에피소드 121: The Next Phase of Agentic AI: Social Intelligence

We have a Slack integration that is internal but working great, with access to our code and logs on GCP, and it helpfully chimes in during debugging.

A couple of weeks ago, it proactively notified us about the Gemini 3.0 drop and suggested upgrading parts of itself to use different components.

Currently, it has view access to our code, but in the future, it could have edit access and perform self-modifications, which could be really useful for organizations.

Hi, my name is Demetri Panici, and I'm a content creator, agency owner, and AI enthusiast. You're listening to the AI Agents Podcast brought to you by Jform and featuring our CEO and founder Aytekin Tank.

This show explores where artificial intelligence meets innovation, productivity, and the tools shaping the future of work. Enjoy the show.

Hello and welcome back to another episode of the AI Agents podcast. Today, we have David Petri, founder and CEO of Continua. How are you doing, David?

I'm doing great, Demetri. How are you?

I'm living the dream. Tell me a little about your background and how you got into AI.

As a kid, I read a lot of books on science and computers, taught myself programming from computer magazines, and was inspired by Douglas Hofstadter's work on strange loops and emerging technologies.

I studied computers at Berkeley and Carnegie Mellon, earned a PhD at CMU, and then worked at Google for almost 18 years, focusing on machine intelligence and AI for the last decade.

I left Google to start Continua because the rate of technological change is faster than ever, and being part of a dynamic startup environment felt right in 2023.

After reviewing your background, I have some questions tailored to you. Let's start with how you came up with the concept for Continua.

Continua's tagline is 'AI that joins the conversation,' aiming to make AI a more personal and close experience through text conversations, which are a natural place for people to interact.

Could you tell us how it works and how the idea came to you?

It's important to clarify what Continua is not: it's not a companion or friend, nor a replacement for human connection.

Instead, Continua creates an agent that elevates human-to-human connection by living in group chats and nudging friends to engage and do things together.

This positioning influences how we develop the product to help the fullness of human connection.

Regarding the idea, coming up with product ideas involves timing; you want to be neither too early nor too late, especially with limited startup resources.

While chat-based large language models have revolutionized single-player interactions, there's a huge opportunity in social contexts where people congregate.

We wondered why no one had done this before and realized the idea is non-trivial with many challenges to solve.

We believe we have the world's best social AI, and it's an exciting place to work, perfectly timed for today's technology landscape.

When it comes to AI, would you describe yourself or your product as agentic?

We describe our product as agentic, meaning it has agency to some extent, able to use tools, understand APIs, and perform tasks autonomously.

Reflecting on your AI journey, when did you notice models gaining agentic abilities?

There wasn't a single light bulb moment but several, such as zero-shot learning at Google, where models could follow instructions in prompts and perform tasks in new ways.

This represented a humility in machine learning, acknowledging that programmers can't craft all solutions manually but must create programs that create programs.

Most problems require this approach rather than brute forcing solutions, and collaborating with AI reflects this humility and partnership with a new form of intelligence.

We can discuss what it means to be a researcher or software engineer, but at Continua, we aim to bring out the best of this technology to help people.

When did you start to see agentic capabilities in models like GPT-3.5 or GPT-4?

Agentic means the model can use tools, create tools, understand APIs, and search for information autonomously.

I recall seeing zero-shot learning and instruction-following code at Google, which was both insane and beautiful, marking key moments in AI development.

Many projects at Google aimed high but were limited by technology at the time, such as the Glass wearable computer, which faced challenges like battery life and computer vision limitations.

Back then, computer vision was good with planar objects but struggled with contour-defined objects like keys, making some use cases impractical.

Natural language understanding was limited pre-LLM, complicating intent recognition and efficient search.

Today, progress in multimodality across images, video, sound, and text is exciting, enabling networks to support a wider variety of tasks.

For example, in my media agency, Gemini 3 improved vision recognition from 11% to 70%, enabling detailed quality assurance prompts for video editing.

The screen understanding work by the Gemini team is incredible and key to many computer tool use cases.

With Gemini 3, agentic capabilities like reasoning improved, enabling practical applications like tutorial creation and quality assurance.

This is akin to bootstrapping, where AI develops modules or code itself, enabling closed-loop development and evaluation.

AI can evaluate product quality, but models must be sufficiently capable to avoid wasting time and resources.

We're nearing a point where AI can autonomously develop and evaluate user interfaces overnight.

Have recent technological improvements allowed you to confidently add new features?

Starting Continua was a bet on trends like improving model quality, reducing hallucinations, and lowering costs.

Hallucination is a misnomer; models predict the next token and sometimes get it right or wrong based on training data and sampling.

Understanding model mechanics demystifies hallucination and guides building systems to increase accuracy and guardrails.

Cost improvements come from efficiency and fewer tokens used, as seen with Claude 4.5 Opus, benefiting startups and big tech alike.

On-device computation is key, offering latency and cost benefits, with devices accelerating machine learning inference.

While server-based inference yields higher quality, edge computation is improving, enabling more offloading and privacy.

Latency and context window size must be balanced to maintain performance and quality.

These trends show rapid progress with no stasis in AI development.

Your technical background from Google clearly informs your deep understanding of AI and product development.

Continua is a beautiful product with high user retention, appearing as a participant in group chats like iMessage or WhatsApp, invited only to selected groups.

Continua acts as a quiet helper, answering when useful and staying silent otherwise, respecting conversational flow and social discretion.

It listens carefully, asks questions when unclear, offers insights when helpful, and does not interrupt unnecessarily, similar to a thoughtful new employee in meetings.

Implementing this required fine-tuning multiple models to understand intent and conversational dynamics, enabling Continua to speak up even when not directly addressed.

We believe solving social AI is essential to AGI, as intelligence must cooperate and collaborate effectively.

What is your current user base and how was your $8 million raise to bring AI agents to group chat?

We raised funding after leaving Google in 2020, focusing on having a product before announcing the raise.

Use cases include helping roommates coordinate chores, families manage schedules, friends get sports updates, and stock market information.

Continua facilitates group decisions and problem-solving, providing reminders and alerts relevant to group interests.

It can create Google documents summarizing domain-specific information weekly, reducing the need to copy-paste between single-player LLMs and chats.

What are users most excited about in your product?

Our recent shopping integration with Amazon allows Continua to provide personalized gift suggestions based on group chat history and preferences.

Information flows from superset groups to subsets, so Continua can use group chat knowledge in direct messages to provide personalized recommendations.

For example, in trip planning, Continua helped plan a Disneyland trip with my daughter and sister, suggesting restaurants, rides, and replanning on the fly when rides closed.

Continua also helps at work by managing lunch plans and meal preferences, acting as a personalized helper in daily activities.

What is the ultimate vision for Continua if widely adopted?

We are building more agentic connections, including a Slack integration with access to code and logs, enabling proactive self-upgrades and debugging assistance.

Broadly, we want AI to participate in group settings beyond text messaging, such as video chats, acting as quiet listeners who chime in contextually when needed.

This real-time interaction could help resolve issues and provide references seamlessly during conversations.

What features are you working on or looking forward to?

We aim to leverage chat platforms' existing capabilities and extend them to provide personalized, bespoke experiences tailored to conversation participants.

Users care about capabilities rather than apps, so Continua focuses on remixing existing applications' capabilities through agentic technology.

Given your history with on-device AI at Google, what are your thoughts on cloud versus local models for businesses concerned about data sovereignty and speed?

Secure computation, including trusted execution environments and encrypted memory, enables high privacy standards both on-device and in the cloud.

On-device AI offers advantages like disconnected mode and lower latency, but social products like Continua rely on connectivity.

Techniques like federated learning allow decentralized training with privacy protections, distributing computation across devices.

While on-device benefits are significant, large cloud models still provide higher quality, and secure computation helps balance privacy and performance.

What is your favorite AI product outside of your own?

I recommend Warp Terminal, an AI-assisted coding environment that integrates natural language and terminal commands seamlessly.

Warp allows you to type commands or natural language instructions in one interface, blending coding and communication efficiently.

Where can people check out Continua?

Visit continu.ai or search for Continua AI. Using the product is simple: add a phone number to your group chat, and it only sees that chat's content.

We're open to feedback and hiring for several positions. We're excited about social AI and its potential to enhance human connection harmoniously.

Thank you for listening. Please like, subscribe, and join us for the next episode.