Dec 11, 2024 · 55m · big-technology
AI Predictions for 2025: Geopolitics, Agents, and Data Scaling — With Alexandr Wang
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of the Big Technology Podcast, host Alex Kantrowitz interviews Scale AI CEO Alexandr Wang to examine his key AI predictions for 2025. Wang details the geopolitical race to export foundational tech stacks, the operational deployment of autonomous agents across military and consumer domains, and the tech sector's strategic pivot toward frontier data scaling.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Alex holds 31.9% of the talking time here. How this is scored →
Alex as informed peer 5.0 Guest teaching 4.9 Guest disagreement 1.4 Alex pushing back 2.3
speaking balance: gold is Alex, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 27:42 Wang dismisses separation of consumer and enterprise workflows
When the host pushes him to exclude business use cases from consumer agent predictions, Wang retorts dismissively that everyone works before proceeding with his point.
Hardest push from Alex ▶ 4:41 Kantrowitz questions battlefield utility of LLMsKantrowitz directly challenges Wang's thesis on military AI dominance by bluntly pointing out that a superior conversational chatbot will not secure victory in a conflict over Taiwan.
Biggest teaching moment ▶ 31:58 Wang introduces the dual-web architectural paradigmWang educates the host on how technical infrastructure must adapt to autonomous bots, introducing the structural concept of two parallel webs operating for humans and agents.
Alex holds their own ▶ 38:40 Kantrowitz details hardware scaling roadmaps from AWS and xAIKantrowitz demonstrates sharp industry expertise by quoting AWS CEO Matt Garman from re:Invent and contrasting it with Elon Musk's planned million-GPU cluster in Memphis.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Alex as informed peer | Guest teaching | Guest disagreement | Alex pushing back | Why |
|---|---|---|---|---|---|---|
| Geopolitical AI Rivalry and Global Export Dominance | 3 | 4 | 1 | 1 | Alex Kantrowitz invites Alexandr Wang to introduce his prediction on geopolitical AI rivalry. Wang outlines how international competition will center on exportable infrastructure in swing states like the UAE. The exchange is collaborative and introductory. | |
| Military Defense Applications and National Security Stakes | 5 | 6 | 2 | 5 | Kantrowitz pushes back by asserting that having a superior chatbot will not win a war over Taiwan. Wang counters with the historical precedence of military tech paradigms, citing drone warfare in Ukraine and China's rapid facial recognition rollout. | |
| AI Infrastructure as Cultural Power and Democratic Soft Power | 4 | 5 | 1 | 1 | Wang elaborates on the ideological implications of AI infrastructure as a cultural export, arguing democratic values must underpin global foundational models. Kantrowitz synthesizes this point under the rubric of American soft power. | |
| Domestic Regulatory Barriers and Open-Source Model Exploitation | 6 | 5 | 2 | 3 | Kantrowitz highlights the contradiction between China's state resources and its reliance on US open-source models like Meta's Llama in military research. Wang explains Beijing's startup clampdown, DeepSeek's rapid replication of OpenAI's o1, and US chip export restrictions. | |
| Global Infrastructure Alignment and 2025 Battlefield Agent Deployments | 6 | 4 | 1 | 2 | Wang predicts the deployment of AI agents in active combat operations during 2025. Kantrowitz links this dynamic to enterprise B2B data workflows and military supply chain management. | |
| Three Operational Pillars of Military Autonomous Systems | 4 | 6 | 1 | 1 | Wang breaks down military agent utility into three operational pillars: multi-source data ingestion, supply chain optimization, and lethal drone autonomy. Kantrowitz listens attentively as Wang structures the operational breakdown. | |
| Warfare Automation Risks and Strategic Deterrence Parallels | 5 | 4 | 2 | 4 | Kantrowitz expresses visceral alarm regarding autonomous lethal targeting, arguing that battlefield agent deployment lacks a self-limiting nuclear taboo. Wang counters with deterrence theory, hoping advanced capability deters wider conflict. | |
| Prediction Two: Consumer AI Agent Breakthroughs in 2025 | 4 | 5 | 1 | 1 | Wang delivers his second prediction, forecasting a 2025 ChatGPT moment for consumer agents driven by UI innovation and delegation of multi-step workflows. Kantrowitz facilitates the transition to consumer use cases. | |
| Designing Ubiquitous Context-Aware Personal Agents | 4 | 4 | 3 | 5 | When Kantrowitz insists on hearing personal consumer use cases rather than workplace email automation, Wang pushes back with a quick quip before outlining personal itinerary and holiday planning workflows. | |
| Healthcare Diagnostics, Enterprise Execution, and Web Defenses | 7 | 3 | 1 | 4 | Kantrowitz references insights from Mark Benioff and GE Healthcare regarding cancer history summarization, then challenges whether personal agents can bypass anti-bot defenses and captchas across the web. | |
| Architectural Division Between Human and Machine Web Systems | 3 | 6 | 1 | 1 | Wang responds to the bot security dilemma by proposing a bifurcated internet: a visual layer for humans and an underlying programmatic infrastructure for autonomous agents handling transactional utilities. | |
| Ethical Implications of Agent Automation and Social Friction | 6 | 4 | 1 | 3 | Kantrowitz raises ethical concerns about agents burdening human customer service agents and eroding human resilience, citing a Wall-E style future. Wang agrees and shares an anecdote about university admissions officers screening AI essays. | |
| Mid-Show Transition and Compute Scaling Hardware Limits | 7 | 5 | 1 | 2 | Kantrowitz cites firsthand reporting from AWS re:Invent and Matt Garman alongside Elon Musk's million-GPU cluster at xAI Memphis. Wang responds with his third prediction: compute scaling will hit walls without high-complexity frontier data. | |
| Generating Frontier Expertise and Hybrid Data Methodologies | 7 | 5 | 1 | 1 | Kantrowitz brings in Aiden Gomez's perspective on advancing from crowdsourced data to PhD annotators. Wang explains Scale AI's hybrid methodology of combining domain experts with synthetic data generation. | |
| Multistep Reasoning Reliability and Autonomous Research Horizons | 5 | 6 | 1 | 2 | Kantrowitz asks whether ingesting all human knowledge makes AI complete or if new discovery benchmarks are required. Wang details the progression from five-nines reliability to autonomous hypothesis generation guided by human advisors. | |
| Evaluating Multi-Agent Systems Against End-to-End Reasoning | 6 | 6 | 2 | 2 | Kantrowitz evaluates Moody's 35-agent voting architecture for portfolio analysis. Wang critiques pre-programmed multi-agent hierarchies in favor of unified end-to-end models learning dynamic problem-solving through trial and error. | |
| Rapid Fire: Hardware Clusters and Quantum Computing Horizons | 5 | 5 | 2 | 2 | In a rapid-fire exchange, Kantrowitz queries Wang about the outcome of Elon Musk's million-GPU cluster and Google's quantum breakthrough. Wang asserts that data remains the binding bottleneck and notes quantum computing's long-term utility for natural sciences. | |
| Rapid Fire: Benchmark Saturation and Frontier Lab Leaderboard | 3 | 5 | 2 | 2 | Kantrowitz asks Wang to crown a frontier lab champion for 2025. Wang refuses to pick a single winner, explaining that existing industry benchmarks are saturated and demanding more rigorous evaluation suites. |
Statements from this episode (28)
Wang: Incoming US administration will help US surpass China in AI
“I do expect that the new admin will come in and help accelerate things to enable the US to compete more aggressively with China and ultimately come out ahead on the technology.”
Wang: Biden administration forced UAE to choose Microsoft over Huawei
“I think one of the best examples of this was in the past year, the Biden admin posed to the UAE, hey, which way are you going to go in terms of AI technology? You can either go into the sort of Huawei China stack, or you could go into the Microsoft United Stat…”
Wang: Exporting AI stacks will define the next decades of geopolitics
“I think this is going to be One of the under the line battles that really defines the course of the next few decades of geopolitics.”
Wang: US must possess superior AI to prevail in conflict with China
“If you believe that there's some potential of some kind of conflict over Taiwan or other, some kind of other, like, hot conflict between the US and China then we really, the United States needs to ensure that we have the best possible AI technology to ensure t…”
Wang: Drone warfare in Ukraine is increasingly enhanced by generative AI
“By the way, I think that the drone warfare in Ukraine is becoming more and more Enhanced by generative AI and more advanced autonomy.”
Wang: US leads China in AI algorithms and compute, but data contested
“If we were to rack and stack versus China, we're ahead on algorithms. We're ahead on compute computational power, thankfully due to a lot of the export controls that the commerce department has put in place. And then on data, it's a little bit of a jump ball.”
Wang predicts China will deploy military AI faster than the US
“So, my expectation is that they will actually deploy AI to their military faster than the U.S., even though the U.S. Is ahead on the core technology.”
Wang: Capital flows into China's startup sector have collapsed precipitously
“One undeniable trend over the past, let's call it five years has been the sort of the collapse of the Chinese startup sector. And this is really driven by policies from the CCP to significantly, you know, they killed certain startup industries. They really lik…”
Wang: China's DeepSeek produced the first replication of OpenAI's o1 model
“OpenAI released O-one and released the O-one preview a number of months ago... Yeah, this is OpenAI's advanced reasoning model, which is great at sort of scientific reasoning and mathematical reasoning and reasoning and code, et cetera. And the very first repl…”
Wang: US chip export controls successfully hamper China's frontier AI development
“There is a very real hamper in a lot of their progress too, which is the chip export controls. And this has been an incredible effort, I think, from the U.S. Department of Commerce and the, you know, the Biden administration in general to sort of hamper the ab…”
Wang: Militaries will deploy AI agents in active combat in 2025
“20, 25 will be the year where we start to see several militaries around the world start utilizing AI agents in active warfighting environments to great effect.”
Wang: AI agent experimentation will yield more lethal autonomous drones
“And, you know, I think this is an area of active experimentation for a lot of militaries, but I think if you start to see that happen, then you're gonna, you will have more autonomous drones that are able to be more and more lethal, more and more effective. An…”
Wang: Military AI will function as deterrence and prevent conflicts
“My hope certainly is that while AIs Application into military is something that is, is very concerning and potentially extremely powerful. It is the sort of same overall effect, which is to ultimately deter more conflict than create it.”
Wang predicts AI agents will have a 'ChatGPT moment' in 2025
“Yeah, I do think that, yeah, I think that 20, 20 25 is really going to be the year where we start to see some kind of very basic primordial AI agents really start working in the consumer realm and creating sort of real consumer adoption. You know, another way …”
Wang: The Chat Paradigm Is Too Constrictive for AI Agents
“I mean, right now we're so stuck as a I think tech industry still on the sort of like chat paradigm and, you know having everything be a chat with one of these models. And I think that's a constrictive paradigm to enable agents to actually really start working…”
Wang: Ideal AI agents must passively observe all digital communication flows
“An ideal AI agent is one that, that I think is observing and naturally in all the sort of like core flows of information and core flows of context that you are in digitally. So, you know, it's in all your Slack threads. It's only your email threads that like, …”
Wang: Designing UX for imperfect outputs is 99% of the agent challenge
“Having a product experience, which Where you don't expect it to be perfect, but you expect it to be pretty good. I think that's like, 99% of the challenge, and that's why we haven't seen it yet, despite the fact that the models already can do a lot of this stu…”
Kantrowitz: GE HealthCare deployed gen AI cancer dashboards for doctors
“I was just speaking with GE Healthcare about how they've now put in dashboards for doctors sort of summaries of cancer patients, medical histories, which would run thousands of pages and the doctors never had a chance to read the whole history. And now the gen…”
Wang: The internet will split into human-facing and AI agent webs
“We will have to sort of fundamentally reformat how the internet works to be able to support it. And I think that like the you know, we're going to need in some senses, like there will be like two webs. There will be the web that that, that, that humans use whe…”
Wang: Consumer AI agent adoption will start with toy-like applications
“I mean, I think ultimately, I think agents are gonna start In an area that they'll feel pretty it'll feel like a toy, just like with any technology. So maybe you know, we'll all start with like a language learning agent, or we'll start with a cooking aid agent…”
Wang: In 2025, AI focus will shift to compute and data equally
“And I think what's going to happen in 25 is we're going to we're going to only be focused on who can create newer, better chips or bigger data centers with more chips, but also who can create newer and better data. And one of the things that I think we're goin…”
Wang: Pure synthetic AI training data has underperformed industry expectations
“One of the things that we've seen over the past few past year in particular is that synthetic data has not worked as well as I think everybody had hoped. You know, pure synthetic data, just using data generated from the models to try to train future models, th…”
Wang: Current AI models invariably fail when chaining multi-step actions
“One of the things that really is true in all the models today is that they're not that good at, you know at taking multi-step actions. Whenever it has to take a few hops, whenever it has to take, chain a few things together, it'll invariably make mistakes alon…”
Wang: AI will eventually formulate hypotheses, run tests, and conduct research autonomously
“Eventually It'll be able to start making its own hypotheses, running those tests on its own, and sort of ultimately making its own sort of discoveries or realizations or sort of conduct its own research.”
Wang: Dynamic single-model reasoning is superior to rigid multi-agent pipelines
“It's a very sort of regimented way to try to enable the systems to do multi-step reasoning. Cause ideally what you want the model to do is to just like how a human does, be able to sort of go through and figure out what are the bits of pieces it needs to know …”
Wang: AI development is currently more bottlenecked by data than compute
“I honestly think right now at where we are in AI development today We are more bottlenecked by data than we are compute.”
Wang: Quantum computing will be really impactful in 5 to 10 years
“I think quantum computing is on kind of like the way AI was back in 2018. It's on a few scaling laws where you can definitely sort of squint and see that, you know, in five to 10 years, this is going to be A really, really impactful technology.”
Wang: AI benchmarks are saturated, making model leaders hard to distinguish
“One thing that we see today with the models is that because all the benchmarks that were used today are what's called saturated, i.e., you know, in other words, like all the models do really well at the benchmarks, it's really hard to discern actually which on…”