GenAI Thought Leadership

In today’s fast-changing tech landscape, one of the biggest challenges is navigating the rise of generative AI. We researched on how generative AI is transforming the way organizations write and document code. Then we sat down with Peter Guagenti, President and Chief Marketing Officer of GenAI code development platform Tabnine, in a two-part webinar to deep-dive into the technical architecture behind tools like Tabnine. Below is part 1 and check back for part 2.?

ETR hosted a Generative AI Thought Leadership dinner in partnership with Guagenti. Many organizations struggle with relentlessly high growth in the number of applications they create and must maintain. The complexity of those applications and the surrounding pool of technical debt has also dramatically increased over time. In addition, the number of open software development roles continues to outnumber the available talent dramatically. As a result, software engineering teams are struggling to keep up, and although AI is emerging as a potential solution, significant fears, uncertainties, and doubts around the nascent technology remain.

As technology and business facilitators, we need to help development teams of every size use AI to accelerate and simplify the software development process without sacrificing privacy and security. All of this must be done with Privacy, Protection, and Personalization in mind. As one of the originators of the AI Coding Assistant space, Tabnine?was in a unique position to help us explore the complexities and opportunities we face at this dinner.

ETR’s Thought Leadership dinner with Tabnine, a pioneer in AI code generation, explores AI's impact on productivity, data privacy, contextual understanding, code validation, and IT security. We also discussed how companies can leverage AI to identify and correct errors and security vulnerabilities, support legacy technology, and measure ROI, all with an emphasis data privacy and regulatory compliance. Tabnine is poised to streamline project onboarding, improve developer workflows, and ultimately auto-generate functional applications from plain language requests, which will anoint a new era in software engineering.

Vendors Mentioned: Anthropic / Cohere / Elastic / Microsoft / Mistral / OpenAI / Oracle / Pinecone / Redis / Tabnine

Overview

Tabnine. Peter Guagenti, President of Tabnine, emphasizes that despite popular misconceptions, AI’s capacity to automate software development did not emerge overnight. “Tabnine launched the very first AI code assistant in 2018, an AI code assistant built on a very basic LLM for Java. It’s even kind of stretch to call it an LLM; it was a small language model, as we think now about them.” Language models excel in highly structured, clear, and well-documented languages—like software. Initially adopted by individual developers for personal use to enhance productivity, AI code assistants soon “exploded,” swiftly becoming mainstream. “We see them truly in every step of the life cycle.”

With a million monthly active users and over three million installations, Tabnine is second only to Microsoft Copilot. “Thousands of organizations use it. We only started selling an enterprise product about a year ago in response to market demand.” Tabnine expects AI-authored code to be as big or bigger than the transition from?

monolith to microservices, or from on-prem to cloud, in particular for project onboarding. “It explains all of the functions, how it's structured, what it's integrated with, the APIs it calls, and how it works. If you forget that, you just hit the onboarding agent again and it explains it to you.” AI is transforming developer workflows by automating substantial portions of the process, from documentation to code generation, with 80% of documentation and over 80% of unit tests now being generated autonomously.

Ultimately, Tabnine imagines a streamlined three-phase process: requirements management, code generation, and code validation, to the point of code generation prompted directly from project management tools. “The SDLC is going to shift dramatically. We're seeing folks like Atlassian and others really getting into, how do we get into ideation and requirements management in a more thoughtful way? I have an AI agent right now that can take a JIRA ticket and just deploy an application.”

Within security, by setting plain language expectations for vulnerabilities and behaviors, AI will autonomously verify code against these standards, eliminating the need for traditional scanners. AI will further autonomously identify and correct errors by analyzing observability data and code changes. “AI has the observability, it sees where the code push was, and it saw what happened. It actually runs through all the checks and says, ‘Somebody pushed bad code. Here's the correct code.’”

The ultimate goal is to reach a level of functionality where developers can make plain language requests to generate applications. “We think we are at most 18 months away from some of these things being fully autonomous, and just a handful of years away from, you can do plain language asks and get functional applications—at least functional microservices. You will see this.”

Discussion Topics

Data Privacy + Ownership. The prevailing narrative from Big Tech is that unrestricted access to user data is essential for building advanced AI systems; Tabnine recognizes that clients today are more focused on privacy, license compliance, and copyright compliance. “Data privacy is the number one issue for us. You just cannot take your customer's data.” Our guest is particularly concerned about a lagging regulatory framework. “The law before AI was a decade behind the technology, if not two decades by the technology. Now with AI, we're legitimately at risk.”

By default, Tabnine uses only compliant models that are trained exclusively on licensed open source, ensuring that any code generated is entirely user-owned. “We have a fine-tuned version of Mistral for software development that also runs privately, so you can run it on-premise. Then we have switchable models. I support OpenAI, Cohere and OCI, and Anthropic. Mistral released something they call Codestral, and we were the first to deploy it.” Over 80% of their customers opt for on-premise or virtual private cloud solutions, meaning that Tabnine never sees their users’ data.

Data, Personalization + Context. Many current tools lack contextual understanding, a distinction between sterile, search engine-like responses from those capable of providing nuanced, context-rich assistance. Tabnine believes the future of AI is personalization and context-awareness, via a blend of LLMs, prompt engineering, and context management via RAG and semantic memory. “It's going to be like a 10-year onboarded engineer who knows all of the details, all the standards, and all the expectations, because we can read the codebase and we read non-code sources of information, extract everything out of JIRA, out of Confluence, and out of your documentation, use it as context, and apply it.” Data will be a competitive edge in various operational contexts, beyond traditional analytics and modeling. “The next trillion dollar value companies are all going to be data and infrastructure. Somebody is going to solve for consolidating around a single tool chain for all of data infrastructure, and then move everything away.”

Code Validation + Version Control. As innovation plateaus and code generation becomes commoditized, Tabnine anticipates a shift towards higher-level models that can, for example, apply standards and guidelines. “DevOps and code push is where we spend most of our time. I can rewrite all of the code if it doesn't meet the Google Java coding standards. We can automatically clean it up.” As AI grows to accurately regenerate code from prompts, version control may become redundant. “Requirements change, and so you have audit logs, absolutely, but you don't have to go and revert back to some version of the software from three months ago.”

IT Security. AI promises to streamline the identification and resolution of security vulnerabilities. Existing tools often overwhelm developers with irrelevant alerts; Tabnine plans to integrate expert guidance directly into developers' IDEs, to focus their attention on actionable insights. “It's actually an in-context fix. In the IDE, I don't have to tell them there's a security vulnerability. I need to just show them the code is wrong and show the recommendation, and at the pull request I can add more color.”

Supporting Legacy Tech. Not all legacy tech can be addressed with off-the-shelf solutions. For niche programming languages, Tabnine curates models as needed and implements client-specific RAG and vector databases, all while vetting to model on only high-quality production code. “For some of these other GenAI use cases, we're going to have to start getting data out of legacy data systems and into more modern systems, to access them in real time.” Here, specialized vector databases such as Pinecone are gaining traction.

ROI. Studies from Carnegie Mellon, McKinsey, and IBM estimate that current AI tools can save 20% to 25% of a developer's time. The ability to prove ROI with concrete data and independent validation will be key to widespread adoption; Tabnine internally is developing more robust tracking mechanisms to measure ROI and ensure code quality as code ships quicker. “What we're able to do inside of all of our tools is report on, how is it being used? How much is making it past the code review and actually being pushed in production?”

Future of AI code generation. Past disruptions, such as outsourcing and cloud migration, have also caused job losses. In the GenAI disruption, future engineers will need a stronger grasp of architecture and data flow given the increasing complexity of software engineering. “We need to generate more of these functions autonomously, if we think we're crazy right now with microservices, data flows, they're going to go up in order of magnitude. We're going to have even more autonomous functions working together.” AI may soon begin to develop code only machines can interpret. “We might get to a place where there are more performing languages that the human doesn't have to read. It may go straight to binary and we're going to have to all become architects at that point.” Tabnine will ultimately be capable of transforming written requirements into functional applications autonomously. “It will be multiple agents who interact with each other, including some human fact-checking, that we string together seamlessly into something that feels like Tony Stark’s Jarvis.”

Navigating Cybersecurity Consolidation

This week, we highlight a clip from an ETR Insights Interview with a VP of Cybersecurity where the discussion explored platform consolidation trends. Our guest expresses skepticism about further integration with Microsoft, favoring best-of-breed solutions. Listen to the conversation which highlights CrowdStrike's declining sentiment and adoption following the July outage.

SASE Adoption Remains an Opportunity in Security Life Cycle Management

In conjunction with the ETR Observatory for Secure Access Service Edge (SASE), the ETR Insights team assembled a panel discussion between senior IT executives who are taking pragmatic, incremental approaches to SASE, marked by challenges within hybrid multi-cloud environments and shifting operational needs. Read on to learn more about how these companies manage SASE, how AI will enhance anomaly detection and operational efficiency, and why SASE alone may not necessarily lead to lower IT headcounts.