As I skimmed the AI headlines from the past two weeks, one thing that jumped out was Alibaba’s Apsara Cloud event, which covered extensively. The headline takeaways: (1) token usage is doubling every two to three months; (2) Alibaba plans an additional ~RMB 150–200B in annual investment on top of the previously announced RMB 380B over three years; and (3) global data center power consumption could be 10x higher by 2033 versus 2022.

Taken together, this signals that BABA is going all in on AI across the stack, aiming to become a true full-stack AI services provider. For context, the big four U.S. cloud providers are collectively on pace to invest roughly US$400–$500 billion in 2025–2026. Scale still feels like what is in the headlines.

Subscribe now

Today is a long and thoughtful piece, as I’d like to think. So please stay with me.

Now, every day brings something new in this sector, and time feels slippery. Yet in my own house, time has slowed. With my newly arrived two-week-old at home, the rhythm is just feeding, diapers, and naps on repeat. But in only 16 days, she’s gained nearly a third of her body weight. Growth feels exponential when the base is small.

Share

Agentic AI Is the Future

Increasingly, I don’t see AI as “a tool” or merely “a tech thing,” and as I think about how AI will transform our society, it evokes Marc Andreessen’s famous essay, “Why Software Is Eating the World.” Now, the question is: how will AI practically eat the world?

If you’ve been following AI Proem, you know two threads have dominated my attention (obsession) recently: (1) launching the podcast, which features a stellar set of journalists, investors, and policy analysts covering China’s AI, and (2) trying to understand how agentic AI will reshape knowledge work. Greg Isenberg put it starkly: “What the app store did for software, agent builders could do for intelligence”.

The current underlying AI fundamentals remain very strong, and adoption is accelerating. While pre-training scaling has slowed, reinforcement learning (RL) is a genuine breakthrough, driving another step-function improvement in AI model capabilities.

At this point, it is clear that the current AI model capabilities are far ahead of applications. Some in Silicon Valley say that even if we pause frontier model development today, we probably have 10 years of application innovations ahead. But as application proliferation no longer needs to rely on pre-training and for the scaling law to continue to improve model performance. We can predict that the next 2 years, AI will likely become more economically useful and valuable through applications, and especially agentic AI.

Beyond the fact that you can tell I have been reading, listening, and resting with my little one. I have found some time to actually sit with my thoughts and ponder about a bigger theme that is behind these currents: does the scaling-law paradigm still make sense, or are we shifting toward something more grounded with reinforcement learning?

So here we go…

Reinforcement Learning Changes the Game

Back in January 2025, on AI Proem, we asked whether OpenAI’s o3 changed the scaling-law debate. Ten months later, I find myself finally coming to a more concrete framework on how the future of AI will be progressed in real-life use cases that may not necessarily require the biggest and latest LLM.

First, let’s understand some basic concepts.

Scaling means making the base model bigger, more data, more compute, more parameters, to unlock general capability (this we know, right?).

Reinforcement learning (RL) takes a different route: learning by doing. An AI agent attempts a task, receives a reward or penalty, adjusts, and tries again. When that feedback comes from people doing the job, that’s RL from human feedback (RLHF), the system optimizes for what actually matters in context: accuracy, tone, safety, and helpfulness.

Consider a sales follow-up. After a call, someone has to draft the email, log the right fields in Salesforce, and schedule the next touch. With RLHF, the model does reps on that workflow; experts grade each attempt; the system learns which phrasing lifts reply rates and which field mappings reduce rework.

Now, let me piece together some ideas/ concepts a16z Marc Andreessen, Professor Ethan Mollick, and Sequoia’s Pat Grady shared and how they have come together in my head to make sense of the future of AI.

Think “Balloon with Spikes” as AI’s New Path to Usefulness

So in an investor meeting a month ago, Pat Grady shared a metaphor. He offered a twist to all of this scaling law vs. RL conundrum. Think of a model as a ball. Scaling is pumping more air to make the ball larger. RLHF is adding spikes, and each spike is a skill sharpened through practice. Add enough spikes and the ball’s reach grows where it matters, even if the core isn’t much bigger.

Scaling = pumping air (data, compute, model size) → bigger, flashier capabilities (all the PhDmath, physics, and benchmarks)

For two years, the tech and AI industry focused on inflating a giant beach ball, more data, more compute, more parameters, so it could roll into more use cases. That’s the scaling-law worldview: if the model isn’t good enough, pump in more air.

What if, instead of fighting for more air, we attached spikes to the ball, and then each spike represents a specific skill sharpened through practice? Add enough spikes and the ball doesn’t just get bigger, it also becomes more useful. It grips the ground.

I didn’t love this analogy at first, tbh it sounded a little gimmicky, but now I get the genius. Hear me out now.

If you look back at AlphaGo, it is the classic example. It didn’t memorize the game; it learned winning strategies by playing, getting rewarded for good moves, and being penalized for bad ones.

This matters because the scaling-law story has always been expensive and uncertain. Where does the following order of magnitude of “air” come from, think clean data, cheaper compute, bigger budgets?

From Child Prodigy to Applied Expertise

The original obsession/ thinking/ dream for many was that a single model would be able to write a sonnet, draft a contract, and debug a build script without breaking a sweat. But now, researchers can continue to chase the frontier model and test themselves against benchmarks, and business builders can stop arguing about whether the scaling curve is flattening and shift their thinking to how to better adopt RL.

What RL will do is that it doesn’t promise that a larger brain will magically generalize to your use case. It shows you how to teach the brain a job.

Picture the base model as a brilliant new graduate: breathtaking recall, broad knowledge, impeccable logic, and zero job experience. Left alone, it writes smart-sounding emails (with big SAT words) that miss the point or fill your CRM with beautifully wrong fields.

RL is the apprenticeship. A top salesperson role-plays with the agent, noting which phrasing disarms a skeptical buyer and which kills momentum. A nurse manager reviews clinical notes, rewarding clarity and penalizing omissions that create risk. A revenue-cycle lead grades denial letters on tone, compliance, and outcome. Providing the guidance needed to train an intern up, essentially, with knowledge, instinct, industry experience - a lot of things that are taught through time and mistakes. But now, each RL cycle leaves a trace. The agent builds the kind of muscle memory we usually associate with people who’ve done the job for years. That’s a spike.

And this is the direction that real-life AI use is happening all around us. captures the pivot in his recent essay Real AI Agents and Real Work. “Generative AI has helped a lot of people do tasks since the original ChatGPT, but the limit was always a human user. AI makes mistakes and errors, so without a human guiding it on each step, nothing valuable could be accomplished. “

Ultimately, “Agents, however, don’t have true agency in the human sense. For now, we need to decide what to do with them, and that will determine a lot about the future of work,” as Mollick had written here.

Bottlenecks Reversed

This is why the bottleneck has flipped. The limiting factor is no longer raw model IQ; it’s the scaffolding around it, data pipelines, reward design, expert time, and safe tool access.

Last year, many of us held our breath to see whether model capability would keep leaping forward and whether scaling laws would hold. That uncertainty has partly dissipated, replaced by a clearer roadmap for integrating AI into existing digital infrastructure. Today, the constraint isn’t raw intelligence. It’s scaffolding. (We touched on this in April in “The Future of SaaS Is a Hybrid of Agent & Platform + Vertical AI Innovation.”)

Do you have data pipelines that feed the agent a fresh, permissioned context? Can you define success tightly enough to prevent reward hacking or optimizing for “shorter” when you meant “clearer”? Are human coaches available and calibrated, or juggling two jobs and three dashboards? Does the agent have tool access with guardrails to do the task rather than merely describe it?

These aren’t mysteries of science. They’re the kinds of engineering and process questions strong organizations solve every day. That’s why the spiky-ball lens is optimistic without being naive. Scaling laws asked us to trust that if we inflated the center, usefulness would trickle down. RL gives us a map: pick a job, define a reward, run the reps, wire it into production, measure the lift, repeat. The uncertainty narrows from “Will we get another emergent capability?” to “Can we build and maintain this feedback loop?” That’s a better kind of risk.

Of course, RL has its own traps. Agents optimize what you measure, not what you intended, and some things cannot be conveyed through language: a gut feeling, an experience, an ick/a red flag. So do not expect an agent to be able to do that. But also, if you reward brevity, don’t be surprised by confident nonsense under 100 words.

In corporate terms, ambiguous domains such as mission, culture, and brand voice sometimes do not have a single correct answer, so you’ll need robust rubrics and periodic adjudication by senior experts. And as agents touch more systems, the blast radius of a permissions mistake grows; governance must be part of the design, not an afterthought. None of these are reasons to retreat; rather, they are reasons to treat RL as an ongoing operation rather than a one-off project.

Now, finally, you could think about the spikes as a kind of institutional IP. Every feedback run encodes know-how that competitors can’t download from a model card. Spend tilts from one-time training to repeatable programs: expert time, tooling, integration, and telemetry. So, in the future, when examining a business, maybe start asking, which workflows did RL sharpen? What’s the measured lift? How often does the agent relearn? Who are the coaches (are they credible and legit)?

Hot Take? Train those Spikes

So, did scaling laws hit a wall? Probably not, but who am I to even comment on this? I think in many ways, let’s borrow the words that people liked to use to describe DeepSeek’s R1 model when Chinese businesses started mass adopting it across products, “it’s good enough.” Echoing what I said at the top of the article, growth feels exponential when the starting point is small, but at what point does size stop mattering, and do we start building out our spikes through expertise and skills?

So the models we know today have hit a usefulness boundary. Bigger models will still matter, and that's for researchers and scientists to continue pursuing. But for businesses, the next mile is paved with reinforcement, which will allow AI to be more useful in practical ways.

Nearly ten months ago, I wrote with a kind of uncertainty, waiting to see if “more air” would carry us across the gap from just the chatbot and all these math/physics capabilities to something that could be more business-useful/commercial. Now, I feel like I really see a clearer vision with agentic AI: teach the job, wire the loop, sharpen the spike, move the metric, and then do it again for the next job.

RLHF teaches a model what “good” looks like for a specific job using human ratings and examples. And then an agent is trained model given access to your apps, data, and rules so it can complete multi-step tasks end-to-end. So in short, the relationship is that RLHF provides judgment and agents apply it in production.

To quote Mollick one last time, “Agents are here. They can do real work, and while that work is still limited, it is valuable and increasing.” So my point is, it’s on us to learn how to prompt, control, and identify AI work, because the next stage of knowledge work is in Agentic AI.

It was a pleasure speaking to a group of 40-50 industry watchers on a industry webinar panel organized by iMpact last week.

The panelists included Tom Nunlist from policy think tank Trivium, of AI-focused investment firm Interconnected, Cameron Johnson (who was just on Differentiated Understanding, see here) of Tidalwave Solutions, , former TechCrunch Editor and now founder of a PR consultancy for AI and tech companies, Bibhu Pattnaik at Benzinga, and yours truly.

We had a super vibrant discussion, so here are some key takeaways to share.

Subscribe now

1) AI policy

Speakers said current policy thinking is being shaped by urgency and, at times, fear—both of falling behind and of unintended consequences. One speaker walked through the U.S. posture (accelerate innovation, build infrastructure, and lead internationally on security), adding that the “lead” piece often manifests as restricting China; they also worried the U.S. could overbuild data-center capacity without solving power and execution bottlenecks, calling this the “stresses of success” problem.

Turning to China, another speaker outlined the new “AI+” plan: to push for rapid practical integration across multiple domains, with an emphasis on science/fundamental research, industrial efficiency, and consumption. They argued that those first three domains matter most and, importantly, pay off on different time horizons (basic science the slowest). The policy goal, as they summarized it, is to boost total factor productivity to counter structural headwinds, such as slowing growth and demographic pressure, while acknowledging constraints, including limited funding compared to a decade ago.

Several speakers noted cooperation is getting harder at both extremes of the stack: frontier research and large-scale deployments. Participants mentioned institutional desires to collaborate across borders, but also cited tightening restrictions (e.g., agency access limits and visa concerns) that make even benign technical exchanges harder.

2) Business and commercialization

Speakers said China’s commercialization path is skewed toward consumer integration rather than enterprise software. One speaker contrasted the U.S. “AGI-first/enterprise-led” mindset with China’s “deployment-first/consumer-led” approach, adding that domestic firms view LLMs increasingly as utilities to be embedded in everyday apps.

Multiple participants pointed to super-app distribution as a decisive advantage. One speaker cited WeChat’s approach: rather than insisting on a proprietary model that leads the benchmarks, integrate a strong external model into a chat-native, walled-garden environment where users already search, share, and transact. Another speaker called this “functional adjacency”: drop the AI into behaviors users already have—typing, searching, mini-programs, so the value shows up immediately. They contrasted that with platforms where a short-video context makes jumping into a chat-style agent feel less natural.

Speakers also said monetization tactics reflect this consumer tilt: improvements to ads, payments, logistics, and content are prioritized, while enterprise SaaS appetite is more mixed. Several participants tied the consumer emphasis to historic willingness-to-pay patterns, the existing mobile distribution rails, and the speed at which large platforms can trial features at scale.

3) Founders and startups

Speakers described a pragmatic startup mood. One participant noted that last year’s “everyone train a base model” moment has given way to “don’t rebuild what you can rent,” with many founders pivoting toward vertical apps, agents, or toolchains built atop stronger models. Another speaker characterized China’s founder culture as “prove first, raise later,” noting the big internet platforms often bankroll or incubate model efforts, which raises the bar for standalone labs and pushes startups to ship quickly and monetize early.

One speaker added that Chinese founders are intensely curious about the U.S. scene and constantly visit the Bay Area, attend meetups, and benchmark product choices. At the same time, the reverse flow into China is thinner (though major conferences do spark spikes of interest). Their point was less about superiority and more about information asymmetry: Chinese teams actively study U.S. best practices to accelerate their own execution.

4) Implementation and use cases

Speakers highlighted deployments in city management, manufacturing, mobility, and healthcare. One participant walked through Hangzhou’s “city brain” work, where it is seen in traffic optimization, faster emergency response, and predictive incident management—as an example of public-private coordination at the urban scale. They also described highly automated factories (e.g., Dongguan) using robotic arms, machine vision, and predictive maintenance, with AI orchestrating inventory and quality control.

On mobility, speakers contrasted U.S. autonomous rollouts that are technically strong but constrained by regulatory fragmentation with Chinese pilots that have scaled more rapidly under defined programs. In healthcare, participants noted collaborations using AI for earlier detection and triage, while acknowledging that scaling nationwide runs into compliance, data, and systems-integration hurdles. The common thread was the speed of deployment, where ecosystems are integrated end-to-end.

On the consumer side, speakers mentioned AI-enhanced search, assistants, and creative tools appearing inside mainstream apps. The takeaway was that not every use case is flashy—some wins are quietly compounding in the background (recommendations, logistics, customer support), while others are visible (AI chat in a messenger; creative/video generation tools).

5) Capital and investment

Speakers said the funding climate in China today is more “proof-oriented” than “research-led.” One participant argued that the domestic VC market has been structurally less exuberant than Silicon Valley recently, and that scrutiny following prior internet-sector crackdowns also shaped behavior. The result: model companies often sit inside or alongside big tech balance sheets, while independent startups are pushed to show traction fast: users, retention, and a credible path to revenue before bigger checks arrive.

Another speaker framed it as selection pressure: when platform incumbents bankroll major model development, startups either find a differentiated wedge (vertical expertise, agentic workflows, domain data) or risk being commoditized. Participants generally agreed that distribution and adjacency to real transactions beat abstract “parameter count” as a fundraising story.

6) Talent and education

Speakers repeatedly pointed to talent density as a competitive lever for China. One participant emphasized the size and practical bent of China’s engineering base, as well as a deep bench of implementation-oriented talent that can turn policy directions into products quickly. Another noted that entrepreneurs are “hungry” and execution-focused, a cultural factor they viewed as an advantage in the deployment-first cycle.

At the same time, a media participant described a knowledge-flow imbalance: Chinese technologists invest heavily in understanding U.S. trends firsthand (frequent Bay Area visits, hacker houses, meetups), while fewer non-Chinese founders do equivalent on-the-ground learning trips in China. They also observed momentum in areas like video and image models and, anecdotally, widespread use of particular Chinese robotics hardware in U.S. labs due to cost and service considerations. These points were presented as field observations rather than formal statistics.

Speakers concluded that talent flows both ways, but the near-term edge lies with teams that can fuse research with fast-cycle engineering and product iteration, where local data, distribution, and vertical context can be leveraged to create differentiated experiences.

Bottom line from the session

The webinar painted a picture of an ecosystem that is messy, fast-moving, and full of contradictions. Policy is cautious but enabling. Business models are diverging from the U.S., leaning on consumer engagement rather than enterprise subscriptions. Startups are pivoting, sometimes out of necessity. Use cases are spreading from chat apps to hospitals and factory floors. Capital is still flowing, but more selectively. And talent — abundant, ambitious, sometimes overstretched — remains the foundation of it all.

If there is a single takeaway, it’s that China’s AI path won’t resemble Silicon Valley’s. It will be its own story, shaped by its scale on the consumer internet, industrial backbone, and regulatory context.

According to Bernstein’s channel checks, they estimated that the BBAT made up 80% of the demand for H20s and that the rough breakdown of demand will remain consistent looking ahead once restrictions are lifted. But as this uncertainty continues, and now the fear of not being able to reaccess them grows, Bernstein also expects that the localization ratio of demand will increase. Currently, the stall on China’s end is likely more of a negotiation tactic than a genuine request to halt purchases, and the short answer to the question above is that China’s domestic solutions are likely still not good enough.

also wrote about why Chinese AI labs won’t want to buy from Huawei purely from a business perspective here as well.

The thing is, what this debacle made me think about more is what the government’s role is in AI development. Is it really just high-level import/export restrictions and industrial policies, or are they actually putting money where their mouth is?

According to Bernstein, the projected total AI capital expenditure is to reach USD 91 billion in 2025 in China, representing a 51% year-on-year increase. And half of that money will be going to chips, thus making up about 15-20% of global AI chip demand - a conservative estimate and a result of some constraints caused by restrictions. So who is spending on AI?

And what efforts is Beijing putting forward to achieve its goal of making AI a $100 billion industry by 2030?https://www.rand.org/pubs/perspectives/PEA4012-1.html

Subscribe now

Government-Led Investments

We’ve touched on the government’s embrace of open source and encouragement to integrate AI across sectors. We’ve also covered provincial and municipal governments luring businesses and talents with tax incentives and ease of doing business across the nation, especially looking at Hangzhou’s efforts. How effective it is to have like a few dozen cities competing to be the next “AI Silicon Valley” is a different discussion, but maybe competition will drive better results. Alas, this piece is about how exactly the government is helping AI beyond policies and talent incentives.

Infrastructure

Since last year, I’ve been writing about how China’s energy capacity is a considerable strength for its AI demand. And according to the International Energy Agency, China’s clean energy investment was more than USD 625 billion in 2024, almost doubling since 2015. Schroders estimates that China has become the world's largest market for energy transition infrastructure. In 2024, global energy transition investment reached US$2 trillion, with China investing more than US$800 billion, accounting for two-thirds of investment globally.

As the country continues to spend heavily on infrastructure, China’s new power generation capacity is expected to exceed 500 gigawatts in 2025, with new renewable energy capacity reaching approximately 400 GW, a result of China's accelerated green energy transition and increasing investment in grid construction, according to the latest July data from the China Electricity Council.

Based on the central government’s estimates, total installed power generation capacity is forecast to reach around 3.9 terawatts by the end of 2025, a year-on-year increase of approximately 16.5 percent. And nonfossil fuel sources are expected to account for 2.4 TW, or about 61 percent of total capacity.

State-led Investments

What I realized that I have overlooked is the government-driven capital injection into the sector. It is reported that the Chinese government is leading AI infrastructure development, as well as actively backing startups through VCs that have state-owned enterprise ties. All in all, China is expected to deploy $84 billion to $98 billion in AI capital expenditure this year, according to a new Bank of America report (similar range as Bernstein estimated). And a big chunk of that is state-backed capital in institutional investment vehicles.

Albeit still small compared to the $500 billion Stargate project, in China, there is the National Venture Capital Guidance Fund, a RMB 1 trillion (~$138 billion), 20-year initiative targeting AI, quantum technology, and hydrogen energy storage, which is not just being deployed to energy but across the whole stack.

There is also the newly launched National AI Industry Investment Fund, which is an RMB 60 billion (~$8.2 billion) fund launched in January 2025, is also going all in.

Regional funds like Shanghai State-owned Capital Investment (SSCI) and Pudong Venture Capital have financed AI leaders such as Zhipu AI and Biren Technology (which have resulted in these companies being added to the entities list). These funds prioritize early-stage AI firms, often acting as first investors before private VCs follow, and reportedly about 23% of China’s $912 billion government VC investments (2013–2023) are directed toward AI.

Beyond the national level, there are now at least 17 provinces offering up to $300,000 per company to offset AI computing costs, according to Bernstein. And many RMB-denominated funds are taking on LPs of the likes of state-owned banks and fund of funds.

Misconceptions of State-Capital and Involvement

Fundamentally, China is a techno-optimistic nation in both the private and public sectors. People believe that more technological advancements will lead to economic prosperity, resulting in better lives. It may seem in some ways naive and straightforward, but that is what the masses experienced over the last two decades as the internet improved the quality of life, access to resources, and increased income.

Now, one misconception many have in the West is that the state is somehow against entrepreneurship or capitalism. But that’s just not the case. China, in some ways, is the most capitalist society ever, acknowledging the irony.

In a recent nuanced conversation between Professor Keyu Jin and Lex Friedman on his podcast, Professor Jin makes the point that I’ve been reiterating on AI Proem, which is that it is because of the grassroots innovation and entrepreneurial mentality that we have the likes of Alibaba, Tencent, ByteDance, DeepSeek, MiniMax, and Moonshot today. The regulatory crackdown on the internet sector was not because the companies were too “rich”. It was because they started to influence societal discord and thus potential political influence.

Essentially, “capital controls the politics in the U.S., in China, politics controls capital,” she said on the podcast. China's approach to innovation is pragmatic, as I’ve written before here. And Professor Jin echoed that from an economist’s perspective as well.

Private Sector: Still Main Driver

Professor Jin said: “China Innovates first, then regulates, so you get P2P platforms, but in Europe, you regulate first, then innovate.” And I believe that also is why the Chinese capital market wasn't active until AI was essentially proven to be “hot” and had seen potential ROI. Scale-based leading tech cost-cutting solutions and pragmatism are still the foundation of China’s AI diffusion strategy.

And as China moves away from just manufacturing “low-cost goods,” as we’ve now seen time after time proven by the Apple saga, and as Patrick MacGee explains in his book, purely just having state support is not enough to replace a whole industry’s supply chain. It takes talent, culture and innovation still.

And as such, analysts at Rand put it eloquently: “the AI industry relies heavily on intangible inputs, such as talent and data, which are less responsive to capital subsidies and harder for the state to control.” So, beyond state support, which helps a lot in the long-term planning of an economy, true innovation and value creation ultimately still depend on individual innovators and corporate investors.

And the big tech, just like in the U.S., still makes up the most significant percentage of that AI capital spending and incubation. Because at the end of the day, they have the most to spend.

Based on our previous coverage, we can see that big tech investment surged in 2025. ByteDance is allocating RMB 150 billion ($20.6 billion) for AI infrastructure, with approximately half designated for overseas expansion. Alibaba is planning to spend RMB 380 billion ($53 billion) over three years on AI and cloud infrastructure, exceeding the company's total spending from the previous decade. And finally Tencent quadrupling its AI-related capital expenditure to 36.6 billion yuan ($5.1 billion) in Q4 2024, representing 12% of total revenue. Just those three companies’ capex added together make up half of the ~$91 billion estimated really.

“For us, it’s about exploring the unknown. Just like AGI, you usually only see the illuminated side of the moon, but the dark side remains mysterious. It’s challenging, yet full of potential. That aligns with our mission.” — Zhilin Yang, Founder and CEO of Moonshot AI (Kimi)

Hi Guys,

Today, we finally look into Moonshot, the company behind Kimi K2. The company has said the model, especially, is performing better than some mainstream open-source models, such as DeepSeek’s V3 and Anthropic’s Claude models, in areas such as coding, completing general agent tasks, and tool integration.

What is this big exhilaration about?

On the evening of July 11, 2025, Moonshot AI released and open-sourced Kimi K2, making it China’s first open-source trillion-parameter model. Its popularity among domestic open-source models surpassed DeepSeek-R1 for the first time. Within 48 hours of release, visits to Kimi’s official website surged by 3.6 billion, Hugging Face downloads exceeded 100,000, and GitHub-related projects skyrocketed by 200%.

On OpenRouter, K2’s token consumption quickly surpassed Elon Musk’s Grok-4, topping the global API call rankings. Perplexity CEO announced plans to fine-tune K2 for further training, and Hugging Face’s co-founder also praised it highly.

Devansh dove into the technicalities of the model here, especially breaking down the MuonClip Optimizer, and Tony broke down the cost of integrating the model’s API here.

But today, we will look at who the founder is and how the research lab became one of the most important teams in the open-source AI model space today. The founder’s background, his values, his mission, and the company’s potential viable business model.

Subscribe now

Moonshot AI's Founder: His Pursuit of AGI and the Company’s Potential Business Model

Amongst the “four tigers,” only MiniMax and Moonshot are still charging ahead at full speed, with MiniMax doubling down on its strength in companion bot/ entertainment bot, given its legacy product Talkie, and focusing on multi-modal models.

And then Moonshot AI is pursuing the most advanced LLM. Zhilin Yang 杨植麟, the founder of Moonshot AI, said he and his team are stubborn AGI purists who are focused on the vision of the future, rather than the product or short-term profit - a contrast to the big tech players in many ways.

Who is Zhilin Yang?

Zhilin Yang has long been a subject of public curiosity. He graduated from Tsinghua University and Carnegie Mellon University, ranking first in his class at Tsinghua’s Computer Science Department and completing his Ph.D. at CMU in under four years.

During his academic career, he collaborated with multiple Turing Award winners, publishing over 20 influential papers in the field of artificial intelligence. Compared to the other founders of the four tigers, he seemed to have the most ‘western’ experience, as before founding Moonshot LLM in 2023, he had also worked at Facebook AI Research and Google Brain. The Chinese AI unicorn Moonshot AI has been backed by Alibaba, Tencent, and Sequoia China, reaching a ~$3.3 billion valuation in late 2024.

Zhiling Yang is a fascinating character. Over a year ago, the founder of Moonshot AI recorded their first podcast appearance on First Push “第一推动” with Co-founder and Partner at Monolith Management Cao Xi, a former Sequoia China Executive. 月之暗面杨植麟：创新、长期、第一性. This was one of his few interviews and first podcast appearances, and here is how he described the founding of the company, his views on talent, and his pursuit of AGI. (Condensed for brevity while retaining key insights. Direct quotes remain in “”)

Zhilin is not your typical “nerd.” As we now know, Pink Floyd’s album inspired the name of the company, but what many don’t know is that Zhilin is a musician himself.

Zhilin is a drummer and was in a band called Skip List. Again, I’m not the technical person here, but what “Skip List” means is a multi-level express lane for searching a sorted list. The bottom layer has all the elements in order, while higher layers act as shortcuts, skipping over large chunks of data. When searching, you start at the top layer, move right until you overshoot your target, then drop down a level and repeat—like taking faster trains and switching to slower ones as you get closer to your destination. This makes searches much quicker than checking every single element, balancing simplicity and efficiency. Thus, it can get the best features of a sorted array (for searching) while maintaining a linked list-like structure that allows insertion, which is not possible with a static array.

He’s a man of reasoning and science, but also a man of arts and aesthetics, which will influence his design and management of the company going forward. During the interview, he said he thinks back on his time with the band fondly and mentioned that, although the direct lessons from forming a band and running an LLM company now might be minimal, it introduced him to a group of people.

The company name Moonshot comes from Pink Floyd’s album Dark Side of the Moon, and in its literal Chinese translation, 月之暗面 - Dark Side of the Moon.

The English name Moonshot reflects something ambitious and difficult, he said to the podcast host, “like landing on the moon.” And that’s fundamentally what they’re trying to pursue here - a seemingly impossible but possible task. His poetic nature comes through while explaining how he came up with the name with the team.

“For us, it’s about exploring the unknown. Just like AGI, you usually only see the illuminated side of the moon, but the dark side remains mysterious. It’s challenging, yet full of potential. That aligns with our mission. “

The Team

DeepSeek’s all-star, domestically educated team has propelled Chinese AI talent to the forefront. But for Zhilin, he did not have this requirement; he himself was educated in the U.S. after obtaining his undergraduate degree in China. While the commonality between him and his co-founders/ key members seems to be Tsinghua graduates: Yutao Zhang, Xiyu Zhou, and Yutong Zhang.

Ultimately, he said it’s not just the technical capabilities, it’s finding the right group of people with the same mission and values. He describes his team as innovative and long-term thinkers, as well as just fundamentally believers in first principles. They are the people “who believe AGI is the only thing worth doing in the next decade,” and that “vision matters most.”

In his eyes, things are evolving daily within the space, but to be successful, it requires three core traits: learning agility and adaptability. Adding that “even my views evolve daily.”

Of course, many of his team members are also music fanatics, and their office meeting rooms are named after various albums of Pink Floyd. He said that his team could form several bands, and in fact, it is this pursuit of aesthetics that subtly shapes their products. His personal hero? Steve Jobs, whom Zhilin calls a man who was able to scale taste.

Pursuit of AGI

With a better understanding of who he is, we now look at what drives him.

“I witnessed a massive shift. AI has developed over 70 years, but never before has there been such a breakthrough. Previously, AI companies focused on B2B. But since last year, applications like GPT have emerged, reaching non-technical users. Hitting 100 million users was a clear signal.”

That’s why Moonshot’s Kimi has not only pursued the most cutting-edge model research but also tried to prioritize a good consumer experience (in the beginning, we talked about how that has shifted).

“Unlike the internet, which connects things, AI creates new productivity. I believe AGI could be the most valuable endeavor of the next decade,” said Zhilin.

When asked about the positive response to Kimi’s launch (2024), Zhilin said, “I’m cautiously optimistic—prepare for the worst but strive for the best. Timing also matters. Years ago, assembling such a team was harder—capital and talent flows were different. Now, the market recognizes AGI’s potential.”

But he believes it’ll be a long journey ahead, as over the last decade in AI, he’s seen its potential to multiply productivity, which has been able to unlock brainpower beyond human limits. And if AGI is pursued correctly, it’ll propel society positively, allowing more people to be free from having to work for money and freeing up creativity. If done right, he believes AI can democratize creation so that anyone could direct a film or compose music. But if done wrong, only a few will control AI, and risks will rise. Thus, he believes that choosing open source will help decentralize the technology, and personalized, open models could prevent dystopia.

Share

Kimi, Open Source Leader?

After DeepSeek, has written about how Kimi K2 was another awakening moment for Western developers, and this trend could potentially continue, where Chinese labs continue to lead in open source models.

Globally, AI labs are feeling the heat as open-source models are increasingly recognized for their role in democratizing AI development. OpenAI just released two open-weight frontier models: gpt-oss-120b (120 billion parameters) and gpt-oss-20b (20 billion parameters) under the Apache 2.0 license, marking their first major open-source release since GPT-2 in 2019. Still, the models are competitive mid-tier options at best, designed to counter rival ecosystems and expand adoption, not to rival their flagship products.

But Chinese labs seem to be normalizing open-source as the industry expectation now. And just at the recent World Artificial Intelligence Conference, more so than ever, the Chinese government essentially countered the US AI Action Plan by saying it will be diffusing AI technology in the global south rather than controlling the technology. The open source embrace for the first time at a national level was being vocalized publicly, so as my analysis has evolved, it's not only a philosophical choice, a choice out of necessity, but now a top-down driven movement as well.

Former Benchmark Partner, Bill Gurley, on the BG2 podcast recently made a very valid point that with Chinese LLM labs choosing open source/ weight, this strategy allows them to work on top of each other like an open-weight remix that can accelerate improvements.

Essentially, using one model to distill another, and in some ways, it might mean there will be less likely to be a monopolistic breakout winner in the long run. Still, for the big techs that are more trying to stay relevant and capture value in the distribution channels of their existing apps, it makes more sense to be open source than not. He said if U.S. big tech companies like Microsoft and Amazon were being “smart,” they should follow what the Chinese big tech companies are doing. They’re commoditizing LLMs, investing in the leading startups, and capturing the value through that, but letting the real frontier model startups do their thing - in this case, DeepSeek and Moonshot in China.

Business Model

So, as foundational AI models become increasingly commoditized, labs around the world are adopting markedly different strategies to generate revenue. ROI is still unproven, but divergent playbooks are being played out.

On one side, we have OpenAI, Anthropic, and leading closed-source LLM labs building a vertically integrated business. Where OpenAI owns the whole value chain, and is trying to become the “Apple of AGI” — an end-to-end product company with a platform and distribution. What it means is that the company is monetizing at the consumer, enterprise, and developer levels.

• Consumer - With over 100 million users of ChatGPT and a $20/month subscription plan, the consumer tier alone could represent a billion-dollar revenue line.

• Enterprise - ChatGPT Team and Enterprise, with per-seat pricing targeting knowledge workers, and an API platform powering thousands of developers and software vendors. If OpenAI captures even a fraction of the productivity software market, its total addressable market (TAM) stretches into the hundreds of billions globally.

So we’ve written extensively on how big tech does it because they can monetize potentially through their existing distribution and reach, essentially, eventually selling value-added services to their existing users, optimizing ad revenue, and enhancing their existing products.

Anyway, I think the divergence highlights another key strategic tension: labs like OpenAI that own the full value chain, from model to interface, can experiment with pricing, upsell across customer segments, and retain rich user data. But labs like Moonshot must rely on being picked by platform gatekeepers, competing on model performance, cost, and fine-tuning flexibility. As open-source alternatives proliferate and customers grow more price-sensitive, value will increasingly accrue to those with proprietary distribution, whether that’s through a household-name chatbot or deep integration into enterprise workflows.

In theory, the most profitable labs will be those that productize their models, not just build them. That means monetizing attention (consumer), action (enterprise), and infrastructure (developers), at scale.

And from where we sit now, only a few players will be able to manage to do all three eventually. However, I still think it’s too early to say companies like Moonshot cannot be profitable (unlike some skeptics).

But then, if that’s the “perfect business model,” how do open-source models such as China’s Moonshot AI want to reach profitability? What can they do?

Moonshot is pursuing a more modular approach. With no equivalent to OpenAI’s global reach or Microsoft partnership, obviously, Moonshot is positioning itself as an AI infrastructure vendor in China (for now), powering third-party apps, mini-programs, and enterprise deployments across China’s tech ecosystem.

Monetization could come from API licensing, model-as-a-service deals, or consumer freemium upsells. But in this approach, Moonshot doesn’t control the user interface, which is basically limiting its ability to extract high-margin recurring revenue. The theoretical TAM is still large, given the scale of China’s consumer internet and SaaS markets, but likely will cap out in the low tens of billions unless Moonshot moves up the stack.

I think this goes back to the questions we keep asking ourselves at AI Proem when analyzing these AI companies:

1) Is everyone trying to be the next trillion-dollar business? Maybe not.

2) Will the Chinese AI ecosystem remain largely separate from the ROW like the internet ecosystem did? Tbh, it might.

3) If Moonshot- and equivalent companies were to diffuse the technology to the global south, will it be limited to the global south, thus is that a big enough market anyway? [Chinese internet and EV companies seem to be proving this feasible.]

Subscribe now

Kimi and DeepSeek’s business model, in theory, is trying to become the “NVIDIA of LLMs in China” — a key supplier to everyone else’s AI products.

If Moonshot can lock in key consumer app partnerships (like being embedded in WeChat, Alipay, or ByteDance’s enterprise tools), instead of competing with them, it can build a defensible distribution edge. But without that, they risk being replaced by cheaper OSS (like DeepSeek or MiniMax) and potentially even incumbent tech giants (Baidu, Alibaba, Huawei) with their own models. And in late 2024, Zhilin realized that by cutting some of its 2C consumer business ambitions, and reshifted resources and attention back to model development.

Thus, if Moonshot/Kimi wants to become profitable, they likely need to

1. Monetize mobile app scale fast — convert general users to premium.

2. Lean into verticals, which is the path Zhipu and MiniMax have taken.

3. Offer white-label LLMs to Chinese tech firms who don’t want to build their own or cannot reach the most frontier level- seems like what Moonshot is trying to do.

4. Differentiate on context window + RAG UX as open source catches up.

So, I think in the mid-term, it would make the most sense for Moonshot to leverage its Data Flywheel and offer Training-as-a-Service, as it could eventually monetize Kimi’s usage data to improve models and offer fine-tuning services, like “bring your own data” for custom copilots.

Then it’s doubling down on solidifying its position as a foundation model lab, but not really as a consumer-facing chatbot company. Capturing consumer mindshare will be very challenging for Moonshot, especially since the BBATs have an existing reach of over 1 billion and access to endless cash piles to fund consumer marketing campaigns. (I guess unless somehow they get crazy funding and pull an OpenAI?)

Subscribe now

But in the words of Zhilin when comparing Moonshot to OpenAI, he said,

“In addition to the technical level, our values differ somewhat from OpenAI: we hope that, in the next era, we can become a company that combines OpenAI’s technology idealism with the business philosophy of ByteDance. I believe the Asian mindset towards commercialization has certain merits. If you don’t care about commercial value at all, it’s actually very difficult to create a truly great product, or to make an inherently great technology even greater.”

Aware of the challenges ahead, he realized that to stand out, Moonshot will need to offer something different. In this case, personalization.

“The ultimate, core value of AI-native products is personalized interaction, which is something previous technologies haven’t implemented well. So this question is actually about personalization — how to enable users to gain highly personalized interactive experiences, the more they use your product. For many products today, the degree of personalization is almost zero,” said Zhilin Yang, in the same interview with Overseas Unicorn on Feb 21, 2024.

See the complete translation of the interview by ’s ChinaTalk in March: Moonshot AI's AGI Vision.

Global Investor Perceptions: China’s AI

There is no real conclusion on how profitability will be achieved at this point tbh, it’s all speculation as the industry slowly matures, but it is interesting to see the rapid speed of progress in the space and how the views on what is worth pursuing, how to build AI, and the ROI of AI are all evolving constantly.

I recently had interesting calls with three different early-stage investors based in San Francisco, Singapore, and Shanghai (for context, none of them were Asian funds). It’s interesting to see how the appetite for Chinese startups and perception of Chinese AI players could vary so much. I’m not sure if this could only be more representative of personal views or the environment that shapes them, but I’m keen to hear your thoughts:

• For the U.S.-based investor, the view on China is that there is so much happening and we need to know what, how, why, and who NOW. A genuine fear of FOMO took over the conversation.

• For the Singapore-based investor, he said that the fund is not so concerned about U.S. regulatory constraints; we believe North Asia, especially China, has a massive advantage in energy efficiency technology and in implementing AI into workstreams to increase efficiency.

• For the Shanghai-based investor, ironically, the one in China is the most bearish; he believes China’s AI players won't be that relevant globally in ten years. Anthropic and OpenAI will be the trillion-dollar players, capturing everything from consumer and enterprise to API market share, because you can only win this race if you’re at the forefront of your models.

Let me know your thoughts. No matter if you’re a startup founder, investor, or enthusiast, get in touch. Dont be shy 🙂

Lastly, if you have found value in my work, I would appreciate it if you could share it with more people. It is word-of-mouth referrals like yours that help me grow. And please feel free to tag me in your post so I can see/share it.

And reach out to me:

Email: aiproem@substack.com

New to AI Proem? Start here

LinkedIn: www.linkedin.com/in/gmzshao/

Instagram: graceshao_proem

It was a pleasure working with and on this piece as a guest author for Exponential View last week. Azeem is a British technology entrepreneur and author. He founded the research group, Exponential View, and web technology company, PeerIndex.

He wrote "The Exponential Age: How Accelerating Technology Is Transforming Business, Politics, and Society" and hosts the podcast "Exponential View," distributed by Harvard Business Review, as well as the Bloomberg Originals television show and podcast "Exponentially with Azeem Azhar.” He writes about the topics of AI, renewable energy, and tech more broadly.

This piece delves deeper into the divergent strategies adopted by Chinese and US AI companies, driven by differences in their capital structures and existing internet ecosystems.

And like what Azeem wrote, “The US-China AI rivalry has become the defining geopolitical lens through which many view the future of technology. And for good reason: it’s shaping export controls, industrial policy, and innovation trajectories on both sides. However, when a single narrative becomes dominant, it can overshadow more nuanced analysis.”

See full post below.

Subscribe now

China and the US are running a different AI race

On Wednesday, the White House published the AI Action Plan, a playbook for building an AI innovation ecosystem. It’s a bet on destiny that intelligence, once summoned, will reorder the world.

China isn’t chasing destiny. It’s deploying fast, frugal, open-weighted models and wiring them into the economy.

That pragmatism unnerves Washington. Chinese labs now ship foundation models faster and cheaper, and – crucially – they publish the weights. Silicon Valley reads this as “open the weights, kill the moats” – a threat to the revenues that depend on keeping the model layer proprietary.

Inside China, the logic is flipped. Once the models are treated as commodities, profit shifts to the application layer. Publishing the weights speeds that shift. Free forks and vertical fine-tunes multiply, each one funneling demand back to the riginator.

However, this isn’t some grand strategy being directed by the governments on both sides; it is market-driven, shaped by three structural forces – chips, capital, and distribution – that make open‑weight releases the logical on‑ramp to value.

Today, I will break down why I believe that China and the US are not running the same race and hopefully offer a different perspective.

Deployment vs Destiny

The real split is over where each country’s tech companies believe the profit will land. China bets on applications; America bets on the model itself. Palo Alto obsesses over model‑led destiny: ever‑bigger parameters, safety benchmarks, and a near‑cult-like obsession in the pursuit of AGI. As Karen Hao recounts in Empire of AI, Sam Altman and his peers see AGI as a world‑changing force capable of solving humanity’s most significant challenges. The wager is long, deep‑pocketed and proprietary: pour venture billions into loss‑making models today; own the platform that reorganizes industries tomorrow.

China’s approach is more pragmatic. Its origins are shaped by its hyper‑competitive consumer internet, which prizes deployment‑led productivity. Neither WeChat nor Douyin had a clear monetization strategy when they first launched. It is the mentality of Chinese internet players to capture market share first. By releasing model weights early, Chinese labs attract more developers and distributors, and if consumers become hooked, switching later becomes more costly.

Entrepreneurs then have the opportunity to utilize these models as free scaffolding. Taking the EV industry as an example, over twenty Chinese automakers, including BYD, Geely, and Great Wall Motors, have integrated DeepSeek into their in-car AI systems to enhance smart assistants and autonomous driving capabilities. In healthcare, it is said that nearly 100 hospitals across the country have now integrated DeepSeek for medical imaging analysis and clinical diagnosis support. Every new integration expands the model’s footprint, tightens switching costs, and shifts margins to the services sitting on top.

China’s so‑called “open‑source strategy” is a corporate pragmatism play, not a grand geopolitical scheme. But to see why pragmatism took this particular form, you need to see the three forces driving it – chips, capital, and distribution – turn openness into the default.

Chip Scarcity

America has had export controls on China since October 2022, preventing Chinese model makers from accessing the latest Nvidia GPUs. Overnight, the brute‑force “scale‑to‑AGI” playbook – whether China wanted it or not – was off the menu. Chinese labs were left with a patchwork of last-gen Nvidia GPUs and local chips, at least one to two years behind the US hardware frontier.

Because chip access is capped, efficiency has become the main event. Chinese researchers have been focusing on extracting the most from their hardware. DeepSeek‑V3 delivered GPT-4o performance at 18x cheaper cost, while Moonshot’s Kimi K2 used MuonClip, which likely halved the FLOPs used during a training run. America’s chip curbs have unintentionally turned China’s models into performance‑per‑yuan champions. As many have said, “necessity is the mother of invention.” Now, Chinese models are so competitive that both local and international companies want to build on them.

Capital Drought

The same efficiency drive also aligns with corporate reality. With venture money scarce and a lingering urge to prove they’re innovators – not copy‑cats – Chinese founders must signal value fast.

The more the models are used, the bigger their moats and their name. This is an essential characteristic of Chinese AI, stemming from the fact that China’s funding environment is scarce compared to the US. US venture capital plowed $100 billion into AI in the first half of 2025. Meanwhile, Chinese startups across all sectors raised barely $11 billion from VCs. Since China’s leading ride‑hailing app DiDi delisted from the NYSE in 2022, a regulatory storm has swept across the country’s internet sector. Some American funds pulled out of China, and even Chinese ones became skittish. The US government implemented a rule that disallowed US investors from investing in AI and chips. Chinese start-ups in AI now, compared to ten years ago during the internet era boom, raise capital only after they can demonstrate a live product and real usage figures.

Chinese models may be efficient, but DeepSeek is still estimated to have spent over $500 million in total R&D for R1. Luckily, they were self-funded by hedge fund billionaire Liang Wengfeng. But how do the other startups get funding? If you look at the “four AI tigers” – Baichuan AI, Zhipu AI, Moonshot AI, and MiniMax – most had to secure backing from the BBAT quartet. For Baichuan and Zhipu, releasing strong open-weight checkpoints helped persuade Alibaba/Tencent that the teams could ship and attract a developer ecosystem, unlocking nine-figure cheques only after the code was dropped.

Distribution Choke-Points

Once a model maker has a reputation, what keeps them open-source?

Not every Chinese model is open‑weight – ByteDance’s Doubao 1.5 Pro keeps its parameters under lock and key – but most still publish. As mentioned earlier, this reflects the Chinese mobile-first digital economy. A typical Chinese digizen opens fewer than ten apps a month, almost all of which are funnelled through WeChat or Alipay; in the US, the figure is closer to thirty. Scale belongs to whoever can seed a model across those few choke‑points fastest. It’s led to some unorthodox decisions. Look at what Tencent did. The WeChat team wanted a good LLM fast. Yuanbao, Tencent’s AI division, wanted its proprietary engine showcased. The WeChat team went ahead and plugged in an external open-weight rival, DeepSeek R1, because it was better than anything Yuanbao had made. This is a typical move by Tencent, where product leaders trump company management in product deployment sway.

But that isn’t how Alibaba or ByteDance operates. Top company executives rallied top engineers to get their heads in the game. Beyond what is known as the 996 work culture, seven-day stints in the office to refine consumer-chatbot models are, to be honest, quite normal. In that environment, open-source software acts as growth hacks: zero licensing friction, instant plug-in, and viral forks. The BBATs want market share, and openness is the shortest route.

In truth, though, it’s not all just about corporate strategy either. It’s the chip on the shoulders of Chinese entrepreneurs that is driving their embrace of open-source, too. For decades, Chinese goods were called “copycatters.” Many AI founders and leading researchers are choosing to open-source their research to demonstrate to the world their innovation and capabilities. The psychological element shouldn’t be overlooked, as this open-source vs closed-source debate has long existed in the technology industry. For those who are pro-open, it means more scrutiny and potentially faster advancement of their research, as well as broader adoption. For those who are pro-closed, the argument is based on proprietary knowledge and protectionist thinking.

The Chinese Government doesn’t particularly care about the open-source strategy in itself. All they want is for AI to be in every pillar of its economy through its five-year plans. Instead, it’s the entrepreneurs’ philosophy, combined with structural drivers, that have pushed the open-source flywheel to the centre of the Chinese AI ecosystem.

[Though you can learn more about the government’s efforts on pushing forward AI and robotics through its “Made in China 2025” plan, “East Data West Compute” here. These top-down-led initiatives are mostly to drive 1) Accelerated Innovation and Self-Reliance; 2) Global Competitiveness; 3) Sectoral Transformation - manufacturing, transportation, healthcare, agriculture; 4) May have socio-economic impacts - bolster economics but may widen regional disparities.]

It’s Not a Race.

I want to reiterate that China and the US are not running the same race. Deployment is China’s dividend, and destiny is America’s dream. Each is chasing what it values most. Chinese companies integrate open-source models into daily life because speed-to-market incentives pay off fastest at the application layer. Silicon Valley pours capital into ever-larger proprietary models, hoping to reach AGI first, whatever that means at this point.

Chinese model makers are responding rationally to their environment. Chip constraints reward efficiency, while funding structures and market-share chasing dynamics push them toward open-source. Make no mistake, China’s current open‑source “strategy” is market‑driven, not state‑driven. It is pragmatic; as a Chinese saying goes, “兴趣不可以当饭吃,” literally meaning “interests can’t be eaten as food.” In China, even AI must earn its keep.

I was in Seoul, South Korea, for a few days last week and had the opportunity to meet with both old and new friends. My J-school classmate, who has been a badass foreign correspondent, covering everything from the Japanese stock market to the Thai political unrest, has recently moved to the Asia bureau of the New York Times in Seoul, so it was such a treat to meet up with her and learn about her journey. I also met up with a few investor contacts in Korea, in the early VC space as well as later public sector, to get a better sense of the overall ecosystem.

Overall, it was a super fruitful trip, filled with kimchi and beef, and a wealth of knowledge from people on the ground about the shifts and changes in the economy, administration, and obviously, the AI & tech space.

Thus, I think today, we can start with the macro context to really understand why Korea is where it is with its AI strategy.

Macro Background

For those of you who might not have been following, Korea experienced a tumultuous few years with its changing administrations. Two presidents were charged with corruption and misuse of power and were then impeached. It was a scandalous decade, and thus, domestically, people were a bit battered and disappointed. Internationally, K-pop continued to make waves with Blackpink performing at Coachella, and Korean films and shows continuing to push boundaries.

However, its economy struggled, and diplomatic and trade relations between South Korea, Japan, and China were also somewhat rocky, given the shifts in approach and philosophy by the three major economies in East Asia. Now, with a new president who just resumed office in June this year, we’re seeing some major efforts being made to stabilize its economy and push forward its technological capabilities. This is all critical information to better understand the investment space and the technology/ AI ecosystem in South Korea.

Governance

South Korea's political landscape over the past decade has been marked by significant presidential scandals, impeachments, and shifting economic and foreign policy priorities. The country witnessed two presidential impeachments — Park Geun-hye in 2017 for corruption and embezzlement, and Yoon Suk Yeol in 2025 for misuse of martial law and illegal arrest of judges. These incidents have reshaped governance. Economically, South Korea faces challenges including slowing growth and high household debt, while foreign policy navigates complex U.S.-China competition. The 2025 snap election brought progressive leader Lee Jae-myung to power, signaling a pragmatic recalibration in both domestic and international arenas.

The new president and government were only officially instated in June this year and thus have said that, as the administration is putting together their agenda and priorities, they had to skip the recent NATO meeting in The Hague, Netherlands, where South Korea has traditionally been a major participant.

South Korea-China-U.S. Trilateral Relations

Historically, South Korea adopted an approach called “anmi kyeongjung,” which translates to "security with the U.S., economy with China." For context, the U.S. has had a security alliance with South Korea since the 1950s, whereas China is South Korea's largest trading partner, and the two nations have long had a bilateral trade agreement.

Citing the Asia Pacific Foundation of Canada, the new South Korean President Jae-myung is “ Lee is expected to depart from Yoon’s U.S.-centric, Japan-friendly regional security initiatives based on shared values and instead resume cautious engagement with China and Russia to maximize South Korea’s manoeuvrability in an increasingly polarized regional order.”

While I walked from the Imperial Palace to Myeongdong, a shopping district just a ~7km walk away, along the streets, I encountered small-scale protests against U.S. interference in Korean politics and businesses, as well as gatherings supporting a stronger U.S. alliance. It seems that the country is undergoing an internal reassessment of pragmatism versus idealism and reevaluating its diplomatic strategy.

Software

But this conundrum doesn’t seem to affect the startups, at least not for the AI developers. From people on the ground, some South Korean AI startups and tech companies are taking an apolitical approach to LLMs, and they’re open to using everything from DeepSeek to OpenAI. While others are choosing not to use either Chinese or U.S. LLMs due to compliance and internal security considerations altogether.

For most businesses, they are leaning into what think tank Carnegie Endowment for International Peace coined as “shifting from engagement-centered idealism to strategic pragmatism” rather than picking one over the other, given that there is currently no strong domestic contender to replace international LLM options.

In fact, for startups, it’s less of a political choice but a strategic business case. Robbie Oh, tech investor at leading VC Altos Ventures, said to AI Proem during a meeting.

“Korean founders are clever and highly strategic. While they may publicly highlight their use of US-based LLMs, their actual decisions tend to be much more practical. Instead of focusing on brand or origin, they compare performance, price, customization flexibility, API accessibility, and service reliability.

In this context, Chinese LLMs are emerging as competitive alternatives. Models like DeepSeek, Zhipu, and Moonshot offer high token throughput, low cost, and relatively open fine-tuning capabilities. These features are attractive not only for experimentation but also for real-world service deployment.

U.S.-based models such as GPT-4, Claude, and Mistral (French) still lead in technical credibility and have stronger communities for support. However, many startups now use hybrid strategies. For example, some run basic operations through a Chinese model and send premium tasks to GPT-based systems.

Startups we meet at Altos rarely rely on just one model. They are building multi-model workflows, optimizing for cost and performance, and designing user experiences tailored to specific markets. In Korea, where the speed of iteration is high and technology adoption is fast, this approach works especially well.”

Hardware

Meanwhile, the U.S. government is reportedly considering revoking export control waivers for major chipmakers operating in China, which includes Korean semiconductor companies such as SK Hynix and Samsung. This move means that the U.S. will require license approvals for future equipment imports and technology updates, which will significantly impact Korean chipmakers' operations, as their Chinese fabs contribute a sizable share of their DRAM and NAND production.

For context, SK Group is a conglomerate that has businesses leading in electric vehicle batteries, semiconductor technology & materials (SK hynix that you’re probably most familiar with), and pharmaceutical development & manufacturing.

Many know of the major players like Samsung Electronics and SK Hynix, which dominate the global memory chip market. Together, they hold close to 70% of the overall memory chip market, with SK Hynix recently overtaking Samsung as the world's top DRAM chipmaker with a 36% market share compared to Samsung's 34%. But these are all within the infrastructure layer of the stack.

According to the latest report from Futurum Group, SK Hynix faces the most substantial risk as its Wuxi fab facility produces approximately 40% of the company’s DRAM, and the Dalian fab contributes 25% of NAND output. License delays or denial could disrupt node migration to 1a nm in DRAM and limit production capacity in DRAM and NAND in China. Meanwhile, Samsung’s Xi’an fab contributes an estimated 40–50% of its total NAND output and ~10% of global NAND supply. The restriction may delay the firm’s ongoing technology transition for its NAND products from V6 (128-layer) to V8 (236-layer) and V9 (286-layer) NAND, as well as the future expansion of production capacity. [See full report here, and shout out to ]

The Building of Sovereign AI

In the early months of this year, due to the political uncertainty, the aftermath of dealing with long-COVID impacts, and increasing international competition in key economic drivers, consumer sentiment was low, and everyone looked to the new potential presidential candidates for clear guidance and direction to usher the country forward.

Both the left and right parties emphasized their “shared perception that AI is a future growth driver for Korea,” though different during the election campaigns, according to Korean news outlet ChosunBiz.

Coming up on top was the Democratic Party’s Lee Jun-seok, who became the 14th president of South Korea. Despite President Lee’s political philosophy that advocates for a smaller government, he has actively pushed for the development of technologies like AI in the role of corporations, and reforming regulations that limit technological development is the responsibility of the new government.

Lee has consistently emphasized the importance of AI policy and made a pledge of 100 trillion won (USD 735 billion) into the sector as one of his signature campaign pledges. His AI platform also included a proposal to develop a “sovereign AI” tailored specifically to South Korea’s cultural and historical context.

All of this is unfolding in front of our eyes. On June 15, the presidential office tapped Ha Jung-woo, head of the AI Innovation Center at Naver, as the country’s first-ever senior presidential secretary dedicated exclusively to AI policy. For context, Naver is one of Korea’s largest tech companies, and its products consist of a powerful search engine, popular blogging and community platforms, a digital comics service, and the nation’s most popular messaging app (LINE). You can think of it as Korea’s equivalent to China’s Tencent plus Baidu.

According to Altos Ventures, Naver began developing its own large language model, HyperCLOVA, in May 2021.

“Korea has a strong sense of technological sovereignty, and there is considerable public resistance to government use of foreign AI services. Naver has aligned its strategy with this sentiment and positioned itself accordingly. As a key player in Korea’s digital rise, Naver was the first private company to advocate for AI sovereignty. Since the recent change in government, two former Naver executives have been appointed to cabinet-level roles, increasing the company’s political influence.

Naver has since upgraded its model to HyperCLOVA X and expanded into multimodal capabilities similar to GPT-4. While not officially part of the government’s Sovereign AI initiative, Naver is its closest private-sector partner from a technical and infrastructure perspective. Naver’s unique position as one of the only domestic firms with both cloud and search technology in-house has made this possible.

If HyperCLOVA becomes widely used by startups in Korea, it would mark a deeper shift beyond just model performance. It would represent a fundamental change in the country’s platform strategy.”

This newly established AI-focused unit within the government will be at the core of leading the nation’s sovereign AI efforts, according to Chosun Daily. And on the same day, SK Group unveiled plans to build South Korea’s largest AI data center in Ulsan’s Mipo National Industrial Complex in partnership with Amazon Web Services, indicating clear top-down efforts to push forward Korea’s AI capabilities.

And just a few days ago, according to a Bloomberg report, Korea Investment Corp’s CEO, Park Il Young, stated that the nation’s sovereign investment arm, which manages US$206.5 billion in assets, is exploring tech investment opportunities in China, specifically in the AI sector.

And he highlighted, in particular, plans to increase allocations to tech startups and venture capital (VC) funds as part of its broader push to deepen exposure to artificial intelligence (AI) and other disruptive technologies.

VCs and AI Startups

So this finally brings me back to my meeting with Altos Ventures. This VC was referred to as “the Sequoia of Korea” by one of their LPs when I spoke to him, and it has backed some of South Korea’s most successful internet companies. It was founded in 1996 by three founders: Han Kim, Ho Nam, and Anthony Lee, two of whom are of Korean heritage. Although its headquarters are based in Silicon Valley, the Korean ties have meant that the fund introduced this Silicon Valley capital playbook to Korea, in some ways bringing back their expertise.

The VC has been widely recognized for being the first investor in fintech unicorn Toss, and an early investment into Danggeun Market, also known as Karrot in some regions, a large second-hand marketplace.

I met with Robbie Oh, an investor at Altos in Korea who focuses on Web3 and AI investments. Below are some of his insights into the space.

“Korea’s AI industry is advancing at a remarkable pace, especially when it comes to the speed of adoption and real-world consumer impact. The B2C sector is particularly responsive, where AI adoption tends to happen quickly and with significant user engagement. For instance, soon after ChatGPT’s launch, Korea became one of the fastest-growing markets for OpenAI’s paid subscription service, second only to the United States.

This mirrors the rapid uptake of smartphones in the past and reflects a broader cultural tendency in Korea: new technologies are embraced not as passing trends, but as tools for real productivity. The result is a tech-savvy user base that is highly open to experimentation and quick iteration. Even under the previous administration, there was growing interest in AI at the government level. Under the new administration led by President Lee, this momentum has accelerated. The government has announced plans to invest what amounts to nearly 100 trillion Korean won(nearly 77 billion USD) by 2027, marking an unprecedented national commitment to building AI infrastructure and securing technological sovereignty.”

He highlighted that, overall, Korea has had very strong B2C startups, and consumer adoption of new technology is very high. Despite the understanding that overall East Asia populations have been very receptive to new tech adoption, I was still shocked to hear that OpenAI’s second-largest market (enterprise and pro) is South Korea.

So far, most of this investment has been directed toward infrastructure, such as data centers, AI semiconductors, and national LLM development. While this tactical, infrastructure-led approach has laid important groundwork, creative productization and consumer-facing services are still in the early stages.

B2C AI

From Altos’s perspective, this is a pivotal moment. The priority now is not just to develop the core technology, but to enable fast cycles of product validation and user feedback. Korea is uniquely positioned as one of the few markets where such rapid experimentation is not only possible but culturally encouraged.

Confirming my research, he stated that there is an expected USD 1-2 billion injection into GPU purchasing for the nation’s sovereign AI research. While AI vertical startups are still in their nascent stages, Korea has yet to see a leading LLM startup rise domestically. He’s seeing more and more innovative startups in the space that are integrating DeepSeek and OpenAI alike, trying out new creative ways to engage with consumers.

One interesting example, reflecting the current societal and cultural shifts in Korea, is the success of AI in the avatar business. As Oh shared, one of the most widely adopted AI apps is a companion chatbot that acts like one's boyfriend. As Korea's marriage rate has significantly decreased in recent years, and fertility rate has dropped to as low as 0.78 births per woman as of 2022, in contrast to 1.66 in the U.S. and 1.80 in Japan, according to the World Bank, one of the lowest amongst developed nations globally.

“The rise of companionship AI shows that it addresses emotional and social needs. For example, AI boyfriend chatbots and avatar-based chat services have gone beyond demos and secured loyal user bases. These products reflect broader social trends, including low marriage and birth rates, and respond to growing emotional gaps,” said the early-stage tech investor.

Automation

Meanwhile, as we know, Korea is a large manufacturer of cars and automation, and he’s seeing AI-driven simulation use cases being explored by startups where AI is being utilized in the design stage, largely boosting efficiency. Robbie Oh added,

“This has enabled the quick adoption of AI simulation tools and AI-integrated design software in industrial settings. These are not tech showcases, but practical tools that solve real problems.

We believe Korea’s greatest opportunity lies in its cultural sensitivity and boldness in experimentation. The market may be small, but it is one of the fastest environments for turning ideas into products with global potential.”

Finally, he added, it’s a talent shortage in South Korea that is the biggest bottleneck right now. Most of the smartest young people want to become doctors, as it's still seen as the most prestigious and stable profession in the country. But eventhough the eventual AI market isn’t as big as China or the U.S. (for context the most widely used app KaKaoTalk has ~49 million MAU that is nearly the whole population), he is confident in consumer adoption and innovation in a future AI-driven world, especially in niche categories because East Asian consumers as a whole has always been very receptive and willing to try new technology, just see everything from Line to Kakao, Groupon to Toss.

China, Japan, and South Korea’s consumer-facing internet apps are significantly ahead of American consumer apps in terms of creativity and functionality, and this trend is likely to continue in the AI era as well.

Today, I want to share my learnings from investors in the Southeast Asia region directly - some on record, some on background.

BTW, I’ve been in Seoul for the last few days and met with some investors and reporters to learn more about the South Korean AI ecosystem - from hardware to software to the interesting AI startups emerging in the consumer space. SUPER interesting. A piece to come on my learnings from this trip.

Subscribe now

Chinese Founders Received Positively in SEA?

According to an Indonesia-based USD fund investor, a trend is emerging where funds are recruiting Mandarin-speaking investors and scouts. The reason is that there is a bias or common belief that Chinese founders are very likely to succeed, and the goal of these investors is to court them.

The VC investor told me that many Southeast Asia-based investors are trying to court Chinese founders over local founders, but PR is the biggest issue.

His reasoning is that there is a common belief/ trend among investors in SEA that they prefer to back Chinese founders over local founders, with the belief that they're more competitive. But this is obviously an overgeneralization, and not shared by everyone, and in no way a diss of local founders.

However, it is underpinned by some truth in demographic shifts and the overall trends in the tech ecosystem in SEA (specifically in Indonesia) over the last decade.

Just as we’ve written about before, many of the major internet players in SEA - especially in major markets like Indonesia are Chinese, such as Alibaba’s Lazada, AliExpress, and Alibaba Cloud, all of which have a considerable presence locally.

Others, such as Tencent’s Gaming, Shein, and Temu, also have operations in the most populous middle-class market in SEA. While local SEA leaders such as Shopee/SEA (SG), Grab (SG), GoTo (ID), and J&T Global Express (ID/HK) dominate, many have also chosen to partner up with Chinese firms in some capacity, such as the Tokopedia/TikTok deal.

[And to add to the point of Chinese founders being technically good founders, this VC investor highlighted that even Shopee/ SEA and J&T’s founders are originally from Mainland China]

Caution Remains as Challenges Persist

Another investor I spoke to, who had only just relocated from the U.S. to Singapore in the last year, works for the VC arm of one of the most significant funds in the region and holds a much more cautious attitude.

He notes that it is difficult to draw a clear conclusion about whether Southeast Asian VCs systematically favour mainland Chinese founders, especially amid persistent public relations sensitivities around Chinese technology, mainly for companies that target Western markets.

He observes that, in global markets, Indonesian founders are often perceived as less technically sophisticated. In contrast, some believe that Chinese founders —frequently alumni of Alibaba, Ant Financial, or ByteDance — benefit from the depth of China’s mature start-up ecosystem and experiences. However, the reality is that this technical gap does not automatically translate into an advantage within Indonesia, given the size and complexity of the market.

Success ultimately hinges on nuanced local knowledge and relationship-driven commerce (to be honest, just like how American companies couldn't simply take their exact roadmap and make it work in China two to three decades ago as well), so even the most capable Chinese technologists struggle without strong domestic networks and a deep understanding. So the ones that are being backed by SEA funds have mostly led teams and demonstrated success at Chinese firms in the local SEA market(s) - such as J&T’s founder who was a former executive at Oppo Indonesia before starting his own business .

In fact, recent encounters with Chinese entrepreneurs in Singapore have made him cautious. Some appear primarily motivated by lifestyle considerations and by the perception that capital is easier to raise in Singapore at a time when China’s current start-up environment is weak. Others cite that their main reason for setting up shop in Singapore or the Southeast Asia region is due to immigration pursuits, rather than strategic or business operational reasons. Even among the credible teams, many import a distinctly “Chinese” operating style that raises governance concerns and may not align with Southeast Asian work norms.

Lastly, he argues that these companies will face significant headwinds in the United States, where investors and customers remain wary of vendors with Mainland Chinese roots, regardless of their Singapore domicile. This is more reflected in some of what we have written about as well, with the “China-shedding” strategy as a priority for many.

Indonesia, the Underdog Market

The fact is, for most realists, let’s be honest, it’s often just what works better and what can potentially generate a higher return in the long run. Whether they are Chinese or not is not a key consideration in whether they financially back them or not. And talent is heritage agnostic.

For many in the West, it is not realized that Indonesia is often the single largest profit lever for Singaporean tech players, such as Sea and Grab. So, I spoke to a leading Indonesian VC firm, Alpha JWC Ventures. The fund has over 70+ active portfolio companies and USD$650m of AUM.

I asked them whether they were specifically courting Chinese founders, given the rumors about the trend, and they said that they work with founders across markets and heritages. And again, proven that it’s their business and work ethic that speak louder than anything else.

Ray Djajadinata, Technology Partner at Alpha JWC Ventures, said to AI Proem: “We are excited by more Chinese founders exploring SEA's potential, in particular in sectors of consumer and healthtech. Specifically regarding AI technology, with more challenging market access to the U.S. and EU, startups in China will have to diversify into other markets, like LATAM, Africa, and Southeast Asia. They have also taken to establishing an entity in Singapore, which is viewed as more neutral, and can facilitate easier access to the U.S. market later on. Then, when a startup is AI heavy or AI native, it will need access to advanced GPUs, where Singapore facilitates easier access to uncrippled NVIDIA GPUs.”

In fact, most of the big Chinese tech companies have become their LPs over the years, and as the fund gets ready to launch its new fund IV, the team is “naturally speaking with new partners, Chinese or otherwise,” according to a company representative.

Externals might not know that one of their unique edges is their China reach, as it has long had venture partners on the ground in Shanghai. However, in no way is the backing more Chinese founders over SEA ones. It’s all about whether the tech is superior, but heritage agnostic.

And the Chinese presence in Indonesia is not a new phenomenon.

Djajadinata added, “Since 2021, we have seen a healthy influx of Chinese entrepreneurs building businesses to capture this market. Some examples from our own portfolio include those of Skintific, Cornerstone Robotics, Peak3, etc. However, a lot of entrepreneurs may simply look at exporting China-based models to apply to Indonesian markets, and successful ones spend more effort in localizing their propositions and channel partners. Alpha JWC is happy to partner with such founders to help them accelerate market adoption through our unique value creation capabilities and resources.”

And case in point, for many sectors, such as healthcare tech, Chinese technology is naturally a leader, and when they expand, Indonesia is often a natural choice, given its significant population and growing middle class.

Chinese Internet and EV dominate SEA, But What About AI?

The reality check is that Chinese companies and products are dominating the roads and the internet sector across SEA, and thus Chinese capital has a large sway locally.

According to various sources, I’ve heard that Chinese EVs have flooded the local premium market. Notably, the seven-seater Denza, a premium marque of BYD Auto, has replaced the Toyota luxury minivan Alphards, as it is more fuel-efficient, cheaper in price, and has become a new status symbol for upper-middle-class families.

While newer players, such as Shein and Temu, have quickly eaten up market share, legacy players like Alibaba’s Lazada and TikTok’s acquisition of Tokopedia over time have solidified Chinese internet players' hold on the e-commerce sector.

So, when it comes to AI, what will be the future?

SEA (ex-Singapore) has sometimes been viewed as a cheap data center destination for U.S. and Chinese big tech, as I’ve covered here re. Thailand and Malaysia, and that is in many ways a fair consideration from a business perspective (the pressure put on local infrastructure and communities is an issue, though).

“SEA's AI ecosystem is not suited to compete on AGI / Superintelligence domains, and for a good reason,” said the Tech Partner at JWC Alpha Ventures.

He explained to me that, indeed, there are innovations and opportunities in the application layer. “SEA AI startups also innovate in building more trustworthy AI, thanks to regulatory pressure (ex. Monetary Authority of Singapore's FEAT principle, Indonesia's Otoritas Jasa Keuangan's Trustworthy AI guidance, etc.,” he added.

So, how will AI play out in SEA in the long run? We do not know yet. But if we look at how even the major locally grown SEA internet players took a lot of consumer playbooks from China to SEA and localized them successfully, then maybe that is one future possibility roadmap to follow as AI consumer apps continue to mature.

In addition, I dont know if you guys remember, when I met with Zhipu’s COO and he said at the Beyond Expo that they mostly work with price-sensitive governments in markets that are less “geopolitically sensitive”? Well, one can wonder if he meant markets such as Southeast Asia and Latin America, where officials have welcomed Chinese FDI and consumers have traditionally been less skeptical of Chinese technology and products.

With a lower average purchasing power, cheaper goods, such as cheaper LLMs, may make logical sense for these markets over American alternatives.

Chinese AI startup Manus reportedly secures funding from Benchmark at a $500M valuation, and it's reportedly Benchmark’s first “China-China” deal.

I actually heard about this deal last week from an industry source, and it’s definitely boosting some confidence within the space in China. The latest funding led by Benchmark Capital is valued at ~$75 million in a funding round led by Benchmark at a roughly $500 million valuation, according to Bloomberg.

Some of you may ask why Benchmark is allowed to invest in Manus AI. As many of you know, a rule that went into effect in January this year prohibited U.S. investors from investing in Chinese semiconductors, quantum computing, and AI. (which low-key led to some scares across China-based U.S. nationals/ investors and institutions)

So I was curious why Benchmark was able to complete this deal. Now, an explanation on this:

The rules prohibit or restrict U.S. investments in Chinese companies engaged in certain activities related to semiconductors, quantum computing, and AI systems that have military, surveillance, or other national security implications (you can read more in the link above to the law firm’s detailed explanation).

This actually explains why China’s internet players all have HUGE U.S. investor ownership - think Tencent, Alibaba, ByteDance, etc., whereas the AI companies do not. [FYI, Patrick O’Shaughnessy’s episode with Matt Huang was quite good and talked about how Huang got into ByteDance really early on]

According to an investor in the space, Manus received the capital in its offshore entity for non-China development. That’s also likely why this investment went smoothly forward. The Manus privacy page shows that the details were last updated just a few days ago in April. BUTTERFLY EFFECT PTE. LTD was mentioned as the parent company, instead of monica.ai as previously reported.

With this precedent, more and more U.S. VC/PE will likely try to enter this space with a few maneuvers or tweaks to the investee’s HQ/ ownership structure.

However, the prohibitions are narrowly defined and do not blanket-ban all AI investments. Many commercial AI applications, especially those not directly linked to military or surveillance uses, may still be permissible or subject only to notification requirements rather than outright bans. In the case of Manus AI, the AI agent is entirely for commercial use.

According to Manus AI’s own research on this topic, Benchmark was able to invest in Manus because of its intentions of moving its HQ out of China, despite announcements of partnerships with Alibaba’s Qwen. [See this link for more context on the background of Manus AI]

But the explanation offered by Manus itself said: “In summary, Benchmark's investment in Manus AI's parent company, occurring after the US restrictions came into effect, was likely made possible by the startup's strategic consideration to relocate its headquarters outside of China. This move potentially places the relevant corporate entity outside the direct jurisdictional scope of the US Treasury's Final Rule, allowing the US venture capital firm to invest while navigating the complex geopolitical and regulatory landscape surrounding US-China technology investments.”

Subscribe now

In addition, according to reports, the company will use the money to expand to new markets, including the U.S., Japan, and the Middle East, Bloomberg noted, citing people familiar with the matter.

Previously, Manus’ investors have been mainly from domestic backers such as Tencent and Hongshan (Sequioa’s China spin-off). Despite receiving high praise from the likes of Twitter founder Jack Dorsey, calling it “excellent,” the Information and TechCrunch both said they did not think it was as good as it was hyped.

Now, I finally got my invite to try it out, and played around with it for various research topics. My sense is it is not that different from OpenAI’s DeepResearch for research purposes, and maybe I’m not using it to its best capacity. The overall reasoning “thinking” speed (~1.5 -2 min for a relatively easy question) does seem like it lags behind OpenAI for the same kind of tasks. Still, the results are pretty comprehensive and impressive. I definitely will keep using it for cross-checking research, and to get a primer research done on new topics I’m not familiar with, but not for simple tasks (seems like a waste of time/ compute power).

What’s the difference between an AI Agent vs. an AI chatbot?

On the surface, the interface or usage of AI agents or chatbots may still not look that different. Although there has not been a standardized definition of an agent, a few factors define the difference. There are more nuances to this, and some chatbots are becoming more advanced, but a high-level breakdown looks like this:

1. Functionality and Complexity

   1. Chatbots are designed to engage in Q&A-style interactions. Mostly providing quick answers to questions or providing users with a predefined workflow. Conversations are usually simple, repetitive, and structured in their interactions, but they cannot complete multi-step tasks in a streamlined way.

   2. Agents handle complex, multi-step, and multi-faceted tasks that may span platforms and services, such as booking plane tickets or making reservations for the user. They can break down more complex tasks into steps, as you can see in the screenshots above, how Manus thought through how to answer the questions. They can orchestrate workflows and execute actions autonomously in the background, and completing the tasks usually takes longer.

2. Autonomy

   1. Chatbots typically operate on a pre-trained rule-based system, keyword matching, or decision trees, that follow predefined scripts. Reasoning was only introduced later.

   2. Agents are said to have autonomous decision-making capabilities and can adapt to new information in real-time and solve problems independently without human intervention. It can initiate actions without prompting such as help monitor email inbox or flag important news.

3. Memory and Planning

   1. Chatbots often have limited memory (if any) of prior conversations. They are mostly reactive, responding based on the current input and limited to what they know in the existing chatbox.

   2. Agents usually have long-term memory, task management, and planning capabilities. They can handle multi-step goals (e.g., book a flight, reserve a hotel, send a follow-up email) and adjust their behavior if something changes.

4. Tools & APIs

   1. Chatbots are usually limited to the webpage or the phone interface.

   2. Agents can use external tools such as APIs, databases, webscraping, file systems, CRMs, etc. They can gather new information or find tools they need to complete a task (much more like how a human would go about solving a problem).

      

And to better understand what a general AI agent can do, I asked “the best agent out there right now” - Manus itself, what else I could use it for besides deep research tasks, and it told me:

• Data Processing and Analysis: I can help you gather, clean, analyze, and visualize data from various sources, creating meaningful insights and reports.

• Content Creation: I can assist with writing long-form content, including multi-chapter articles, technical documentation, and comprehensive guides. I can also summarize information and format content according to your needs.

• Development and Automation: I can build websites and web applications, develop custom tools and scripts, and automate processes like form-filling and data collection across websites.

• Collaborative Problem-Solving: I can break down complex problems, suggest alternative approaches, and adapt to changing requirements as we work together.

Must listen/ watch this week

By the way, I highly recommend listening to this BG2 episode. I find Bill Gurley’s (former Benchmark partner) take very fair and balanced. He said that there is an oversupply of commentaries in the U.S. that are advocating to “win at AI,” but he doesn’t even know what that means, given that some say one party reaches AGI first, then what? He also talked about containing China’s AI development through export bans, which has seemed to be counterintuitive, and how the “arms race” really is more about humanity vs. technology and not China vs. the U.S. [I filed a op-ed for Fortune last week that largely aligns with this view, will share when pubbed]

He highlights that there will be so much progress and potentially more risk, especially as we’re still in the technology’s infancy stage. It’s refreshing to hear from someone who is thinking about this pragmatically and stepping back from all the geopolitical noise. He emphasizes that it’s better to be collaborative than not. Knowledge, research, capital, and talent have been so intertwined that a decoupling isn’t going to be possible, nor is it productive for anyone. Please do give it a listen.