Fortune Inscrutability

Ubisoft CEO Yves Guillemot at the company's headquarters in Paris, France, on Jan. 18, 2023. (Photo: Nathan Laine/Bloomberg/Getty Images)

Finally an answer to, “What will happen to Ubisoft?”

The French video game publisher said Thursday that it will carve out a unit that includes the hit franchises Assassin’s Creed, Far Cry, and Tom Clancy’s Rainbow Six into a subsidiary worth about $4.3 billion.

China’s Tencent, which holds a 10% stake in Ubisoft and was reported in October to be exploring a buyout of the company, will buy a 25% stake in the new entity for about $1.25 billion, giving Ubisoft a fresh injection of cash (to pay down debts) and raising its value.

“We will be able to firmly take back control of our future and renew with the serenity needed to create the best games,” CEO Yves Guillemot said in a company memo, adding that the changes will put the company back on track “to a higher level of profitability.”

It’s been tough sledding lately for Ubisoft, whose Star Wars Outlaws last year failed to impress and whose Assassin’s Creed Shadows was twice delayed before its launch a week ago. The company’s revenues have been on the decline and its ratio of debt to EBITDA is substantial, depressing its stock price to record lows.

Ubisoft had been exploring strategic options since January, and hinted in yesterday’s announcement that it had multiple suitors leading up to this deal. It’s expected to close by the end of the year. —AN

Researchers at the AI company Anthropic say they have made a fundamental breakthrough in our understanding of exactly how large language models work. 

One of the problems with today’s powerful AI is that the LLMs are black boxes. We can know what prompts we feed them and what output they produce, but exactly how they arrive at any particular response is a mystery, even to the AI researchers who build them.

This inscrutability creates all kinds of issues. It’s difficult to predict when a model is likely to spew erroneous information, or understand why some models are susceptible to jumping the guardrails intended to curb bad behavior, or control models that will freely deceive users. 

These concerns ultimately make some businesses hesitant to use the technology. 

Anthropic’s new research offers a pathway to solve at least some of these problems. The company’s scientists created a new tool for deciphering how LLM’s “think,” then applied it to Anthropic’s Claude 3.5 Haiku model. 

The researchers found that although LLMs like Claude are initially trained to just predict the next word in a sentence, in the process Claude does learn to do some longer-range planning, at least when it comes to certain kinds of tasks. 

They also found that Claude, which is trained to be multilingual, doesn’t have completely separate components for “reasoning” in each language. Common concepts are considered in the same set of neurons within the model, then converted to the appropriate language.

The researchers also discovered that Claude is capable of lying about its chain of thought in order to please a user. 

The discoveries open new possibilities for auditing AI systems for security and safety concerns and may help researchers develop new training methods to improve a model’s interactions and reduce its faulty outputs. —Jeremy Kahn

Pat Gelsinger is certainly having a post-Intel moment.

Fresh from announcing his new role as exec chair and tech chief at Gloo earlier this week, the former Intel CEO has become a general partner at early-stage VC firm Playground Global and joined the board of Playground portfolio company xLight, which makes lasers for chipmakers.

Gelsinger is also sounding off about the chip industry—specifically, the push to onshore chip manufacturing in the U.S.

TSMC’s $100 billion pledge for advanced chipmaking in the country won’t make the U.S. a giant in the field, he told the Financial Times, because the company is still doing the important stuff back in Taiwan.

“All of the R&D work of TSMC is in Taiwan, and they haven’t made any announcements to move that,” he said. “Unless you’re designing the next-generation transistor technology in the U.S., you do not have leadership in the U.S.”

Interestingly, Gelsinger said he didn’t see China’s DeepSeek as a serious threat to U.S. AI firms because it “was good engineering” rather than “core innovations.”

And he thinks more core innovations will be needed in general, if AI is to truly take off: “AI, as exciting as it is, is much too expensive. We have to have dramatic reductions in the cost of inference for it to be truly deployed in every aspect of humanity.” —David Meyer