Peter Guagenti has spent almost 30 years working in technology, and what drives him is the opportunity for positive change and transformation. He believes that on their best day, tech innovators and entrepreneurs have the power to make the world a better place. “The rapid advancements in tech, particularly in AI, have opened up endless possibilities,” according to Peter. His passion stems from the fact that technology, when applied correctly, can drive significant societal benefits, which keeps him motivated to push the boundaries of innovation.

Peter currently serves as President of Tabnine, a company dedicated to transforming the software development process through AI. Their mission: to accelerate and simplify every stage of the development life cycle with AI-powered solutions. In this interview, he sheds light on the highly competitive landscape of AI in software development and the groundbreaking work being done at Tabnine.

Check out Peter’s insights on AI’s role in software development. Watch the video (or read on) as he covers Tabnine's AI code assistants, the importance of privacy in enterprise AI, and the current challenges facing developers, along with a look at the future of AI in software engineering.

0:00

/5:41

Tabnine's Peter Guangenti on AI code assistants that work.

Below is an edited version of the conversation.

How has AI evolved since you started working with it?

It's hard to say exactly when I started working with AI because the lines have been really blurry between machine learning, deep learning, and now generative AI. I started working with algorithms and leveraging what I would consider the earliest waves of AI nearly a decade ago. Back then, as a sales and marketing professional, I used AI and algorithms to identify my ideal customer profile, to change targeting algorithms, and to optimize recommendations. AI solved problems like delivering the right message to the right audience at the right time, tasks that humans couldn’t accomplish at scale.

Today, I’m both surprised and unsurprised that it took the emergence of ChatGPT for people to realize that AI was already here and capable of so much. I believe we're entering a new digital transformation, where AI will be central to business processes. This change is as significant as what we went through over the last 25 years. But we’re at, or just slightly past, the peak of the hype cycle, so it’s time to get real about what AI can actually deliver. I'm excited about the potential for positive change that AI will bring.

What role does Tabnine play in the AI code assistant space?

Tabnine created the first AI code assistant using LLMs about seven years ago, initially focused on code completion for Java. Fast forward to today, and we are a clear #2 in the category, with millions of users and thousands of organizations using our product.

Two years ago, we saw how generative AI was becoming increasingly important to CIOs, so Tabnine evolved into a more enterprise-focused product, offering features like fully private deployments, robust security, and multi-user management. Our strength lies in being the AI code assistant that you control, providing organizations with control over where AI is deployed, what models it leverages, and personalization through fine-tuning and context awareness. Our focus on privacy has allowed us to serve industries like finance, pharmaceuticals, and defense, where data security is critical.

What are the key components required for an effective AI code assistant?

We've discovered that generative AI applications operate best when they follow what we describe as a “three-legged stool” model. The first leg is the LLM itself, which is the core component. But you can't achieve success with just the model. You need to add two other elements: prompt engineering and context.

Prompt engineering interprets the user’s intentions and shapes what gets sent to the LLM and what comes back. For example, if a developer asks for a function, documentation, or a test, even with vague wording, we adjust the prompts to be more specific and provide the optimal outputs for the use case. Good prompt engineering ensures that we always get the highest quality response.

The third leg is context. This is critical. By layering in the context of the codebase, including global and non-code sources, the quality of the output improves dramatically. For instance, when we applied context awareness solely from what was visible in the IDE to Tabnine, we saw a 40% higher acceptance rate for AI-generated code. That number jumps even more dramatically when we layer in awareness of a company’s wider codebases and non-code information like requirements and documentation. It’s not just about generating code. That’s easy. It’s about generating the right code that fits within a company's standards and practices, and of a high enough quality to make it to production.

0:00

/5:12

Tabnine's Peter Guagenti on the key challenges with AI coding assistants and the future of software engineers.

What are the biggest challenges in deploying generative AI in software development?

The biggest challenge is that every development team is unique. Their architectures, standards, and coding practices are all different. One-size-fits-all generative AI doesn’t work because if it can’t address each team’s specific needs, then the code will not make it past the pull request. That’s why many generative AI tools are failing. At Tabnine, we built a robust Enterprise Context Engine, which pulls relevant insights from a company’s codebase and other software development tools to tailor the AI’s behavior to that company’s specific requirements. As I said earlier, simply generating code isn’t enough; it needs to fit the company’s standards, or developers will reject it.

How are developers engaging with AI code assistants today?

We've seen two types of developers: the ones who have fully embraced these tools, making the most of AI by using them like an Ironman suit for their minds, and those who are more fearful and resistant to change. Experienced users view AI as a natural extension of their work, including fully automating tasks like documentation, test generation, and code fixes. But some developers are afraid that AI will replace their jobs, and so they are hesitant to incorporate it fully into their workflows. In reality, AI is automating mundane tasks, so that developers can focus on more creative, higher-order work. Developers have little to fear here. AI isn’t a replacement; it’s an accelerator.

Why is security and private deployment so important for AI in today’s enterprise environment?

If you want to understand how important privacy is, you have to think about it from two perspectives: how IT decision makers feel today, and how vendors have been behaving that reinforce that. Recent studies show that over 70% of CIOs state that privacy is their number one concern with generative AI. They don't want proprietary company data exfiltrated to third parties.

We’ve seen this story before — companies moved proprietary data to third parties, and it ended up being collected and/or stolen. Even though vendors claim that they don’t use private data to benefit their businesses, the sensitivity is still incredibly high as there have been documented cases where the behavior of some of these tools appeared to differ from what the vendors have stated. As such, many companies seek out comparable capabilities via private deployments of their generative AI tools and the underlying models, especially in regulated industries like finance, pharmaceuticals, and defense.

At Tabnine, we offer a full range of deployment options to address these concerns, from cloud-based solutions with zero-data retention to fully air-gapped systems where data can never leave the company's infrastructure. This flexibility is why we’ve been able to work with customers across industries who require strict adherence to their policies and protocols.

0:00

/2:23

Tabnine's Peter Guagenti on the importance of private deployments for AI.

How do different models work within Tabnine’s platform?

Tabnine can run on top of any large language model. Our prompt engineering and context awareness tools ensure that each LLM performs at its best. This gives our customers maximum flexibility. We support third-party models from OpenAI, Cohere, and Anthropic, allowing users to switch between models depending on their needs. We’ve found that each model behaves differently, so it’s essential to optimize for specific use cases. In addition, we’ve built a proprietary model, Tabnine Protected, which is trained exclusively on permissively licensed code for use cases with a high degree of sensitivity to copyright and license compliance.

How does Tabnine collaborate with Cohere?

We’ve been exposed to Cohere from the very early days, and we’ve worked with the team for some time. When Cohere released Command R this year, we were impressed with its performance, and as part of our multi-LLM strategy, it made sense to integrate Cohere into Tabnine.

We’ve now deployed both Command R and Command R+, which has shown even better results. I can't share the adoption data, but I can tell you that among the seven or eight models we deploy from third-party providers, Cohere ranks in the top third for usage and acceptance rates. The models appear to do particularly well in code fixes. This collaboration has been very positive, and we’re excited to see more innovation coming from Cohere.

The partnership with Oracle is also critical. We have several customers who choose to run Cohere on Oracle Cloud in private deployments. Thanks to our collaboration, we have been able to serve a whole new cohort of enterprises that prioritize security and control over their AI infrastructure.

0:00

/2:23

Peter Guagenti on how Tabnine works with Cohere.

Want more? Check out this demo showcasing Tabnine Chat using Command R+ for code generation, code explanations, documentation generation, and AI-created tests. Boost every aspect of your software development starting today. Watch now.

What are your hopes for AI-powered coding?

I feel like we've always been limited by our horsepower as engineering teams — our capacity always limits us, not our imagination. I envision a future where developers will be able to input complex asks into an AI system, which will respond in real time, generating applications or components based on requirements instantly. AI will be able to interpret requirements, ask clarifying questions, and come back with fully realized components of an application. With that kind of capability, software creators will be able to iterate dramatically faster. The interaction will be much more fluid and dynamic, like what you see in movies with virtual assistants. This would allow us to go from idea to production 10x or 100x faster than we do today.

I think this will be like the leap we saw when we moved from hand-drawn models to 3D CAD in manufacturing. We’ll be able to focus on creativity and innovation rather than the mundane aspects of coding. Ultimately, AI will free developers from the mechanical tasks of coding, allowing them to concentrate on more strategic, architectural decisions. This is what will unlock the next level of innovation in software development.

About Peter Guagenti

Peter Guagenti is President at Tabnine. He is an accomplished business builder and entrepreneur with expertise in strategy, product development, marketing, sales, and operations.

Building enterprise AI agents

A comprehensive blueprint for developing scalable AI agents in regulated industries.

Download the eBook

The future of AI code assistants: Insights from Tabnine