The LLM cooldown isn't bad news — a decades-long technology cycle has barely started
Large language models are just machines ruminating on knowledge humans have already digested; real machine intelligence will learn by observing the world directly. After the cooldown, only those who keep building will make it to the end.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Large language models are just a cow chewing its cud
Chen Xun uses one metaphor to explain what large language models are doing: humans first observe the world and sediment their cognition into linguistic symbols, and large language models then learn from that corpus that has already been digested once — so they are ruminating. He stresses this is already astonishing, but machine intelligence will not stop there: machines can observe the world directly, observe the universe, and even invent their own symbols, which might be zeros and ones that humans cannot read. If machines can observe society directly and distill intelligence from it, then combine that with the capabilities humans have already summarized, the imaginative space is immeasurable.
— Chen XunAI 2.0 winners relied on a single scalable market
At SoftBank Vision Fund I, Chen Xun invested in AI 2.0, and he sums up that generation's biggest challenge as variable costs being too high: from collecting data, cleaning, governing and labeling it, developing models, fine-tuning, to inserting them into the business, every real application had to be redone from scratch, requiring large amounts of manpower, time and top-tier talent, and depending on enterprises changing how they operate. The good companies that broke out shared one trait — they found a single market that could be scaled, so those upfront costs could be amortized away. The most successful application was facial recognition, which can be used in many scenarios, but overall commercialization was not good.
— Chen Xun3.0 turns variable costs into fixed costs
The biggest change in the large-model era is doing all those things that had to be repeated in AI 2.0 in one go, turning variable costs into fixed costs. Doing tuning, industry-ization, enterprise-ization and professional application-ization on top of foundation models makes the cost threshold very low, the speed very fast, and the demands on people much lower. Chen Xun says 3.0 has produced signs of AI for everyone — any enterprise with even a shred of technical self-respect is now working on large models, trying, exploring, learning, and some have even pushed into real application scenarios.
— Chen XunDeployment is stuck on the open-versus-closed contradiction
Chen Xun points out that the large-model era must solve a fundamental problem: users have no choice but to use foundation models, but they must solve information security and trade-secret protection, because once you hand secrets to a large model in the cloud, nobody knows when, in what form, or by whom it will be called back out — so large enterprises broadly and explicitly forbid employees from using cloud services like ChatGPT for company work. At the same time, foundation-model providers have their own IP protection problem — the derivative large model you optimized is no longer mine, so what commercial interest do I have going forward. This open-versus-closed contradiction did not exist in the small-model era, because then one thing was done from start to finish within a controllable scope.
— Chen XunWhether OpenAI can run away with it depends on commercial moats
Chen Xun does not believe OpenAI can run away with it and keep widening the gap forever. His logic: OpenAI showed everyone this can be done, and once it is known to be doable, the world — especially the Chinese — have enormous capability, skill and determination to do it well, and many number twos will emerge, and number twos find it hard to pull apart from each other because they start the same and their technical architectures are the same. The key question is whether OpenAI can quickly convert its technical lead into other commercial moats, so that when others catch up technically it is no longer competing on technology. He draws the analogy of Google: Google's dominance was not because nobody else could walk the technical path, but because it formed a commercial ecosystem.
— Chen XunThe cooldown was predictable; calm after frenzy matters most
Chen Xun says the cooldown was predictable, because humans are usually too optimistic about the short term and too pessimistic about the long term. After OpenAI demonstrated feasibility everyone was hugely shocked, and excessive expectations were normal; now it is simply a gradual rationalization. He is not anxious at all — on the contrary, he thinks everyone is now doing deployment work. He gives the example of text-to-image for ad images: none of the generated photos of a bag is actually that bag — color, shape, proportions and dimensions can all change — so the advertiser faces the question of which bag it is actually advertising. Foundation models face one problem or another wherever they are deployed, and people are needed to solve them.
— Chen XunThe transition from training to inference is not yet done
Chen Xun says today's moment is better described as the starting phase of a decades-long AI cycle. Citing OpenAI's call volumes, he points out that if you go ask every other large-model company or open-source organization about inference volumes, you find inference volumes are actually not large — everyone is making models, developing models, tuning models, training models, but the scenarios where they are actually put to use are not many. He judges that the most anticipated thing in the next three years is large models moving from training to inference, and the inference ecosystem will become very rich, because inference is about applying capability in different scenarios — consumers, enterprises, governments — and the forms will be very different, as will the computing requirements they carry.
— Chen XunDon't make your personal abilities the objective function
Chen Xun's advice to native AI founders: startups are always the same — first, think clearly about whom you serve and what value you provide; don't casually assume you are Steve Jobs; the vast majority of people need to understand, get to know and discover the needs of a locked-in user group. He recently talked with a friend about to start a company, and the friend said he chose this direction because it maximizes years of accumulated experience, lessons and abilities. Chen Xun says that is a young person's thinking; he doesn't think that way now — maximizing your personal abilities only gets you so far; what you most need to figure out is where the world is heading and what you can do within that big trend. Even if some of your abilities are useless in that big trend, so what — an arm can never twist a thigh.
— Chen XunIn their own words · checked verbatim
The biggest change in 3.0 is that it takes all those process steps that had to be done before and does them in one go, turning variable costs into fixed costs.
3.0的最大的变化就是他把前面要做的这个 所有的那些过程当中的事情 能够把它一次性做完 把可变成本变成固定成本
Chen Xun13:16
But I think in the end it won't be like that, because today it's already very clear with large models: no matter how good a technical large model you build, when you face an industry's problem, it still falls short.
但是我觉得最后不是这样 因为大模型今天已经看得很清楚 做的再好的技术大模型 当你面对一个行业的问题的时候 它还是有所欠缺的
Chen Xun24:26
Whether it can quickly convert its technical lead into other commercial moats, so that when others catch up technically, it is no longer competing on technology.
他能不能够把技术的领先 迅速的 转换成其他的商业的壁垒 这个时候在别人在技术上赶上你的时候 你就不靠技术竞争
Chen Xun26:29
The lesson is that calm after frenzy is the most important moment, because the work that truly puts this thing to use and produces real value is all done at this time — only those who persist at this time.
心得是说狂热之后的冷静 是最重要的时刻 因为之后真正能够把这个东西用起来 产生真正的价值的工作 都是在这个时候做出来 只有在这个时候坚持的人
Chen Xun34:39
My own judgment is that the most anticipated thing in the next three years may be large models moving from training to inference; the inference ecosystem will become very, very rich — inference is far richer than training.
我自己的判断是 未来三年可能最值得期待的一件事情 大模型从训练走向推理 推理的生态会变得非常非常的丰富 推理要比训练丰富的多
Chen Xun36:39
Even if some of your abilities are useless in this big trend, what does it matter? Which is more important — that thing, or your abilities? In the end, an arm can never twist a thigh.
即使你有一些能力在这个大趋势里头 没有用 有什么关系呢 是那个东西重要 还是你的能力重要 最后胳膊永远是拧不过大腿的
Chen Xun1:00:56
Figures
| Foundation model development cost | hundreds of millions of US dollars, tens of billions of RMB | 18:00 |
| XGBT visits month-over-month growth (April-June) | 12.6%, 2.8%, -9.7% | 32:37 |
| XGBT visits month-over-month growth (January) | 131.6% | 32:37 |
| Internet bubble peak to trough | peaked in 2000, bottomed in 2002 | 34:39 |
| Machine vision frame rate | the most advanced can reach 1000 frames per second | 41:44 |
Glossary
- Foundation Model
- A large model that provides a high-quality starting point for a rich model ecosystem; in China it is also called a large language model.
- society of minds
- The concept of multiple models competing, complementing and collaborating with each other to ultimately form a community-style intelligence.
- EV/EBITDA
- A valuation metric commonly used by mid-to-late-stage investors, using EBITDA in place of complex cash flows.
- bottom-up
- Investing is a micro-level business: you must understand how each company differs from the others, rather than investing by macro track.
How to listen
Investors and founders watching large-model deployment, the inference ecosystem and AI startup directions — especially those trying to decide whether to keep investing through the current cooldown.
The opening personal résumé, and after 45:16 the parts on switching between investing and founding and on mid-to-late-stage investor taste can be fast-forwarded.