Llion Jones had a big role at Google, where he worked for almost 12 years. He was one of eight authors of the pivotal Transformers research paper, which is central to the latest in generative artificial intelligence.
However, like all of his co-authors, Jones has now left Google. He’s joining fellow ex-Google researcher David Ha to build a generative AI research lab in Tokyo called Sakana AI. Jones said that while he has no ill will toward Google, he realized that the company’s size was keeping him from doing the kind of work he wanted to pursue.
“It’s just a side effect of big company-itis,” Jones told CNBC in an interview. “I think the bureaucracy had built to the point where I just felt like I couldn’t get anything done.”
Jones, who studied AI in college and has a masters in advanced computer science from the University of Birmingham, is at the center of the action. The 2017 paper he helped write at Google laid out innovations that played into OpenAI’s creation of the viral chatbot ChatGPT. The T stands for Transformers, an architecture behind much of today’s frenetic generative AI activity.
“We’re kind of crazy,” Jones said. “We’re looking at nature-inspired methods to see if we can find a different way of doing things, rather than doing a huge, humongous model.” Sakana isn’t announcing any investors.
Jones became a software engineer at Google’s YouTube in 2012. According to his LinkedIn profile, he started “researching machine intelligence and natural language understanding” at Google in 2015.
Google is one of a number of large tech companies that hired hordes of researchers in recent years, some straight from universities, to construct AI models aimed at enriching their products. Over time, Jones said he encountered questions about why the software was malfunctioning and whose fault it was. He found it all to be a distraction from the research.
“Every day I would be spending my time trying to get access to resources, trying to get access to data,” Jones said.
Now, after many years building products in labs, Google is rushing to incorporate generative AI, including large language models (LLMs), into its search engine, YouTube and other products. The models can summarize information and come up with human-like responses to written questions.
In Jones’ view, Google is focusing “the entire company around this one technology,” and innovation is more challenging “because that’s quite a restrictive framework,” he said.
Ha said he and Jones have spoken with others who want to work on LLMs, but they haven’t finalized their plans.
“I would be surprised if language models were not part of the future,” said Ha, who left Google last year to be head of research at startup Stability AI. He said he doesn’t want Sakana to just be another company with an LLM.
Both Jones and Ha have unflattering things to say about OpenAI, which has brought the concept of generative AI to the mainstream but raised billions of dollars from Microsoft and other investors to do so. Ha described it as “becoming so big and a bit bureaucratic,” no different really than groups within Google.
Jones said he doesn’t think OpenAI is all that innovative. He said that for OpenAI’s two biggest successes, ChatGPT and the DALL-E service for creating images with a few words of text, the startup took research he performed at Google and applied it on a large scale, making refinements along the way but holding off on sharing the developments with the community. While OpenAI has released neither of the technologies under an open-source license, it has published papers on some of the underlying systems.
Representatives from Google and OpenAI didn’t respond to requests for comment.
Ha said Sakana has brought on a part-time researcher from academia, and the company will eventually hire more people. Asked if they’ve added any other Google employees, Ha said, “Not yet.”
China’s artificial intelligencedevice market is already booming, and in the advanced technology race against the U.S., the country’s expertise in hardware could give it an edge.
“The advantage comes from the fundamental root that China is a nation of manufacturing,” Dr. Kai-Fu Lee, CEO of 01.AI and chairman of Sinovation Ventures, told CNBC. “Today, the competition is on the software, the models, the agents, the applications. But soon it will move to devices.”
Meta has sold millions of its smart glasses since introducing the specs in 2023, and the Chinese have caught on, with more than 70 Chinese companies creating competing products in the space.
Eyewear from companies such as Inmo and Rokid are sold worldwide. Xiaomi and Alibaba‘s are found only in China and are embedded with the tech giants’ own AI.
Alibaba’s DingTalk, a messaging platform for the workplace, this year released a credit card-sized AI gizmo meant for note-taking on the job.
The DingTalk A1 can record, transcribe, summarize and analyze speech from as far as 8 meters (26 feet) away, about the length of a large boardroom.
The device is similar to the Plaud Note, which is available in the U.S.
The device experimentation in China spans from the practical to the unconventional.
Chinese startup Le Le Gaoshang Education Technology released a “Native Language Star” brand translating gadget aimed at Chinese parents with limited English to teach English to their own children.
Read more CNBC tech news
The contraption, which is looped around the back of a user’s neck like a travel neck pillow and comes down toward the chest, has a sort of muzzle unit that goes over the mouth and mutes the user’s own voice.
The unit is embedded with Tencent and iFlyTek AI and is billed as a way to turn an English-speaking Chinese parent into a “laowai,” or foreigner. It retails for $420.
Having so many hardware touchpoints helps with adoption and with getting people used to the technology. It’s also a boost for companies to gather a war chest of data compared to other countries, analysts say.
“When you still hear people outside of China talking about what the future of the AI device might be, the market is full of AI devices here already,” tech consultant Tom van Dillen of Greenkern said at his office in Beijing. “This creates this feedback loop again to make the AI even better.”
Yet an edge in hardware is far from a guarantee to win the AI race, especially if China’s AI lacks appeal with global customers due to privacy or other issues, or if it falls well behind its counterparts in the U.S. or elsewhere.
“You really have to be that Apple iPhone to reap the most of the reward,” Lee cautioned, referencing late entrepreneur Steve Jobs’ invention that is often seen as one of the most transformative consumer products ever. “I think the China advantage for building the Apple iPhone for the AI age is that the capabilities are there — engineers and entrepreneurs, and so on. But it will still be a race.”
President Donald Trump on Monday said Nvidia will be allowed to ship its H200 artificial intelligence chips to “approved customers” in China and elsewhere, on the condition that the U.S. gets a 25% cut.
The policy “will support American Jobs, strengthen U.S. Manufacturing, and benefit American Taxpayers,” Trump wrote.
“The Department of Commerce is finalizing the details, and the same approach will apply to AMD, Intel, and other GREAT American Companies,” he added in the post.
The H200 is a higher-grade chip than the H20, but not the company’s top-of-the-line product.
Nvidia shares climbed earlier Monday on news that the Commerce Department was set to approve the China sales, but later pared those gains. The stock rose about 2% after hours.
Stock Chart IconStock chart icon
Nvidia (NVDA) and Advanced Micro Devices (AMD) stock prices
“We applaud President Trump’s decision to allow America’s chip industry to compete to support high paying jobs and manufacturing in America,” a spokesman from Nvidia told CNBC in a statement.
“Offering H200 to approved commercial customers, vetted by the Department of Commerce, strikes a thoughtful balance that is great for America,” the spokesman said.
Semiconductors, which are key components in nearly every category of electronics, are at the center of the AI race between the U.S. and China.
They have also played a role in the tumultuous trade relationship between the two economic superpowers.
Read more CNBC tech news
When Beijing imposed export controls on rare-earth minerals, which are used in the production of some high-end chips, the Trump administration threatened to massively increase tariffs on U.S. imports from China.
After meeting in South Korea in late October, Trump and Xi struck a tentative trade truce in which China committed to end “retaliation” against U.S. chipmakers, according to the White House.
Trump said after that meeting that he discussed the export of Nvidia chips with Xi.
Broadcom shares hit an all-time high during Monday’s trading session after the emergence of another encouraging sign that the company’s custom chips are all the rage on the AI scene. The newest development comes from the tech website, The Information, which said Microsoft could be looking to move its custom chip business from Marvell Technology to Broadcom. The report is the latest in a string of recent good news for Broadcom, which delivers quarterly earnings after Thursday’s close. Shares of Marvell were understandably falling more than 7%. Also weighing on Marvell stock was a note from Benchmark, in which the analysts call out with a “high degree of conviction” that Amazon may also be looking to move the development of future generations of its Trainium chips away from Marvell to AIchip, a Taiwanese designer. Taken together, Broadcom shareholders should feel good about the company’s standing in the custom AI market, as specialized silicon emerges as a competitor to Nvidia’s all-purpose AI chips, which have been the gold standard in running artificial intelligence workloads. At the same time, the weakening position of Marvell amplifies Broadcom as the go-to company for custom chips. The Information report, as it relates to Microsoft, comes after the success of Google’s tensor processing units, which were co-developed by Broadcom. The TPUs have been praised in recent weeks following the release of Gemini 3, the latest large language model from Alphabet ‘s Google. Gemini 3, which has leapt to the top of the app leaderboards, was trained and runs entirely on Google’s custom TPUs. A couple of weeks ago, The Information reported that Meta Platforms was thinking about using Google’s TPUs for its data centers in 2027. AVGO YTD mountain Broadcom YTD While it’s great to watch Broadcom’s share price climb, we don’t love it when a stock runs into an earnings release, as it indicates high expectations. We do understand the move, though, because all this news has made it clear that Broadcom’s custom silicon business is primed for further gains. We don’t expect to hear much about these latest two developments on the post-earnings call. We do, however, suspect that talk about custom chip demand will center around the interest Broadcom has been seeing following the launch of Gemini 3. Outside of custom chips, there will be high interest in Broadcom’s networking business, which has seen incredible growth over the past year, given the increased need for high-bandwidth networking solutions resulting from the explosion of AI adoption — especially with the introduction of reasoning models and agentic solutions. On the legacy front, we expect to see some gradual improvement, thanks in part to seasonality as the company’s wireless revenues are tightly linked with the iPhone sales cycle, given that Apple is the company’s primary wireless customer. As for software, we continue to expect strong growth and margin performance driven by VMware, and we will be interested to hear about any additional synergy and cross-selling opportunities the team has been working on. (Jim Cramer’s Charitable Trust is long AVGO, MSFT, AMZN, META. See here for a full list of the stocks.) As a subscriber to the CNBC Investing Club with Jim Cramer, you will receive a trade alert before Jim makes a trade. Jim waits 45 minutes after sending a trade alert before buying or selling a stock in his charitable trust’s portfolio. If Jim has talked about a stock on CNBC TV, he waits 72 hours after issuing the trade alert before executing the trade. THE ABOVE INVESTING CLUB INFORMATION IS SUBJECT TO OUR TERMS AND CONDITIONS AND PRIVACY POLICY , TOGETHER WITH OUR DISCLAIMER . NO FIDUCIARY OBLIGATION OR DUTY EXISTS, OR IS CREATED, BY VIRTUE OF YOUR RECEIPT OF ANY INFORMATION PROVIDED IN CONNECTION WITH THE INVESTING CLUB. NO SPECIFIC OUTCOME OR PROFIT IS GUARANTEED.