Technology

➚ MOVE FAST: Throw down. Usually staid Intel is stirring things up in the AI chip battles. CEO Pat Gelsinger is taking on Nvidia’s dominance in the industry, describing his rival’s competitive moat as “shallow and small.” Most Nvidia challengers have been a bit more subtle.

➘ BREAK THINGS: Slow down. AI-powered Cruise, the first company to offer paid driverless taxi rides in San Francisco, is laying off 900 workers and revamping the firm, TechCrunch reported. Cruise ran into regulatory issues after a much-publicized accident, though its cars are probably still safer than human-driven taxis.

It sounds like something Elon Musk might name his offspring, but “8x7B” is actually a new large language model from French startup Mistral and it’s creating buzz in the AI world. In various industry benchmarks, it’s about as good as OpenAI’s GPT-3.5, which was the underlying model used by ChatGPT when it launched. Nicknamed “Mixtral,” this new model is open-source, free, and small enough that it could run on a powerful laptop.

The figure most often cited to describe the size of an AI model is the number of parameters used to train it. Mixtral has 46.7 billion parameters, but it employs a trick called (and this is where it gets its nickname) “mixture of experts” to effectively lower the compute power needed to run the model.

For each token generated, only about 12.9 billion parameters are used. It’s estimated that GPT-3.5 has anywhere from 175 billion to 380 billion parameters and does not use this “mixture of experts” technique. In other words, Mixtral is much more economical to run and is super capable, but also it’s free and could run on a souped-up laptop. It shows how quickly the capabilities of AI are increasing and the power needed to run them is decreasing.

There’s a kind of Moore’s Law at work. But AI is very different from semiconductors. When chips were new, in the late 1950s, the private sector didn’t need them. There was basically one customer: The U.S. government, which used them in the space program and elsewhere. The private sector later bought older chips when the costs came down.

So here’s an interesting question for the AI industry today: What customer is willing to pay more for the best models, and can’t wait for the costs to come down later? And is it worth it for companies like OpenAI to spend billions to create cutting-edge models if they can be quickly duplicated, are free, and cheap to run?

Moore’s Law was driven in part by the difficulties in copying semiconductor manufacturing, and that basically holds true today. But right now, it seems there’s no moat that can stop this French AI invasion.

— Reed

Inside the hottest event for recruiting AI talent

THE SCENE

Chinese tech companies and Wall Street trading firms were among the most prominent participants at the hottest annual machine learning conference this week, when more than 10,000 of the world’s top artificial intelligence researchers gathered in New Orleans.

Their presence shows how intense the competition for AI talent has become, with many PhD students making the trip specifically in the hopes of landing a job at a place like Google or OpenAI, where they could potentially earn upwards of $500,000 a year.

Several attendees at the Conference on Neural Information Processing Systems, or NeurIPS, told Semafor they arrived in Louisiana knowing their skills were more in demand than ever, thanks to ChatGPT and the broader frenzy over generative AI tools. The chatbot and its competitors have proven that open-ended machine learning research can lead to lucrative breakthroughs, giving companies a compelling reason to pour more resources into it.

That’s one reason why Chinese tech companies were so visible at the event, despite growing tensions between Washington and Beijing over the development of advanced technology, especially artificial intelligence. Ant Group, an Alibaba subsidiary responsible for the payment platform Alipay, had a large booth advertising its latest tech, though the word “China” didn’t appear anywhere.

A recruiter from Tencent hung up flyers around the conference advertising jobs working on “fundamental AI techniques” that would be applied to video games. The ads were written in both English and Mandarin, and instructed interested applicants to reach out on WeChat. Newer Chinese companies were there too, including Zhipu, a startup working on large language models that has raised over $340 million this year.

Tech firms like Amazon and TikTok, which one California professor said both pay among the highest starting salaries to newly minted PhDs, competed for talent with another industry: finance. Trading shops like Citadel had big booths at NeurIPS too, where they pitched AI researchers on using their skills to develop trading algorithms. Jane Street Capital boasted that new employees would be helping to “keep financial markets transparent and efficient.” (Not to mention making enormous profits.)

Semafor’s 2024 World Economy Summit, on April 17-18, will feature conversations with global policymakers and power brokers in Washington, against the backdrop of the IMF and World Bank meetings.

Chaired by former U.S Commerce Secretary Penny Pritzker and Carlyle Group co-founder David Rubenstein, and in partnership with BCG, the summit will feature 150 speakers across two days and three different stages. Join Reed, Louise, and other Semafor colleagues for conversations with the people shaping global technological shifts.

Moderna and Merck announced Thursday that they’re one step closer to an AI-enabled skin cancer vaccine. Results from a three-year study showed the vaccine prevented recurrence or death in high-risk melanoma patients by 49%. That shows it has some staying power.

The treatment uses individualized neo-antigen therapy, and is based on the same mRNA technology that powered some Covid vaccines. It employs AI to determine which cells are the problem in patients who have been treated for skin cancer. Then, a personalized vaccine is created to train the body to go after only those specific cells. Moderna and Merck hope they can be the first to get FDA approval for a treatment in this space.

Silicon Valley bristled when Andreessen Horowitz announced it was putting $350 million into a new startup founded by Adam Neumann, the co-founder of WeWork, which embarrassed the tech industry when it collapsed.

His new company, called Flow, is a real estate play that some speculated was tied to crypto. But according to Business Insider, which was granted a media tour of a Flow-owned complex, it’s not really a tech startup. Flow is basically taking over apartment buildings and sprucing them up with Millennial amenities, like snappy mobile apps and fitness classes.

A key detail is that Flow owns the buildings. That’s a less risky bet than WeWork, which committed to long-term leases. Neumann, for all his faults, made co-working spaces cool. WeWork went bankrupt, but it left behind many copycats. From speaking with people at a16z, I think the hope is that Neumann has learned to execute better at turning his vision into an actual business.

If Flow can build up a critical mass of desirable apartment complexes for young professionals, there are a lot of interesting things it could do to monetize that customer base. If that bet fails, at least the real estate assets will still be there.

— Reed

• The White House believes it is closing in on a border-Ukraine deal. But many Democrats are already outraged that Biden may accept restrictive immigration reforms in exchange for Ukraine funding.
• Ghana’s goal for 2024: Go from containing Malaria to eliminating it. A newly-approved vaccine has raised hopes of stopping local transmission altogether.
• Beijing makes plans to revive China’s flagging economy. The country has struggled with weak demand, a deepening housing crisis, and widespread youth unemployment.