113: 闵可锐 1: Back to the Beginning of the Story, Those “Prophecies”
Summary
闵可锐’s core bet was not to build a universal large model, but to let applications determine where models, data and compute should go. Models without a clear application focus easily become “seemingly able to understand everything, but unable to do anything to a basically usable standard”; Metaso therefore moved step by step from legal translation and correction to generation and dialogue, prioritizing professional use cases where “machines may be expensive, but people are more expensive.”
In his view, the real paradigm shift was large-scale pretraining, not the specific Transformer architecture. Transformer solved the engineering problems of parallelization and training at scale, but even without it GPT might only have arrived two years later; GPT-3’s breakthrough was not a new architecture, but taking the industrial assumption that roughly 100 million parameters were already “too large to deploy” and pushing it past 100 billion.
Model capabilities ultimately have to clear the unit-economics hurdle, and general-purpose search was far less viable at the time than high-value professional work. 闵可锐 cited 360’s estimate: generating roughly 200 Chinese characters cost RMB0.04, implying about RMB40M a day and more than RMB10B a year if applied to every search query, against annual advertising revenue of roughly RMB8B; by contrast, law-firm legal translation could cost more than RMB1 per character, while a senior lawyer might charge RMB5,000 an hour, leaving ample room to “trade compute for labor.”
Metaso had already demonstrated a meaningful user base and revenue, but it also exposed the ceiling of AI application monetization in China. 写作猫 had more than 7 million registered users, a paid conversion rate of roughly 2%-5%, and still gave free users an 8,000-character daily checking allowance; legal translation had penetrated roughly 30%-50% of the limited pool of China-focused cross-border lawyers, but demand would fall alongside cross-border M&A and international business, meaning a limited ceiling and limited competition were two sides of the same coin.
AI applications will first enter a Warring States period of “everyone versus everyone,” with moats emerging only through dynamic iteration. A product built by an individual developer in 3 days using the OpenAI API could directly challenge features that major companies had spent years accumulating; the realistic opportunity for startups was not to take GPT-4 head-on, but to occupy edge markets that giants would not staff heavily for 0.1% incremental revenue, then go deep in one writing or legal workflow.
Open source will sharply lower the barrier to entry, but it will not automatically erase gaps in scale, data or feedback speed. Fine-tuning from open-source weights could cost two orders of magnitude less than training from scratch; but if an OpenAI-like player could spend 100x what another startup spent and compress one feedback cycle from 3 months to 1 day, the strongest closed-source model could still lead, while much of what was called open source amounted to “fan-club models” modified from a handful of base models.
The episode’s most investment-relevant judgment was that Chinese VCs might collectively “miss” genuinely ambitious large-model entrepreneurs by applying the traditional hot-sector playbook. 闵可锐 argued that backing from tycoons, an academician-led team and tens of millions of dollars in funding would not necessarily replicate OpenAI; the key was to find people still writing code on the front line, who knew how to spend compute, and give them sufficient resources. Teams without a product flywheel that discovered their in-house models performed worse than open source could exit rapidly within 6-12 months, creating a “grow fast, die fast” shakeout.
Deep dive
1. The 2023 Interview Captured Views Just as China’s Large-Model Boom Was Getting Underway
The interview took place in mid-March 2023: Light Years Beyond had just been founded, Moonshot AI was still being assembled, the “Six Little Giants” label had not yet taken off, and DeepSeek had not yet been formally established.
The value of revisiting the episode is that 闵可锐 had already identified several themes that would recur repeatedly: pretraining mattered more than any specific architecture, a dense model arms race was unsustainable, and applications would enter a Warring States period.
His positioning of Metaso was unequivocal: “We will definitely be application-driven.” Without a clear application focus, a model could easily remain in a state of “seeming to understand everything” without doing any one thing that was truly usable.
2. 闵可锐’s Path Ran Through Algorithm Competitions, Academic Training and Two Startups
闵可锐 encountered algorithms and data processing through algorithm competitions in middle school, was admitted directly to Fudan University’s computer science department, then earned a master’s degree in mathematics at Oxford before pursuing a PhD at UIC and eventually returning to China to start a company without completing it.
He described the path plainly: he had been working on AI for nearly 20 years, and did not turn to large models only after ChatGPT appeared; his work had long centered on machine learning and NLP products.
When Metaso was founded in 2018, he also wrote “The Past and Present of Machine Learning.” The program’s commentary specifically noted that the piece already discussed an early version of the scaling loss phenomenon Dario had identified while working at Baidu.
3. Boson and Cheetah Let Him See Both the Value of AI Products and the Limits of Big Tech
闵可锐 founded Boson Data around 2013, ran it for 4-5 years, and ultimately brought its last product into finance; Ant Group considered it a rare team that was “really building products and doing AI” and could sell to financial institutions.
During acquisition talks in 2017, he left early because he did not want to be tied to Ant’s system. He then led the AI team at Cheetah, working on recommendation systems, OrionStar and a smart-speaker partnership with Ximalaya.
Smart speakers quickly fell from more than RMB1,000 to tens of yuan, eventually becoming giveaway items sold at little more than cost. That price war gave him a direct view of the difference between technology being usable and a product generating sustainable profits.
He stayed at Cheetah for only about a year because he had already seen the larger opportunity in the next generation of AI technology. That judgment did not depend on any particular Transformer paper, but on the point when pretraining genuinely began to work.
4. The Opportunity in Large-Scale Pretraining Came Before Transformer Itself
When 曼琪 asked whether Transformer had directly triggered his decision to start a company, 闵可锐 rejected that simplified narrative: the opportunity came from “deep learning plus large-scale pretraining,” while Transformer was simply a highly effective engineering implementation that followed.
Before BERT, work such as ELMo had already attempted pretraining with RNNs and achieved good results across multiple tasks; the problem was that RNNs’ sequential dependencies made large-scale parallel training extremely difficult.
Transformer made parallelization and scaling easier rather than creating the idea of pretraining from scratch. 闵可锐’s view was: “Even without Transformer, GPT might have arrived two years later—but at most, two years later.”
The distinction matters: pretraining changed how machine learning solved problems, while network architecture was a technical choice for implementing that direction, not something to turn into a matter of faith.
5. Pretraining Changed Each Task from Fresh Labeling to Knowledge Transfer
Under the old paradigm, sentiment analysis required separately collecting 10,000 comments and labeling each one positive, neutral or negative; every new task was effectively a fresh start for the model.
Pretraining first taught the model language from vast quantities of text. 闵可锐 resisted simply calling this “unlabeled” learning: writing a passage already expresses a judgment, so the process is closer to self-supervised learning.
BERT learned language by masking and reconstructing text, while GPT predicted the next word from a prefix; after that, a small amount of fine-tuning data could take a task from 70 points with 10,000 examples to perhaps 85 points with 1,000.
The bigger gain was not saving 90% of the labeling, but acquiring transfer capabilities that had previously been out of reach: “I haven’t seen this sentence, but perhaps you have,” because similar structures may have appeared 10 times across internet text.
6. The Problem with Rule Systems and Knowledge Graphs Was Not Intelligence, but Consistent Scaling
Older systems often amounted to “as much human labor as intelligence”: a seemingly simple judgment might be backed by a manually written rule, and a sentence the system had never seen simply could not be handled.
Even 10,000 rules could not guarantee consistency—rules written today and tomorrow, or by A and B, would conflict, while the conflicts themselves would require additional rules to resolve.
闵可锐 cited Google’s investment in knowledge graphs as an example: making symbolic data update in real time, remain valid and cover every industry was “basically impossible” as a scaling task, even with substantial capital and headcount.
7. Legal Translation Was the First “High-Value, Logically Structured, Machine-Replaceable” Wedge
Metaso chose legal translation in 2018 not because people considered it simple; on the contrary, legal translation requires simultaneous command of law, language, terminology and practical experience.
闵可锐’s view was that a task could be complex and involve multiple reasoning steps, but as long as it was “logical and governed by rules,” data-driven methods could potentially bring it to a very high level.
That became Metaso’s initial screening framework: users had a clear willingness to pay, labor costs were high, AI could gradually replace the work, and the company had enough resources to cover the required training and productization costs.
He acknowledged that AGI still looked far away at the time, so the approach was to “do as much as the resources allow”; a general model, by contrast, was like integrating 1,000 common tasks into one model at once.
8. Before ChatGPT, Almost Any Application-Layer Technology Could Be Reverse-Engineered
闵可锐 summarized the early industrial landscape in one forceful judgment: “Before ChatGPT, there were no secrets in this industry.” Once he saw the output of a Google or Microsoft application, he felt comfortable estimating whether it could be replicated in a few months and for a few million yuan.
Some of the algorithms and engineering tricks Metaso designed for translation appeared in Google publications about a year later. The team did not publish papers simply because it considered product performance more important than publication.
GPT-4-class industrial-scale models were what broke that sense of certainty: for the first time, he felt that knowing the output quality did not mean a team could reproduce it with limited resources.
9. GPT-3’s Breakthrough Was Pushing Scale into Territory Industry Would Not Consider
Around 2020, industry still viewed BERT base’s 100 million parameters as too large, with research focused on quantization and compression and on shrinking models to 3 million or 5 million parameters.
GPT-3 did not introduce much new architecture. Aside from some details around attention, it was largely based on the 2017 design; what it really did was push scale beyond 100 billion parameters.
Even 闵可锐 found the experiment merely “pretty fun” at the time and did not think it could become a product: it was roughly 1,000x more expensive than the 100-million-parameter model industry already considered too costly, making it look more like scientific exploration.
ChatGPT and GPT-4 later presented that scale of capability in product form, creating the real industrial barrier.
10. To Him, the BERT-versus-GPT Debate Was Like Arguing over a Red Pen or a Blue Pen
The original Transformer contained both an encoder and a decoder; BERT extracted the encoder for pretraining, while GPT used the decoder to predict word n+1 from the first n words.
闵可锐’s analogy was: “It’s like a red ballpoint pen and a blue ballpoint pen. You insist on saying which one is better—does that matter?” The task determines the architecture.
曼琪 argued that after the iterations through 3.5 and 4, GPT had indeed demonstrated stronger generality. 闵可锐’s response was that the market had never conducted a fair, large-scale comparison at the same model size.
Comparing a 100-billion-parameter decoder-only model with a 10-billion-parameter alternative and then declaring the former stronger does not establish architectural superiority; encoder-decoder models such as T5 had also gone head-to-head with GPT at comparable scale.
11. Decoder-Only’s Victory May Have Started with Machine-Cost Economics
闵可锐 believed one basic reason OpenAI adopted decoder-only was that the architecture was simpler, easier to optimize and better able to keep GPUs fully utilized.
Encoder-decoder models involve two computational chains—sometimes A waits for B, sometimes B waits for A—making resource scheduling materially more complex; decoder-only may therefore have delivered a 20%-30% throughput advantage.
The expansion of this route may thus have reflected engineering and machine costs making it the easiest design to scale, rather than a philosophical commitment to next-token prediction.
曼琪 cited Ilya’s detective-fiction analogy, in which predicting the murderer at the end already entails reasoning. 闵可锐 responded that from the standpoint of theoretical computer science, sequence-to-sequence models can simulate any computable function, which is a “comparatively trivial conclusion.”
12. Multimodality Is Converging Different Domains into the Same Computational Problem
In the past, researchers working on speech, NLP and images each had their own feature-engineering and signal-processing skills. Deep learning has increasingly relied on unified architectures and compute, and less on domain knowledge.
闵可锐 used Tesla’s autonomous driving as an example: raw signals from cameras and sensors can all enter one unified model, which then outputs predictions such as object locations; in essence, the problem can still be written as sequence-to-sequence.
The problem of excessively long image sequences can also be compressed through engineering: grouping 32×32 pixels into one larger pixel can reduce a sequence of millions to roughly 1,000, after which a standard Transformer can be used.
The core idea behind Vision Transformer can be explained in one sentence, but completing the experiment and proving it works still has value. He simply rejected mythologizing that work as an impossibly brilliant breakthrough.
13. Truly Brilliant Work Often Requires a Semester of Knowledge to Understand
The major breakthrough in 闵可锐’s mind was taking network flow and maximum-flow problems—studied for more than 40 years and widely thought to require n² or n³ complexity—and bringing them down to near-linear complexity with a proof spanning dozens of pages.
Understanding such a paper might require 20 sessions of lecture notes. Even with all the tools in place, the final step would remain non-trivial; only after understanding it could one appreciate that the technique was “really very strong.”
The media, by contrast, find researchers’ legendary backstories easier to spread. 曼琪 believed ordinary people could only empathize with the story; 闵可锐 agreed, but warned that a formidable person without a compelling story might solve major problems and still go unnoticed.
14. Deep Learning Blurred Disciplinary Boundaries While Amplifying the Risk of Opacity
The breakthrough in NLP cannot simply be attributed to vision researchers crossing into the field: deep learning was itself a new paradigm for vision researchers, while Transformer expanded from translation into images, audio and video.
Machine-learning research in the 2000s focused heavily on graphical models. Theoretical researchers questioned deep learning for years, with a major reason being that it was substantially less interpretable than prior methods.
闵可锐 summarized the shift this way: before, it was “you are not interpretable, and I don’t think you work particularly well either”; now it is “you are not very interpretable, and somehow you work extremely well.”
Once capabilities exceed expectations, regulatory demands naturally rise. When a model makes a mistake, it may not be a minor error but a serious consequence of the model’s involvement in more consequential decisions.
15. The Product Line Ran from Translation and Correction to Generation and Dialogue—but No Single Model Did Everything
Metaso launched legal translation in 2019. 写作猫 first offered Grammarly-like checking and correction, added generation in November 2022, and then dialogue in February 2023.
The team followed a simple rule: use whatever works. Cost, online user volume and inference capability would all determine the model mix; there was no insistence that every feature run on the same large model.
写作猫 did not merely change A into B. It also tried to explain why something was wrong, aiming to be “a professional teacher sitting beside you all the time” and checking every line in real time.
闵可锐 believed new entrants often overestimated large models: some “got carried away and insisted on going all in,” while others knew the economics did not work but used the concept first to push up a stock price.
16. Search Economics Did Not Work, While Professional Services Could Absorb Expensive Inference
闵可锐 cited the algorithm discussed on a 360 earnings call: generating roughly 200 Chinese characters cost RMB0.04, implying annual costs of more than RMB10B if GPT capabilities were applied to every search query, with daily costs around RMB40M.
Compared with 360’s annual advertising revenue of roughly RMB8B, that meant scaling a generative model across all search did not work at the cost level of the time, unless a new business model emerged.
Jasper’s important contribution was finding the wedge where “machines may be expensive, but people are more expensive,” keeping each inference materially cheaper than the labor it replaced.
17. The Difference between Legal and News Translation Was Not Quality, but Willingness to Pay
Human legal translation at a top law firm could cost more than RMB1 per character. A senior lawyer might charge RMB5,000 for an hour with a client, making a machine subscription costing tens or hundreds of dollars still look cheap.
Metaso believed it could also translate news better than Google, but chose not to: “Who would be willing to pay us for news translated just a little better?”
High-value work can support trading compute for labor. Low-value content remains difficult to monetize even when quality improves; this was a harder constraint on Metaso’s application choices than any model benchmark.
18. 写作猫 Used Free Allowances to Buy Scale; Paid Conversion Was Not a Standalone Optimization Target
At the time, 写作猫 had more than 7 million registered users and a paid ratio of roughly 2%-5%. The team considered that reasonable and was not rushing to push the rate higher.
The product gave free users an 8,000-character daily checking allowance, allowing low-frequency users to keep benefiting. Reducing free benefits to zero would certainly lift paid conversion, but many users who did not convert would also leave.
“Several tens of percent” of new users used up at least their free allowance for generative features, showing that generation and dialogue remained a first real experience of a new interaction for many Chinese users.
闵可锐 defined the target customer pragmatically: someone who writes only 500 characters a month might remain free forever, while high-frequency writers such as integrated-media professionals and civil servants would pay to reduce errors and save proofreading time.
19. Legal-Translation Penetration Was Not Low, but Demand Was Locked to the Cycle of Cross-Border Work
Legal translation was mainly sold as a standardized B2B product to law firms, lawyer studios and partner teams. Within China’s limited pool of cross-border lawyers, Metaso estimated penetration at roughly 30%-50%.
Lawyers would never “open this product for fun when they had nothing to do.” Usage depended entirely on whether they had cross-border M&A or other international matters; related business fell in 2022, and usage fell visibly with it.
The product was also sold to Asia-Pacific offices of UK and US law firms, but after China-US trade tensions, US firms subjected Chinese software to more rigorous IT reviews, and the onboarding cycle for a new vendor could stretch into years.
The market had “a very limited ceiling,” but that also made it unattractive for new players to invest in for the long term. For law firms willing to pay for high-quality service, there were not many alternatives.
20. Subscriptions Are Harder in China, but Heavy B2B Is Not Automatically Easier
闵可锐 believed Chinese users might have weaker payment habits than overseas users, with significant differences in revenue per customer; Jasper could reach $100M ARR quickly, but reproducing that pace in China with the same product would be difficult.
RMB100M in subscription revenue was still possible, while a $100M-scale subscription business was rare. But switching to B2B did not automatically solve the problem: “Doing even B2B in China is not easy.”
Legal translation could not suddenly give lawyers 3x as many orders; it could only make existing work more efficient. Metaso therefore knew from the beginning that a specialized niche could not independently support unlimited growth.
21. Standardized Light B2B Preserves the Technology Edge; Customization Can Drag Both Sides into a Lose-Lose
Metaso preferred selling one standardized product to enterprises and absorbing the higher business-development costs itself. 闵可锐 was wary of customers’ tendency to turn a standard product into a customized service.
He explained the mechanism with the example of a client revising a document 10 times and then returning to the first draft: clients often wanted to go through the revision process, not simply receive the final result.
Heavy B2B sales chains are too long, with technology only one link in the chain. A team of top talent might spend 6 months landing a RMB200,000 or RMB500,000 project, with revenue well below the cost of the people involved.
The result was “mutual harm”: the client felt RMB500,000 entitled it to offer endless opinions, while the vendor knew it was losing money but still needed the client logo to tell a story to the next round of capital.
22. The Next Market Had to Be an Order of Magnitude Larger, Not a Copy of Several Small Projects
闵可锐 did not want to replicate multiple specialized tools generating RMB1M a year, because two projects of the same scale would have limited strategic significance for the company.
He wanted the next opportunity to scale at least one order of magnitude beyond the existing applications, preferably two. That was also why 写作猫 targeted a broader writing population.
But a larger market also meant stronger competition. Pressed on the source of differentiation, he refused to offer a static answer and simply said: “Wait and see.”
23. A Startup’s Moat Has to Grow from the Edge Markets Giants Ignore
“Wait and see” was not a lack of judgment, but an acknowledgment that technology, interaction and user feedback were changing rapidly. The team had to keep identifying which problems were worth optimizing and do them better than others.
After using GPT-4, users would only ask why your output was worse than GPT-4; they would not care about the cost gap. That made a broad “total war” unfavorable for resource-constrained startups.
The opportunity lay within 0.1% of a giant’s revenue: RMB100M could be strategically important to a startup but not worth assigning a giant’s main forces. A startup could occupy the position first.
Writing was not one market. Civil-service writing, e-commerce copy, email marketing and SEO required entirely different knowledge, interaction and workflows, and could not necessarily be solved by one general-purpose model.
24. The Application Layer Would First Become a Melee among Individual Developers, Startups and Big Tech
闵可锐 called the application ecosystem of the time a “Warring States period”: juniors and part-time developers using the OpenAI API could launch a feature in 3 days that challenged years of accumulation at a major company.
Individual developers faced extremely low trial-and-error costs: “The worst that can happen is that it doesn’t work.” Once they saw that someone else had merely wrapped an API, they could quickly copy the product and compete on price.
The market had therefore not yet reached a period of business-model convergence. Subscriptions and pay-per-use were merely charging formats; the real question was what value the product offered that a new entrant could not immediately copy.
He did not believe there was a “brilliant business model” that could remain valid forever regardless of changes in open source, model prices and the competitive landscape. In a world that changed every day, companies had to adjust dynamically based on feedback.
25. AI Native Will First Change Interaction, Not Create a Clean From-Zero-to-One Break
闵可锐’s understanding of AI native was literal: major technological breakthroughs would come with changes in user interaction, while migrations by large companies would be constrained by legacy interfaces and existing user habits.
Once a long-established product changed its interaction model, it would inevitably alienate users accustomed to the old version. The burden on a large company to innovate was like forcibly taking back users’ old phones and handing them an immature replacement.
曼琪 defined AI native as “something that was impossible before but is possible now.” 闵可锐 rejected that binary split, arguing that AI mostly occupied a gray spectrum from “doing poorly” to “doing better.”
Multi-turn dialogue within a defined scope could have been simulated by expert systems 20 years ago; the scope was simply narrow, the cost high and the conversion poor. The change he emphasized was the expansion from specific scenarios toward more general capabilities, together with improvements in performance and cost.
26. Healthcare AI’s Biggest Obstacles May Be Power, Data and Regulation Rather Than Model Capability
闵可锐 was unwilling to draw a technical conclusion about IBM Watson’s failure, but suspected that healthcare’s strict regulation, high cost of errors and organizational barriers mattered more than the performance of any single model.
Famous doctors might not be highly dependent on a hospital, while rare-case data could affect papers and promotion. Doctors often wanted to publish first before placing “private data” into a public system.
Departments also competed with one another. Administrative orders could not easily eliminate concerns about data sharing, while a system claiming to be more intelligent would eventually face the response: “I’m the one making decisions here. You’re telling me how to do my job?”
曼琪 called this social engineering. 闵可锐 went further: if a system eventually told doctors, “You’re no longer needed,” why would doctors voluntarily help train it?
27. Self-Training, Fine-Tuning and Prompting Form a Continuum of Control and Investment
闵可锐 arranged the three routes as a control chain: training from scratch required the most investment and offered the most control; fine-tuning an open-source model sat in the middle; prompting was cheapest but depended most heavily on capabilities developed by others.
If the base model did not possess a capability, prompting could not create it from thin air. “If it has the capability, you can use prompts to bring it out. If it can’t do it, don’t do it yet.”
Metaso wanted to find high-value problems that the third tier could not solve and the second tier could not solve well, then return to data, architecture and pretraining sources to address them, rather than spending first on training and looking for a use afterward.
The team had trained models from scratch since 2018, initially without a suitable open-source base. Translation and 写作猫 used different models, with a new version iterated roughly every 2-3 months on average.
28. Parameter Scale Is a Cost, Not Capability Itself
In its early days, Metaso mainly trained on self-purchased consumer gaming GPUs, accumulating several dozen, with each card usually costing less than RMB20,000. For solving problems in a specific domain, more parameters could actually be a burden.
Around the end of 2022 and the period before ChatGPT, the team concluded that writing generation required a larger model and began using more cloud GPUs. Even then, the investment could be covered by its own revenue; it did not need to raise a separate round to rent a dozen cards.
闵可锐 stressed that there was no one-line “brilliant trick” explaining how to achieve good results with limited compute. It was the cumulative effect of data, architecture, training stability and engineering optimization.
The model team had fewer than 10 people, many of whom had worked together since Day One, and was not made up entirely of the most decorated résumés. The key was: “I’m someone who knows how to spend money, so that’s enough.” He still made the major investment decisions.
29. The Real Compute Bottleneck Is Not Just GPUs, but Power and Interconnect across an Entire Data Center
In 2022, monthly cloud rental for a single A100 could be below RMB10,000. By the time of the interview, renewal prices had not changed much, but new capacity was harder to secure, and buying a single card cost around RMB100,000.
Metaso believed it could still source up to 50 or even 100 cards through multiple channels, including Alibaba, Kingsoft and smaller suppliers. The real difficulty began at four-digit card counts.
Training with thousands of cards also required InfiniBand networking to link them at high speed. Many Chinese data centers had been built to standards from 10 years earlier: gaming-card servers might occupy only one-tenth of a rack while consuming the entire rack’s power budget.
New H-series cards could bring a single server close to the rack’s power ceiling. Supercomputing centers built by local governments were often idle because software and hardware were mismatched, showing that the bottleneck ran through the entire chain from defining demand to building infrastructure.
30. Open Source Will Flourish, but a Noisy Release Cycle Does Not Mean More Genuine Large Base Models
闵可锐 believed there were still few genuinely large open-source language models at the time, mainly a handful of bases such as OPT, BLOOM and LLaMA. Many releases were “fan-club models” that simply swapped in data and fine-tuned an existing base.
At more than 100 billion parameters, OPT still performed noticeably worse than the results reported in the GPT-3 paper, which was why Meta continued training LLaMA. Equal parameter counts did not mean equal training quality.
A model’s loss might fall only slightly, producing little change in its overall average score, while certain long-tail capabilities could jump from near-random to clearly usable. Deciding how to allocate compute and judge convergence was itself know-how.
China had also released models with tens or hundreds of billions of parameters, but 闵可锐’s assessment after trying them was that they were undertrained. “Getting it to run” and actually training a model well were not the same thing.
31. Open Source Cut Iteration Costs by Two Orders of Magnitude, While Closed Source Could Still Lead on Feedback Speed
Traditional open-source software had clearly defined features: when users found a bug, they could submit a patch. With language models, even what was being optimized was hard to define; improving one task could make another worse.
The core of an open-source model was its weights, not just the code. A genuinely complete release would provide the training code, inference code and parameters, allowing developers to use the model directly and fine-tune it with data such as 10,000 medical examples.
Training a 20B-30B-parameter model from scratch on a single machine could take at least a year, while fine-tuning might take 3 days. 闵可锐 estimated that this reduced the cost of innovation by roughly two orders of magnitude.
But if an OpenAI-like player could spend 100x what another startup spent and obtain in 1 day the feedback that others needed 3 or 6 months to see, the strongest closed-source model would still be more likely to lead.
32. The Boom Raised the Benchmark and Could Reorder Capital, Organizations and the Ultimate Winners
Metaso launched AI writing 1-2 weeks before ChatGPT, potentially giving itself a 6-month window to educate the market. ChatGPT instead raised the public benchmark from “smarter than Siri” to ChatGPT and GPT-4 within 2-3 months.
At the time, 写作猫 had more than 7 million registered users, a team of several dozen people and a company that was broadly self-sustaining. Revenue was below RMB20M, while 闵可锐 said total costs were several tens of millions of yuan but did not separately disclose labor costs. Financial investors were more concerned that it would face direct competition from large tech companies worth hundreds of billions of yuan.
闵可锐 concluded that Chinese VCs might “collectively miss” the genuinely ambitious large-model entrepreneurs: traditional capital preferred tycoon-backed teams and senior academicians, while OpenAI’s key figures were young, still writing code on the front line and genuinely knew where the money should go.
曼琪 cited MiniMax, 杨植麟 and others as counterexamples, arguing that capital was not investing only in traditional winners. 闵可锐 nevertheless stressed that public attention was far greater than actual investment, and that whether founders remained on the front line mattered more than résumé labels.
Model startups could also get stuck in the middle: unable to match GPT-4’s investment above, yet not necessarily strong in applications and user acquisition below. Over the next 6-12 months, every company would claim to have reached GPT-3.5, but he expected that “90% of users would still be on ChatGPT.”
The episode used education to discuss industry transformation: a robot costing $20-$30 a month could answer questions one-on-one and practice spoken English with a student 24 hours a day. Who would make the money remained uncertain, but the product could profoundly change the existing way of working.
The dense investment of the boom might not last. If core revenue fell and a team’s self-developed model costing hundreds of millions of yuan still performed worse than a newly released open-source model, it could exit within 6-12 months, and the market might once again see large quantities of second-hand A100s.
Metaso still chose to take legal work deep enough to be “better than 99% of people in the field,” while expanding beyond translation. Laws and regulations updated daily, and real-time performance, differences among legal systems, on-premises deployment and user privacy would all become long-term capabilities.
The company would use users’ active feedback—actions such as “this is inaccurate,” “ignore” and “add”—but would not touch private user data. The feedback also had to be distilled, because one person’s judgment could conflict with those of 9 others.
For Metaso, financing was icing on the cake. It targeted roughly $10M, which could materially accelerate its existing path. 闵可锐 refused to claim “full-stack deployment across 20 industries” merely to appear ambitious; his conclusion remained: “Do a good job with what’s already in our hands.”