uniquely rotten behavior from Apple misses a wider truth: every major AI player dishes out hype. In this industry, hype is practically a feature, not a bug.

Consider a few examples from Apple’s peers:

• Google’s Bold Claims & Early Missteps: Google I/O 2023 was essentially one big AI flex, with the company declaring it was infusing AI into everything. Demos showed things like “Help me write” magically drafting emails and Project Tailwind summarizing notes. Ambitious! But when it came to their core product, Search, Google got a little ahead of itself. The hurried launch of Bard – clearly to counter ChatGPT – led to that high-profile demo fail we mentioned. The inaccurate Bard answer in Google’s own ad made headlines and had people questioning if Google had rushed out an AI that wasn’t ready for primetime .

• Meta’s Galactica Fiasco: In November 2022, Meta (Facebook) unveiled Galactica, an AI model it touted as a sort of scientific research assistant that could synthesize academic knowledge. The hype was that Galactica would revolutionize how we lookup scientific info. What happened? Within three days of launching a public demo, Meta had to pull Galactica offline because it was spewing bunk – authoritative-sounding but completely incorrect and even biased nonsense . Experts and users were appalled at the results . It turned out Meta’s hype about an “AI scientist” was far ahead of the reality; the model wasn’t ready and the use-case wasn’t well thought-out. Meta ate crow and went back to the drawing board. If Apple’s concept video was “bullshit” in Gruber’s words, Meta’s Galactica was a whole manure truck – they hyped it, then had to admit it really wasn’t working as advertised.

• Elon Musk / xAI’s Bravado: Elon is never one to shy away from hyperbole, and his new AI venture xAI is no exception. Just recently he proclaimed their latest model, Grok 3, to be “the smartest AI on Earth,” claiming it “outperforms anything that’s been released” . Subtle, right? 😏 Grok might be good, maybe even great, but calling your own creation the smartest on the planet before independent benchmarks is peak hype. It’s meant to grab headlines – and it did. Whether it truly lives up to that label in real-world use is another story.

The pattern here is clear: tech companies routinely market future capabilities to get developers, investors, and customers excited, even if the tech isn’t fully baked yet. Sometimes it’s necessary to keep pace with competitors’ narratives. Sometimes it’s to stake a claim of leadership (“we’re on this too!”) to avoid looking behind. And yes, sometimes it’s to distract from current shortcomings with visions of what’s coming.

Apple is actually less guilty of this than many. For years, Apple’s reputation was that it under-promised and over-delivered – they’d release stuff when it was ready, and it would often exceed expectations. The AirPower charging mat saga (promised, then quietly canceled) and now this Siri delay are notable precisely because they were rare in Apple’s history. Meanwhile, Apple’s rivals have been out there hyping moonshots and occasionally eating crow. IBM’s Watson was sold as an AI doctor that’d cure cancer – that sure didn’t happen, and IBM had to drastically scale back Watson’s healthcare push after wasting billions. Microsoft had its Tay chatbot go rogue within 24 hours on Twitter in 2016, which was an embarrassing “whoops, we hyped AI and it turned into a racist troll” moment. Amazon has hyped drone deliveries and Alexa revolutionizing our lives; the former is still pilot-only and the latter, well, Alexa’s future is reportedly uncertain as the hype died down.

This is not to excuse Apple, but to normalize the situation. Hype is the norm in tech. Everyone’s trying to sell a bit of the future before it fully arrives. Apple’s misstep with Siri is news precisely because Apple generally tries to avoid vaporware. It’s fair to call them out on it – and Gruber did, fiercely – but let’s not pretend Apple’s alone in this sin. If something is rotten in Cupertino, it’s a smell that wafts through all of Silicon Valley. 

The key is what comes next: does the promise eventually materialize, or does it stay rotten? In many cases, the initial hype does eventually lead to real product (Google did deliver solid AI features later; Meta’s open-source LLM work post-Galactica has been stellar). The onus is on Apple to turn their hype into tangible results – and I believe they can, given a bit more time.

Vision Pro: Iterate, Iterate, Iterate

To see why giving Apple some breathing room can pay off, look no further than the Apple Vision Pro. Here’s a product that, on day one, nobody could call perfect – yet it’s perhaps Apple’s boldest bet in years, and a showcase of the company’s perseverance.

Apple unveiled Vision Pro in June 2023 with great fanfare. The device is ambitious: an entirely new spatial computing platform, breaking the boundary between digital and physical spaces. Reviewers who tried it were impressed by the potential. Eye-tracking that lets you navigate menus with just a glance, a display so sharp it’s like a 4K screen per eyeball, the ability to have multiple virtual screens floating in your living room – it’s the stuff of sci-fi. As WIRED put it, “The Vision Pro is a big leap for spatial computing” .

But (and this is a big “but”) – Version 1.0 of Vision Pro also has serious drawbacks. It’s literally heavy-handed in its approach. The headset is bulky and heavy – about the size and weight of a scuba mask strapped to your face. Testers reported it’s too heavy to wear for more than a couple hours without neck strain, to the point Apple is devising extra straps to help distribute weight. It also has an external battery pack that only lasts about 2 hours, which is hardly “all day computing.” And did we mention the price? $3,499 – a small fortune.

Remember the first iPhone? Revolutionary, yes, but it lacked 3G and copy-paste and was stuck on one carrier – limitations that seem laughable now. Apple methodically improved each generation (3G, then 3GS…) and in a few years the iPhone conquered the world. Or the Apple Watch: the 2015 Gen 1 watch was slow and its purpose unclear (was it for notifications? fitness? fashion?). Early critics called it an expensive bauble. Yet Apple kept refining the software and hardware; by Series 3 and 4, the Watch found its stride in fitness and health, and today Apple Watch absolutely dominates the smartwatch market (it even outsold the entire Swiss watch industry by 2019!). The pattern is clear: the tech industry often starts with a bold but imperfect 1.0, then relentlessly iterates until it owns the category.

Vision Pro looks to be following this playbook. Apple doesn’t expect to sell millions of $3499 headsets in year one. What it expects is to get the technology out there, into the hands of developers and early adopters, and learn from real-world use. They’re likely already working on Vision Pro 2 and a lighter, more affordable “Vision Air” (just guessing the name!) that fixes the Gen-1 pain points. But Apple needs that time to do its thing – to cook, if you will, the recipe for spatial computing greatness.

Apple could have waited on Vision Pro until it was as small as ski goggles and half the price – but they knew they’d learn more by launching the best they could now and improving it. 

Similarly, one could argue they should have waited until they had that GPT-4-level Siri brain running smoothly. But by announcing early, Apple signaled both to users and (importantly) to its own teams internally that this is the future direction. They put a stake in the ground: Siri will evolve from a simple voice assistant to a proactive, context-aware “do engine”. That kind of clear vision can galvanize development internally (even if it also raises external expectations).

Conclusion: High Expectations, Higher Ambitions – Let Apple Cook 🍎👨‍🍳

It’s fair to feel let down when a company with Apple’s reputation stumbles. Gruber’s critique came from a place of high expectations – expectations Apple earned by usually delivering. Yes, Apple set a high bar for itself, and yes, we should absolutely hold them accountable when they fall short. Complacency and “mediocrity… and bullshit” should not be given a free pass . Apple’s leadership should feel the heat to get this right (and no doubt, internally, they do).

However, there’s a difference between accountability and cynicism. Critique the delay, demand excellence – but don’t write off the whole effort as doomed. We, as tech enthusiasts and consumers, should maintain a culture of technological adventurism. That means encouraging companies to take big swings, even if they might whiff a few times before hitting the home run. It means tempering our disappointment in the short term with a bit of optimism and patience for the long term.

Apple took a big swing with its AI strategy: rebranding Siri under Apple Intelligence, teasing an ambitious vision of a truly smart assistant woven throughout your life, all while sticking to their privacy guns. The first crack at it hasn’t been smooth. But the game isn’t over – in fact, it’s just the first inning. The ingredients for a comeback (or rather, a cook-through) are mostly there: Apple now has access to the foundational tech it was missing, it has a proven ability to iterate hardware/software over time, and it enjoys an ecosystem integration (hardware + software + services) that others can’t easily replicate. If Apple can get its own LLM or equivalent integrated and deliver those Siri features in the next year or so, the narrative will swiftly change from “rotten” to “ripe.”

So, let’s let Apple cook – both figuratively and literally (we’re talking about the house that Cook runs, after all). That doesn’t mean blind faith or ignoring missteps. If you barge into the kitchen mid-way through meal preparation, you might see a mess of pots, odd smells, and doubt the meal will be any good. But a great chef can turn a mid-cook mess into a masterpiece given a bit more time, and Apple has proven to be a pretty great chef in the past.

High expectations are a compliment to Apple – it means we believe they can do incredible things. Personalized, privacy-preserving AI that seamlessly works across billions of devices is a moonshot worth aiming for. No one else has quite pulled that off yet. If Apple does, all the early frustration will be forgiven in a heartbeat.

In the meantime, call out the delays, sure – but also, keep some faith in Apple’s process. They’re in the kitchen, the heat is on, and the recipe is complicated. I say: let them cook. The feast, when it’s finally served, could be one for the ages.

Sources:

• John Gruber, Daring Fireball: Apple’s delay of “Personalized Siri” features and critique of Apple’s overreach .

• Reuters: Google’s rushed Bard launch and $100B market cap drop after a demo error .

• InfoQ: DeepSeek open-sources R1 model, matching OpenAI’s top model on key benchmarks .

• Business Insider: Elon Musk/xAI’s claim of Grok 3 as “smartest AI on Earth” and context of DeepSeek’s impact .

• VentureBeat: Meta’s Galactica model pulled after 3 days due to inaccurate (hallucinated) output .

• Apple Support documentation: Apple’s on-device approach to Apple Intelligence and Private Cloud Compute, preserving privacy of user data .

• WIRED review of Apple Vision Pro: describing it as a “big leap” with impressive tech but noting it’s heavy/expensive .

• MacRumors/UploadVR: Reports of Vision Pro being too heavy for extended use, with Apple considering additional straps .

• Meta (Mark Zuckerberg) Newsroom: On the rapid progress of open-source AI models like Llama 2 to Llama 3 closing the gap with frontier models .