Turing Post

Can AI-generated worlds remember what happened off-screen? How do dual-agent setups reduce costs? In this episode of Attention Span, we break down the biggest shifts in AI’s ability to understand, reason, and act in the physical and digital world.

We dive into Alaya’s programmable world model, real-world vs. simulated robotics with Telexistence and Skild S1, Cognition’s Devin Fusion and SWE-2, and the heated frontier slowdown debate between Dario Amodei and David Sacks.

Hosted by Ksenia Se (founder of Turing Post). Covering September 7–13, 2026.

👇 Which topic should we cover in a deep-dive episode? Let us know in the comments!
Also, this is the first episode of Attention Span Weekly News – let us now if you like us to continue doing this. 

🔳 SPONSORSHIP
Partner with Turing Post to reach top AI engineers, researchers, and builders: ks@turingpost.com

🔷 WORLD MODELS
Alaya PWM paper: https://arxiv.org/abs/2609.10540
Demonstrations: https://alaya-lab.github.io/pwm/
World Labs - Atlas: https://www.worldlabs.ai/blog/atlas
World in World: https://arxiv.org/abs/2609.11548

🔶 ROBOTS & SIMULATION
AWS and Telexistence - DreamZero experiments: https://aws.amazon.com/blogs/physical-ai/bringing-a-frontier-world-model-to-the-convenience-store-inside-telexistences-dreamzero-experiment-on-aws/
NVIDIA - Skild S1: https://blogs.nvidia.com/blog/skild-ai-s1-physical-ai/
Skild’s S1 research: https://www.skild.ai/blogs/s1
Mila and Worldmodeldata: https://mila.quebec/en/news/worldmodeldata-and-mila-partner-to-prove-scaling-law-for-world-models

🔹 DEVIN FUSION & SWE-2
Cognition - Fusion: https://cognition.com/blog/local-fusion
SWE-2: https://cognition.com/blog/swe-2

🔸 PERSIMMON & PERSONAL AGENTS
Persimmon: https://persimmon.humansand.ai/blog/persimmon.html
Grok Bot: https://x.ai/bot
Meta Muse: https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/
Muse security architecture: https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse
Attention Span episode on Muse: https://www.youtube.com/watch?v=1pqzii7di7w

🔺 SAFETY, EVALUATORS & THE SLOWDOWN DEBATE
Anthropic - cybersecurity alignment assessment: https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents
Dario Amodei - We Must Pace the Frontier: https://darioamodei.com/post/we-must-pace-the-frontier
Hugging Face Open Alignment Initiative: https://www.techmeme.com/260912/p13
David Sacks’s response: https://x.com/DavidSacks/status/2098973625252708460

🔻 MORE FROM TURING POST
Newsletter, research coverage, and AI Builds AI: https://www.turingpost.com/
Instagram: https://www.instagram.com/turingpost_tv
TikTok: https://www.tiktok.com/@turingpost_tv

Subscribe for Monday news digests and Attention Span deep dives into how AI works.

#WorldModels #AIAgents #TuringPost

Meta just gave its new personal agent, Muse, a remarkable amount of freedom. It can browse, work across your accounts, write code, build tools and run subagents. But Meta made one part of the system deliberately difficult for Muse to control: its own authority.In this episode of Attention Span, I look inside Muse Secure VM and the security architecture surrounding the agent.

We get into Sentinel, the separate system that decides what Muse is allowed to do; why Muse can use your accounts without ever seeing the real credentials; how “tainted egress” changes permissions after a process touches private data; and why reading an email can give an agent considerably more power than “read access” suggests.

And somehow, all of this takes us back to computer-security ideas from 1975.
The larger question is one we are going to encounter everywhere as agents become more capable:
How much freedom can we give an agent to discover new ways of doing things without allowing it to expand its own authority?

*Watch it*, and tell me where you would draw that boundary.

*IN THIS EPISODE*
→ What “rogue agent” actually means technically
→ How Muse Secure VM isolates the agent
→ Why Sentinel sits outside the environment Muse can modify
→ How an agent can write a new tool without granting that tool permission
→ How surrogate tokens keep real credentials away from the model
→ What “tainted egress” means and why permission may need memory
→ Why “read my email” can unlock much more than email
→ Least privilege and complete mediation, 51 years later
→ Where Muse's security architecture still has unresolved problems
→ Secure from whom? The agent, other users, or Meta itself
→ DeepSeek Harness + Muse: capability can become fluid; authority cannot

*META MUSE*
→ Introducing Muse Meta's product announcement and overview of Muse Secure VM. https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/
→ How We Built Safety Into Muse The technical deep dive. This is where Meta explains Sentinel, Linux isolation, surrogate tokens, credential insertion, eBPF-based taint tracking and network egress controls. https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse

*THE SECURITY IDEAS BEHIND IT*
→ Melanie Mitchell: Misleading Metaphors and Real Risks A useful grounding discussion of what we actually mean when we say an agent “escaped” or “went rogue.” https://aiguide.substack.com/p/misleading-metaphors-and-real-risks 
→ Saltzer & Schroeder: The Protection of Information in Computer Systems The 1975 paper behind principles such as least privilege and complete mediation that suddenly look very current again. https://web.mit.edu/Saltzer/www/publications/protection/Basic.html 

*RELATED ATTENTION SPAN*
→ Everything Is a Plugin: DeepSeek Harness Our previous episode on agents that can create and modify their own tools. https://www.youtube.com/watch?v=jtyV7O4Pt0s 
→ AI Escaped? The OpenAI–Hugging Face Incident What actually happened when cybersecurity agents found routes outside their intended environment. https://www.youtube.com/watch?v=RGeZ2moLkIc 

*MORE FROM TURING POST*
→ We Don't Know What They Know Our deeper look at the problem of understanding what increasingly capable models know and how they will use it. https://www.turingpost.com/p/we-don-t-know-what-they-know
→ Turing Post https://www.turingpost.com 
→ Instagram @turingpost_tv https://www.instagram.com/turingpost_tv 
→ TikTok @turingpost_tv https://www.tiktok.com/@turingpost_tv 
→ Interviews: @realturingpost

What if a model finds an optimization that surprises the engineers who have spent years working on that system? What, exactly, has it understood?

I brought that question to OpenAI’s Phil Tillet and Matt Ferrari, whose work involves making AI cheaper and more accessible. They’re increasingly doing that work with the models themselves.

Matt talks about research ideas his team used to dismiss because the engineering would be too complicated. Now they can give a model years of earlier research and ask it to explore what might work. They’re using models to help improve the smaller models behind speculative decoding, including the training process itself. So I ask whether the model is also helping choose the ideas, and how much that expands what they’re willing to try.

I also ask something familiar to anyone who uses these systems: why does the same model sometimes feel different? Phil explains that even changing the order of floating-point calculations can introduce differences in its behavior. That puts a very concrete problem behind our conversation about understanding: an optimization can make the system faster and still change something you wanted to preserve. We get into how they catch those changes, and what happens when the failure is something nobody thought to test for.

*We talk about:*
When models became useful for engineering decisions.
Why OpenAI’s models needed more control than Triton gave them.
What happens inside the system after you send a request.
How model-assisted kernel improvements helped cut Sol’s serving costs by 20%.
Letting models investigate bugs independently, and deciding when to step in.
Why successfully optimizing something can still be a waste of effort.
Whether models need an internal representation of how a computer system behaves.
How AI assistance opens up experiments that engineers previously couldn’t justify attempting.

*Chapters:*
 
*Follow on*: https://www.turingpost.com/

*Did you like the episode? You know the drill:* 
📌 Subscribe here and here (https://www.turingpost.com/subscribe) for more conversations with the builders shaping real-world AI. 
💬 Leave a comment
👍 Like it 
🫶 Thank you for watching and sharing!

*Guests:* 
Philippe (Phil) Tillet created Triton, a programming language that makes efficient GPU programming more accessible. He joined OpenAI as an intern in 2019, before it had an API or a product, and spent years improving training efficiency. His interests extend from compilers and kernels to Bertrand Russell and philosophy of mind.
Matthew (Matt) Ferrari works on inference efficiency at OpenAI, across request routing, load balancing, debugging and speculative decoding. His fascination with optimization began in school, when GPU programming changed his understanding of how fast an algorithm could run. Today, he brings that curiosity to the entire system serving a model.

#openai #inference #optimization

NVIDIA was once so hostile to open source that Linus Torvalds gave the company the finger. Today, it maintains open Linux modules, releases hundreds of models and datasets, and has reportedly agreed to buy Hugging Face for $12.9 billion. WHAT?!

The change makes sense once we examine what NVIDIA learned from nearly dying with NV1, spending years searching for CUDA’s market, and watching researchers discover deep learning on gaming GPUs.

In this episode, we follow that strategy from NV1 and CUDA to NVIDIA’s reported $12.9 billion acquisition of Hugging Face, and ask whether the company has found the most profitable model for open-source AI.

*Watch it.*
👉 Subscribe for high-signal AI analysis
👉 Instagram https://www.instagram.com/turingpost_tv
👉 TikTok https://www.tiktok.com/@turingpost_tv
👉 More analysis: https://www.turingpost.com/
👉 Interviews: @realturingpost
Attention Span is here to explain the technical and business choices shaping AI.

Links:
NVIDIA FY2026 Form 10-K https://www.sec.gov/Archives/edgar/data/1045810/000104581026000021/nvda-20260125.htm 
Interview with Spencer Huang (Nvidia) https://www.youtube.com/watch?v=NEv9EnD7JVU&t=1231
Interview with Clem Delangue (Hugging Face) https://www.youtube.com/watch?v=DfJV722V1WY 
NVIDIA Open Source https://opensource.nvidia.com/en-us
NVIDIA on Hugging Face https://huggingface.co/nvidia
Reuters on the reported $12.9B agreement https://www.reuters.com/technology/nvidia-talks-acquire-hugging-face-13-billion-deal-business-insider-reports-2026-08-27/
NVIDIA Open GPU Kernel Modules https://github.com/NVIDIA/open-gpu-kernel-modules
Turing Post’s history of computer vision and AlexNet https://www.turingpost.com/p/cvhistory6

#NVIDIA #OpenSourceAI #HuggingFace #AI #GitHub

Demis Hassabis, Yann LeCun, Fei-Fei Li – they all talk about “building a world model.” Some of them are dedicating their professional lives to it!

But do they mean the same?

So before joining the World Models workshop at Chicago Booth, I wanted to answer a basic question: what do researchers mean by a world model, and how many different ideas are sitting under this name?

World models are absolutely fascinating area of research with its GPT moment still in the nearest future. 
This episode is based on the current research and provides a comprehensive overview of three broad approaches: generating future observations, predicting inside learned representations such as JEPA, and learning only what a planner needs to make decisions.

*Watch it.*
👉 Subscribe for high-signal AI analysis
👉 Instagram https://www.instagram.com/turingpost_tv
👉 TikTok https://www.tiktok.com/@turingpost_tv
👉 More analysis: https://www.turingpost.com/
👉 Interviews: @realturingpost

Attention Span is here to show you AI isn’t magic. Sometimes the best way to understand a model is to change the background to purple and see what breaks.

*Links:* 
Demis Hassabis on world models
https://www.youtube.com/watch?v=sZaM6MadDZU
Yann LeCun on world models
https://www.youtube.com/watch?v=8sS9UJzb_t4
Fei-Fei Li on large world models
https://www.youtube.com/watch?v=pNYVckbCFuk
Beyond LLMs: JEPA and the Road to AGI – the main milestones so far https://www.youtube.com/watch?v=z0fh0SY3VWc
stable-worldmodel https://github.com/galilai-group/stable-worldmodel/issues/153
VideoPhy-2, a benchmark https://arxiv.org/pdf/2503.06800 
Physion-Eval https://arxiv.org/html/2603.19607v1 
What Is JEPA? LeCun Architecture & World Models https://www.turingpost.com/p/jepa 

#WorldModels #AI #MachineLearning #YannLeCun #FeiFeiLi #DemisHassabis #JEPA #PhysicalAI #TuringPost #AttentionSpan