Google Cloud drops an AI coding assistant; IBM likes the optics

Welcome to Runtime! Today: Google makes a flurry of year-end AI announcements, IBM brings fiber optics inside the data center, and the quote of the week.

(Was this email forwarded to you? Sign up here to get Runtime each week.)

Ship it

Jules, fix my code: Google waited until AWS re:Invent was done to drop several new enterprise AI services this week, including the agentastic Gemini 2.0 model. But it also jumped into the AI coding assistant market Wednesday with the introduction of Jules, "an AI-powered code agent" that looks a lot like GitHub Copilot.

Jules was designed to handle the grunt work of coding, such as fixing bugs, and Google said developers would be able to try it out in early 2025. “Jules might be late to the party but has the largest developer and code base to tap into and transform how they code not only for Android but also for AI-centric code bases such as Python and JavaScript,” said Neil Shah, partner and co-founder at Counterpoint Research, told InfoWorld.

Agents in spaaaaaace: Google wrapped up its AI week on Friday with the unveiling of Google Agentspace, which basically wants to be the home page for AI apps used inside businesses, likely ones that have adopted Google Workspace. "It unlocks enterprise expertise for employees with agents that bring together Gemini’s advanced reasoning, Google-quality search, and enterprise data, regardless of where it’s hosted," Google Cloud's Saurabh Tiwary, wrote in a blog post.

Google didn't say when anyone would be able to actually try out Agentspace, which it said would allow companies to search for information in corporate documents, upload the results to its NotebookLM tool, and launch internal AI agents to carry out tasks based on that data. A future version will allow customers to build their own custom AI agents using "a low-code visual tool."

Warp One: Fiber-optic cables are used all over the world to connect data centers but data moving around within data centers tends to travel over copper wires, which is starting to become more annoying given the immense computing power required to train AI models. IBM said Monday that it has developed a chip-interconnect technology that promises to move between chips at light speed.

In order to bring optical networking technology to chips themselves, the company developed a new polymer material that is more reliable and less expensive than glass, which is traditionally used inside fiber-optic cables. Should it be widely adopted, this technique could produce an 80x improvement in interconnect speeds between chips in data centers, IBM's Mukesh Khare told Network World.

Fake it until you make it: AI researchers trying to improve the performance of their models keep running into one persistent problem: There's just not enough real-world data out there anymore to keep feeding into the machine. For some applications, synthetic data can step in and save the day, and Databricks introduced a new API Monday that can make it easier to use ersatz data to improve the performance of AI agents.

Customers can upload real corporate data to Databricks' Mosaic AI Agent Evaluation tool and tell the API which questions and answers it wants the agent to generate. Databricks then uses that information to "generate evaluation sets tailored based on that proprietary data and your unique use cases," it said in a blog post.

Need for speed: While many businesses are content to use general-purpose LLMs from companies like OpenAI and Google, there are lots of others that are willing to invest the time and money it takes to train a custom model. But they still have budgets, and on Tuesday ServiceNow released an open-source software library that it thinks can speed up the process.

Fast-LLM does pretty much exactly what you might think: It "reduces the time needed to train a model so that researchers can complete more experiments and product engineering teams can get specialized models to market faster," ServiceNow's Joel Lamy Poirier wrote in a blog post. ServiceNow might not be in the model business itself, as its new triple-threat executive Amit Zavery told me last month, but the company does plan to release future versions of Fast-LLM to help customers work with those models.

As the year winds down, now is a great time to sponsor Runtime and get your message in front of the enterprise tech industry leaders and decision makers that are looking for new solutions in 2025. Extend your 2024 budget into next year: Book a weekly sponsorship of Runtime before 12/31 and get a second week free! See our rate card and formats here.

Stat of the week

Companies that use open-source software in some portion of their infrastructure (which is an enormous number of companies) face increasing threats from attackers bent on distributing malware disguised as legitimate updates to that software. The number of malicious packages in the wild rose 200% in 2024 compared to last year, according to Sonatype, and nearly two-thirds of attacks wielding that malware were targeted at governments.

Quote of the week

"In regulated industries, hallucination is a real big fear. They actually have something almost 99% there, and then they used to tell me the last one percent turns out to be the longest, because I can't afford to get this wrong." — AWS AI czar Swami Sivasubramanian, relating in an interview at re:Invent how a lot of enterprises are struggling to get their generative AI applications into production.

The Runtime roundup

The Biden administration wants to put new caps on sales of advanced GPUs to countries in Southeast Asia and the Middle East in hopes of preventing those chips from winding up in China, according to The Wall Street Journal.

AWS and Marvell signed a new partnership agreement that will see Marvell providing data-center chips to AWS while running its chip-design workloads on AWS's cloud services, according to DCD.

Thanks for reading — see you Tuesday!

This post was updated to correct the percentage increase in the number of malicious packages circulating as reported by Sonatype.