Google expanded its generative AI efforts this week with sneak peeks of new services coming for software developers on Google Cloud. But its large language model is still playing catch-up, according to experts.
The foundation for the rest of the updates unveiled at the Google I/O conference is the second version of the Pathways Language Model (PaLM 2). It underpins the new Duet AI for Google Cloud, a virtual assistant and pair programming tool that overlaps significantly with Microsoft's GitHub Copilot through a new API called Codey.
Codey was one of three new generative AI foundation models released to Google's trusted tester program this week. The others were Imagen, a text-to-image model, and Chirp, a speech-to-text foundation model. PaLM 2 is also now the basis for Google's ChatGPT search competitor, Bard, which became available without a waitlist for the first time.
Google has been training Bard to write software code over the last month or so, and execs demonstrated some of the results using Python on the Google I/O keynote stage. But one IT expert who has experimented with coding in Bard said PaLM 2 still has far to go to catch up with its chief rival, OpenAI's ChatGPT-4. A code-specific large language model (LLM) from OpenAI named Codex powers GitHub Copilot.
"Over the last few weeks, I have been facing off Bard and ChatGPT-4, and the results were incredibly lopsided in favor of ChatGPT," said Torsten Volk, an analyst at Enterprise Management Associates. "ChatGPT code worked right away 90% of the time, while Bard-generated code was much below 50%. I found the solutions provided by ChatGPT-4 a lot more creative compared to Bard."
Volk was also puzzled by Google's product strategy as it rolled out these new features, which positioned PaLM and its associated services for a broad range of users -- provided they are users of Google Cloud Platform (GCP). Duet AI will integrate with Microsoft's Visual Studio Code and JetBrains IDEs. While example workflows given by Google in keynote presentations and blog posts ranged from cloud engineers provisioning GCP resources to HR managers tracking travel requests, all were centered around GCP.
"Focusing on its 'home turf' means taking off some of the competitive pressure that has grown over the past few months, based on ChatGPT performing so much better than Bard in almost all areas," Volk said. "Carving out your own niche is the logical response when you get beaten in the open marketplace. But a company like Google needs to be able to go head-to-head with the other big players in this multi-cloud world."
Google bangs drum on privacy, model access
Despite the significant gap between Codey, still not yet generally available, and Copilot, which became generally available nearly a year ago, Google has clout in AI that has attracted some early generative AI partners, such as GitLab.
At Google I/O, company execs continued to emphasize the differentiators cited by GitLab in its partnership announcement: data privacy and direct access to training models. By comparison, the ChatGPT API, released in March, requires users to host data on OpenAI's cloud and doesn't offer access to the underlying model, although users can opt out of having their data used for general training.
"You have sole control of your data and the costs of using generative AI models," said Google Cloud CEO Thomas Kurian during a keynote presentation. "In other words, your data is your data and no one else's. You can also choose the best model for your specific needs across many sizes that have been optimized for cost, latency and quality."
Torsten VolkAnalyst, Enterprise Management Associates
Besides GitLab, companies such as Accenture, Box, Cognizant, Deloitte, KPMG and Snap have signed on to partner with Google on generative AI, according to a slide Kurian showed during the presentation.
Healthy competition between large IT vendors in generative AI is good for prospective customers, according to one software development pro.
"PaLM will have big adoption as companies start to think about a potential dependence on OpenAI," said Chris Riley, senior manager of developer relations at marketing tech firm HubSpot in Cambridge, Mass. "Alternatives are going to be critical to create your own proprietary models and reduce dependence as we hit the inevitable trough of disillusionment."
Being first to market isn't a total advantage for GitHub and OpenAI, either -- those companies and GitHub parent company Microsoft are being sued over copyrights on AI-generated code. At the same time, some enterprise IT organizations are concerned about potential licensing issues if AI-generated code gets embedded into proprietary apps.
These worries have at least partially frozen the enterprise market for generative AI tools until legal questions have been resolved, according to Michele Rosen, an analyst at IDC.
"My sense is that enterprise customers, particularly those in highly regulated industries, want vendors to solve the legal and data privacy issues before they adopt a coding assistant," Rosen said.
There's also plenty of room in the market for multiple generative AI models and tools, she said. Further generative AI competitors among large IT vendors are also emerging -- notably IBM, which launched its IBM Watsonx platform this week.
"We are definitely seeing developers using AI assistants -- we have survey data coming out on this very topic in the next few weeks," Rosen said. "More than half of the developers who responded to our survey have used at least one coding assistant, and many of them have tried multiple such tools."
Beth Pariseau, senior news writer at TechTarget, is an award-winning veteran of IT journalism. She can be reached at [email protected] or on Twitter @PariseauTT.