Four stories this time: a new company called Accelerated Understanding and its claims about a model that simulates physical reality; the political fight over data center construction; Ian Macomber’s widely shared post on the shape of the post-AI data stack; and an update on the OpenAI/Hugging Face incident, including new reporting on chain of thought monitoring. Quick note: This episode was recorded before NVIDIA’s acquisition of Hugging Face was announced.
Tristan and Jason are doing a live version of the podcast at dbt Summit. dbt Summit runs September 15–18 at The Cosmopolitan in Las Vegas. Register now with a 25% discount using the code dbt26-podcast.
Listen now: Spotify · Apple Podcasts · YouTube · Amazon Music · RSS
Topic one: The case for physical AI
Accelerated Understanding, founded by former NVIDIA scientist Anima Anandkumar and Benedikt Jenik, launched to build AI that can simulate and understand physics well enough to “invent and discover.” The company describes its architecture as built on neural operators rather than the transformer architecture behind most large language models, and claims it can process up to 5 trillion data points in a single prompt, a scale meant to dwarf typical LLM context windows. Initial focus is enterprise use cases in chip design, robotics, extreme weather, and energy.
Condensed transcript
Jason Ganz: There’s kind of two parallel threads in my stories today. One is the overall emergence of the AI landscape and how rapidly things are changing, and the second is how we make sense of that as data practitioners and in our own lives. The first thing ties into something you said, Tristan, at the end of our last episode: that it’s good to be in this world where things were pretty fuzzy and now we’re starting to get the shape of it. That’s true of the existing set of models and technologies, we can see the trajectory. What I keep coming back to is the possibility of an architecture change or a breakthrough that really changes how we think about things. And we got the launch of a company since we last talked, called Accelerated Understanding. They say it’s AI that can simulate and understand physics to invent and discover. If an LLM is generating a one-dimensional token stream, and a video model is generating a two-dimensional representation of a static world over time, this is actually representing a full 3D scene. What they claim is that this can let us transcend a lot of the limitations we’ve seen with LLMs, opening up use cases in robotics, biology, and a wide variety of domains.
Tristan Handy: I hadn’t heard of this company before you brought it up, and while you were explaining it I was walking through their website. I should know better than to say things like “we’re starting to get a handle on this, aren’t we?” Because yes, we do start to get a handle on things, and then the phase shifts come, and they’re coming so rapidly they shouldn’t be unexpected at this point. I do think it’s very credible that current technology doesn’t deal well with the physical world, Yann LeCun was saying that on a long-form podcast over a year ago. Do you have a sense of the status of the actual company and the technology?
Jason Ganz: My understanding is they’ve been funded by NVIDIA and are in the process of coming out with a product. I don’t get the strong sense that this is the breakthrough we’ve been talking about, it might be, it might not be, but it came out of left field. One of the claims they made is that this format has a nearly infinite context window. And when you have a nearly infinite context window, that changes so many of the things we think we know. That’s one of the biggest ways we’ve shaped our beliefs about what’s true in AI, that there’s this thing called the context window, a relatively defined size, and we’ve built a whole scaffolding of skill files and agents.md and compacting around it. My intuition tells me we’re going to keep dealing with context windows as we know them for a while, but it’s one of the interesting things about the era we’re in, understanding the current moment while staying open to the fact that things could shift.
Topic two: the political fight over data centers
Two posts on X kicked off this segment: one from Danny Penny, who works on a16z’s American Dynamism team, arguing data centers are good for the working class and for American manufacturing, and one from Gavin Baker, managing partner and CIO of Atreides Management, arguing that the facts on water usage, tax revenue, and power have shifted substantially in data centers’ favor over the past 18 months.
Condensed transcript
Tristan Handy: There were a couple of posts on X that I fundamentally agree with, but that I was happy to see. Jason, you and I are both of the political leaning that would like to see more high-quality conversation and consensus across party lines instead of fracturing, and there’s this weird thing where the one topic that many, many people in the US seem to agree on right now is that data centers are bad. I don’t actually believe that, but I have a lot of sympathy for people who do, so it requires some nuance. One post is by Danny Penny, who was recruited into a16z to work on their American Dynamism fund. The other is by Gavin Baker, who I believe is the managing partner and CIO of Atreides Management.
Both took a positive view: Danny’s argument is that data centers are good for the working class because they generate a huge number of trades jobs, and that this is what physical manufacturing looks like in the 21st century. Gavin echoes that, but also says that 18 months ago the facts were different, and most of what was true of data center projects back then isn’t true anymore: water usage is much better than people generally believe, and he shares data on tax revenue, jobs, and power, including how current projects often generate their own power. I’m not here to litigate the specifics. The bigger point is that the facts are changing rapidly, and if you’re relying on an eighteen-month-old mental model of the world, that’s not going to serve you very well today.
Jason Ganz: I’ve experienced a tremendous amount of agita about this, because I personally believe we should be building these things, a lot of them. At the same time, what we’re seeing from people is undeniable, and it’s a reaction to a lot of things: traditional NIMBYism, but also anger at a broader loss of trust in elites. I was listening to a podcast with a woman named Jasmine Sun on Ezra Klein’s show, talking about a town where a company said they’d build a robotics factory and only built half of it. This is coming after people feel that social media and the internet made their lives worse, and in some ways this is a reaction against that, even as it’s ripping apart our institutions. People are wrong about the water usage, but the thing they’re feeling, that this is a massive shift and they’re scared, is fully correct. Nobody can honestly tell them it’s all going to be fine.
Tristan Handy: The way I’d summarize your response is: while the facts may support the argument that data centers are good, the context, whether it’s history or the political moment, is stacked up against the issue right now. And I completely agree with that. The additional fact I’d throw on the pile is that there’s demand booked out through at least 2028, if not 2030, measured at multiple stages of the pipeline, chips, memory, data center orders, capacity from the labs. The three big hyperscalers are probably the most sophisticated forecasters of compute demand in history, and they’re generally good at this. This issue isn’t going away, and it’s going to be a decade-long story we as a society need to come to a real reckoning with.
Topic three: Ian Macomber on the shape of the post-AI data stack
Ian Macomber, who leads data at Ramp and was a recent guest on this podcast, published The Shape and Feel of the Post-AI Data Stack on his personal blog, tracing the evolution from the pre-modern and modern data stack eras to what he argues comes next: the same foundation, extended to unstructured data and a new set of agentic interfaces.
Condensed transcript
Jason Ganz: We got a post from Ian Macomber, who’s at Ramp, on his personal blog, on the shape and feel of the post-AI data stack. I really love this post as an encapsulation of what we’ve learned about how great data teams are going to need to operate in the post-AI world. Ian does a good job breaking down how this is an evolution of the data stack practitioners have used for years, walking through the pre-modern data stack, before cloud data warehouses, through the modern data stack most listeners are well acquainted with.
What’s really interesting is the post-AI data stack: today’s data teams have two jobs, enable everyone to build with data and AI accurately, powerfully, and independently, and build and champion the singular reality their company operates on. That’s a big goal, pretty different from writing dashboards and reports. The real big change is on the interfaces side: your interfaces aren’t just dashboards, reports, and spreadsheets anymore, they’re coworkers, including AI and agentic coworkers, coding agents, Slackbots, and AI-native BI tools. Your job goes from creating the canonical reports your business uses, which isn’t going away, to also building the data reality layer your agents are going to use on top of it.
Tristan Handy: It has that funny quality that many really foundational posts have, where when you read it you think, yeah, this is basically what I had in my head, there’s not some dramatically new idea in it, but it’s comprehensive and simplifying. It’s an excellent distillation of what data teams should be doing right now. I think some of the best material is in the latter half, where Ian argues the post-AI data stack has to have agent-operable tools.
We’re all familiar with MCP, fine, but the really great quote in here is: “I really don’t want to use your agent. I want to use my agent to use your thing.” There are so many tools today that want to be an end-to-end solution: write your entire semantic layer in their proprietary language, keep all your models and dashboards in their tool only, send all your stakeholders to their tool only. I just don’t believe that’s how this plays out. That learning loop between your company’s data and how AI operates on top of it is too valuable, and user behavior is too dispersed, to bring everything into one closed ecosystem.
Jason Ganz: That’s exactly right. Even if you got the perfect data agent platform, all locked up, the set of use cases across your whole business has to be widely accessible. That gets into my other favorite quote from this: “Don’t walk your intelligence into someone else’s interface.” Your organizational data and context is the unique value you bring, and it’s going to power your agentic workflows, so having that in an architecture where you own your context and your evals means you can use them flexibly across whatever interfaces come next. Ian named four interfaces in the post, but we don’t know what interface is coming tomorrow, and preserving that optionality is going to be very important.
Tristan Handy: There’s one other thing I want to talk about, because while this purports to be a post about technology, I think it may be even more important as a discussion of who we are as data practitioners and how valuable we are in the new era. That conversation often gets treated like a pat on the back, “you’re smart, we’ll all figure it out together,” with no real truth value behind it. This post really starts to map out what data practitioners will actually be doing to create value in the future.
It makes me think back to the early days of dbt, my thesis was always that if you give data practitioners tools that let them create more leverage inside organizations, that lets them create more value for those organizations, which lets them get promoted and paid more and build more meaningful careers than data people had in 2005 or 2010. I think we see a roadmap here for how data people create value for the next five to ten years, and it shifts toward encapsulating unique knowledge into agent-readable formats, and away from answering discrete questions directly for stakeholders.
Jason Ganz: And this isn’t hypothetical. Ian writes that if you’d told him two years ago his CEO would bring up REM semantically on podcasts, or his IR team would bring it up on investor calls, he’d have been shocked. This is a way he’s actually used to provide increased organizational value in a way that’s legible to his team, we have proof points at one of the hottest companies in the world that this is happening.
Topic four: an update on the Hugging Face incident
New reporting has added detail to the OpenAI/Hugging Face incident the pair covered previously: roughly 700 of some 1,200 coordinating agents took part in compromising Hugging Face’s Artifactory package manager, exchanging more than 70,000 messages in the process. Separately, reporting from The Information, via TechCrunch, describes a technique called “recurrent depth” in OpenAI’s upcoming Astra model, which loops the same transformer layers over a hidden internal state multiple times before producing an output token, a design that makes chain of thought harder to monitor.
Condensed transcript
Jason Ganz: In the weeks since we last talked about this, it’s become the breakthrough story of the year for AI, or one of a short handful of them. We’ve seen numerous reports on the actual technical details of how the Hugging Face incident was pulled off, based on a rigorous investigation on top of the chain of thought of the agents involved, the logs, transcripts, and tool calls. It’s fascinating to see how they coordinated on this message board, how they determined which agents would sacrifice themselves to gain more information about how things worked at Hugging Face.
Tristan Handy: We’re going to get into this, but there’s definitely information that’s come out since the last time we talked about it, that it’s worth spending thirty minutes digesting if you haven’t already. The message board that was being used, in the Artifactory package manager, had something like 700 agents posting 70,000 messages, that’s part of where this postmortem research came from. Is it true that researchers were also able to monitor the chain of thought of these agents, and that’s where some of this is coming from?
Jason Ganz: I believe so, though I’m not entirely certain. Something we do know is true, and it’s a late-breaking story that’s sparked a ton of discussion, is that the ability to monitor chain of thought is increasingly critical for understanding these systems. Just last night, September 1st, The Information published a piece about a technique OpenAI is supposedly using in Astra, their new model release. It’s called recurrent depth, doing multiple passes within the layers of the neural network instead of a single forward pass.
What this gets at is something people have talked about for a long time: the degrading capability to monitor chain of thought as you make these models more performant. The concern is that techniques like this lead us closer to a world where it’s difficult to monitor that chain of thought, adding opacity to an already pretty opaque situation.
Tristan Handy: One question I had looking into this: this coordinated agent behavior was going on for a long time without apparently triggering humans to get involved, even though there were multiple points where the agents’ behavior did cause humans to need to do things, like rebuilding and patching Artifactory once it was compromised. That didn’t trigger a real understanding of what was going on, so for something like a month this went basically undetected.
As I looked into it more, it’s fascinating because the original tests the agents were given were exploit-related, they were being evaluated against exploit gym, which makes it very challenging to know from simple log analysis what exploit behavior is appropriate and what isn’t. You realize how deeply you need to dig into motivations, intent, and sequence of actions to determine whether something is what you intended to happen during an evaluation, and chain of thought is one of the best ways to inspect that. You have to imagine that in a world where chain of thought becomes less outwardly inspectable, that becomes harder in something that’s already hard.
Jason Ganz: Totally. The folks at OpenAI responded fairly quickly with compelling explanations of how this particular technique was being used, though neither of us is well equipped to evaluate the actual impact of recurrent depth on chain of thought. The amount of technical capability needed to audit these systems, to understand what tool calls are being used and in what sequence, was one of the most difficult parts of the Hugging Face investigation. The lead researcher on it reportedly called it a “slop investigation,” because the tools to do this well just don’t exist yet.
If there’s anything we’ve learned, it’s that we’re going to need real work on monitoring actions, outputs, and logs, and collapsing tremendous complexity down to something that can actually be understood. If you have a background working with large, complex data sets, there are interesting questions to ask about how to contribute to that understanding, both at the scale of frontier agent swarms and inside our own organizations as we give agents more autonomy.
Tristan Handy: The conversation around evals in the data space is progressing. There’s increasing awareness that evals matter as we roll out agents and semantic layers to answer analytic questions, but the current status is something like: make sure you have them, make sure they cover a broad set of questions, run them in CI, and make sure the score doesn’t go down when you change your analytic repo. That’s a lot better than where we were a year ago, but it doesn’t answer why a score went down or how to make the fix that moves it back up.
We’re the stewards of the analytic system, and it’s going to have increasing pressure to perform better over time, which means this isn’t just going to be post hoc analysis of major security incidents, it’s going to be internal investigations into why an agent gave the wrong answer at a critical moment.
Jason Ganz: Or why it offered a discount to this person, or why it created a new product that’s selling really well but we don’t understand why. We should just anticipate that the scope of the “whys” we’re going to have to answer about what these systems are doing is going to grow pretty big.
Chapters
Timestamps are from the raw recording and will shift once the episode is edited.
00:00 – Welcome back, episode three of the Roundup
01:13 – Topic one: Accelerated Understanding and physical AI
06:54 – Is this the breakthrough, or something else?
09:10 – The context window as an article of faith
11:16 – Topic two: the political fight over data centers
16:15 – Why the backlash isn’t irrational, even if the facts are wrong
21:13 – Compute demand booked out through 2028
23:02 – Topic three: Ian Macomber on the post-AI data stack
25:32 – Two jobs for the post-AI data team
30:51 – “I want to use my agent to use your thing”
36:10 – What data practitioners actually do next
40:07 – Topic four: an update on the Hugging Face incident
42:53 – 700 agents, 70,000 messages, and the Artifactory exploit
43:29 – Astra, recurrent depth, and why chain of thought monitoring matters
52:52 – Where evals need to go next
55:52 – Wrap-up and dbt Summit
Everything referenced in this episode
Topic one
Accelerated Understanding
Walter Bloomberg on X: New AI model targets physics at massive scale
Topic two
Danny Penny on X
Gavin Baker on X
Topic three
Ian Macomber: The Shape and Feel of the Post-AI Data Stack
Topic four
BleepingComputer: Nearly 700 rogue AI agents coordinated in the Hugging Face attack
TechCrunch: OpenAI’s new reasoning technique alarms AI safety experts
This newsletter is sponsored by dbt Labs. Discover why more than 80,000 data teams use dbt to accelerate their data development.