Is compute the main bottleneck? Or maybe the models just aren’t big enough for AGI? It depends who you ask, because Ulrik Hansen, president and co-founder of Encord – a data development platform – says the real constraint is data.

What does it mean for the development of AI?

And he has a point. In this episode of Inference, we talk about why models are mostly swappable and “good enough,” but without the right data orchestration they don’t get very far. Ulrik – being bullish on Tesla – explains how it’s self-driving edge over Waymo comes from a compounding data advantage – “live human feedback in every car”, and why robotics is harder than digital AI because it lacks the instant feedback loop.

We also dig into why expert feedback is quickly becoming the new gold, the risks of synthetic data “eating its own tail,” the split between cheap intelligence (facts and patterns) and expensive intelligence (creativity, taste, vision), and why trust and brand will become more valuable in an AI-saturated world.

And we touch on the Scale AI/Meta deal, and whether there will be more acquisitions of data-labeling companies in the coming year.

This is a conversation about the real bottlenecks in AI, the strategies that will win, and the connection economy we’re moving into. Watch it!

This transcript is edited by GPT-5. Let me know what you think. And – it’s always better to watch the full video) ⬇️

Ksenia:
Hello everyone, and welcome to the Inference podcast by Turing Post. I’m very happy to welcome Ulrik Stig Hansen, president and co-founder of Encord. Welcome, Ulrik.
And let’s immediately start with a big question. If intelligence is solving tasks under uncertainty, what’s still holding us back? Some say models, some say inference. What’s your take?

Ulrik:
Great question – we think about this a lot. From our vantage point working with more than 200 top AI teams, it really comes down to data. That was the founding insight behind Encord: the main constraint on progress is orchestrating the right data so models work in production.

Models are increasingly swappable and, in most cases we see, already good enough to add tremendous value. What blocks further progress is getting the data enriched and organized to provide sufficient context – for training, for inference, for the systems wrapped around the models – so they make the right decisions. It’s like humans: you can be a bit smarter, but if your information is incomplete, you won’t get far. Data is the hard part.

Ksenia:
You started Encord in 2021 – very different times. What changed for you when ChatGPT was launched?

Ulrik:
We began by automating annotation. Back then, the bottleneck was quantity – producing enough labeled data to train base models. With the rapid improvement of language models and AI more broadly, the problem shifted from bounding boxes to frontier post-training and alignment work. As we’ve climbed the scaling-law curve, it’s moved from data quantity to data quality – getting the right data – and then to post-training for the most advanced areas.

As models get smarter, each incremental step gets harder. You see this most clearly in language models and self-driving – two of the most mature domains. In self-driving, it’s about long-tail edge cases; in LLMs, it’s fine-grained alignment. Robotics is earlier – roughly where LLMs were in 2017–2018 – with baselines improving fast, but an eventual shift toward harder, more esoteric problems. We’ve followed that evolution with our customers since pre-ChatGPT. It’s been exciting to watch up close.

Ksenia:
Give me an example of a recent challenge.

Ulrik:
Take self-driving. Waymo is live in San Francisco, but hasn’t expanded all the way to SFO. Why? Real environments are dynamic. Eliminating every edge case a car might encounter is extremely difficult. You need to design your data orchestration so that when a new scenario appears, the system can either handle it heuristically or route feedback back into training.

Tesla benefits from supervised driving – when I intervene in my own car, that signal goes straight back to Tesla. They’ve built a tight feedback loop to eat the long tail as it happens. Robotaxis aiming for unsupervised operation have a tougher time – they rely on fleet drivers collecting specific data, which slows that loop.

Ksenia:
Those are very different strategies. Tesla leans on pure vision and scale; Waymo is more step-by-step and careful. What’s your bet – does Tesla’s more-data, more-learning approach win, or Waymo’s cautious strategy from a data perspective?

Ulrik:
Yeah – the one thing Tesla really has going for it is a massive, compounding data advantage over every other player. They collect live human feedback at scale. Every Tesla driver is essentially providing a thumbs up or thumbs down, like you’re used to in ChatGPT, but embedded in the car. That feedback loop compounds fast.

Waymo, by contrast, has to collect its own data. They don’t have Tesla’s distribution advantage, so their rollout is inevitably slower. Their choice to add extra sensors beyond RGB vision may be safer in theory, but it’s more expensive and takes longer. Personally, I’m bullish on Tesla getting ahead. They’re starting to roll out a robotaxi fleet, which they can probably scale quickly. The open question is regulation – Waymo has a head start with regulators, and Tesla doesn’t. That dynamic will be very interesting over the next three to five years as both go live in more cities.

Ksenia:
Do robotics have the same bottlenecks as autonomous systems like self-driving, or are they different?

Ulrik:
A lot of the constraints are the same. With robotics you add the challenge of running models on edge devices with limited GPU power. That’s a huge engineering hurdle. Then you need an iterative feedback loop from real-world product usage back into the model, which is harder than in pure digital domains.

Take a robot folding laundry – if it fails, how do you capture that data and feed it back into training? What if the robot isn’t connected to Wi-Fi? Latency and connectivity become big issues. That’s why progress has been fastest in digital-first areas like language models or coding, where the feedback loop is instant. In embodied AI, feedback is slower and harder to collect. Some companies solve it by adding teleoperations to steer robots when they fail, but long-term you can’t rely on that. Clever, scalable feedback loops will determine which robotics companies build lasting advantages.

Ksenia:
And feedback loops – we’ll see more of those in agentic AI systems too, right? How does infrastructure need to evolve to support that?

Ulrik:
That’s exactly the question we’re working on. Over the past couple of years, companies shifted from building models to deploying them in production. Once you go live, human feedback infrastructure becomes essential.

The challenge is that every use case needs its own tooling. If you’re training with LIDAR, 3D data, or video, you need specialized tools. The AI toolchain is still very immature – a lot of it hasn’t been built yet. Over the next 5–10 years we’ll see the same kind of developer lifecycle that the internet economy built over 30 years, but compressed.

I expect we’ll recreate analogues of software infrastructure – version control, networking protocols, and others – but adapted for AI. The twist is that AI systems are dynamic, operating in real environments, which makes building reliable toolchains harder than for traditional software.

Ksenia:
Yeah, a lot of uncertainty to deal with.

Ulrik:
100%.

Ksenia:
You mentioned more PhDs being hired for annotation and expert feedback. Is expert feedback the new gold, what is the role of humans here? Are we just callable functions inside these systems?

Ulrik:
I think humans will primarily provide preference and feedback in different ways. On the consumer side, take your personal AI assistant – it learns from the preferences you give it: when you wake up, what you eat, how you like to work. That feedback makes the model more personalized and useful.

In enterprise, it’s professionals providing feedback through their day-to-day work. An investment banking analyst, for example, is constantly showing the AI how workflows should operate. The AI learns from that. As AI moves deeper into enterprises, humans will constantly provide feedback while doing their jobs.

And then at the frontier research level, experts and PhDs provide feedback to push the boundaries of base models – reducing how much customization is needed later. So the human role is twofold: providing preference on how we like things done, and providing expertise in areas of uncertainty where models struggle.

Ksenia:
From what you’re saying, it feels like we’ve moved from model-centric machine learning, to data-centric, and now to task- or even context-specific machine learning. Do you agree with that evolution?

Ulrik:
Yeah, 100%. Going forward I think the model will be that AI handles the baseline – maybe 80% of the work – while humans focus on the edge cases and policy decisions. So the AI proposes, and the human verifies, escalates, or corrects. That’s especially true in fields like investment banking, where an analyst may own the exceptions and steer the overall policy.

Systems should be designed for complementarity – AI surfaces uncertainty, asks for rationales, allows edits, and is even incentivized to find its own mistakes. To make that work, we need tools and measurement systems that can track cost, risk, and compounding feedback.

Ksenia:
There are a lot of competitors in the data labeling space. How do you see the area evolving? What role will companies like Encord play?

Ulrik:
You can think of the market in two buckets: services-first companies and software-first companies. Services-first companies specialize in supplying human experts – they mostly work with frontier labs that are pushing the boundaries of model research.

Software-first companies provide the tooling and infrastructure that enables this work. That’s where we sit. We work with applied AI teams, Fortune 500s, and enterprises. These companies usually aren’t building their own foundation models; they’re taking off-the-shelf or open-source models and fine-tuning them with their own preference data.

So services companies typically go after hyperscalers and frontier apps, while software companies like us focus on enterprises that are operationalizing and deploying models. Their USP is people, ours is software.

Ksenia:
So the deal between Meta and Scale AI – which category does that fall into?

Ulrik:
Yeah, I’d put Scale in the services-first bucket. They built a fantastic business selling into hyperscalers, and they’ve provided human experts to Meta for a long time. I expect that will continue post-acquisition, though the details are hard to predict – there’s been a lot of chatter. But it’s a different market from what software-first companies serve.

Think of it like Databricks or Snowflake: they rarely sell into hyperscalers, instead they focus on enterprises. Hyperscalers tend to build their own tools in-house. So really it’s about which market you’re going after – some companies provide human experts to frontier labs, others like us provide software for enterprise use cases.

Ksenia:
Do you expect more acquisitions of data-labeling companies in the coming year?

Ulrik:
Yes, I think so. AI is still new, the market is consolidating, and data is one of the three essential ingredients alongside models and compute. Data is where the proprietary value and IP really live. You can’t build a custom AI system without context on the problem you’re solving.

So companies now realize data is vital. Models and compute can be replicated, but data is what makes systems defensible. That’s why acquisitions will continue – securing strong data infrastructure and access to data is strategic.

Ksenia:
So context comes from people. But what about synthetic data? It’s everywhere now – for edge cases, rare classes, bootstrapping new domains. Does it really bridge the gap? Where does it help, and where does it fall short?

Ulrik:
Great question. We see synthetic data work really well in areas where physics-based simulation is possible. You can model the world, introduce randomness, and simulate real environments. Self-driving is a great example – you can simulate long-tail edge cases that would be nearly impossible to capture at scale in the real world.

Where it falls short is when models generate synthetic data that other models then train on – data that’s already been “seen.” That can lead to collapse, like a snake eating its own tail. But there are also effective strategies – for instance, using a larger model to generate data for training a smaller one.

The most successful approach is hybrid: combining human or real-world data with synthetic data. That combination is what we see winning out.

Ksenia:
You work with high-stakes domains – medical AI, robotics, autonomous systems. What are the biggest mistakes these companies make when moving from prototype to production in how they handle data?

Ulrik:
The savvy companies we work with usually understand their shortcomings, and they work hard to close those gaps. Their biggest challenge is simply resources – they don’t have the same infrastructure as frontier labs or hyperscalers. That’s where we come in, providing the tools and infrastructure so they can move at the same pace in bringing performant AI systems into production.

Ksenia:
So you’re their guide in this scenario.

Ulrik:
Exactly. We build products that help them move from prototype to production faster than if they built everything in-house. Often it’s just not feasible – AI systems are only getting more complex. Companies want to solve harder tasks, and the requirements keep evolving.

For example, models have shifted from unimodal to multimodal, which opens up entirely new challenges. Supporting that is out of reach for most enterprises on their own. It’s similar to how companies used to run in-house data centers, then shifted to the cloud so they could move faster. I think we’ll see the same pattern in AI development.

Ksenia:
That’s very interesting. Everything we’ve talked about so far is practical. But are there conversations about AGI with the companies you work with? And what’s your personal take on it?

Ulrik:
Yeah, I think there are of course companies working on AGI. From our side we mainly focus on applied AI – bringing intelligence into production. Treating intelligence as a single construct is difficult because there are different levels.

There’s “cheap” intelligence – raw knowledge, facts, pattern recognition on common tasks like writing boilerplate code, summarizing documents, generating standard designs. Models can produce infinite versions of that at near-zero cost.

Then there’s “expensive” intelligence – human agency, creativity, original insight, taste, vision. The kind of intelligence that feels like genius, because it can’t simply be reverse-engineered from data.

So I think we’ll see a bifurcation. For the cheap types, there will be less need for humans in the loop. For the expensive types, human agency and judgment remain crucial.

Ksenia:
That’s super interesting. For me, applied AI feels much more exciting than the AGI talk.

Ulrik:
Yeah. At the end of the day, most models are good enough. The real constraint is embedding them into products that actually solve problems. The form factor matters a lot. ChatGPT can help you write emails, but it can’t do your whole job. These systems are improving, but there’s a long tail of human tasks that will remain outside the reach of today’s form factors for quite some time.

Ksenia:
You work with data at every stage – training, post-training, context management – and with high-stakes companies where it’s critical to manage data safely. What concerns you most about this world you’re helping to build, and what excites you?

Ulrik:
The concerns are real. AI carries risks, and we’re already seeing them. Mass-produced misinformation is one example. We’ll need reliable ways to verify whether something was generated by AI or created by a human. That’s also why I think the value of brands will increase – if I read the Financial Times or The Economist, I trust they’re doing their best to be objective.

Ksenia: Because they have responsibility.

Ulrik:

Exactly, and investing in trust will become more important across the economy.

On the flip side, the excitement is enormous. We’re likely to see more change in the next ten years than we saw in the last fifty. Models working with humans could help discover new drugs, create treatments we’ve never had before, and bring robots into our homes to handle everyday chores. That frees people to focus on what we value most as humans – time with family and friends, meaningful work, creativity.

I think we’re moving from an “intelligence economy,” where value has been tied to knowledge work, into more of a “connection economy,” where new types of jobs and roles emerge – just like no one 150 years ago could imagine software engineers. We’ll see professions appear that we can’t even conceive of today. It’s a very exciting time to be alive.

Ksenia:
I’m ready for all of that. My last question is about books. I always ask this because I believe books shape people. What is a book or idea that shaped your thinking? It can be related to AI and machine learning, or completely unrelated.

Ulrik:
I’ve read so many good books over the years that it’s hard to pinpoint one. But a recent one that really made me reflect on my own life and career is Daily Rituals: How Artists Work by Mason Currey. It started as a newspaper column about daily routines, and then he expanded it into a book covering about 150 of the world’s great artists and thinkers – people like Mark Twain, Isaac Newton, Ayn Rand.

What struck me is how breakthroughs – whether a new law of physics, a great novel, or a major artwork – didn’t come from a single flash of insight. They came from years of grinding, iterating, and showing up every day. There’s a direct analogy to building a company. People think of startups as overnight successes, but in reality they’re many years in the making.

It reminded me of Aristotle’s line: We are what we repeatedly do. Excellence, then, is not an act, but a habit. That’s true for company building, research, and AI. The big breakthroughs in artificial intelligence – just like every breakthrough in history – come from sustained effort and repeated practice.

Ksenia:
That’s a great lesson to end on. Thank you so much for this interview – it was fun.

Ulrik:
Thank you, Ksenia. Thanks for having me.