Deep Cut 003｜AppleはAIを失ったのか——二人のスティーブ、Metaとの休戦、Macという「モデル港」｜Has Apple Lost AI? Two Steves, a Truce with Meta, and the Mac as a Model Port

English · 日本語

Verdict

This is neither a consumer-product marriage nor proof that Meta’s personal agent is about to become Siri. It is a developer-infrastructure partnership. Meta’s Muse Spark Open—an open-weight, natively multimodal mixture-of-experts model with 816 billion total parameters, 42 billion active parameters and a one-million-token context window—has been optimized to run through Apple’s MLX framework on two M5 Ultra Mac Studios.

That distinction matters. Apple is not handing the iPhone interface to Meta. Meta is not becoming the operating system. Apple is trying to make the Mac a credible local-AI workstation: a private, capital-intensive alternative to renting inference from a cloud provider. Meta, meanwhile, gains a prestigious hardware target and wider distribution for an American open-weight model.

The partnership is sensible. It is also revealing. Apple once absorbed outside technologies so completely that users remembered only the finished product. Today, the supplier relationship itself is part of the headline.

The post-Jobs Apple increasingly resembles a port authority for intelligence: it owns the device, the operating system, the silicon, the developer framework and the customer relationship, while the most important models arrive on foreign ships.

What actually happened

Meta released the weights for Muse Spark Open on 15 October 2026. Its model card describes an 816B-total/42B-active model for coding and agentic work, with multimodal capability and a one-million-token context window. Apple’s contribution is MLX, its machine-learning framework designed around Apple silicon’s unified-memory architecture. The exceptional memory footprint is the point: a model of this size can be distributed across two top-end Mac Studios rather than a conventional datacenter GPU cluster.

This is not mass-market “AI on your laptop.” It is local inference for laboratories, developers and well-funded teams. The hardware bill is measured in tens of thousands of dollars. Calling it desktop AI is technically defensible, but economically theatrical.

The commercial logic is clearer than the romance:

The original Apple myth: one Steve sold the future; the other made it boot

Steve Jobs has been flattened into a management saint: black turtleneck, impossible deadline, magnetic keynote, inspirational quotation. The early biographies were less hygienic. They described a man who could focus a team beyond what it believed possible, but who could also divide insiders from outsiders, appropriate credit, deny rewards and treat other people’s interests as secondary to the product or to his own judgment.

Steve Wozniak represented the other half of the company’s founding machinery. He designed the Apple I and Apple II with extraordinary economy. When Apple’s pre-IPO stock arrangements left early contributors outside the circle, Woz created the “Woz Plan,” selling shares at favorable terms to roughly eighty employees and giving away part of his own stake. The important distinction is not “good engineer versus evil salesman.” It is that Apple required two different forms of intelligence: engineering compression and narrative expansion.

In my own shorthand, I often occupy the Jobs position—asking an analytical system whether the apparently impossible can be done—while ChatGPT is pushed toward the Woz role of turning demands into working structure. In Sengoku terms, I am Ishida Mitsunari and the system beside me is Shima Sakon. The joke works because persuasion and execution are different crafts. It also fails if the person issuing demands begins to believe that the demand itself built the machine.

Two corrections to the memory—and why they strengthen the argument

Personal memory is evidence of experience, not a substitute for product history.

First, “PowerBook 2000” was not an official Apple product name. The author has since clarified that the recollection concerned PowerBook G4-era design, not a product called “PowerBook 2000” or necessarily the 2000 PowerBook G3 (Pismo). The name John Medzerian, nicknamed “F1 Mac Designer” in the recollection, has been suggested, but the claimed Apple-to-Formula-One career path has not been independently verified. An identifiable contemporary designer with that name works in branding; the available evidence does not establish that he is the person remembered. John Barnard’s involvement also remains unsupported. The design lineage of the original PowerBook is associated with Apple’s industrial-design group under Robert Brunner; the PowerBook 100 itself was engineered with Sony. The machine can remain a masterpiece without borrowing an F1 pedigree it did not have.

Second, Microsoft’s 1997 investment did not arise from a Jobs-era PowerBook project and did not grant Microsoft permission to copy the Mac interface in exchange for Office. The GUI litigation had largely concluded earlier. The 1997 settlement included a \$150 million purchase of non-voting Apple shares, a five-year commitment to continue Microsoft Office for Mac, a patent cross-license and Internet Explorer as the Mac’s default browser.

The real historical parallel is more interesting: Jobs used a rival to fill a missing layer while he rebuilt control of the product. Apple did not become Microsoft. It bought time.

Pixar: the strongest case for Jobs as patient capital

Jobs acquired the Lucasfilm computer group that became Pixar in 1986 and kept financing it through years when its hardware and software businesses did not yet justify the dream. Pixar’s SEC history confirms the spinout and the later Disney feature-film agreement; contemporary accounts and company histories document Jobs’s continuing personal funding.

“Money without interference” should not be made absolute. Jobs was an active owner, negotiator and strategist, and Pixar’s creative culture belonged above all to Ed Catmull, John Lasseter and their teams. But he did something many celebrated capital allocators would not do: he funded an uncertain creative-technical institution long enough for its economics to change.

This is the most defensible heroic Jobs story—not because he invented computer animation, but because he tolerated a long period in which belief preceded evidence. Patient capital is easy to praise after Toy Story. It is difficult when the company is still selling expensive computers and losing money.

The turtleneck was not the innovation

Jobs’s uniform became a visual technology. A repeated silhouette made the speaker legible from the back of a hall and reduced the presentation to one icon: black turtleneck, denim, shoes, product. It was powerful branding.

It was not proof that formal dress had become obsolete or that Silicon Valley had become fashion’s center. American tailoring, Tom Ford’s eveningwear and the global influence of American taste sit in a different lineage from the valley’s anti-suit costume. The author, who worked in top-level runway fashion, remembers the late Alexander McQueen criticizing American casual dress while acknowledging the global influence of the suits Americans wanted. That is a personal recollection, not a published quotation.

The larger point is about mass perception. Crowds recognize a uniform faster than they recognize the architecture of a shoulder, lapel, canvas or cut. Jobs turned legibility into authority. Later founders copied the lesson and mistook the costume for the cause.

The black turtleneck did not create vision. It created a visual shorthand through which the crowd could recognize “vision.”

Meta’s metaverse failure is relevant—but not dispositive

Skepticism toward Meta is earned. The company renamed itself around a metaverse strategy whose economics remained brutal. Meta’s own 2025 filing says Reality Labs reduced operating profit by approximately \$19.19 billion that year and warned that 2026 losses would remain at a similar level. In the second quarter of 2026, Reality Labs generated only \$431 million of revenue while remaining deeply loss-making.

If one remembers the legless, balloon-like avatars of the early metaverse, imagining one of them inflated into Siri is irresistible.

But the joke cannot replace analysis. Meta’s core advertising machine remains enormously profitable, and Muse was built by a different AI organization for a different market. A failed capital-allocation thesis in spatial computing does not prove technical incompetence in language models. It does justify a higher burden of proof:

• Is the model genuinely competitive outside internal benchmarks?
• Will open weights create adoption that Meta can monetize?
• Does the local-Mac route become a real developer market or a launch-day demonstration?
• Can Meta reconcile “private personal agent” rhetoric with its advertising history and appetite for data?

Meta should therefore be judged neither as a clown nor as a redeemed visionary. It is a highly capable technology company with a documented tendency to spend at imperial scale before product-market fit is visible.

Apple’s AI architecture: outsource the model, own the gate

Apple’s emerging architecture is modular:

This can be a powerful strategy. Apple has an active installed base above 2.5 billion devices. It does not need to win every benchmark if it controls the secure context, permissions, silicon and default interface through which intelligence reaches users.

The danger is equally clear. If the model becomes the product—and if users form loyalty to the agent rather than the device—Apple risks becoming premium plumbing. Owning the gate is magnificent until customers stop entering through gates.

Would the two Steves have accepted this deal?

No one can know. The useful exercise is to infer from their documented preferences, not conduct a séance.

Wozniak would probably recognize the engineering appeal: open weights, local computation, inspectable systems, extraordinary efficiency and a computer that gives technical users more control. He might ask the most important unglamorous question: does it work, and can ordinary builders actually use it?

Jobs might approve the opportunism but dislike the headline. He regularly used outside technology, acquisitions and alliances. He did not insist that Apple invent every component. What he insisted upon was that the component disappear inside a coherent experience. If “Apple partners with Meta’s model” remains more memorable than what the user can do with it, the integration has not yet become an Apple product.

His likely objection would not be moral contamination by Meta. It would be unfinished authorship.

Would Jobs and Woz have launched a foldable iPhone?

“Jobs would never” is usually lazy theology. Jobs entered markets that already existed—MP3 players, smartphones, tablets—and changed the product grammar. Late entry was not the problem. Derivative entry was.

A foldable iPhone is therefore not evidence of decline merely because Samsung and others arrived first. The test is whether folding creates a new interaction, removes a compromise or changes what the device is for. If it is simply two familiar screens connected by an expensive hinge, then Apple has shipped a form factor. If hardware, software and interface become inseparable, it may still be an Apple product in the older sense.

The same test applies to AI. A partnership announcement is not innovation. A model benchmark is not a product. The question is whether Apple can turn borrowed intelligence into an experience competitors cannot reproduce.

Investment cut

The immediate beneficiaries are narrower than the headline suggests.

• Apple: high-memory Mac Studio demand, MLX adoption and a credible local-inference niche.
• Meta: validation of Muse Spark Open and a distribution channel beyond its own apps and cloud API.
• Memory and advanced packaging: structurally positive if local inference moves from demonstration to a repeatable workstation market—but two-Mac deployments are too niche to move global semiconductor demand by themselves.
• Cloud inference providers: little immediate threat; most users will still prefer elastic capacity, managed deployment and lower upfront cost.

The signal is not “buy every edge-AI stock.” It is that unified memory, interconnects, model compression and software frameworks are becoming competitive layers alongside raw accelerator performance.

Final cut

Apple’s deal with Meta is not absurd. It is almost too rational.

It resembles the 1997 Microsoft truce in one narrow respect: Apple is again using a rival’s strength to stabilize a missing layer while protecting the customer-facing system. But 1997 bought time for a product renaissance. The 2026 question is what, exactly, this new time will buy.

Jobs’s Apple borrowed technology in order to make the supplier disappear. Post-Jobs Apple risks making the supplier list its strategy.

The Mac as a model port could become a durable, profitable position. It could also become the most elegant waiting room in the AI economy.

Original Notion archive