Racks of Cerebras equipment
Photo: Cerebras

Putting the AI in the CIA

Protocol Enterprise

Hello and welcome to Protocol Enterprise! Today: why a new chip from Cerebras has a lot of interesting potential, what developers really think about the tools of their trade, and how Mandiant expects to evolve under Google Cloud.

The CIA wants cheap AI training

Training AI is hard and expensive enough that the CIA has backed a way to make it cheaper.

Through its nonprofit investment operation called In-Q-Tel — which is also funded by several other members of the U.S. intelligence community — the agency has bet that a Silicon Valley startup called Cerebras can make a dinner-plate-sized chip that can drastically reduce the cost of training AI models.

  • AI models are methods of organizing mathematical calculations by breaking them into steps and then setting rules on how the steps talk to one another.
  • But to make AI models better, they get bigger, and then, of course, more expensive to train because they need many more chips to do so efficiently.

Cerebras has set out to change that. Wednesday, it announced it hit something of an important roadside marker on that trip: CEO Andrew Feldman said the company’s engineers had managed to train a 20 billion parameter AI model on its dinner-plate-sized chip.

  • Fitting an AI model on a single chip offers distinct performance advantages for training models that can reduce the time required from weeks down to days, which also makes it cheaper.
  • Cerebras calls the superchips the Wafer Scale Engine, and the idea behind fitting a 20 billion parameter model onto a single one is to make the cost of training AI cheap enough for just about anyone to do it.
  • The cheapest option appears to be $25,000 a week for a Cerebras system that the WSE-2 powers, though customers can rent it for a few hours from a cloud provider if that’s what they require.

Cerebras is backed by the CIA — sort of. Feldman told Protocol that In-Q-Tel is one of the company’s backers but it has likely received the bulk of the funds from regular VCs.

  • In-Q-Tel invests on behalf of several spy agencies, but doesn’t usually drop a ton of cash on each bet.
  • Cerebras has raised $735 million that pegs it at a valuation of $4.1 billion.
  • VCs became interested in AI chip startups when it became clear general purpose computing offered by the likes of Nvidia, AMD and Intel might need some help with specialized chips.

Read the full story here.


— Max A. Cherney (email | twitter)

SPONSORED CONTENT FROM SAP

The competitive edge of digital solutions

For the last 50 years, SAP has worked closely with our customers to solve some of the world’s most intricate problems. We have also seen, and have been a part of, rapid accelerations in technology in response. Across industries, certain paths have emerged to help businesses manage the unexpected challenges over the last few years.

Click here to read more from SAP

Developers speak up

From programming languages to cloud platforms, one of the best places to track enterprise technology trends is the annual Stack Overflow Developer Survey, which had over 70,000 responses this year.

Like it’s been for the past nine years, the most commonly used programming language is still JavaScript. The decades-old language managed to beat out its competitors by a significant margin: 65% compared to 55% for HTML/CSS and a tad over 48% for both Python and SQL. When it comes to new developers, however, HTML/CSS and Python are just as popular as JavaScript.

Across the board, Python and TypeScript are two programming languages that seem to be making serious inroads with developers. Python and TypeScript are two of the most common languages developers want to use — if they aren’t already. Rust is still the most beloved by developers though, with 86% of developers planning to continue using the language.

In the cloud development world, AWS is the most-used platform, with 51% of developers doing extensive work there. Microsoft Azure and Google Cloud consistently fight for second place in the hearts and minds of developers, and this year Azure came out ahead with 29% of developers building on the platform. For new developers though, the story was different: Google Cloud edged out AWS and Azure, but still lost the first place spot to Heroku.

That’s a key trend revealed by the survey: Although legacy programming languages and cloud platforms have a grip on professional developers, there’s still room for other providers to win over new developers.


— Aisha Counts (email | twitter)

Mandiant’s mandate

Google's $5.4 billion deal to acquire cybersecurity powerhouse Mandiant could shake up the industry. At the RSA Conference this month in San Francisco, I sat down with Mandiant CTO Marshall Heilman to discuss what the merger — expected to close later in 2022 — could change for the company and for Google Cloud.

Assuming the acquisition closes, how will Google Cloud customers benefit?

When we sell an integrated solution, with someone else's technology [plus] our intelligence, it doesn't really matter whether it's a Google-specific customer or a Microsoft-specific customer. They're [all] benefiting from it. Now, Google customers may benefit a little bit more over time — because we are natively going to get more experience in GCP [Google Cloud Platform].

What's something significant that will change for both sides?

I think we're going to try and take the best of the Google culture, and what's made them so successful, and try to bring that to Mandiant — while at the same time, take Mandiant's approach to how we handle enterprises and how we work in the security world, and bring that to Google. What I'm hoping, personally, is that we blend the two environments.

What other benefits could Mandiant get from joining with Google?

If you look at our intelligence, Google has a ton more data than we do. That'd be fantastic [to get access to]. I'm sure there are some things we can learn from [Google's] engineering. I think from a consulting perspective, I know Google doesn't really have consulting services the way that we do, in the security aspects. So that's some expertise that we can bring to them.


— Kyle Alspach (email | twitter)

Around the enterprise

Cisco disclosed several software vulnerabilities in its security tools, which happens to lots of security companies but isn’t the best look.

Qualcomm released a unified software package that should help developers write AI applications for its chips.

SPONSORED CONTENT FROM SAP

The competitive edge of digital solutions

When companies invest in maintaining their “green ledger” with the same commitment they have to their financial ledgers, they will be able to connect their environmental, social, and financial data holistically so they can steer their business towards sustainability. At the end of the day, what gets measured, gets managed.

Click here to read more from SAP

Thanks for reading — see you tomorrow!

Recent Issues

The SEC vs. CISA