Apple deserves credit for shaping its AI strategy around its existing brand identity. This was a very Apple take on AI—as my Fortune colleague Sharon Goldman explains in this analysis of yesterday’s news. Heck, if the message that Apple intended to put its own unique stamp on AI wasn’t clear enough, Apple even redefined the acronym AI as “Apple Intelligence.” So in terms of brand strategy—bravo; what Apple announced makes a lot of sense. It also gives consumers a legitimate choice around data privacy and may convince other companies working on AI to do more to protect data privacy too. That’s also a good thing. But, in terms of both technology and business strategy, I think what Apple is doing is risky. Here’s why:

Apple is making three big bets on the future development of AI, any one of which might not pan out. The first is that it will be able to deliver the functions and features that consumers really want from AI primarily on device—that is on your iPhone or laptop. That in turn is really a bet that computer scientists and engineers will continue to find ways to mimic the capabilities of extremely large AI models, such as OpenAI’s GPT-4, with much smaller AI models. So far, this has been the trend—AI developers have done an amazing job of optimizing models and fine-tuning them to do some of what the giant models can but in much smaller packages.

But there are drawbacks. The small models aren’t as capable across different tasks as the largest models. They tend not to be as good at reasoning, in particular, which may be a problem as we want to move from AI assistants, like Siri, that simply execute a command, or like ChatGPT, which mostly just generates content, to ones that will plan and take complex actions on our behalf.

These AI agents will have to work well across multiple apps and tasks in response to a single query. Ideally, I’d want to ask a future version of Siri, “Make a restaurant reservation in town for me tonight,” and have the model be able to check my calendar, understand that my last work call finishes at 7 p.m. so the reservation must be after that time, check my location and also allow for travel time to the restaurant, search for restaurants that are within an appropriate time-distance envelope, and then actually make the booking on my behalf. That is a complex task that involves planning and reasoning and the ability to both pull data and take actions across multiple apps. This is likely the kind of thing consumers will want their AI assistants to be able to do.

Right now, no AI on the market can do this. But it’s the goal that the top AI labs are all running towards as fast and as hard as they can. For this kind of task, a small model probably won’t cut it. It will likely require a very large model running in the cloud. But Apple has said it will try, as much as possible, not to use that kind of model, hoping that we will keep finding ways to cram more capabilities into a small model that can run on device. This on-device approach may quickly hit its limit.

This brings me to Apple’s second big bet: That it will, someday soon, be able to build a large model as capable as any from OpenAI or Google that will be able to handle these more complex reasoning and planning tasks. This large model would sit in the new ultra-secure cloud that Apple has built to handle any AI queries that cannot be dealt with on-device. The ultra-secure cloud is again very clever and could become a real brand differentiator (although security researchers have some concerns about how it will work in practice. You can read more about that particular aspect of Apple’s announcements here).

But there is a big question mark over Apple’s ability to leap out to the bleeding edge of large model development (or really three big question marks, as I’ll explain in a minute). If Apple had been able to match OpenAI’s or Google’s large AI models, we would have heard about it yesterday. Instead, we learned that for the most complex Siri queries Apple is relying on a partnership with OpenAI that will see these prompts passed, if a user allows, to OpenAI’s GPT-4o. This is basically an admission by Apple that it doesn’t have the goods—that its own internal efforts have, despite at least 15 months of effort, failed to match OpenAI’s. And if Apple continues to not have the goods, its dependency on OpenAI is likely to grow, putting the company in a vulnerable position.

Could Apple catch up on large models? That depends on the answer to three other questions: Does it have the right talent? Does it have the right data? And does it have the right compute? Apple can certainly afford to hire top AI researchers and machine learning experts. And the positioning of Apple’s brand around protecting consumer privacy might actually appeal to that talent (beyond whatever cash and stock Apple offers).

But does Apple have the data to train a highly capable next-generation LLM? The most obvious place to get it would be from its own users. But Apple has promised not to use that data to train its AI models. So it’s likely hampered in this regard.

Finally, there’s compute. Training a cutting-edge LLM takes a lot of specialized hardware. Apple has access to a variety of AI chips, and it has decent cloud infrastructure, but it doesn’t have the same kind of data center heft that Microsoft and Google have, nor is it known whether it has the equivalent of the 350,000 Nvidia GPUs that Meta CEO Mark Zuckerberg has boasted his company will have online by year-end.

Currently, Apple uses Google’s AI chips, which are called tensor processing units or TPUs, to train its LLMs. Does Apple have access to enough AI chips to train a model at the frontier of AI capabilities? That’s unclear. Also unclear: If larger models are needed to handle the kind of queries consumers will most want to ask an AI assistant and if these larger models can only run in the cloud, does Apple have enough AI chips and cloud infrastructure to offer this feature to its 2 billion users worldwide? Probably not.

Okay, now to Apple’s third big AI bet. So far, Apple is betting your phone will be the primary way people interact with AI personal assistants. That’s not an outlandish wager, but it might not be right. A lot of people think you ideally will use a device that is some sort of wearable—either a pair of smart glasses, camera-equipped earbuds, or a kind of AI pin (like Humane AI’s idea, but better!) that can see what you are seeing and hear what you are hearing and allow for AI responses that are contextually accurate. It’s not at all clear from yesterday’s WWDC announcements how well Apple is positioned for this future. Apple’s entry into the world of AR and VR so far is its expensive Vision Pro—which is not the sort of thing you’d want to wear around all day. What’s more, wearable devices like glasses, a brooch, or earbuds have even less space to accommodate AI chips than a phone. So it may be even harder to have AI run only on device with these sorts of wearables. And again, it’s unclear whether Apple has the cloud computing infrastructure to handle constant AI inference requests from billions of consumers.

The nightmare scenario for Apple is this: What if consumers discover that the most useful thing about Apple devices is a more powerful AI assistant, and what if those higher-order AI functions are provided by OpenAI’s software and not Apple’s, and what if some other non-phone device winds up being the best way for consumers to interact with AI and Apple doesn’t make that other device? Well, then OpenAI can simply roll out its AI earbuds or smart glasses and take all of Apple’s current users with it. If I were Tim Cook, that is the scenario that would keep me awake at night.

So, yes, at WWDC, Apple proved it can meet today’s AI moment. Meeting tomorrow’s, however, will require a lot of things to break its way.

With that, here’s more AI news. 

Jeremy Kahn
jeremy.kahn@fortune.com
@jeremyakahn

Correction: In last week’s edition of the newsletter, the full name of the law firm Paul Weiss Rifkind Wharton & Garrison was misspelled.

**Before we get to the news, a quick reminder to preorder my forthcoming book Mastering AI: A Survival Guide to Our Superpowered Future. It is being published by Simon & Schuster in the U.S. on July 9 and in the U.K. by Bedford Square Publishers on Aug. 1. You can preorder the U.S. edition here and the U.K. edition here.

AI IN THE NEWS

Mistral valued at more than $6 billion in new venture capital round. The Paris-based AI startup, which has been hailed as France’s homegrown gen AI champion, closed €468m ($502 million) funding round that values it at more than $6 billion, sources with knowledge of the deal told tech publication Sifted. Investment firm DST Global led the new funding round with participation from existing investor General Catalyst, the sources said. It has only been about six months since Mistral, which builds open-source LLMs, raised more than $400 million in a Series A venture capital round led by Lightspeed Venture Partners and Andreessen Horowitz, and about a year since the company launched out of stealth mode with a $100 million seed round.

Tech companies and venture capital firms campaign against California AI bill. Silicon Valley has come out swinging against a bill (SB 1047) that passed the California State Senate last month and would impose a number of provisions designed to guard against catastrophic impacts from powerful AI models. The bill, which was introduced by Democratic State Sen. Scott Weiner, would require AI companies to certify that their models don’t have “a hazardous capability” and also to build a “kill switch” into their AI models. It would require cloud computing companies to ensure AI models running on their clusters comply with the law. It would impose big liabilities for noncompliance and set up a new Frontier Model Division within the state’s Department of Technology to review certifications. Critics say that legislation would be too difficult for tech companies, especially those developing open-source AI models, to comply with, the Financial Times reported. “If someone wanted to come up with regulations to stifle innovation, one could hardly do better,” Andrew Ng, another deep learning pioneer who led AI research at Alphabet’s Google and Baidu and now sits on Amazon’s board, complained on X. The bill is expected to come to a vote in California’s House of Representatives in August.

AI search engine Perplexity accused of copying articles without adequate attribution. A new feature the AI-powered search engine Perplexity debuted called “Perplexity Pages” that curate content on a particular topic came close to copying verbatim passages from publications including Forbes, CNBC, and Bloomberg, while only providing an easy-to-miss logo at the bottom of the page to signal where the information may have come from, Forbes reported. In response to the criticism, Perplexity CEO Aravind Srinivas acknowledged the issue and promised improvements, emphasizing his company’s commitment to proper attribution.

EYE ON AI RESEARCH

A new benchmark for LLMs in African languages indicates serious shortcomings. In previous newsletters, we’ve noted how the seemingly stellar performance of LLMs at multilingual tasks falls off considerably when one looks at “low-resource languages”—those for which there is relatively little digital content available online. Now researchers from Microsoft, Cohere, and a number of universities in both Europe, North America, and Africa, have developed a new benchmark for African languages—IrokoBench—that demonstrates just how much ground there still is for LLMs to make up. The new test judges the AI models on a set of high-quality human-translated tests across 16 African languages for which there is relatively little digital data available. It covers three tasks: natural language inference, mathematical reasoning, and multiple-choice question-answering. The researchers tested 10 open-source AI models and four proprietary ones. They found that the best open-source AI model, Aya-101, only managed to get 27.9% of the questions right on the test in any of the African languages. OpenAI’s GPT-4o was the best-performing model overall, and it still got the questions right less than half the time. On language inference tasks, a model trained specifically for African languages, called AfroXLMR, got about 66% of the questions right, but it is not an LLM that can answer questions or do mathematical reasoning. The research highlights how much more work there is before LLMs can be deployed worldwide. And because of this AI might exacerbate existing inequalities between regions and countries. A paper on the new IrokoBench test was published on the non-peer-reviewed research repository arxiv.org here.

FORTUNE ON AI

Ashton Kutcher sides with the enemy over actor and writer colleagues in AI debate: ‘I can just generate and then watch my own movie’ —by Chloe Berger

We used AI to analyze 24 years of retailers’ SEC disclosures—and found the one factor that would have doubled investor returns —by Jeffrey B. Wegner and George Zuo (Commentary)

AI-generated ‘BBL Drizzy’ started as a Drake joke—but music creators are now terrified it spells the beginning of the end —by Eva Roytburg

CHIPS Act faces talent shortage despite $500 billion investment: ‘We have to make semiconductor manufacturing sexy’ —by Dylan Sloan

AI CALENDAR

June 25-27: 2024 IEEE Conference on Artificial Intelligence in Singapore

July 15-17: Fortune Brainstorm Tech in Park City, Utah (register here)

July 21-27: International Conference on Machine Learning (ICML), Vienna, Austria

July 30-31: Fortune Brainstorm AI Singapore (register here)

Aug. 12-14: Ai4 2024 in Las Vegas

BRAIN FOOD

Could AI enhance democracy? There’s been a lot of talk about how AI is poised to destroy democracy by supercharging disinformation and deepfakes. I share many of those concerns. But I also think there is a more hopeful case for AI, a way in which the technology could reform and enhance our democratic processes. A few weeks ago, I moderated a panel at Viva Tech, the French technology conference in Paris, that was about this very topic. One of the more interesting ideas—discussed by panelists Sonja Solomun, deputy director of the Centre for Media, Technology and Democracy at McGill University, and Audrey Tang, a Taiwanese politician who served as the country’s minister of digital affairs—is having AI moderate something called a “citizens’ assembly.” This is a group of citizens, sometimes selected randomly and other times selected to be demographically representative of the citizenry in a particular place, who hold a discussion on policy topics where the idea is to deliberate and arrive collectively at a preferred policy. In this context, AI systems can be used to present background material and context in a form that the average non-expert citizen can understand. It can also summarize arguments for different policy options, and also to help coax people towards compromise.

Now, a political candidate in the upcoming U.K. national election is taking this kind of idea a step further by running as “AI Steve.” Businessman Steven Endacott is running for Parliament as an Independent in the seaside town of Brighton. But he promises to act as a kind of human puppet for an AI system he is calling “AI Steve.” The idea is to use an AI avatar to interact with voters, who can ask questions and raise issues. The AI software will then summarize these viewpoints and condense them into policy recommendations that Endacott would act on if elected. Endacott tells tech publication Wired that he sees AI Steve as an experiment in a new form of direct democracy. “We are actually, I think, reinventing politics using AI as a technology base, as a copilot, not to replace politicians but to really connect them into their audience, their constituency,” he said. Neural Voice, an AI voice company that Endacott chairs, created the technology behind AI Steve. Jeremy Smith, the company’s cofounder, told Wired that AI Steve can have up to 10,000 simultaneous conversations with voters and then condense all that information into a policy position.

I will be curious to see how AI Steve does, but also if others take up these ideas of using AI to encourage more participation in our democracies and to actually improve our democratic discourse and our ability to reach consensus, rather than worsening it.