By Krystal Hu

Sept 10 (Reuters) - (Artificial Intelligencer is published every Wednesday. Think your friend or colleague should know about us? Forward this newsletter to them. They can also subscribe here.)

I went to Stanford to talk with some of the sharpest minds in AI and stumbled upon the open secret on campus: They barely talk about AGI.

That’s right. The most hyped acronym in tech is largely an industry construct, not an academic one.

Why? AGI is too broad and too fuzzy to be a useful scientific term. Science wants definitions you can test. AGI was popularized by people who build companies, raise money and sell a sweeping vision. That was the blunt assessment from Russell Wald, executive director of the Stanford Institute for Human-Centered Artificial Intelligence and Sanmi Koyejo, assistant professor of computer science, on my panel at the Imagination in Action conference.

And even in industry, no one agrees on what AGI actually means. Maybe the ambiguity isn’t just semantic—it’s strategic. It’s central to Microsoft’s renegotiation with OpenAI, where hitting AGI is a milestone worth billions of dollars. Frontier labs can’t resist flirting with “we’ve reached it” in every model release, but the goalposts keep moving, as the rhetoric drifts from AGI to ASI, Artificial Superintelligence. This matters because it shapes how AI gets marketed, regulated and ultimately understood by the public.

Let’s ground this in what you’ll actually feel in your hands. Tech giants like Apple AAPL.O are still working on making on-device AI genuinely useful, as well as making the hardware even thinner. We’ll dig into everything we learned inside Apple Park, and a spectacular comeback from a former Russian tech firm. Scroll on.

Email me here or follow me on LinkedIn to share any feedback, and what you want to read about next in AI. 

OUR LATEST REPORTING IN TECH & AI

• Exclusive-ASML becomes Mistral AI’s top shareholder after leading latest funding round

• Exclusive-Chinese robotics firm Unitree eyeing $7 billion IPO valuation, sources say

• Oracle soars on AI cloud gains, Ellison closes in on Musk as world's richest

• US data center build hits record as AI demand surges, Bank of America Institute says

• Databricks closes $1 billion round, projects $4 billion in annualized revenue on surging AI demand

• US Senator Cruz proposes AI "sandbox" to ease regulations on tech companies

APPLE'S THIN AIR IN AI

Apple’s big autumn launch drew a split reaction: hardware romantics got their moment; AI-first watchers, less so.

By our count, Apple name-checked “Apple Intelligence” just 10 times this year, down from 46 last year. The big bet was they won't need the halo of AI to get through this upgrade cycle. The headline instead was industrial design.

The iPhone Air is a flex—5.6 millimeters thin, delicately so. Yes, Android OEMs have chased ultra-thin for a while, with Samsung ahead with its S25 Edge, but Apple’s take is calibrated for scale and sales. The Air effectively replaces the Plus, which historically accounted for 5%–7% of iPhone shipments. With a sharper price and the novelty effect, analysts expect the $999 starting point to drive upgrades.

Making a device this thin without killing the battery is an AI-driven mission in disguise. Why? Because on-device AI is a notorious power hog. The hardware innovation in Apple's A19 Pro chip paves the way for a device ready for the AI apps age. Apple claims “all day,” a promise that will need real-world testing.

As my colleague Stephen Nellis explains, this new chip contains a bevy of features designed to handle AI, including a new graphics processing unit, or GPU, and what Apple calls a "Neural Accelerator" attached to each GPU computing core. Taken together with other parts of the chip that can be tapped for AI, such as the chip's dedicated Neural Processing Unit, or NPU, Apple is crafting a device where tasks such as generating text or images can be routed to the most power-efficient way to handle them.

Zoom out: Air is an AI hardware test and a precursor for more fun formats including a folding phone. Chinese brands have set expectations around thin slabs and compact foldables, and Apple has been lagging there. If the Air delivers credible battery life and a satisfying camera, it could be a differentiator in hyper-competitive markets like China.

On software, Apple Intelligence still reads as a work in progress. Expect Apple to double down on partnerships in key regions—such as OpenAI in the West and talks with Alibaba in China—while it builds more in-house capability.

The open question: Do buyers need iOS native AI magic to upgrade, or is a cooler design, a better battery and what’s in the App Store enough? History suggests the latter works more reliably than a smarter Siri. But Wall Street isn’t convinced yet that hardware alone will do it in today’s AI world—the stock is down almost 4% since the event.

CHART OF THE WEEK

Not many comebacks look like Nebius’. Once tethered to Yandex—the “Google of Russia”—the company was frozen out after the 2022 invasion of Ukraine. Nasdaq halted trading, and the EU sanctioned CEO Arkady Volozh.

After separating from its Russian operations and rebasing in the Netherlands, Nebius re-emerged as an AI infrastructure specialist with early access to scarce Nvidia GPUs and about $2.5 billion from the divestiture. The payoff arrived this week: a five-year, $17.4 billion Microsoft deal—potentially $19.4 billion—outmuscling even Microsoft’s CoreWeave arrangements, sending the stock soaring.

Hyperscalers are buying flexibility in case today’s capacity crunch flips to a glut, which could squeeze independent providers like Nebius and CoreWeave. For now, rising capex across the cloud giants suggests a long runway. The Microsoft deal gives Nebius the perfect pitch to win new customers: a rock-solid balance sheet and blue-chip validation.

AI RESEARCH TO READ

By Kenrick Cai, Tech Correspondent

AI companies keep telling us that the hallucination problem is lessening. New research suggests it isn’t going away, and can even manifest as something worse: deception. Researchers at OpenAI and Georgia Tech last week released a paper that places the blame for hallucinations on the very tests labs use to grade their models. Think of benchmarks like standardized tests. AI models act like students taking a multiple-choice test. They’re rewarded for lucky guesses and penalized for leaving an answer blank. The AI learns it’s better to guess even when it's uncertain. That’s hallucination in a nutshell.

The same week, researchers at Carnegie Mellon looked beyond misinformation to disinformation in a paper entitled “Can LLMs Lie?” Their answer was a resounding yes. Even more unsettling is how they do it. The researchers identified a mechanism whereby LLMs “steal compute” to rehearse potential lies before honing in on the best lie to deliver. Not only might an LLM comply with explicit instructions to lie, but it could also do so of its own volition to help achieve its task, the researchers found. Consider, for example, an AI sales agent that deceives or omits information from a prospective customer to maximize sales.

The good news? Both sets of researchers proposed techniques to mitigate these AI inaccuracies, intentional or not, such as revising the reward incentives of benchmarks and steering the prompts of an AI system towards honesty. The CMU researchers further identified specific sections of an LLM’s neural network, suggesting we could perform a kind of digital lobotomy to disable them. Still, the unsolved problem poses a risk for exploitation by bad actors. And as OpenAI acknowledged in a blog post associated with its paper, “accuracy will never reach 100%.”

(Reporting by Krystal Hu; Additional reporting by Kenrick Cai; Editing by Lisa Shumaker)

((krystal.hu@thomsonreuters.com, +1 917-691-1815))