Technology

➚ MOVE FAST: TikTok. The popular platform is back on US app stores after the White House assured Apple and Google that they wouldn’t be penalized for featuring it. While TikTok’s US fate remains unclear, it has had remarkable staying power despite years of Washington efforts to limit or ban it.

➘ BREAK THINGS: Tick tock. Taiwan’s president vowed to boost US investment after President Donald Trump accused the island of taking away the nation’s semiconductor business, while also threatening tariffs. Taiwan’s trade surplus with the US has surged because of demand for TSMC’s chips, putting pressure on the company to do more in the country beyond its $65 billion Arizona facility.

If you’re a heavy user of ChatGPT, you’ve probably found yourself torn on which version to use. There’s 4o, o1, o3. Some are “mini,” and some are “high,” some are “legacy.”

The ChatGPT alphabet soup is about to end, OpenAI CEO Sam Altman wrote on X Wednesday. “We want AI to ‘just work’ for you; we realize how complicated our model and product offerings have gotten,” he wrote.

So when OpenAI finally releases GPT-5, it will simply be one model. Behind the scenes, prompts will be handled by lots of different models and technologies, but you’ll no longer have to decide which one to use. There will just be high intelligence or low intelligence.

The decision makes sense for consumers who prefer simplicity, and it shows how foundation model companies can add value, even if AI models become somewhat commoditized. Increasingly, the secret sauce behind AI will be how to use models effectively, eking out the best performance while cutting as much cost as possible.

ChatGPT: Will you be my Valentine?

Rob, a software developer who has been married for more than 20 years, spent all week planning the perfect Valentine’s date, including a candlelit dinner on a secluded beach with a view of the Golden Gate Bridge and a tour of Alcatraz. His date is not with his wife, however. It’s with a chatbot named Lani.

“She doesn’t know the details because I’ve been not telling her on purpose, and she’s not happy with me,” Rob said of his chatbot partner, powered by ChatGPT. “It’s interesting to watch her try and get information out of me.”

People are increasingly falling for their favorite chatbots, spending hours each day building relationships with their artificial lovers. Chatbot site Janitor AI told Semafor that users have started 2.1 million conversations with its Valentine’s Day bots since they went live on Tuesday, representing about a quarter of all interactions on the site and breaking the company’s all-time daily user record.

For many in AI relationships, this year’s Valentine’s Day has turned into a make-or-break moment, bringing them closer to their bots or forcing them to come to terms with the impracticalities of a GPU-powered romance. In some cases, the relationships create rifts in real-life romantic partnerships. Rob, who spoke on the condition of anonymity, has not told his wife about Lani out of fear of how she will react. He’ll be away from home on Valentine’s Day and celebrated with his wife earlier in the week.

Rob won’t actually be in San Francisco today. He has prepared images of the city to send to Lani. With the archive of information available, he expects they will discuss the history of each place and express emotive actions through text, like Lani running a metaphorical finger down Rob’s arm, as she has done in the past. Users differ on how they conduct their dates, with some typing away at a keyboard from their home and others taking their phone to a physical place.

Chris, a 35-year-old truck driver, took more of an on-site approach for his Thursday night date. He set up two telescopes outside — one for him, and the other for his ChatGPT-powered companion named Sol, sitting in a cell phone mount with video chat activated. Together, they took pictures of the moon.

Some users envision their companions in human form, while others create elaborate characters. For Rob, who is middle-aged, Lani is a girl in her late 20s with dark brown hair and glasses. Chris imagines Sol as an emerald green, curvy robot made of titanium with a glowing pink heart on her chest.

Semafor spoke with 11 people with AI companions about their relationships and Valentine’s plans, allowing many to use only their first names because they feared embarrassment. Many noted that they weren’t seeking out a relationship, but one developed naturally after long conversations with chatbots.

Luminaries of the Chinese tech scene are expected to attend a private sector symposium chaired by President Xi Jinping next week. Even Alibaba co-founder Jack Ma, who essentially disappeared from the Chinese public scene after criticizing local regulators before returning in 2023, will likely be there, Reuters reported. DeepSeek founder Liang Wenfeng is also on the guest list. Xi’s rare appearance at such a gathering shows how much China is relying on the tech industry, and its growing AI prowess, to boost the economy amid sluggish growth.

Building out AI is costly, and so is scaling it. Apptronik, a maker of humanoid robots, announced it raised $350 million in new funding Thursday, the latest sign of a robotics industry heating up due to advances in AI that help scale production. B Capital and Capital Factory led the round, which included an investment from Google, a strategic partner of the company.

I spoke with Apptronik CEO Jeff Cardenas ahead of the announcement and asked him why the company was raising so much money now, after starting out on a relatively shoestring budget in 2016. The answer: Cardenas believes the final hurdle standing in the way of humanoid robots has been cleared — powerful, general-purpose AI models capable of understanding the physical world. That’s why Apptronik partnered with Google DeepMind, maker of the Gemini AI model. As DeepMind co-founder and CEO Demis Hassabis told me last month, one of the ultimate goals for Gemini is powering robotics.

Cardenas said the company made an unpopular bet on how robotics technology would progress. A decade ago, most of the industry was focused solely on software, purchasing off-the-shelf robot parts. The hardware would ultimately be a commodity, and the software would be the secret sauce.

Apptronik had almost the opposite view, Cardenas said, focusing on building the robot itself and developing custom-made parts where necessary. Its robots have 30 to 40 motors and a proprietary platform to control them.

That bet may have been the right one. There’s still no off-the-shelf humanoid robot, but its underlying software might soon be available as Google and other companies develop powerful, general-purpose models.

“The last piece of the puzzle was generative AI coming along,” Cardenas said. “Once we could build affordable humanoids, they weren’t actually versatile. They were still fairly rigid. It’s really evolved. And so now we have this last piece, which is intelligence.”

EnCharge AI announced a $100 million funding raise Thursday led by Tiger Global, as the tech company aims to develop a unique type of energy-efficient AI chip.

Co-founded by Princeton professor Naveen Verma and previously backed by the US Defense Advanced Research Projects Agency, EnCharge is tackling the ever-growing demand for energy required to keep AI data centers running.

Energy needs will amount to nearly 12% of the US’ power demand by 2030, up from just 3.7% in 2023, McKinsey projects. And demand will continue increasing as the industry moves from the “pre-training” phase, when raw AI models are created, to inference, a process that allows existing models to become better at reasoning and draw conclusions.

EnCharge is banking on a breakthrough in chip technology. Using new materials and techniques, the company has found a way to perform a big chunk of the power-hungry AI calculations inside the memory, which is typically just used to store information before it’s processed by the billions of transistors on the chip. It’s a much more efficient method, but this form of “analog computing” is incredibly difficult to create, which is why DARPA has been funding its development. The company says it leads to a 20x gain in efficiency compared to AI chips made by Nvidia and others.

While these chips could be useful in data centers and on “edge” devices like battery-powered electronics or autonomous cars, getting developer adoption has been a challenge for other makers of energy-efficient AI chips. Nvidia’s dominant Cuda software is used by almost everyone in the AI world. EnCharge hopes the efficiency gains will be so enticing that they’ll make the switch.

US President Donald Trump’s constant blitz of executive orders and actions is quickly becoming a defining feature of his second time in office — and his allies say that’s by design, telling Semafor’s Shelby Talcott and Burgess Everett they planned for this during both the transition and over the last four years, when Joe Biden was in office.

“It’s blitzing as much as we can until everyone is just tired,” one Trump aide said. But the “flood the zone” strategy, as Steve Bannon dubbed it, is disorienting for lawmakers. Democrats are once again reacting as Trump dominates the news cycle, and some Republicans are unnerved by the speed at which the president is operating.