After deploying an AI customer service agent for a large client, the first thing I’d do was wait for customer feedback. Most customers never leave a review, or a Customer Satisfaction Score (CSAT), as it’s commonly known in the industry. But for a large enough client, it was only a matter of minutes before the first responses would roll in. Like clockwork, the initial feedback appeared.
The first ones are always the same: star ratings without comments. The comments, however, were what I waited for. These were the gems we shared with our clients, the real stories behind the numbers. They didn’t just highlight customer satisfaction. Comments demonstrated how effective and positive a service like ours was.
For one particular client, we named the AI agent Zoe, after the head of customer service’s dog. To our surprise, customers loved Zoe. They’d leave glowing reviews, not just about their experience but directly addressing Zoe. They wanted her promoted. They joked that she deserved a raise. And they thanked us profusely for using human agents instead of automated systems.
Not only had our AI bot seemingly passed the Turing test, but it had also been embraced as a relatable, intelligent "human" who deserved recognition. And all of this happened through the simple medium of text.
What Is AGI Supposed to Be?
Sam Altman, OpenAI's CEO, says:
AGI is basically the equivalent of a median human that you could hire as a co-worker. They could do anything that you’d be happy with a remote co-worker doing just behind a computer.
This sounds impressive, but what does it actually mean? How would we know if AGI were achieved?
Andrew Ng, one of the pioneers of machine learning, has a more grounded view:
[...] It’ll take technological breakthroughs before we get there, and it may be decades, hundreds, or even thousands of years away.
Even if AGI were achieved, how would we interface with it? Would it still be a chat interface like ChatGPT, just with a more powerful model? Or would it take on entirely new forms? And, most importantly, how would its users know they were working with an AGI?
Specialized Jobs Require Specialized AI
Take Zoe, for example. Zoe wasn’t an AGI. At best, she was an NLP (Natural Language Processing) model tied to a classifier. When a customer sent an email to support@service.com, here’s what happened:
Sanitization: We cleaned the input, stripping out signatures and unnecessary details.
Tokenization: The sanitized input was processed by an NLP model to identify entities and extract meaningful data.
Classification: A classifier identified the type of request, such as "track order," "exchange," or "cancel subscription."
Response: Based on this classification, a pre-written script tailored to the client’s business was used to generate a reply.
Every email signed off with a disclaimer that Zoe was a bot, but customers didn’t care. For their purposes, Zoe worked.
This got me thinking: If OpenAI created an AGI tomorrow, would it really look all that different? Would it just be another chatbot interface? Do we have to read the benchmarks to be convinced it's an AGI behind it?
Interfaces Matter
Google won the search engine battle largely because of its simplistic interface. ChatGPT followed suit with it's simple yet powerful chat user interface. It’s intuitive, clean, and versatile. For writing, the chat-like interface allows for quick research and editing in real-time. But when it comes to coding, ChatGPT’s interface doesn’t fit into most developers’ workflows.
As a software developer, I use tools like VS Code, Sublime Text, and Vim. These are tools designed to fit seamlessly into a programming pipeline. ChatGPT’s interface, by comparison, requires constant context switching. It’s not that it can’t write code (though it does hallucinate sometimes); the problem is that its interface isn’t built for developers.
Contrast this with tools like GitHub Copilot or Claude, which integrate directly into IDEs. They’re powered by similar underlying models but designed with developers in mind. The result is an interface that feels natural and works effortlessly.
The same principle applies to hardware-focused AI like Cerebras. Its web interface screams “designed by hardware engineers.” It prominently showcases metrics like "2,157 tokens per second." Impressive? Yes. Useful for most end users? Not really.
AGI and Job Specialization
To understand why AGI might not be the answer to every problem, consider a specific job: an account executive.
An account executive does much more than answer questions. They:
Build relationships with clients.
Drive sales and identify new opportunities.
Act as the primary point of contact for client concerns.
Collaborate on strategy with internal teams.
Provide regular updates and reports.
Now imagine trying to build an AGI to handle this role. It would need to:
Email, call, or chat with clients to build relationships.
Understand the company’s product well enough to sell it and address concerns.
Communicate client needs to internal teams and work on strategies.
This is doable—not because of AGI but because the job can be broken down into workflows. For example:
A chatbot interface could handle client communication.
An AI model trained on product knowledge could answer client questions.
Another system could facilitate internal team collaboration.
These aren’t tasks that require AGI; they’re tasks that require specialized AI applications designed for the job.
AGI Is About Perception
Even if AGI could perform all these tasks, how would people respond? Most clients don’t want to talk to an AI. Even if it’s as capable as a human. They want to talk to someone they trust.
The problem isn’t whether AGI can perform a job. The problem is how it interfaces with the people it’s meant to serve. If the interface doesn’t fit the task or the user’s expectations, even the most advanced AGI won’t succeed.
This is why AGI isn’t just a technical challenge; it’s an interface problem.
It is often argued that self-driving cars will only reach Level 5 autonomy with the advent of AGI. But their success will still depend on how that AGI’s interfaces with the physical world. The car will require an array of precise sensors to detect, interpret, and navigate its surroundings effectively. One of the key selling points of companies like Waymo is that, when their systems work seamlessly, they appear indistinguishable from AGI. However, their failures often reveal themselves as technical limitations rather than a lack of intelligence.
For example, in 2021, a Waymo car became "stuck" at an intersection in Chandler, Arizona. The car encountered an unexpected traffic cone pattern and could not decide how to proceed. Despite having advanced AI, it required human intervention to navigate the situation. This failure wasn’t because the car lacked intelligence in a general sense; it was a result of a gap in how the system processed and responded to a specific real-world scenario.
This highlights that achieving true autonomy isn’t just about intelligence; it’s about integrating intelligence with the right tools and interfaces to handle the infinite complexities of the real world.
One thing I didn’t share about Zoe are her failures. Not the false positives or false negatives. Customers reacted negatively when they realized they were interacting with an automated system. These reactions are just as important as technical accuracy.
For instance, Zoe might provide a correct and complete response, resolving the customer’s issue in seconds. However, the speed of the resolution itself can trigger suspicion. Some customers will ignore the solution entirely and respond with something like, “Yeah, I’ll wait for a real person to respond.” Even when Zoe has done everything right, her identity as an AI becomes the sticking point.
Perception and trust. No matter how accurate or efficient an AI system is, customers need to feel confident that they’re being heard and understood. If the interface—whether it’s Zoe or another system—fails to establish that trust, its success is undermined, not by technical limitations but by human psychology.
I like to frame AGI as an interface problem because it shifts the conversation. It goes from simply asking whether we can build AGI to how we interact with it in meaningful ways. Intelligence, no matter how advanced, is only as effective as its ability to communicate and integrate with the world around it. Whether that’s humans, machines, or physical environments. Here are a few reasons why this idea resonates:
1. The Role of Context
Intelligence, whether artificial or human, relies heavily on context to make decisions. AGI might theoretically possess broad, human-like capabilities, but without the right interface to access context—be it sensors, APIs, or human feedback loops—it becomes impractical. For example, an AGI driving a car still needs an interface (sensors and processors) to perceive and respond to its environment, just as a chatbot like ChatGPT requires text inputs to operate.
2. Specialization vs. Generalization
Most successful AI applications today are highly specialized. A self-driving car doesn’t need to know how to write poetry, just as a customer service bot doesn’t need to solve physics equations. Even if AGI is achieved, its utility will depend on interfaces designed to channel its general intelligence into specific tasks. These interfaces define how the AGI applies its knowledge and capabilities in specialized domains.
3. User Perception and Trust
The way people perceive and interact with AGI will largely determine its adoption and success. Zoe, the AI customer service agent, illustrates how interface design and framing can make an AI appear not just functional but "human" and relatable. The reverse is also true. A poorly designed interface can make even the most advanced AI feel clunky, unhelpful, or frustrating, undermining its potential.
4. Failures as Interface Gaps
Failures in AI systems, like self-driving cars struggling with unusual traffic patterns or chatbots providing nonsensical responses, often highlight gaps in the interface rather than flaws in the underlying intelligence. These issues underscore the importance of refining how AGI connects to its environment, processes inputs, and acts on outputs.
5. Broader Implications
If we frame AGI as an interface problem, it opens up new ways to approach its development. Instead of focusing solely on creating a "perfect brain," we might prioritize designing better tools, sensors, and communication methods to maximize its utility. This could mean integrating AGI into specialized workflows (e.g., account management, robotics, or education) where the interface bridges its general capabilities with specific needs.
The idea reframes AGI not as an abstract milestone but as something inherently practical. A tool whose value is defined by how well it interfaces with the systems and people it is meant to serve. This perspective grounds the conversation and helps clarify the challenges and opportunities ahead. It emphasizes that the true test of AGI is not an intelligence benchmark. But a system that can connect and integrate seamlessly into the environments it is meant to serve.
Comments
There are no comments added yet.
Let's hear your thoughts