What Really Blocks AI Progress? Ulrik Hansen from Encord thinks it’s…

The data bottleneck holding back AI progress

The constraint on AI progress is data. Ulrik Hansen treats models as largely swappable and good enough, and Encord's business as orchestrating and enriching data with enough context for training and inference. The problem space has moved from annotation automation to hiring PhDs for post-training and alignment. That explains why Tesla's driver data compounds against Waymo's slower self-collection and why robotics trails language models. It is also why enterprises buy tooling instead of matching frontier-lab resources.

Takeaways

• Data is the binding constraint on AI. Encord's founding insight was that data decides who wins in AI. Hansen calls models largely swappable and good enough, and says the work is enriching data with enough context for training and inference.
• Annotation work moved from boxes to PhDs. Encord began by automating annotation to solve data quantity problems for base models. Frontier labs now hire PhDs for post-training and alignment, shifting the market from annotation volume toward expert judgment on uncertain cases.
• Tesla's driver data compounds against Waymo. Tesla collects intervention data from drivers directly, which compounds. Waymo must gather its own data, and real environments are dynamic with edge cases hard to enumerate, so its rollout is slower despite safer sensors.
• Robotics trails language models on feedback. Robots add edge deployment on GPU-limited hardware and feedback that is hard to relay, since a robot folding laundry may fail with no fast upload path. Coding advanced fastest because its feedback loop is instant.
• Experts will train models by doing their jobs. Consumers already give preference data to assistants. In enterprises, analysts and other experts will supply feedback on model uncertainty and preference, with AI handling about 80% of the grunt work while experts own edge cases.
• Labelling splits into services and software. Services-first firms supply human experts to frontier labs and hyperscalers. Software-first tooling companies like Encord serve applied AI and Fortune 500 enterprises, a position Hansen compares to Databricks and Snowflake.
• Synthetic data needs real data beside it. Physics-based simulation generates long-tail self-driving edge cases that are hard to capture in reality. Training models on model output risks collapse, so the winning formula combines real and synthetic data.

What comes next

• Self-driving spreads across US cities over three to five years, though Waymo is live in San Francisco without an airport route because edge cases are hard to enumerate.
• The AI tool chain, including version control and feedback infrastructure, gets built over five to ten years, more compressed than the internet's thirty-year build.