In this episode of the Weekly Tech Update, host RAFIA interviews guest Muhammad Umar to explore a range of emerging AI trends and their broader implications. They begin by discussing DeepMind’s recent study on multi‑agent AI systems, which shows that without a central orchestrator, error amplification can grow 17‑fold, whereas centralized coordination reduces it to about 4.5‑fold, highlighting a coordination bottleneck similar to human team dynamics. The conversation then shifts to model performance, noting Anthropic’s claim of a 2% month‑over‑month improvement on the Software Engineering (SWE) benchmark and DeepMind’s new “Facts” benchmark that aims to evaluate genuine coding ability rather than pattern‑matching. Umar next examines AI‑driven automated material discovery, citing Google’s UK lab focused on semiconductors and solar cells and drawing parallels to AlphaFold’s impact on protein folding. He emphasizes how AI can predict material properties without expensive lab equipment, accelerating hardware innovation. The discussion moves to user interfaces, arguing that AI chat is becoming the primary health advisor and that UX is collapsing into natural language, making chat the universal UI for both consumer and technical products. Umar then highlights the application of software intelligence to physical infrastructure, using Palantir’s ShipOS example that reduced a 200‑step Navy workflow to 12 seconds, illustrating how software can compress manufacturing cycles to machine‑clock speeds. A major theme is the shift in compute economics: memory and data movement now dominate cost over raw logic on GPUs, reshaping hardware design, cloud economics, and national strategy. He connects this to energy considerations, noting the massive power needs of AI data centers and proposing orbital compute as a long‑term solution—space‑based data centers benefit from continuous solar intensity, natural cooling, and vacuum‑based laser communication, potentially alleviating Earth’s energy constraints. Finally, Umar reflects on the broader vision of AI in space, referencing Elon Musk’s comments on AI for orbital applications and the growing satellite infrastructure (e.g., Starlink), suggesting that within a decade AI‑powered space compute could become mainstream. The episode concludes with a recap and thanks.
Chapters
Hi everyone. Welcome to the code Ninja spaces. Today we have really great and interesting news to share with you guys. So starting with AI is hitting the coordination limit. So there's a lot of talk about multi agentic AI system replacing peels. So but recent research suggests that these things are becoming bottlenecked. So what exactly did DeepMind discover here?
So it's a recent report and what DeepMind report is showing that a multi agent system is when multiple AI agents work together, each agent handles a subtask and collectively they try to solve a complex problem. In theory this should scale linearly and in practice it doesn't. So DeepMind found that when agents coordinate without a central authority, small mistakes compound across agents causing errors to amplify over 17x. And when coordination is centralized, what it means is when agent or controller orchestrates the entire system, error amplification drops to 4.4 or 4.5 acts. So what it means is in agentic systems, if you have a centralized authority or an orchestration layer that helps you achieve your sub goals better with agentic system than if you don't have those, if you don't have that agentic centralized agentic orchestrator. So that's what this report shows and that's, I also feel like that's pretty much what how humans operate like someone probably is incredibly good in achieving something individually, but that not necessarily mean she or he is good in coordinating the entire team and managing the team and then achieve the final goal. So probably these digital beings are following the same mantra. So yeah, that's what this report shows.
So perfect. Moving on. Models are improving faster than benchmarks. We keep hearing that AI models are getting better, but this time anthropic shares that 2% per month. Their models are improving against SWE benchmark. These are, this is huge. So what's your take on this?
So first of all, Rafael Sweep benchmark or software engineering Benchmark is a benchmark that tests whether AI models can solve real software engineering tasks, not just toy problems. And here's something very important. A 2% improvement month. So generally what happens is we think of AI as one explosive achievement. But that's not, that's not happening. What's happening is the models are getting better every month. So for example, here, for example, it's, it's about anthropic. A 2% improvement month over month means a lot. When you see how this compounds over a year. And that's, that's, that's what here, what we are seeing is that models probably a year have this 2% improvement month over month. But last two years from last one year they have improved a lot. And if you just plot it for the next two years you'd see where it goes. So that's what DeepMind launch facts. It's a benchmark that designs to test what models actually know, not what they can pattern match their ways to passing tests. So this is about facts is about we. So when we talk about sweep we just have these examples where we just pattern match these models to see if they can solve certain problems not but there's three benchmark was not actually designed to see whether the models can perform real coding tasks. So Now Facts by DeepMind is actually a step towards that. How can we evaluate that these models can solve and actual problems, actual software problems. Think of this like GDP valves. So you know, I mean in the last month we did something on GDP valves where OpenAI introduced how can we measure models performance in actual GDP growth. That's like it's a deep mind fact is something like that but specific to three benchmarks.
Perfect. So we talk a lot about software but science is being automated at the hardware level as well. So Google defined building an automated material discovery lab in uk. Focus areas are around semiconductors, solar cells. So I feel like personally, personally this is a huge shift like AI helping scientists to AI doing science. So what do you think? Is this a fair for me?
You know. Yeah, I think it's accurate, not the fair I think I mean, you know this is my passion. I mean applying into the physical world is something I'm very passionate about. Like for example how certain material works. Now when we talk about material formation, I mean one way to look at is like you use let's say X ray crystallography or some of these techniques, these incredibly sophisticated techniques to understand how certain materials perform in certain conditions, in certain weathers. But with AI, I mean you don't need these sophisticated apparatus. What you can see is you can predict that how this material would have performed. So for example think about protein folding. I mean there's no real. We currently don't know how exactly these protein folds but our guess with Alpha Fold 3 is so good that it's closer to the actual reality. So that's, that's what basically the lens to see when we apply these models into material design. I mean and by the way material design is incredibly important. For example space flight, rocket design, chip design. So then, so this, this, I mean I know DeepMind has a history but I'm pretty sure that that's exactly the lens DeepMind is applying here, even if you see that the term they've used hypothesizes material. So how to, what hypothesis can we take about this material? Right? And then AI is very good at it. It's very good in predicting what probably has been happening under the surface in this material. So that's, I am very excited about this.
That's huge. Like from software to hardware, you know, we really often talk about taking bet against ui, that chat is the new ui. So there was something really interesting. What does it mean when people start using AI chat as a primary health advisor?
So, so if you see, I mean, let's take health. I mean you talk about healthcare or health, think about it, health. I mean when you talk to your doctor, that's still chat, right? Most of the 80% of the, our interaction with the doctors is chat. Right. So I see people talk a lot about that. It's, I mean, it's just a chat bot. I mean your doctor who you pay for example $100 just to visit or if you talk about Pakistan, 5,000 rupees just to see whether what's happening in your, let's say with your skin or not, it's still a chat, right? So our interaction with humans or with other intelligent species is a chat. So in which AI, AIs preferred UI as a chat is a huge, huge moonshot, I would say. And it's not about just replacing, let's say doctors, it's about being the first line of reasoning. And if you talk about for builder, this means UX is collapsing into a language. Just think about it. I would again say it. UX is collapsing into a language and products that don't speak. Honestly Rafia, I mean you've been working on hyper for so long, right? Yeah, I mean, you know, I asked you like the more voice you would include in the product development, the better it would be. So products that don't speak wants to write mock way words they want to write. Voice is the language, is the interface and chat is the interface. Even for technical products, that, that's, that has to be there. So again I would say this is so, so important again because I know you're involved in UX Circle. So again I would say UX is collapsing into a language and products that don't speak once a while, even those.
Designers do fight with us regarding that. But still. So we talked about software, UI hardware. So let's move toward physical infrastructure. Physical infrastructure is actually moving at the speed of machine. I guess this is a huge shift like from software Optimization to the physical world. Optimization, how we look towards that?
I think one of the biggest shifts we would see is building AI systems for the physical world. So for example, if you know Palantir shipos reduced a 200 on Navy production workflow to 12 seconds. Just think about it. I mean at the end of the day it's a software. When we talk about software, what is software? Software is intelligence, compression and correlation. Right? I mean imagine you with, without a software, what does it take to build a ship? Let's say it takes a year and with the software, let's say it takes a month. Do you think it's a good, it's a good idea to deploy software for ship manufacturing? Of course what Palantir has shown us is like that is incredible idea that you can, I mean just imagine reducing a 200 Navy production workload to 12 seconds. That's what software for physical world means. Both in not. I mean of course Palantir involves mostly in defense workflows, but apart from defense too. So I think this means your supply chain, your defense, your healthcare are being re timed to machine clocks, not human ones. So countries that adapt been consisted on of course they would become dependent.
These things are taking us towards the capital stack as well. This is a big statement to say that silicon is eating the capital stack. What does it mean when memory costs more than logic on GPUs?
So I mean what it means is compute economics have flipped. Of course we used to optimize logic, now we optimize data movement and this means hardware design, cloud economics, national infrastructure strategy. So yeah, I mean that is so what the real equation to look at it is like as Elon says, like intelligence produced per gigawatt of energy spent. So at the end of the day the real matrix to look at is I think. In your last podcast I've also shared that the true lens to look at it is like with energy it's an important part software and knowledge work. These are the three levers to drive meaningful growth. Not just chips alone, not just software alone. They just would work across and in, in collaboration to make sure that we would be able to solve some important problems. So that's, that's, that's, that's what it means. Jensen also talked about this a lot. I mean for example Tesla is using Groq to design chips now think Grok is a software and chips, I mean hardware. So it means applying AI to the physical world. Oracle is investing a lot. I think there's more than $500 billion of a backlog that is Oracle is investing in chips. Super micro. I remember super micro. And Xai is building gigawatt scale data centers in Memphis. And when we talk about data center Rafa, what it means is to power a data center a small nuclear reactor is required. Just think of this, this data. So that it's the amount of the. The capital that is going into hardware and to software into knowledge work is just insane.
Do I do agree more in GPUs means more energy, more resources. So what does climate and geo energies are we now engineering the planet itself. I'm curious like where we are going.
So think about it. I'll give you one. I mean of course as I said, it's intelligence. The real equation is intelligence produced per gigawatt spent. That's the equation. So intelligence is related to the energy spent. And what is intelligence? If you ask what is intelligence? Again I would quote Elon. It's information compression and correlation, right? So we need a lot of energy. Think with a, with a nuclear reactor, small nuclear reactor that powers a data center. You can power 1.2 million homes with that. Right? So it means we need a lot of energy to power these data centers. But here's my take on this on energy. Because there's a lot of people who oppose this much of investment in data centers. And the idea is that probably we don't have that much energy for our households. And then why we're spending that much into data centers and billionaires, they're doing it. But let me share one thing with you. So you know, orbital compute. We are bringing data centers into space. Think about it. Space has lots of energy. There is. Sun is 70% more intense in space. This right? It's 24 hours. There's no. You don't. So it means sunlight is available 24 hours. There's no night. So it means there are no batteries, right? And then you know, why do we need this energy here on Earth for the data centers? The reason for that is we need. We need cooling system, right? So we need a lot of energy to cool down these data centers in on space all you need is a radiator at the back of the data centers and you don't need a cooling system. So that's that space is absolute zero already. And also what is the. What. What's the last thing you need in space is. I mean what's lasting you need for. For building these data centers is like how you going to connect these clusters is currently what you do is like fiber optics, which is basically laser in a You know, in a fiber, right? This is on planet Earth. This is the fastest way to communicate within the racks. Now what is the way? But is, is there a faster way even with the fiber optics through which we can communicate tracks? Yes, it is laser through vacuum and that space is vacuum. So I can assure you that if it's an incredibly hard problem, but maybe in next few years if we are able to set up orbital compute, mark my words, we would have this problem of energy on the planet would be solved. No one would be, I mean would have excess energy because that was the first time we'll be harnessing sun energy and powering all that to power these intelligent species. So I think yes, in the shorter run we need energy to power these data centers. So for example, if you see lots of energy is going in Texas. I mean we're building lots of them in Riyadh because they have excess energy. They can produce lots of energy through oil. But real capex is going to space. So yes, in the short run there's a lot of energy demand. In the longer run, I don't think we have energy shortage. We have access of energy. And I would also like to share one important scaling card. Ashal Scaling. What it means is so how do you measure whether civilization is how advanced a certain civilization? So Kardashev was a Russian scientist and he had this like the best way to measure it is like how much of the planet's energy has been put to a useful work. That is basically a scale. You can tell whether the civilization is advanced or not. I think if you see that energy is the most important thing and now we are putting that into useful work. I guess that's my take on energy. So I'm not, I'm not on in on. I'm not looking into from that Doomer's lens that we have, we have less energy, we have a lot of energy.
Oh, that's great. Was listening to Elon few days ago one of his summit in Riyadh related to any investment. He was saying we soon going to have AI from trace, like AI for space as well.
And you would be using it. I mean, yeah, I mean what it means is like you would be processing major intelligent operations in the data center and you'll be communicating directly from space to your device Now. I mean it's nothing. I mean, I mean it looks crazy but I mean, you know, Starlink, I mean it is in more than 140 countries I guess and most of the people who are using it is basically they've been connected to satellites. So I know it looks crazy, but if you know how we are moving towards space, what's the progress? I think it would be crazy in the next 10 years if you don't, if we don't do it, it's going to happen.
So far it's a great conversation talking with you. So that's it for the Today. Thank you. We'll see you next week.
