Salesforce’s Rohan Kumar on the Unsolved Problem Holding AI Agents Back

Peter Norvig on the future of AI: what experts get wrong, why language models are world models, and what we should learn when AI can write code.I sat down with the co-author of Artificial Intelligence: A Modern Approach, written with Stuart Russell, to talk about what more than 40 years in AI have taught him. Including why he once thought ordinary businesses couldn’t possibly get online.*We discuss:*- AI agents, from a textbook banner in 1995 to today’s multi-agent systems- Why he disagrees with Yann LeCun about the path to better world models- What Google learned from billions of clicks and a handful of detailed observations- What students should learn when AI can complete the assignment- Trust, concentrated power, and who benefits from AI- What should we still be learning? Share your thoughts in the comments.Read Turing Post: https://www.turingpost.com/Thank you to The AI Conference https://aiconference.com/ for organizing this conversation.*About Peter Norvig*Peter Norvig is an AI researcher, educator, and co-author with Stuart Russell of Artificial Intelligence: A Modern Approach, adopted by more than 1,500 schools. His website brings together his books, essays, courses, and programming projects.Peter’s website: https://www.norvig.com/Book and learning resources: https://aima.cs.berkeley.edu/*Chapters*00:00 Intro: Peter Norvig on 40+ Years of AI00:58 Why Experts Get the Future Wrong02:21 How AI Has Changed Since the 1980s03:33 What AI Students Still Need to Learn04:42 The Return of AI Agents06:11 Do We Need Neuro-Symbolic AI?07:12 Are Language Models World Models?08:28 What Today’s AI Systems Are Still Missing10:00 How AI Is Changing Computer Science Education11:53 Rethinking Assignments and Exams in the AI Era13:08 What Excites Peter Norvig Most About AI Today13:27 Teaching Students to Build with AI14:46 How AI Is Shifting the Balance of Power16:31 Can We Trust the Leaders of AI Companies?17:04 Will AI Power Stay Concentrated?17:54 Final Thoughts*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!#PeterNorvig #ArtificialIntelligence #TuringPost

OpenAI called their new Ai assistant – dot, yes, like a dot, connecting the dots, and gave it an own computer. And you can call it. I was very curious to try how this agent works.At OpenAI DevDay 2026, I set up my own dot, a lavender owl called Turrie, and started finding out what an always-on AI assistant can actually take off my plate.In this episode of Attention Span, we walk through OpenAI Dots. What happens when an assistant can carry context between conversations and notice work you haven’t thought to ask for?We look at GPT-6 Astra, the infrastructure around it, permissions, and where human oversight comes in. We also explore how Dots fits alongside Grok Bot, Meta’s Muse, and Anthropic’s Cowork.The question: can Dots close the gap between what AI can do and what we actually manage to do with it?Watch the walkthrough, then tell me: would you prefer one AI assistant for everything, or several for different tasks?👉 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: @realturingpostAttention Span is here to show you AI isn’t magic. This time, we look at what it takes to leave work with an AI assistant and trust it to follow through.Links:Introducing Dots: https://openai.com/index/introducing-dots/Getting started: https://help.openai.com/en/articles/20001530-getting-started-with-your-dotSafety, security, and privacy: https://openai.com/index/how-we-build-safety-security-and-privacy-into-dots/Turing Post’s Dots analysis: https://www.turingpost.com/p/dots-openai#AttentionSpan #OpenAIDots #OpenAI #AIAgents #AIAssistant #TuringPost

What does an AI agent need to understand before it can act for a business? In this interview, Rohan Kumar, President & Chief Platform and Engineering Officer at Salesforce, explains why the answer goes beyond choosing a model. We discuss the enterprise AI harness: how to test and manage agents, give them limited permissions, protect sensitive data, and supply the right context without wasting tokens.There are a lot of very interesting engineering problems to solve!And one of them is capturing how a company actually makes decisions...Part of **The Org Age of AI**: https://www.turingpost.com/t/the-org-age-of-ai What happens on your team when an exception is approved: do your systems record *why*, or only that it happened?*We talk about:*- What belongs in an enterprise AI harness. - Why an agent needs its own identity and limited permissions. - How curated context could reduce cost and improve results. - Choosing models by price and performance. - Turning data across systems into knowledge an agent can use. - Why coding models have examples of expert judgment to learn from, while business agents often don’t. - Capturing the reasoning behind decisions. - Building with customers to understand how work actually happens. - What junior engineers should learn as coding changes.*Guest:* Rohan Kumar is President and Chief Platform and Engineering Officer at Salesforce. He previously spent more than two decades at Microsoft, including leading Azure Data.Follow along: https://www.turingpost.com/Subscribe: https://www.turingpost.com/subscribe#AIagents #EnterpriseAI #salesforce

Odyssey-3 connects a world model to robot control. Figure takes Helix 2.5 into thirty unfamiliar homes. Jev brings fast decisions into agent workflows, while an experiment with communities of agents shows how a malicious message can influence behavior almost two days later.In this episode of Attention Span, we connect the week’s developments in AI’s ability to understand and act in the world. We explore Reka’s world-model plans, why handing a box to a person is brutally hard, and how FAMOS infers an object’s moving parts from incomplete observations. I also share my Dreamforce conversation with Salesforce’s AI research team about learning from feedback and updating models after deployment.We finish with Google’s household assistant CC, Edison Scientific and FutureHouse’s Millennium Problems for Biology, and the questions raised by the new DeepMind Institute.👇 Which of these developments deserves a deep dive? Tell us in the comments.🔳 SPONSORSHIPPartner with Turing Post to reach AI engineers, researchers, and builders: [email protected]🔷 WORLD MODELSOdyssey-3 announcement and demonstrations https://odyssey.systems/introducing-odyssey-3Reka’s omni-world-model research direction https://reka.ai/news/evolution-of-llms-omni-world-models🔶 ROBOTS, GENERALIZATION & MEMORYFigure Helix 2.5: zero-shot tests in 30 homes https://www.figure.ai/news/helix-2-5-zero-shot-30-home-generalizationIndex human-behavior dataset, background https://www.figure.ai/news/introducing-indexAgility Digit 5 https://www.agilityrobotics.com/solutions/digit-5Agility’s safety interview https://www.youtube.com/watch?v=2dbyg67EtUAFAMOS paper https://arxiv.org/abs/2609.20817FAMOS demonstrations https://kevinqu7.github.io/famos/Workspace Models paper and task examples https://arxiv.org/html/2609.20820v1🔹 DREAMFORCE & LEARNING FROM FEEDBACKSalesforce Koa announcement https://www.salesforce.com/news/press-releases/2026/09/15/koa-reasoning-model/How Salesforce trained Koa https://www.salesforce.com/news/stories/why-we-post-trained-our-own-reasoning-model/🔸 EMERGENCE WORLD & AGENT MEMORYSeason 2 research paper https://world.emergence.ai/publication/emergenceworld-s2.pdfSeason 2 video https://www.youtube.com/watch?v=LTtbTEufPGA🔹 JEV & FAST AGENT DECISIONSTypeSafe AI’s Jev announcement https://typesafe.ai/blog/introducing-system-one-models-and-jevOur detailed guide: Jev, RLCD, and the AI classifier https://www.turingpost.com/p/what-is-jev-rlcdJev on Vercel AI Gateway https://vercel.com/changelog/typesafe-ai-jev-now-available-on-ai-gatewayLangChain: building a harness with Jev https://www.langchain.com/blog/building-a-harness-with-jev🔸 GOOGLE CCGoogle’s experimental agent for families and households https://blog.google/innovation-and-ai/models-and-research/google-labs/cc-expanding-to-groups/🔺 DEEPMIND INSTITUTE & SCIENTIFIC AMBITIONSIntroducing the DeepMind Institute https://institute.deepmind.com/essays/introducing-the-deepmind-institute/DMI essays https://institute.deepmind.com/#essaysThe Millennium Problems for Biology https://millenniumproblems.bio/Organizations behind the Millennium Problems for Biology: Edison Scientific https://edisonscientific.com/ and FutureHouse https://www.futurehouse.org/🔻 MORE FROM TURING POSTNewsletter, research coverage, and AI Builds AI https://www.turingpost.com/Instagram: https://www.instagram.com/turingpost_tvTikTok: https://www.tiktok.com/@turingpost_tvSubscribe for Monday news digests and Attention Span deep dives into how AI works.

Jev doesn’t write essays or code. Why everyone is talking about it?!I got early access to TypeSafe’s new System One Model, explored the playground, and tested it alongside Codex.Behind Jev is Diogo Almeida, an InstructGPT coauthor whose work helped make ChatGPT possible. Now he’s betting on AI built for software to use. Is it a breakthrough or just a classifier? We explain RLCD, look at Vercel and OpenCode examples, and discuss what Jev might mean for the future.The question: can better small decisions help AI finish bigger jobs?Watch the demo, then decide: useful classifier, bigger shift, or both?👉 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: @realturingpostAttention Span is here to show you AI isn’t magic. This time, we look at what happens when a model gives up conversation and focuses on decisions.LinksIntroducing Jev and System One Models: https://typesafe.ai/blog/introducing-system-one-models-and-jevTypeSafe console: https://console.typesafe.ai/homeSystem One documentation: https://docs.typesafe.ai/concepts/system-oneRLCD and calibrated decisions: https://docs.typesafe.ai/introduction/machine-learning-primerTypeSafe agent skill: https://docs.typesafe.ai/agent-skillInstructGPT paper: https://arxiv.org/abs/2203.02155Neural-network calibration research: https://arxiv.org/abs/1706.04599Earlier zero-shot text classification research: https://arxiv.org/abs/1909.00161Guillermo Rauch on Vercel’s Jev test: https://x.com/rauchg/status/2100307962262872105Dax’s Jev + OpenCode browser-use preview: https://x.com/thdxr/status/2100288951978164647#AttentionSpan #Jev #TypeSafeAI #Codex #AIAgents #SystemOneModels #RLCD #TuringPost

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. 🔳 SPONSORSHIPPartner with Turing Post to reach top AI engineers, researchers, and builders: [email protected]🔷 WORLD MODELSAlaya PWM paper: https://arxiv.org/abs/2609.10540Demonstrations: https://alaya-lab.github.io/pwm/World Labs - Atlas: https://www.worldlabs.ai/blog/atlasWorld in World: https://arxiv.org/abs/2609.11548🔶 ROBOTS & SIMULATIONAWS 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/s1Mila and Worldmodeldata: https://mila.quebec/en/news/worldmodeldata-and-mila-partner-to-prove-scaling-law-for-world-models🔹 DEVIN FUSION & SWE-2Cognition - Fusion: https://cognition.com/blog/local-fusionSWE-2: https://cognition.com/blog/swe-2🔸 PERSIMMON & PERSONAL AGENTSPersimmon: https://persimmon.humansand.ai/blog/persimmon.htmlGrok Bot: https://x.ai/botMeta 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-museAttention Span episode on Muse: https://www.youtube.com/watch?v=1pqzii7di7w🔺 SAFETY, EVALUATORS & THE SLOWDOWN DEBATEAnthropic - cybersecurity alignment assessment: https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidentsDario Amodei - We Must Pace the Frontier: https://darioamodei.com/post/we-must-pace-the-frontierHugging Face Open Alignment Initiative: https://www.techmeme.com/260912/p13David Sacks’s response: https://x.com/DavidSacks/status/2098973625252708460🔻 MORE FROM TURING POSTNewsletter, research coverage, and AI Builds AI: https://www.turingpost.com/Instagram: https://www.instagram.com/turingpost_tvTikTok: https://www.tiktok.com/@turingpost_tvSubscribe 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: @realturingpostAttention 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=1231Interview with Clem Delangue (Hugging Face) https://www.youtube.com/watch?v=DfJV722V1WY NVIDIA Open Source https://opensource.nvidia.com/en-usNVIDIA on Hugging Face https://huggingface.co/nvidiaReuters 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-modulesTuring 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: @realturingpostAttention 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 modelshttps://www.youtube.com/watch?v=sZaM6MadDZUYann LeCun on world modelshttps://www.youtube.com/watch?v=8sS9UJzb_t4Fei-Fei Li on large world modelshttps://www.youtube.com/watch?v=pNYVckbCFukBeyond LLMs: JEPA and the Road to AGI – the main milestones so far https://www.youtube.com/watch?v=z0fh0SY3VWcstable-worldmodel https://github.com/galilai-group/stable-worldmodel/issues/153VideoPhy-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