Warning: The Kimi K3 Stock Market RESET.

요약 보기

Kimi K3 출시 후 미국 반도체 약세장 진입. Kimi는 Claude 증류(distillation) 의혹이 있으나 MuonClip, Delta Attention 등 실질적 효율화 기여도 있음. 오픈 스케일(open scales)이지 완전 오픈소스가 아니며 향후 수익화 목적. 효율화가 모델 품질 개선→사용량 증가로 이어져 AI 하드웨어 수요에 단기 긍정적이나, 모델 개선이 멈추면 AI 랠리 종료 위험. Anthropic·Moonshot IPO는 지속적 자본 조달 목적으로 해석.

다듬은 글

※ 자동 자막 원문(다듬기 전)

01. Today we're going to try something completely different. We'll take a professional look at what's killing semiconductor stocks , including Kimi k3, and everything you need to know about what's really going on and how it's affecting your money in this video. Congratulations to everyone who came in positive on the triple Qs index this morning in our alpha report.

02. We thought we would go from -1 to a positive value . Boom, we just went "in the black" on triple Qs. They drove. If you need these trading signals, be sure to join meetkevin.com. Now that I'm back, we have a coupon that expires tomorrow night. Check it out at meetkevin.com. US semiconductor stocks briefly entered a bear market, and this trend may continue.

03. A lot of this has to do with Kimi. But first, look at the graphs yourself. When you look at the Philadelphia Semiconductor Index, you see that it is down 15% at the time of this writing, but it has entered a “bear” market, down more than 20%. The 3x leveraged version is still almost 50% below its historical highs today.

04. This video is not about the dangers of using leverage. This video is about what Kimi really means. There are five really important components . I think once we understand these five different components, we'll get a much better idea of ​​what this really means for large language models and hardware investments.

05. So, let's keep it simple and take it step by step. First, the initial statement that comes out of Kimi k3's breakout is that this is it. And this is what I call the mindset of beginners. Kimi k3 argues that if we regulate American companies like OpenAI or Claude, we end up allowing countries like China to catch up with us with AI breakthroughs.

06. They come out of nowhere, and that's how we end up losing to China. This is probably quite selfish on the part of venture capitalists who don't like any regulation, and I'm not here to say whether it's needed, but simply to say that this is what I think the newbies are saying . But a professional looks at it and says, “This did n’t happen because of regulation.

07. "What happened here actually happened through distillation." And that's just one of the five components of this video. So let's just dismiss this distillation argument. You may have heard this before, but here's my take, and then we'll move on to the other four points. Look, I think a pro looks and says, "There's a lot of strong evidence that Moonshot AI actually copied Claude." Anthropic itself documented over 3.4 million queries to Claude extracted through fake accounts, IP addresses, and metadata, and they were able to link the actions of Moonshot employees and their profiles to exactly what was being extracted from Claude.

08. So you had a bunch of people taking Claude's files and just constantly photocopying them. Of course, this is a very simplified view. In fact, they purposefully copied agentic thinking, tooling, programming—this is probably the most profitable direction for such LLMs, as well as computer vision. And it was in these areas that Kimi K3 scored the highest points.

09. This does not mean that Kimi just copied Claude, because in some aspects Kimi even surpasses him, and there are reasons for that . So we'll talk about the other four components that really matter here. But we can't start a conversation about Kimi or this discussion without mentioning a fact that Redwood Research found: K3 often identifies as Claude 4.5, and, as they noted, disproportionately often.

10. There have also been independent studies that found that the Kimi K2, an older model, is 82.7% similar to the Claude Sonnet 4.5. The interesting thing here is that you would expect that Claude Opus is quite similar to Claude Sonnet, you know, because they are kind of an evolution of each other. You would expect them to be similar.

11. But a study on arXiv showed that Kimi K2 was more similar to Sonnet than the other Claude models were similar to each other. This is actually a way of saying that they copied Claude in the past. They probably copy Claude even today. So, point number one: when it comes to Kimi, the newbie says, “We let China win because we tried to regulate AI.

12. Now they have a better model with open scales than we have in America.” Forget the fact that they haven't even released the scales yet and it's not open source. These are open scales. This is different. "They win, we lose because we tried to regulate them." The pro says, “No. Kimi just blatantly copied American companies with K2, and probably did the same with K3.

13. She calls herself Claude. "They almost certainly used our American data after training." So, while it's not exactly a copy of Claude, there's no doubt that in this case the Chinese AI looked at our work like a student taking an exam and thought, "Oh, I'm late for my exam. " Come on, I'll quickly copy all the answers from that smart guy, and I'll do the last 5% myself to pretend like I did something too ." Another analogy can be made : it's like horse racing.

14. Can we just give a horse a dope, inject it in the buttocks, to catch up with the leaders in a race? They are all still competing, but someone has gained an unfair advantage. And let's put it this way , this is kind of China's usual method of operation. If we spend $10 billion on something, China often comes along and says, "We'll do the same thing in a year for 1% of the cost ." We will do it for $100 million.

15. This is the nature of how China has historically always caught up with American innovation. They steal and copy, then reproduce it cheaper. But this is unfair, because it is clear that the costs of understanding what works do not fall on Chinese companies. They fall on American companies and American investors.

16. And so this "Timmy moment," point number one, isn't as significant as it's made out to be. Now point number two . To be fair , Moonshot aren't stupid guys. It's not like some idiot just copied Claude and it's all pure imitation. That is not what we are claiming here. So, to be clear: a newbie will hear what I said and ask, "So you're saying Timmy is just a copy of Claude?" " No." They actually did real work." And here a professional will say: “ Look, they really have some good revolutionary ideas that the entire industry can use.” For example, they use the Muon Clip optimizer.

17. This is essentially a sophisticated way to make learning more stable. I like to use this analogy. Imagine someone throwing you a newspaper in Spanish, and you don't know the language and you just "fall into a rut" because there's too much to learn. LLMs can also "drop out " during their studies, and this is called a spike.

18. Really, and I'm going to make it very simple now, okay? They can go the wrong way and mess up their training data. Now imagine that you have a Rosetta Stone instead. We're not even at Duolingo level yet, right? These are interval repetitions, that's pretty good. Now we're like Rosetta Stone, where we give you this newspaper in a more understandable form to help with your learning, keeping your pace without unnecessary stress.

19. They really innovated, didn't they? This is Moons Club. They did a great job helping to train large language models. So they deserve recognition. They are a good laboratory . They do useful things, just like Anthropic does useful things for LLM, as does GPT, and Gemini, and Grok somewhere around there .

20. They also use a strategy invented by DeepSeek, a “mix of experts” model or strategy that actually activates only a fraction of the trillion models, engaging only 3% of the model at a time for clearer and more efficient results. When I say "sharper," you really do get more accurate information when you use smaller parts of the model, only the ones you need.

21. You get less noise. So they really developed good strategies for creating LLMs. And we don't want to just call Kimmy a copycat. We want to be very clear on this issue. The newbie says : “Okay, that’s enough. I've already listened. Kimmy AI is just a copy.” But a pro knows they actually did good things too.

22. First of all, yes, they copied Claude very well. This is bad. Probably, at least. Second, they perfectly copied DeepSeek's expert mix strategy , optimizing the amount of LLM work for each query. Of course? But they also added a little “ sugar”—or whatever you want to call it, salt and pepper—and invented the MoE on clip engine to make LLM learning more efficient.

23. So we can't say that Kimmy is just a mindless copycat or a lazy student who skipped half the school year and copied from others to catch up. They are actually smart students who may be ruthlessly copying American models, but still add their own value after being "bad students" and getting dropped, right?

24. As with everything, there are nuances, and that brings us to point number three. You should know the difference between open source and open scales, because this is where we get into the monetization, investment, and money we all love so much. Look, the newbie says that Moonshot's open-source release of the K3 scales makes the model open, like Linux, on top of which anyone can freely build.

25. The pro replies: " Wait, not so fast." Open source means you get everything. Model weights, training code, training data, full recipe. Anyone can recreate this from scratch. It would literally be like getting all the data to train Tesla cars. For example, cameras, people in the car, accidents, dangerous situations.

26. Each individual input from users. So, someone else can replicate Tesla's autopilot system. This is technically open source. Open scales are completely different. Open scales mean, “Hey, you get full autopilot. You get a Kimmy K3, a full LLM, a Tesla that drives itself, but you customize the style the way you want it.” Again , I’m simplifying, but this is essentially a way of saying, “How far do we want to deviate from the data?

27. How creative do we want to be? How deterministic do we want to be? Which is very useful because, let's say, if you're writing a legal document, you don't want to be creative. You want to be very precise, which is different from when you're telling a story . But maybe if you're just forming an idea for a court case or how you're going to present it, you can add a little creativity.

28. And you can adjust the weights for what your company ultimately does without telling others how you're building your AI recipe, which gives you a certain proprietary advantage because you're not giving away all your weights, which, as Palantir's Mr. Karp says, is a big concession to companies like Anthropic, because it's your proprietary information.

29. You shouldn't give away how you customized your LLM to get the results you're trying to achieve in software, law, business, accounting, or anywhere else. This should be your own information. And so that's another area where people are somewhat obsessed with Kimmy. Like, yeah, okay, let it be. It's not open source, but open scales—that's pretty cool.

30. And they're giving it away for free, so it must be good, right? Well, this is where people who understand say, "Of course." Kimmy is essentially catching up with others by creating this product to give it away for free and build a relationship with you . Now that you have a relationship and Kimmy installed, you'll want Kimmy K4.

31. You will want new models. You will need corporate support, customization, and everything Moonshot AI has to offer in your direction. So let's simplify this for a moment. A newbie comes in and says: “Kimi has open scales, meaning it’s actually open source. They have no goal of making a profit. Wow, they are so wonderful.

32. American LLMs like Anthropic and GPT are a thing of the past.” A professional comes and says, " Stop, stop, stop, not so fast." They were the last to show up. This is an advertising technique. They literally give out free samples at Costco, hoping you'll sign up for a membership in the future. And this is open scales, not open source.

33. You can download Kimi, customize it for your software, and run it yourself, but when new products or corporate support are released, they will make a profit. That's why they're looking to go IPO, just like Anthropic. So did Kimi really kill American LLMs? No. Rather, they simply appeared, caught up with others, and added new ideas and features to the LLM discussion.

34. A professional knows this and understands that they are simply trying to dump prices to capture market share and business. Quite logical. Now let's move on to money. And when I say " logically," I mean the free- to-play part, not the copying part. I want to emphasize this clearly. We do not condone copying or distilling Kimi.

35. Good. The fourth point in all of this. The Kimi K3 model is collapsing shares of American semiconductor companies. In fact, Deep Seek Trade 2.0. Let's break this down. Newbie says: “Kimi K3 launched and semiconductor stocks plummeted, just like with Deep Seek. The same strategy, the same logic, nothing more is needed.

36. We need less computing power. That's all. "Stocks of companies like Nvidia and AMD have burned." A professional sees it differently. He knows that Deep Seek was focused on computational efficiency, for example, using a mix of experts to run part of the LLM rather than the entire model. K3 is a bit different.

37. A professional understands the difference. Kimi also introduced something called Kimi Linear, also known as Delta Attention. Basically, while learning or chatting, instead of saving everything, LLM only pays attention to the “delta ”—a fancy word for changes. This works too. As I said, when studying, if you read a book, you don't have to read every page completely, you just study the information that has changed.

38. This helps you study more effectively. This is really important when studying. This is very important in chats. Because standard memory works in such a way that when the context is doubled, the memory requirements increase by a factor of 4 . So, your chat context window becomes twice as large, and your memory requirement is 4 times larger.

39. This gets expensive very quickly. That's why Kimi introduced an innovation with Delta attention. Which, again, is useful. So Kimi really adds value to AI development, and that's cool. The danger is that some people say, “Oh, well, DeepSeek’s efficiency meant more usage,” but that’s not entirely true.

40. People say, "Oh, it's Jevons' paradox again." If costs fall, we will use it more. I think these are related things, but that's not what's actually happening. You see, DeepSeek's effectiveness didn't automatically mean more usage. DeepSeek's effectiveness in "mixture of experts" models has allowed for the creation of better models.

41. And as the models got better, more people started using them. This is another stage. I really think this is very important to consider. Let's erase everything from the board and solidify this point, because this is probably one of the main takeaways from this video that many people misunderstand. So, just saying, "Hey, something has become more efficient, so we're going to use that more." Of course?

42. This does not reflect what is actually happening. In fact, there is an increase in efficiency. As a result, the models become better, retain more contextual memory about the conversation, and hallucinate less. And this leads to increased usage. In other words, a bad, ineffective product doesn't become more popular.

43. A better product gets more use. So, in my opinion, quality is related to usage. If as quality increases, the product becomes more efficient—that's great. But it is the improvement in quality that leads to increased usage and new breakthroughs. So the bottom line is that Kimi's breakthroughs in improving models are driving the development of LLM.

44. This is actually optimistic for AI and iron stocks in the near term. This means that if we can learn from DeepSeek and Kimi to do certain things more efficiently, and as a result improve the quality of the models, then more people are likely to use LLM for more tasks and more often. And what will happen then?

45. The AI ​​trend will actually continue. So, when models stop getting better and computing power reaches its limit, that's when the AI ​​boom will begin to die. And you see, this is really important to know. The AI ​​boom will continue as long as models are improved. It will stop when the development of models stops.

46. Simply put , the newbie says, "The launch of Kimi 3 killed the shares of equipment manufacturers." " The newbie also says, ' Well, you know, I mean, if there's more efficiency, more people will use it, and so maybe there 's still a case for AI stocks and hardware stocks, right ?'" "The pro is actually much wiser on this issue.

47. Profi says, “The greater efficiency of AI models, like what DeepSeek brought us or what Kimi brings, actually allows us to create better models. Better models equal more use. More usage and better models mean the rally in AI hardware stocks could continue. The danger to watch out for is that when the model improvements run out , when the S-curve peaks and everything starts to slow down, that's when AI stock trading will really end.

48. That's when the collapse will come. So, no , the release of Kimi 3 is not bringing down hardware stocks. They actually help equipment promotions because you implement more efficient tools to work better with LLM. That's good. This is good for equipment stocks. But once these models stop getting better, that's your red flag to pay attention to.

49. Now, number five. Anthropic and the Moonshot IPO. A newbie comes here and says, “ Look, both companies are rushing to IPO because they are at the top of their game. Moonshot in China, Anthropic here. It's really a circle of honor, and they want to go public because they're winning." A pro looks at it a little differently.

50. The pro says, “Look, the value of any model we have today won’t be the same in 6 months. In other words, the value of the model that exists today, whether it's Fable or Opus 4A or whatever you're using, depreciates over time as other technology becomes better. No one is excited about using ChatGPT 3.5 anymore, when there are much better models available today.

51. Remember what Mark Zuckerberg talked about during Meta's latest earnings report. It's a vicious circle. You have to keep spending money to stay relevant in the AI ​​field. It actually describes a treadmill. So, yes, the incentive for IPOs is monetization, but what they're really trying to do is raise capital that can be used to continue to succeed, which means even more equipment costs.

52. Big and ...That's why stocks like MiniMax's IPO rose 109% on the first day. Although IPOs are usually intentionally undervalued for marketing hype. So, the important thing to remember here is that a newbie buys the hottest AI IPO just because it's popular. And the professional comes along and says, " Wait a minute.

53. What if in two years the numbers are, well, five out of ten? What price am I willing to pay now if there's a recession in the future?" The newcomer thinks that Anthropic wants to go IPO as soon as possible because they are eternal winners, the same with moonshot projects. But really, a professional looks at it and says, "These companies are raising money to stay relevant.

54. When they're no longer relevant, it's because the models aren't getting better. And if no one is improving the models anymore, that's the biggest risk to our AI hardware sector. It has implications for the entire economy." So let's simplify this for better understanding. And yes, if you want more of my fundamental analysis, you can always join me at meetkevin.com.

55. We have a coupon, you can it. For now, let's analyze what we just went through. There is almost no doubt that Kimi is a copy. But they also brought us real efficiency. MoU and clip optimizer are great. This is useful for large language models. This is fantastic. In addition , their delta attention is useful when learning .

56. This makes us more effective in the field of AI. And this allows the models to get better. As models improve, AI stocks could continue to rise. Is Kimi open source? No. Are they a charity? No. They use a smart marketing strategy to gain market share. But they actually introduce real efficiencies that help the broader market expand and continue to create new models.

57. The biggest risk is not the Kimi moment. The biggest risk is actually that the models will stop improving. So, I hope you enjoyed this breakdown. Let us know what you think in the comments below. .