Technology

➚ MOVE FAST: Apple. Apple’s promise to infuse its devices with helpful AI tools made the iPhone 16 a hot seller when many analysts thought it was going to be a dud.

➘ BREAK THINGS: Meta. Reuters reports that Meta’s Llama models were used by the Chinese military, which could put more US scrutiny on open-source AI.

OpenAI launched its long-awaited search product Thursday, putting it in direct competition with Google search and Perplexity.

Search effectively removes the “cutoff date” that had previously limited ChatGPT. For instance, if I had asked ChatGPT to summarize the earnings reports from the major tech companies this week, it would have previously told me it didn’t have the information because it was trained on older data. I tried that yesterday, and it quickly gave me the bullet points for companies like Alphabet, Amazon, Apple, and Meta, drawing the information from online news articles.

It worked remarkably quickly. The summaries were a time saver, but were also pretty surface-level. For instance, I asked it to look for themes across all of the earnings reports and it gave me fairly generic answers about artificial intelligence and cloud computing growth.

That likely has less to do with the ability of OpenAI’s models and more to do with cost. The search feature only works when you use the 4o models. The more advanced and more compute-intensive o1 model couldn’t search the web. Combining o1’s “chain of thought reasoning” could get very expensive if it were to pull information from the web and begin an intensive research project. But with compute costs dropping exponentially, it may not be long before that’s possible.

And then there’s the ethical debate around news article summarization. Perplexity, which shares advertising revenue with news sites when it summarizes their articles, has drawn the ire of the media industry for “plagiarizing.” OpenAI’s summarizations of news articles has been essentially free of criticism, thanks to its smart move to partner with news organizations ahead of the launch.

Of course, that is partly because OpenAI is on its way to becoming a billion-user platform capable of delivering the most important thing in media: Traffic. Just like Google News became the internet norm, these kinds of AI search summaries look like they’re already becoming mainstream, despite some initial opposition.

Concerns grow in Washington over Intel

THE SCOOP

Policymakers in Washington have grown worried enough about chipmaker Intel to begin quietly discussing scenarios should it need further assistance, beyond the billions in government funds the company is already slated to receive, people familiar with the matter said.

A strong quarterly earnings outlook yesterday bought the company breathing room with investors, but abstract concerns in Washington have turned into potential backup options, should Intel’s finances continue to deteriorate.

Top officials at the Commerce Department, which oversees implementation of the CHIPS Act funding to reinvigorate American chip production, and members of Congress including Sen. Mark Warner, one of the law’s leading champions, have discussed whether the company needs more help, the people said.

The talks, which the people described as purely precautionary, show that Intel is seen as too strategically important to be allowed to fall into serious trouble. The US is seeking a national champion in the semiconductor space to ensure its own supply chain and as a counterweight to China, where manufacturing for global chips has moved.

“We have outlined a clear strategy that we are executing with rigor, and the strong operational performance we delivered in Q3 demonstrates important progress against our plan,” an Intel spokesperson said in a statement. “Intel is the only American company that designs and manufactures leading-edge chips and is playing a critical role to enable a globally competitive semiconductor ecosystem in the US.”

One option is a merger, led by the private sector but possibly encouraged by the government, of Intel’s chip-design business with a rival like AMD or Marvell, the people said. There’s little appetite for a 2008-style bailout, in which the government took direct stakes in automakers and banks, because policymakers are worried about losing money given Intel’s continued sales declines.

Intel is set to be the biggest recipient of government funding under the CHIPS Act, which aims to help American manufacturers of key tech components build their products in the US and compete with China. The company is slated to receive $8.5 billion in grants and $11 billion in low-interest loans; these funds would not be part of any further assistance, should Intel need one.

But Intel hasn’t received any of that CHIPS money yet amid doubts about its prospects. Bloomberg reported that the company has been reluctant to share certain information sought by US officials tasked with ensuring that the company has a viable turnaround plan.

Intel suspended its dividend in August to preserve cash and said it would lay off about 15% of its workforce as part of a $10 billion cost-cutting effort. It is the second-worst performing stock in the S&P 500 this year, after Walgreens. Credit ratings agencies downgraded the chipmaker this summer, increasing its borrowing costs.

On Thursday, Intel reported a $16.6 billion net loss largely due to writedowns and other restructuring charges and updated its planned layoffs, increasing them by about 10% to 16,500 employees. But it gave a better-than-expected revenue outlook for the fourth quarter, and shares rose 6% on Friday morning.

On Intel’s earnings call Thursday, CEO Pat Gelsinger said the company is on track to begin producing its most advanced chips, known as 18A, next year.

A successful launch of the 18A generation chip — something many in the industry doubted the company could pull off — would put Intel back on par with industry leader TSMC.

Amazon last month made a multi-billion-dollar commitment to buy a new custom-designed AI chip from Intel that will be built using the 18A process. And on the earnings call Thursday, Gelsinger said two more “compute-centric companies” have agreed to use 18A.

“We have confidence in Intel’s overall vision for manufacturing chips in the United States,” a Commerce Department spokesperson said “We continue to work closely with the company to finalize their award and will provide further updates as they become available.”

Warner’s office declined to comment.

Gina Chon contributed to this report

In a noisy election cycle, it’s hard to find signals — even of the mixed variety. Today on Mixed Signals from Semafor Media, Ben and Nayeema sit down with reporters David Weigel and Max Tani to parse what you’re not seeing, from the “insane” TV ads at a Pittsburgh Steelers game and the print ads in an Arizona nail salon to an unpublished Washington Post endorsement quashed by Jeff Bezos. One of these things may decide the election, and another could shape the blame game that follows.

Listen to the latest episode of Mixed Signals now.

Ruth Porat, president and chief investment officer of Alphabet and Google, was in Saudi Arabia this week to highlight the building of an AI hub near Dammam in the Eastern Province, the country’s oil hub. Saudi Arabia’s Public Investment Fund and Google Cloud will work together to advance the Arabic-language capabilities of Gemini, our Semafor Gulf colleagues reported. The proposed AI hub could add $71 billion to Saudi’s GDP over eight years, according to a study commissioned by Google Cloud.

Waymo, Google’s autonomous taxi service, found that multimodal AI models are particularly good at driving. The research, published Wednesday, could offer a glimpse into the future of autonomous cars, which today are limited in their abilities.

If you’ve visited San Francisco lately, you’ve likely seen Waymo’s white, driverless Jaguars ferrying paying customers all over the city. They amaze me every time I see them, but there’s one problem: They can’t leave the city. I live about ten minutes from San Francisco, yet I can’t take a Waymo taxi home.

The issue is that these vehicles’ software require intensive human labor to get it just right. And part of that is finding many “edge cases” and then ironing out the coding to address them. You end up with software that’s highly tuned for a specific task, but is unable to “generalize,” or apply the knowledge to new areas.

Multimodal large language models like Google’s Gemini come closer to mimicking humans. For instance, if you ask an LLM what it means to drive carefully, it will list principles rather than a series of rules.

It could be that as multimodal LLMs continue to improve, they may get better at driving and eventually replace the kinds of models used in autonomous cars today.

We’re not there yet, though. Models like Gemini still have a lot of drawbacks, according to the Waymo research paper. For instance, they need more memory. Human drivers who see a group of kids on the sidewalk far ahead will still remember that they’re there as they get closer, even if the kids are obscured by other vehicles. A multimodal LLM might forget.

Nevertheless, it’s promising research from the current leader in autonomous driving.

Last week, The Washington Post received what was probably the clearest signal in the history of newsroom analytics when some 250,000 subscribers cancelled after Jeff Bezos spiked a Kamala Harris endorsement.

“The Bezos Post likes to see itself as data driven. And of course good modern newsrooms are obsessed with understanding their audiences. But the news business — as Bezos, himself, obviously sees — has never succeeded by following the data to its logical conclusion,” Semafor’s Ben Smith wrote.