De 7 Extra | Maken wereldmodellen AI eindelijk echt intelligent? - De 7

Summary4 min read

De 7 Extra | Maken wereldmodellen AI eindelijk echt intelligent?

Host: Bert Rymen
Guests: Vincent (AI Professor at Harvard), Stefanie De Smid (TED Colleague)
Date: January 16, 2026
Podcast: De Tijd – De 7 Extra

Overview

In this episode of "De 7 Extra," host Bert Rymen explores the question: Are world models finally making artificial intelligence truly intelligent? With insights from AI experts (including Vincent, professor at Harvard, and Stefanie De Smid), the discussion unpacks the hype and reality around recent breakthroughs in world modeling for AI—focusing on robotics, language models, virtual simulation, and their limitations.

Key Discussion Points & Insights

1. What Are World Models in AI?

Robotics Needs Common Sense
- Vincent: “In order for robotics to be smart about the physical world, it has to understand things like gravity, friction, object permanence so that the AI has world common sense.” (00:12)
- Emphasis on world models as essential for grounding AI in physical reality; without them, machines can’t interact robustly with their environments.
Simulating the World for AI Planning
- AI builds an internal simulation or “model” of the world to plan actions—much like in a video game.
- Vincent: “It’s as if you bend in a video game... The world updates conditional on the action you undertake.” (02:36)
- Gaps still exist: These virtual worlds have rules that don’t always match real life.

2. From Language Models to Simulated Worlds

The Role of Language Models (LLMs)
- Discussion on advancements by OpenAI and Sam Altman, building language models that begin to incorporate some form of “world understanding.”
- Vincent: "The pivotal moment is spatial random correctness in the first instance... information about the world is added through Tao and similar approaches.” (04:50)
- There is debate over whether these models truly understand the world or just mimic understanding through data.
Graphics and Virtual Reality
- Nvidia’s advancements highlighted: Using virtual reality engines as a testbed for world models.
- The idea: Simulate embodied AI acting and learning within rich virtual spaces, making future deployment in the real world smoother.

3. Limitations & Challenges

Simulation vs. Reality
- The leap from virtual models to real-world performance remains a huge hurdle.
- Vincent: “Physique AI. Robots study and learn in virtual experiments but data handling and transfer to the real cosmos has gaps.” (07:42)
- There's an “Alphabet” of approaches, but so far, no silver bullet.
Energy Paradox
- As models become more capable and complex, energy consumption remains a major bottleneck.
- Vincent: “The paradox of progress: the efficiency increases, but so does the energy demand—Jefferson’s paradox.” (13:04)

4. Chips and Hardware for New AI

Breakthrough Hardware
- Introduction of new AI-specific chips for real-time, low-latency processing—e.g., Nvidia’s new 'superchip'.
- Bert: “Chips for AI are breaking new ground... a new superchip.” (11:34)
- Powerful chips are critical for the next stage of AI, as they handle the computing needs of world models.

5. The Future: Autonomous AI Agents

Full autonomy for AI agents is becoming more realistic, but fully generalized agents—able to represent and interact safely with the world—remain an aspiration.
Bert: “The key point of world modality is what makes it more effective. The field is now breaking open for autonomous agents...” (14:10)

6. Societal Impact & Expert Caution

Expert guests emphasize that while technical breakthroughs are dazzling, the impact—both positive and negative—requires careful societal debate and feedback from multiple disciplines.
Vincent: “In the bigger scheme of things, there’s still impact yet to be measured.” (16:02)

Notable Quotes & Memorable Moments

Vincent on Embodied AI:
“In order for robotics to be smart about the physical world, it has to understand things like gravity, friction, object permanence...” (00:12)
On Simulation Limits:
“You can build worlds in video games, but real-world unpredictability is another story.” (paraphrased, 02:36)
On Energy Consumption:
“The paradox of progress: the efficiency increases, but so does the energy demand—Jefferson’s paradox.” (13:04)
On the Next Wave:
“The field is now breaking open for autonomous agents—all eyes on the next epicenter of fabricen.” (14:10)
On Societal Feedback:
“...Another element where experts stress impact—feedback and consequences will be huge.” (16:02)

Timestamps for Key Segments

00:12 — Understanding World Models and Common Sense in AI
02:36 — How AIs Use Models: Virtual Games vs. Reality
04:50 — The Language Model Debate: OpenAI’s Path
07:42 — Sim2Real Challenge: From Virtual To the Cosmos
11:34 — The Hardware Arms Race: New AI Chips
13:04 — The Energy Paradox in AI Progress
14:10 — Move Towards Autonomous Agents
16:02 — Societal Impact & Expert Reflections

Conclusion

This episode demystifies the state of “world models” in AI: impressive virtual achievements are getting closer to bridging the gap with real intelligence, but major hurdles—particularly in real-world effectiveness, hardware, and social impact—remain. The field is on the verge of enabling truly autonomous agents, but experts urge continued critical scrutiny and conscious debate as these technologies advance.

Loading summary

Transcript17 lines

[00:00]
A
Startje dag wol ambisi mette sewe tilje digitale platforme en marketing. Naren hohernivo met IO mackenz ob IO digital puntkomm.
[00:13]
B
In order for robotics to be smart about the physical world, it has to understand things like gravity, friction, object permanence so that the AI has world common sense.
[00:25]
C
Jensen Huang wasdit CEO from Nvidia over berold modelen van Derb. World's models of werolt modele igbe mertreme and welcome to Deceiver Extra.
[00:48]
A
Deceiver extra.
[00:56]
C
Vincent hunis professor artificial intelligence harvard and ted collega stefanie desmid over the world model in AI.
[02:37]
B
This is effective and sort of an external handling an axi D in the world and nonprobery d world update to vorspeller build you can inbuild is you bend in a video game on its bell video game with natyrluk obscurb and with the linkersen in building well, the world updates conditional of the axidi under nominees that is early quad.
[03:36]
A
You had before build you can't perfect and start plan in AI stoppen and the.
[04:35]
B
The was the wurmallage hooftran AI be meta. These are new since engkleman vertrocken and heft and startup.
[04:48]
C
Okay and unhappy at Sam Altman model.
[04:51]
B
And then in the dat Alik Nimnus and Altman as a build The OpenAI Noalted CEO is and the heil Erich Vastada and Tao model who will discuss and viral model of nimmer and is there the under ideologie and tal model hev the pinenkel moment spatiale ramdli kerf so explicit Anglic eglic corrected in erst instance in welted helvel informasif on the world door to Tao and eglik is there the hook. The gemaktworth is also an older model in the tal model worlds. Who will adnoit active ingebakagbestis. Offbuilding and on interdat unrendering can mark three dimensional video game. An effective anset. Nvidia model of building virtual reality engine.
[07:42]
A
Physique AI. Internet. Hilvrand who hired data OK for Zamlin who hired robot Darup Trenin the robots the world in student and latin experimentator as well as in butter damach does the data hand on ergism to come and does in the data where in video met cosmos von et had elsewhere zone. Absolute Alphabet. For games.
[10:29]
B
Were spread over obverse hilland. Overall.
[11:35]
A
Stop or stop bislasing and Neiman Vonnikker start setting Anderson real time. Chips for AI breaking in Lancertr a new superchip. In chips.
[13:05]
B
The paradox van de herald. The aiva van dyk is on set and fail efficient and tukh is the energi consumption of Jeffson's paradox.
[14:10]
A
Stopped at ons herm and in on dialects Leven. The high prunt world modality is what marked it more and effective. The word field has broken over AI agents, the strax autonomous. Well on fabricen epicenter.
[16:03]
B
Intra here. And that is not. Feedback. In the bigger scheme of things. Another element where experts is. Impact.
[20:07]
A
This was the seife extra. From Duxidor. Laura Alegardt one of the reductive on the tet. Zam meh info op digital for you Bunbe I O. Experience is everything.