1 Oct 2026
|
Mike Boland
AWE Talks: Defining & Demystifying World Models
AWE USA 2026
Welcome back to AWE Talks, our series that revisits the best AWE conference sessions. With AWE USA recently concluded, there's a fresh batch of stage footage to keep us occupied for weeks.
We continue the action and insights this week with a deep dive on the new hotness: world models. The broad term is being used for several approaches, so what are world models, and what are their key ingredients?
See the summarized takeaways below, along with the full session video. Stay tuned for more video highlights each week and check out the full library of conference sessions on AWE's YouTube Channel.
Speakers
Matt Miesnieks, CEO, Stealth Mode AI Startup
Sharon Lee, Co-founder, Moonlake AI
Ahmed Ahres, Founding Team - Product & GTM, Reactor
Key Takeaways & Analysis
- AI's tremendous growth and advancement have been mostly confined to the web.
- The action has resided with large language models to process human knowledge.
- But greater anticipation lies in AI's potential to be unleashed on the physical world.
- This brings in the momentous areas of physical AI and world models.
- World models could do for the physical world what LLMs have done for the web.
- For example, training models on 3D models and spatial geometry can empower robotics.
- Rather than training industrial or household robots in physical spaces, it can be simulated.
- Advantages of this approach include lower cost and other bits-vs-atoms streamlining.
- World models also play into everything from gaming to filmmaking to industrial production.
- However, the magnitude of the opportunity is matched by its challenge and complexity.
- In other words, world models aren't easy to build, as the physical world is so nuanced.
- This can end up with big models that are challenging and expensive in compute terms.
- Matt Miesnieks also notes that there are too many "corner cases" in the physical world.
- This makes it difficult to build models that can reliably understand and predict outcomes.
- Technical approaches also vary in the field's early days, including 3D model creation.
- For example, there aren't enough 3D models to adequately train world models.
- Workarounds include creating 3D models out of 2D images (which are plentiful).
- Meanwhile, LLMs could be insufficient in their ability to understand the physical world.
- Miesnieks meanwhile champions a different approach, deviating from LLMs and imagery.
- Training models on concepts lets them extrapolate and scale physical-world understanding.
- This approach models itself after how the human brain understands and processes the world.
- There's of course a lot more to it, and the early-stage approaches continue to diverge.
- Despote question marks, one thing is certain: world models will be a massive area of development.
Want more XR insights and multimedia? ARtillery Intelligence offers an indexed and searchable library of XR intelligence known as ARtillery Pro. Learn more here.
