Deepmind is now an alpha coder
Image: Boris SV/Moment/Getty Images

Deepmind is now an alpha coder

Protocol Enterprise

Hello and welcome to Protocol Enterprise! Today: Google’s Deepmind rolls out a new AI-powered programming assistant, npm enhances its link in the software supply chain and new ideas for the complicated world of enterprise billing.

Spin up

Google’s decision to get a little thriftier with its cloud infrastructure replacement cycles paid off in a big way during 2021. By extending the expected life of its servers from three years to four years, and the life of its networking equipment to five years, Google saved $2 billion over the course of the year, it said during yesterday’s earnings announcements.

Shallow code from Deepmind

The brains at Google’s Deepmind are known throughout the world for their research advancements in deep learning and neural networks. Nowadays, as AI techniques that were once off-limits to anyone without a computer science PhD are slowly becoming more accessible, the latest project from the team at its U.K. research lab involves a way to ease the burden of writing code for performing simple tasks.

  • While Deepmind hasn’t necessarily cracked the code on democratizing the complex programming involved in crafting a neural network, the new system uses transformer-based language models to automatically write code, then spin it into programs that can win in a coding contest.
  • Deemed AlphaCode, the system writes computer programs at a competitive level that Deepmind says is as good as the top 54% of human participants in programming competitions on Codeforces, a platform hosting speedy-fingered coding battles.
  • Deepmind has scored accolades in the past when using games to illustrate the power of its algorithmic creations: The research outfit built AlphaGo, the AI model that beat human champ Lee Sedol in 2016 in a tense match of the ancient game Go.

Building AI that works to solve real-world problems often entails more human creativity than what winning a game might involve, said Amir Kashani, director of AI and Digital Products for Stanley Black and Decker’s industrial division, and a veteran of college programming competitions.

  • Coding contests “don’t show up that often in everyday programming jobs beyond the job interview day for most programmers, but come up often for those doing academic research work to push the boundaries of computer science,” he said.

Here’s how Deepmind built AlphaCode:

  • The company pre-trained a model using publicly available code from GitHub, then fine-tuned it on a small set of coding contest data.
  • Then for each contest problem, they built “a massive amount of C++ and Python programs.”
  • From there, they used debugging and testing to automatically filter those programs to find 10 worth assessing externally and possibly submitting.

Software engineers said the new automated coding model has limited value:

  • “The primary focus of most programmers is assembling pieces of infrastructure and utilizing frameworks and libraries to deliver workable solutions to the business or organization,” said Marty Nelson, CEO of Alchemy Code Lab, a software developer training school.
  • Writing small-scale algorithms is just a fraction of the process, he said. “If AI solutions like this could also be trained for highly comprehensible code I could see them being a useful piece in the programmer toolbox, but they won't be doing the whole job.”
  • Solving coding challenges isn’t easy, said Jeevan Kalanithi, co-founder and CEO of OpenSpace, which uses computer vision AI to visualize and archive work on construction sites. Still, he said, “It’s also something kind of rote.”
  • The sort of automated coding enabled by Deepmind’s new system can help alleviate some menial tasks for programmers but, “I don’t think it means that we can just automate software engineers out of existence,” Kalanithi said.

It’s also worth noting that Microsoft’s GitHub is working on a similar project called Copilot that acts as a coding assistant to a human developer using publicly available code to train an AI system. Life could be much easier for future generations of cloud software developers.

— Kate Kaye (email| twitter)


Dataiku is the only AI platform that connects data and doers, enabling anyone to transform data into real business results — from the mundane to the moonshot. Because AI can do so much, but there's no soul in the machine, only in front of it. Without you, it's just data.

Learn more

npm tries to seal its packages

There’s more than one way to secure the software supply chains that run the world, as yesterday’s announcement from OpenSSF showed.

GitHub’s npm, which manages dozens of popular software library packages used by developers around the world, now requires maintainers of those packages to use two-factor authentication when logging into their npm accounts. The new policy will first apply to the maintainers of the top 100 projects on npm but will eventually roll out to all users, The Record reported.

Software libraries are bits of code that developers like to reuse as part of their own software, and are made available to them via packages stored at repositories like npm’s. But that popularity and ease of use makes those packages big targets for hackers to insert malware into popular packages and gain control of remote systems once they are downloaded and installed.

The move should give developers a little more peace of mind that the libraries they need to build their applications contain exactly what they say they do, and nothing more.

— Tom Krazit (email | twitter)

When the bill is due

Enterprise billing is notoriously complex: It encompasses everything from taxes and chargebacks to refunds and credits. Throw in popular SaaS-based pricing models like seats, subscriptions and consumption-based usage, and you have quite an infrastructure challenge.

Metronome is one startup trying to tackle the messy world of scalable and flexible billing systems. Flush with cash after a $30 million fundraising round led by a16z, Metronome wants to build the underlying infrastructure for enterprise billing data.

The startup’s technology is designed to let businesses experiment with billing models without needing to rewrite code. This ability to flex your billing structure at will could be a competitive advantage for the next generation of SaaS companies.

While the potential is great, billing still remains unsolved. It’s an area we’ll be watching closely in 2022.

— Aisha Counts (email | twitter)

Around the enterprise

VMware will go ahead with the hire of a former AWS executive who left the company after Protocol reported on allegations of discrimination and harassment against the executive, but in an individual contributor role.

Coupa introduced a new AI-powered buying system for purchasing managersthat promises to make sourcing deals and pricing information much easier.

JetBlue announced plans to use Salesforce’s sustainability software to track its climate impact, as the market for software that can help companies measure that impact surges.

The Department of Defense said the new role of chief digital and artificial intelligence officerhad achieved “initial operating capability,” as only a military memo would put it.


Dataiku is the only AI platform that connects data and doers, enabling anyone to transform data into real business results — from the mundane to the moonshot. Because AI can do so much, but there's no soul in the machine, only in front of it. Without you, it's just data.

Learn more

Thanks for reading — see you tomorrow!

Recent Issues