Machines Learn Faster with Human Feedback: An Interview with Daniel Langkilde, CEO of Kognic

Autonomous driving represents one of the most ambitious challenges in AI today. While billions have been invested in getting machines to navigate the real world safely, the industry is still learning how to efficiently integrate the human judgment required to train and validate these systems. In this blog post, we explore the current state of autonomous driving technology and the critical role of human feedback, based on an interview with our CEO, Daniel Langkilde.

The Current Approach to Autonomy

According to Daniel, today's autonomous driving systems break down a complex problem into sequential components—perception, prediction, and planning. While each component is individually sophisticated, this sequential processing can lead to information loss between stages.

As Daniel explains, "The trend lately has been that people feel we are sort of throwing away too much information doing this as sequential steps. And instead, some argue we should try and solve this using one giant neural network, which is sort of known as an end-to-end model or a fully differentiable stack."

The Role of 'World Models' and Human Feedback

A key component of emerging approaches is what Daniel calls the 'world model'—the common sense understanding that machines need to make sense of the physical world. However, building world models that can accurately predict real-world scenarios remains a significant challenge that requires continuous human feedback.

Daniel discussed Tesla's approach, noting that "Tesla, obviously also should be included, that are trying to train very large quote-unquote world models based on observing large quantities of driving." However, he emphasized that this approach is still evolving, and success depends heavily on the quality of human feedback used to validate and align these models with real-world expectations.

Looking Towards the Future: The Human-in-the-Loop Imperative

When asked about the future of autonomous driving, Daniel highlighted that expanding the operational capabilities of autonomous vehicles remains challenging and expensive—precisely because machine learning still requires human judgment to handle ambiguity and edge cases.

He stated, "Machine learning still struggles to build sort of common sense and generalize to new situations. That's our biggest challenge dealing with all that ambiguity." This is why human feedback remains essential: no matter how powerful models become, they must be guided, validated, and aligned with human judgment to ensure safety and reliability.

Daniel also emphasized the importance of remaining open to new paradigms. "If there's evidence that there's a new paradigm, all it takes is a few like bold entrepreneurs. And then it doesn't matter if existing companies hold on to their old ideas. Someone else is going to try the new one and maybe it works, maybe it doesn't."

The path forward for autonomous driving isn't just about bigger models or more data—it's about integrating cost-efficient, scalable human feedback into development pipelines. At Kognic, we believe that machines learn faster with human feedback, and we're building the platform that makes this human-machine collaboration productive and reliable for the autonomy industry.

Listen to the full interview here.