Apple Won the AI Race

English
BBetter Stack
컴퓨터/소프트웨어경제 뉴스가전제품/카메라스마트폰/모바일

스크립트

00:00:00Apple had a rocky road when it comes to AI.
00:00:02After years of embarrassment and even a lawsuit for underperforming,
00:00:05they finally released the beta containing updated Apple Intelligence and Siri AI.
00:00:11Early testers are saying this lives up to the hype of earlier demos
00:00:14and it's the best Siri has ever been.
00:00:16And I think this may be Apple's moment to dominate again.
00:00:20Because when it comes to AI, no single company has even maintained a lead.
00:00:25Anthropic, Google, OpenAI, they all released the best model
00:00:29for just a few weeks or months to pass,
00:00:31then the lead is matched by everyone else.
00:00:33A lead only works if you can maintain it
00:00:35and Apple's no stranger to sitting back and waiting for others to figure out hard problems.
00:00:40It's not even the models they want to compete on anyway,
00:00:42it's everything else.
00:00:44And Apple has the perfect combination of advantages to pull this off.
00:00:48So today I want to explore those advantages
00:00:50and why I think Apple may just win the AI race.
00:00:58Everyone keeps saying that the new Siri runs on Gemini,
00:01:02but that's not true.
00:01:03They do have a deal with Google that's rumored to cost 1 billion per year.
00:01:07So the speculation was that Gemini would be running all of the inference for Siri.
00:01:12But Gemini was actually used to distill the models that run Siri.
00:01:16See, Apple announced five different models.
00:01:18Two which run on device,
00:01:20AFM 3 core and AFM 3 core advanced.
00:01:22And three which run on the cloud in Apple's own private cloud compute.
00:01:27AFM 3 cloud,
00:01:28ADM 3 cloud image and AFM 3 cloud pro.
00:01:31Now Apple's own wording is that these are custom built in collaboration with Google.
00:01:36Craig Federighi is Apple's senior vice president of software engineering
00:01:39and he made it crystal clear that these do not run on Google.
00:01:43We of course don't use Google search or anything like that as the foundation of our system.
00:01:48So I hope that's clear.
00:01:49The amount of Google assistant we use is none.
00:01:52So Apple's deal with Google was to be able to use Gemini for distillation and training.
00:01:56A move that allows Apple to achieve much better intelligence,
00:01:59but still do things the Apple way.
00:02:02But just a quick one,
00:02:03if you enjoy videos like this,
00:02:04then you do us a huge favor by subscribing to BetterStack
00:02:07so we can keep creating content you enjoy.
00:02:09Now, back to the video.
00:02:11Models can be distilled,
00:02:12legally or not,
00:02:13and with researchers and engineers constantly moving between companies,
00:02:17the knowledge of how to build these models is constantly being shared.
00:02:20And this leads on to the next point.
00:02:22No single company has ever been able to maintain a lead in LLMs.
00:02:26Anthropic, Google, OpenAI and now XAI,
00:02:29they constantly release the best model
00:02:31and then just a few weeks pass before the other Frontier Labs catch up.
00:02:35We see this happen constantly.
00:02:37Grok fell behind for a year,
00:02:38then released 4.6,
00:02:39which is now only one point behind Fable on benchmarks.
00:02:43And when Anthropic released Fable,
00:02:45the public version of Mythos,
00:02:47you know, the model too dangerous to be in the hands of the public,
00:02:49OpenAI released Sol Ultra just a few weeks later.
00:02:53Not to mention the open source models,
00:02:55DeepSeq, Kimi, Minimax,
00:02:57are all trailing closely to closed source models
00:02:59at just a fraction of the cost.
00:03:01And look, labs like Anthropic are the first to complain
00:03:04when these models are being distilled,
00:03:05but they were caught pirating.
00:03:07So Apple used Google to help distill their models.
00:03:10And as I mentioned earlier,
00:03:11two of those models run on device.
00:03:14And this is a big deal because local AI has seen massive improvements recently.
00:03:18Particularly, mixture of experts models have seen orders of magnitude
00:03:22improvement in performance,
00:03:23able to run direct on device.
00:03:26Projects we've already covered on this channel,
00:03:27like Turbo Fieldfare, Calibri, and Free Token,
00:03:31all rethink the architecture of these models.
00:03:34Streaming data direct from SSD,
00:03:36caching frequently used experts in memory,
00:03:38and many more optimizations,
00:03:40all to run massive models on device
00:03:43with incredibly low memory footprints.
00:03:45Apple will take advantage of their specific hardware
00:03:48to squeeze the absolute most out of these models
00:03:51because now they own the entire stack
00:03:53from the model itself to the silicon it runs on.
00:03:56And of course, Apple go hard on privacy
00:03:58and local AI plays perfectly into this.
00:04:01Apple's next advantage is their distribution channel,
00:04:04which has two benefits.
00:04:06Users are already locked into Apple's ecosystem,
00:04:08including multiple products like AirPods,
00:04:10Apple Watch, Photos, Password Manager,
00:04:13and with over 2.5 billion active devices worldwide,
00:04:17most users are never going to switch.
00:04:19Competitors would need an astronomical improvement
00:04:22to their own ecosystem
00:04:23to fight the friction of switching.
00:04:25Realistically, that's never going to happen.
00:04:28So their models don't even need to be the best.
00:04:30They just need to be good enough to be useful
00:04:32and Apple has all of the advantages to make that true.
00:04:36Apple also has the means to train their models
00:04:38on massive amounts of data.
00:04:39Those 2.5 billion active devices
00:04:42are the perfect source of interaction data to train from.
00:04:45At one point, for example,
00:04:46it was reported that 40 to 50% of Anthropix revenue
00:04:50came from Cursor alone
00:04:51and that data surely contributed heavily
00:04:54to their dominance in software engineering.
00:04:56Millions of real users interacting
00:04:58with millions of lines of code
00:04:59and the data from those interactions
00:05:01is what makes those models better.
00:05:04Apple has the same advantage
00:05:05and they could do the same,
00:05:06but on a much larger scale.
00:05:08And talking of Apple's ecosystem,
00:05:10they can also lead on tight integration
00:05:12by locking third parties out
00:05:14with claims of privacy or security.
00:05:17Zuck highlighted that Apple were able to do
00:05:18this exact same thing with AirPods.
00:05:21They build stuff like AirPods, which are cool,
00:05:23but they've just thoroughly hamstrung
00:05:26the ability for anyone else to build something
00:05:29that can connect to the iPhone in the same way.
00:05:32So, I mean, there are a lot of other companies
00:05:33in the world that would be able to build
00:05:34like a very good earbud.
00:05:35And whenever you push on this, they get super touchy
00:05:38and they basically wrap their defense of it
00:05:41in, well, if we let other companies plug into our thing,
00:05:45then that would violate people's privacy and security.
00:05:47It's like, no, just do a better job
00:05:49designing the protocol.
00:05:50And that makes you think,
00:05:51does Apple really care about privacy
00:05:53or is it just a convenient excuse
00:05:55to lock out competitors?
00:05:56Either way, it works in their favor.
00:05:59Now, the big argument is that Apple
00:06:00are too late to the party,
00:06:02but Apple are well-known for sitting on things for years,
00:06:05letting others innovate, work out the wrinkles
00:06:07and then slide in with a better product.
00:06:09Whether this has worked has been a mixed case for Apple.
00:06:12iPod, iPhone, iPad watch, AirPods and Apple Silicon,
00:06:16all late to the party,
00:06:17but all eventually leading their respective markets.
00:06:20But Maps, HomePod, arguably Vision Pro
00:06:23and the car they canceled,
00:06:25all failed to gain dominance.
00:06:26And which camp Apple's intelligence sits in
00:06:29is the big question.
00:06:30Up until now, they've fallen behind.
00:06:32But in areas they'd failed previously,
00:06:34there was always a clear market leader.
00:06:36Alexa dominated the smart speaker market
00:06:38and Google dominated with Maps.
00:06:41But with no single company able to maintain a lead with AI
00:06:44and Apple having the advantages we've discussed,
00:06:47the possibility of them taking the lead
00:06:48is now very real.
00:06:50A model that's just good enough,
00:06:52focusing on everything else that makes a difference.
00:06:54The deep integration, the harness, the tools, the user base.
00:06:58Apple has all of this.
00:07:00And if you wanna see how Apple Silicon
00:07:01specifically can take massive advantage
00:07:04of running local AI,
00:07:05then check out this video
00:07:06where we explore the topic in deeper detail.

핵심 요약

Apple dominates the AI race not by having the single best foundational model, but by combining on-device local execution, hardware control, and a 2.5 billion user ecosystem.

하이라이트

  • Apple uses Google Gemini exclusively for model distillation and training rather than running inference for the new Siri.

  • Apple announces five AI models consisting of two on-device variants and three private cloud compute versions.

  • Over 2.5 billion active Apple devices worldwide provide a massive interaction data source for continuous model training.

  • Apple leverages custom silicon and full-stack control to run mixture of experts models directly on-device with low memory footprints.

타임라인

Apple Intelligence Architecture and Google Deal

  • Apple releases updated Apple Intelligence and Siri AI beta versions.
  • A rumored one billion dollar deal with Google utilizes Gemini strictly for model distillation and training.
  • Apple introduces five custom models split between on-device execution and private cloud compute.

Initial reports misattribute Siri inference to Google Gemini, but Gemini only distills the foundational models. Apple senior vice president Craig Federighi confirms zero reliance on Google search or assistant for the final system. Five custom models emerge, comprising two on-device versions named AFM 3 core and AFM 3 core advanced, alongside three cloud-based variants.

Unstable Lead in Large Language Models

  • No single company maintains a permanent lead in large language models.
  • Frontier labs like Anthropic, OpenAI, and Google repeatedly catch up to competitor releases within weeks.
  • Open source models from DeepSeek, Kimi, and Minimax closely trail closed source options at a fraction of the cost.

The artificial intelligence landscape features constant leapfrogging where leading models hold dominance for only weeks or months. Grok falls behind for a year before releasing version 4.6 to narrow the benchmark gap. Frontier labs frequently complain about model distillation while practicing similar methods themselves.

On-Device AI and Hardware Integration

  • Mixture of experts models achieve orders of magnitude performance improvements for local execution.
  • Architectural updates stream data directly from solid state drives and cache frequently used experts in memory.
  • Apple controls the entire stack from the underlying model down to the custom silicon.

Local artificial intelligence advances rapidly through architectural redesigns that minimize memory footprints. Projects like Turbo Fieldfare, Calibri, and Free Token enable massive models to run directly on hardware. Apple maximizes these efficiencies by owning both the silicon and the model architecture while prioritizing user privacy.

Ecosystem Distribution and Data Advantage

  • Over 2.5 billion active devices create a massive barrier to switching ecosystems.
  • Interaction data from millions of active users fuels continuous model training.
  • Cursor software engineering dominance derives largely from revenue and interaction data generated by massive user bases.

Distribution channels provide an insurmountable advantage against competing ecosystems. Users locked into hardware products like AirPods, Apple Watches, and password managers face high switching friction. Interaction data harvested from billions of devices offers training resources comparable to commercial coding tools.

Integration Control and Market Timing

  • Tight ecosystem integration restricts third-party hardware access under security and privacy justifications.
  • Apple historically enters markets late by allowing competitors to solve initial wrinkles before releasing refined products.
  • Past late entries like iPod, iPhone, and Apple Silicon eventually dominated their respective markets despite early skepticism.

Apple restricts third-party peripheral integration, citing privacy and security protocols while effectively locking out competitors like custom earbud manufacturers. Unlike past failed late entries into maps and smart speakers where single market leaders existed, the AI market lacks a permanent leader, creating a viable path for Apple dominance.

커뮤니티 글

모든 글 보기

Apple Won the AI Race

makedream11일 전1480
이 영상에 대해 글쓰기