In this edition, Google’s Demis Hassabis discusses building more cost-effective “light chips,” an em͏ ͏ ͏ ͏ ͏ ͏ |
| Reed Albergotti |
|
Hi, and welcome back to Semafor Tech. With Davos behind us, I’m reflecting on a profound conversation I had with Google DeepMind CEO Demis Hassabis at the conference this week that gave better insight into the ongoing AI race between large tech firms.
In any race, your strategy is based on the length of the course. If you’re running 100 meters, you train a lot differently than you would for 400 meters. But in the AI race, everybody has a different idea of where the finish line is — and it’s not here yet.
When ChatGPT first kicked off, Hassabis and Google made a couple of methodical calculations. First, text alone would not get you to the finish line faster. They made a bet that reasoning would come through multimodal training data. Another part of that strategy was focusing on increasing the “context window” of models, or the ability to process more data in an exchange.
That meant starting a little bit slower. The second was that raw compute power alone was not enough. Predicting that AI inference would become important, they bet that efficiency at the chip level was a key metric and began building custom silicon specifically for that purpose, as Hassabis detailed in his interview below.
Inference and AI capability are now tightly linked — a concept that was fuzzy until the last quarter of 2024. Therefore, if one company is able to run state-of-the-art AI models at a significant discount, they will have a massive advantage.
As we watch AI companies make major, multibillion-dollar moves in the coming months and years, a key question I’ll ask myself is: What does this say about where the finish line is in their minds?
Also below, the Gulf is facing its own AI race and Meta struggles to reach the finish line with its chatbot stating the wrong US president.
Carlos Barria/Reuters ➚ MOVE FAST: Sam. The $500 billion Stargate AI infrastructure project touted by Donald Trump will solely serve OpenAI, the FT reported. Despite public criticism by archrival Elon Musk, OpenAI’s chief hailed the Texas investment in terms the president would appreciate, describing it as “big. beautiful. buildings.” ➘ BREAK THINGS: Elon. Musk’s dunk of the Stargate initiative has drawn White House anger, according to Politico, which quoted unnamed Trump aides saying they were furious with the First Buddy. Still, the president said he wasn’t bothered, noting Musk “hates one of the people” in a nod to the entrepreneur’s feud with Altman. |
|
Ricky Carioti/Pool via Reuters We all make mistakes, chatbots included. This week, Meta is cleaning up its bot’s latest faux pas in naming the incorrect US president. When asked who the current commander in chief was, Meta AI answered Joe Biden, despite Donald Trump being sworn in on Monday, Reuters reported. “Everyone knows the President of the United States is Donald Trump,” a Meta spokesperson told Reuters. “All generative AI systems sometimes return outdated results, and we will continue to improve our features.” The error comes at a delicate time for Meta, whose CEO Mark Zuckerberg has cozied up to Trump and made moves that align with his views, like scrapping fact-checking on Facebook and Instagram. The blunder is also a reminder that even as Big Tech takes steps towards AGI, there is still a way to go in building accuracy and trust in these emerging systems. As of Friday morning, Meta AI was giving the correct answer, according to a Semafor test. |
|
TT News Agency/Pontus Lundahl via Reuters Google has found a cheaper way to run AI models, one of the tricks up its sleeve that could give it a long-term edge in the high-stakes race between the largest tech companies, DeepMind co-founder Demis Hassabis said in an interview with Semafor. For years, the compute power used in generative artificial intelligence was concentrated in the “pre-training” phase, when a raw AI model is initially created. But as models have evolved, the demands of running them — known as inference — have grown. If an AI model were a brain, inference would be akin to thinking. And it turns out, thinking longer can drastically increase the model’s capabilities. That means the compute power available to AI companies today isn’t sufficient to extract the full value of the technology. Hassabis said new processors, known as “light chips,” are in the works that could make it more cost-effective to run the models. “Sometimes you have the ‘victim of your own success’ problem,” Hassabis said. “If you build a very performant model, like [Google’s Gemini] 2.0 Flash, everyone wants it, which is great. But then suddenly, you only have a set amount of chips. You need more for serving.” He said the new Google chips are based on the same architecture as the company’s Tensor Processing Units, a custom-designed AI chip that the company has been working on for around a decade. |
|
Dado Ruvic/Reuters Silicon Valley-based payroll platform Deel has requested a Florida court dismiss charges that it facilitated money laundering transactions, contending the case is actually a hit job from its biggest competitor, likely indicating Rippling. Earlier this month, court-appointed receiver Melanie Damian filed a class action against Deel, alleging it processed payments without the proper licenses and enabled money laundering related to dealings with former customer Surge Capital Ventures. Surge is part of a separate US Securities and Exchange Commission action alleging a Ponzi scheme that defrauded church members out of $35 million, and Damian is tasked with recovering assets. Though the SEC has not accused Deel of any crime, Damian’s complaint against the company claims it facilitated at least $2.27 million in illegal payment transactions on behalf of Surge, as well as facilitating payments to Russia, violating US sanctions. Deel denied any wrongdoing and pointed to its biggest competitor, Rippling, calling the lawsuit “a coordinated effort by a major investor in Deel’s primary competitor seeking to tarnish Deel’s stellar reputation.” While Deel does not name Rippling in the lawsuit, that’s the likely competitor. The plaintiff’s lawyer, Thomas Grady, helped set up Waveling Insurance Services, now known as Rippling, according to documents filed to the Florida Department of State. He was reportedly an investor in the venture. |
|
Firebird Films Cloudflare CEO Matthew Prince thinks the conventional wisdom about the Gulf region’s AI prowess is upside down. Abundance of capital, cheap energy, and innovative regulation are often cited, in that order, as the top reasons why the biggest tech companies are flocking there. But Prince said it will actually be the Gulf’s approach to AI rules that will determine its place in the development of the technology. “My hunch is that when the history books are written about this 50 years from now, it’s going to be that experimentation around regulation, which will probably turn out to be the thing that is hugely beneficial,” Prince said at a Semafor Davos event. “Or if they do it wrong, what holds it back.” |
|
Andrew Kelly/Reuters Few CEOs know the double-edged nature of dealing with US President Donald Trump as viscerally as Pfizer’s Albert Bourla, Semafor’s Andrew Edgecliffe-Johnson writes. In an interview on the sidelines of the World Economic Forum in Davos, the pharma executive said he is optimistic about the new US administration but is also realistic about the risks that come with “radical change,” and discussed how he plans to work with Trump. |
|