半沐
508 Posts
半沐
慢就是快，相信复利的力量 币安100%返佣邀请码: BM168
Open Trade
High-Frequency Trader
8.1 Years
10 Following
124 Followers
272 Liked
Posts
Portfolio
All
Quotes
Today’s Korean stock market and China’s A-shares in 2015 can’t be said to be similar—they’re basically exactly the same.
ChangXin started a fight with Huawei. ChangXin wanted to raise the price of Huawei’s chips, but Huawei couldn’t take it anymore and directly demanded a price cut. Without saying a word, ChangXin kicked Huawei’s on-site team back to their mother’s house.
As for Alibaba: it holds 5% of ChangXin Storage shares, so it directly gets a long-term contract price—cheaper and more favorable. Who told you all to have voted back then? Only you, Huawei, were stubborn: $BABAB
As for Alibaba: it holds 5% of ChangXin Storage shares, so it directly gets a long-term contract price—cheaper and more favorable. Who told you all to have voted back then? Only you, Huawei, were stubborn: $BABAB
Alibaba comes in so strong right away; it looks like the most powerful domestic inference chip besides Huawei.$BABAB
MarsBit News
·
--
Alibaba Cloud Zhenwu Super-Node Adapted to Kimi K3
Mars Finance news: according to Alibaba Cloud, its Zhenwu M890 super-node instance has been adapted to Kimi K3 on Day 0. This is the first super-node in China to run a model with nearly 300 billion (3 trillion) parameters. The two sides will further expand domestic compute-power cooperation. The Qianwen AI platform and Alibaba Cloud Bailian will also provide Kimi K3 model APIs.
Made in China is rising—burst the US stock bubble. ChangXin goes public, burst the US storage bubble. Domestic lithography machines go public, burst ASML’s hegemony. Everyone, come back—Big Brother is back.
Alibaba invests 7.9 billion to acquire 5% of Changxin Memory shares; now its value is 180 billion RMB—just say how strong that is.
ChangXin Storage’s market value is 3.66 trillion, and Alibaba directly earned 180 billion RMB
The EU recently has gone crazy trying to get money, even through fines
西雅图夏至
·
--
Antitrust fine of nearly one billion euros: the EU penalizes Google for violating the Digital Markets Act
On the 23rd, the European Commission issued an announcement stating that it imposed a total fine of €890 million (about US$1 billion) on Alphabet, the parent company of the U.S. tech giant Google. The decision found that the company violated relevant provisions of the EU’s Digital Markets Act (DMA) in the operation of its search engine services and app store operations. This is the latest enforcement action by the EU regulatory authorities against market-monopoly behavior by multinational tech giants.
The penalty details published by the European Commission show that the above fine consists of two targeted sanctions: Google was fined €460 million for improperly favoring its own services, such as Shopping, Hotels, and Flights, in Google search results; and it was fined €430 million for restrictions in its Play Store app store rules that limit Android developers from directing consumers to third-party payment channels with better options.
The penalty details published by the European Commission show that the above fine consists of two targeted sanctions: Google was fined €460 million for improperly favoring its own services, such as Shopping, Hotels, and Flights, in Google search results; and it was fined €430 million for restrictions in its Play Store app store rules that limit Android developers from directing consumers to third-party payment channels with better options.
Who is the most profitable company this summer? It’s not “who”—it’s Alibaba. The Qwen 3.8 preview has been released; its capabilities have already surpassed Opus 4.8 and GPT 5.5. When the official version comes out, it must not be able to be outclassed. As for Kimi K3, it’s already been crushing tests on overseas networks, but domestically, some idiots are playing both sides—yet Kimi’s official coding plan is sold out. And for the Moon’s Dark Side, its largest external shareholder is still Alibaba, holding 36%. That means they’re directly raking in money.
Kimi is about to go public. With Kimi K3, it directly blasts out a $1 trillion market cap—Alibaba holds 36%, directly raking in huge profits: $BABA
Binance News
·
--
AI Trends | Mian Plans a Hong Kong IPO Within 6 Months
Mian told investors that it is preparing to go public within the next six months at the earliest, using the latest model to upend industry perceptions and capitalize on the opportunity to trigger a global tech-stock shakeup.Insiders say this Chinese AI pioneer has distributed shareholder resolutions to investors to seek support for a Hong Kong listing. Starting the notification process means the IPO could occur within the next six months.Mian is wrapping up a funding round that may value the company at over $30 billion. The company believes the timing is right: its annualized recurring revenue reached $300 million in June, up sharply from $200 million in April.The company began preparations for an IPO before releasing Kimi K3 last week. The model has 28 billion parameters, and its overall capabilities are second only to Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6. In some leading frontier benchmarks, Artificial Analysis ranks Kimi K3 above Anthropic Opus 4.8, marking a first milestone for a Chinese open-weight model to achieve this.
Only next to Fable 5, a model that can evolve every day—terrifyingly so. Probably preparing for Qianwen 4.
半沐
·
--
On how Qwen3.8 achieves evolution “on a day-by-day basis”: based on currently available public information, its core mechanisms mainly manifest in three areas: high-frequency automated updates, dynamic activation of the underlying architecture, and overnight validation of Agent (intelligent agent) skills. The details are as follows:
1. High-frequency automation for weight and link updates
Qwen3.8 breaks the traditional single-shot large-scale release model and establishes a high-frequency automated iteration mechanism. The team has disclosed that the model automatically updates weights every day at midnight, refreshes inference chain-link optimizations every 48 hours, and pushes brand-new multimodal alignment capabilities on a weekly basis. This turns model upgrades into a continuously ongoing “live broadcast.”
2. Dynamic expert routing and a hierarchical memory cache architecture
At the underlying architecture level, Qwen3.8 discards the traditional sparse-activation playbook and instead adopts a hybrid architecture combining “dynamic expert routing + hierarchical memory caching.” This design allows the 2.4T parameters to operate dynamically. In runtime, the model is “not a pile of brute-force monsters,” but rather an intelligent agent that can breathe, reflect, and know how to make trade-offs. This architecture-level optimization provides the technical foundation for high-frequency iteration.
3. The “overnight evolution and validation” mechanism for Agent skills
At the level of concrete Agent task execution, the Qwen series (such as Qwen3-Max) applies a “collective skill evolution workflow.” This mechanism converts real user interaction conversations into structured evidence. Evolver analyzes patterns and generates candidate updates. These candidate skills are not deployed directly; instead, they enter an “overnight validation stage”: in a real environment, both the old skills and the new candidate skills are executed simultaneously. Only when the new skills truly outperform the old ones in overall task success rate and execution stability will they be accepted and deployed. This ensures that the deployed skill pool does not degrade over time, enabling continuous and stable performance improvements.
4. Ongoing testing of preview builds and post-training
Currently, the released Qwen3.8-Max preview model is evolving continuously on a day-by-day basis. Analysts point out that continuous updates to small versions of the base parameters suggest the model may be undergoing intensive post-training in preparation for the eventual release of the final 4.0 official version.
1. High-frequency automation for weight and link updates
Qwen3.8 breaks the traditional single-shot large-scale release model and establishes a high-frequency automated iteration mechanism. The team has disclosed that the model automatically updates weights every day at midnight, refreshes inference chain-link optimizations every 48 hours, and pushes brand-new multimodal alignment capabilities on a weekly basis. This turns model upgrades into a continuously ongoing “live broadcast.”
2. Dynamic expert routing and a hierarchical memory cache architecture
At the underlying architecture level, Qwen3.8 discards the traditional sparse-activation playbook and instead adopts a hybrid architecture combining “dynamic expert routing + hierarchical memory caching.” This design allows the 2.4T parameters to operate dynamically. In runtime, the model is “not a pile of brute-force monsters,” but rather an intelligent agent that can breathe, reflect, and know how to make trade-offs. This architecture-level optimization provides the technical foundation for high-frequency iteration.
3. The “overnight evolution and validation” mechanism for Agent skills
At the level of concrete Agent task execution, the Qwen series (such as Qwen3-Max) applies a “collective skill evolution workflow.” This mechanism converts real user interaction conversations into structured evidence. Evolver analyzes patterns and generates candidate updates. These candidate skills are not deployed directly; instead, they enter an “overnight validation stage”: in a real environment, both the old skills and the new candidate skills are executed simultaneously. Only when the new skills truly outperform the old ones in overall task success rate and execution stability will they be accepted and deployed. This ensures that the deployed skill pool does not degrade over time, enabling continuous and stable performance improvements.
4. Ongoing testing of preview builds and post-training
Currently, the released Qwen3.8-Max preview model is evolving continuously on a day-by-day basis. Analysts point out that continuous updates to small versions of the base parameters suggest the model may be undergoing intensive post-training in preparation for the eventual release of the final 4.0 official version.
On how Qwen3.8 achieves evolution “on a day-by-day basis”: based on currently available public information, its core mechanisms mainly manifest in three areas: high-frequency automated updates, dynamic activation of the underlying architecture, and overnight validation of Agent (intelligent agent) skills. The details are as follows:
1. High-frequency automation for weight and link updates
Qwen3.8 breaks the traditional single-shot large-scale release model and establishes a high-frequency automated iteration mechanism. The team has disclosed that the model automatically updates weights every day at midnight, refreshes inference chain-link optimizations every 48 hours, and pushes brand-new multimodal alignment capabilities on a weekly basis. This turns model upgrades into a continuously ongoing “live broadcast.”
2. Dynamic expert routing and a hierarchical memory cache architecture
At the underlying architecture level, Qwen3.8 discards the traditional sparse-activation playbook and instead adopts a hybrid architecture combining “dynamic expert routing + hierarchical memory caching.” This design allows the 2.4T parameters to operate dynamically. In runtime, the model is “not a pile of brute-force monsters,” but rather an intelligent agent that can breathe, reflect, and know how to make trade-offs. This architecture-level optimization provides the technical foundation for high-frequency iteration.
3. The “overnight evolution and validation” mechanism for Agent skills
At the level of concrete Agent task execution, the Qwen series (such as Qwen3-Max) applies a “collective skill evolution workflow.” This mechanism converts real user interaction conversations into structured evidence. Evolver analyzes patterns and generates candidate updates. These candidate skills are not deployed directly; instead, they enter an “overnight validation stage”: in a real environment, both the old skills and the new candidate skills are executed simultaneously. Only when the new skills truly outperform the old ones in overall task success rate and execution stability will they be accepted and deployed. This ensures that the deployed skill pool does not degrade over time, enabling continuous and stable performance improvements.
4. Ongoing testing of preview builds and post-training
Currently, the released Qwen3.8-Max preview model is evolving continuously on a day-by-day basis. Analysts point out that continuous updates to small versions of the base parameters suggest the model may be undergoing intensive post-training in preparation for the eventual release of the final 4.0 official version.
1. High-frequency automation for weight and link updates
Qwen3.8 breaks the traditional single-shot large-scale release model and establishes a high-frequency automated iteration mechanism. The team has disclosed that the model automatically updates weights every day at midnight, refreshes inference chain-link optimizations every 48 hours, and pushes brand-new multimodal alignment capabilities on a weekly basis. This turns model upgrades into a continuously ongoing “live broadcast.”
2. Dynamic expert routing and a hierarchical memory cache architecture
At the underlying architecture level, Qwen3.8 discards the traditional sparse-activation playbook and instead adopts a hybrid architecture combining “dynamic expert routing + hierarchical memory caching.” This design allows the 2.4T parameters to operate dynamically. In runtime, the model is “not a pile of brute-force monsters,” but rather an intelligent agent that can breathe, reflect, and know how to make trade-offs. This architecture-level optimization provides the technical foundation for high-frequency iteration.
3. The “overnight evolution and validation” mechanism for Agent skills
At the level of concrete Agent task execution, the Qwen series (such as Qwen3-Max) applies a “collective skill evolution workflow.” This mechanism converts real user interaction conversations into structured evidence. Evolver analyzes patterns and generates candidate updates. These candidate skills are not deployed directly; instead, they enter an “overnight validation stage”: in a real environment, both the old skills and the new candidate skills are executed simultaneously. Only when the new skills truly outperform the old ones in overall task success rate and execution stability will they be accepted and deployed. This ensures that the deployed skill pool does not degrade over time, enabling continuous and stable performance improvements.
4. Ongoing testing of preview builds and post-training
Currently, the released Qwen3.8-Max preview model is evolving continuously on a day-by-day basis. Analysts point out that continuous updates to small versions of the base parameters suggest the model may be undergoing intensive post-training in preparation for the eventual release of the final 4.0 official version.
Will it directly reach a limit-up on Monday?
MarsBit News
·
--
Alibaba Qianwen Qwen3.8 Large Model to Be Released and Open-Sourced
Mars Finance News, (Science and Technology Innovation Board Daily) learned that Alibaba Qianwen Qwen3.8 large model is about to be released and open-sourced, with up to 2.4T parameters. Currently, the Qianwen 3.8-Max preview version has already been showcased on Alibaba’s Token plan, Qoder, and QoderWork. (Reporter Huang Xinyi, Science and Technology Innovation Board Daily)
Domestic AI is going to go wild; Alibaba is awesome
MarsBit News
·
--
Alibaba Qianwen Qwen3.8 Large Model to Be Released and Open-Sourced
Mars Finance News, (Science and Technology Innovation Board Daily) learned that Alibaba Qianwen Qwen3.8 large model is about to be released and open-sourced, with up to 2.4T parameters. Currently, the Qianwen 3.8-Max preview version has already been showcased on Alibaba’s Token plan, Qoder, and QoderWork. (Reporter Huang Xinyi, Science and Technology Innovation Board Daily)
Alibaba is set to launch Qianwen 3.8, and its capabilities are only second to Claude Fable 5. Alibaba is going to step up its efforts.
It’s all Alibaba’s investment—what a mess, didn’t expect that, did you? Pull it out and scare you to death.
链研社lianyanshe
·
--
Alibaba has also been lucky recently: Zhongxin has a floating profit of 130 billion yuan. Today, Kimi K3 is making the charts, and Alibaba holds 36% of the shares. Part of the investment also involves participating in it in the form of providing compute power services.
If Kimi were listed in Hong Kong, with limited shares available for trading, Alibaba, Meituan, Tencent, and Xiaohongshu have all invested. Its market cap might exceed Zhipu. Then the money Alibaba earns would be more than what it earned on Changxin—just from these two, it could make 20% of Alibaba’s market value.
Source: @lianyanshe on X
#Crypto #Web3
If Kimi were listed in Hong Kong, with limited shares available for trading, Alibaba, Meituan, Tencent, and Xiaohongshu have all invested. Its market cap might exceed Zhipu. Then the money Alibaba earns would be more than what it earned on Changxin—just from these two, it could make 20% of Alibaba’s market value.
Source: @lianyanshe on X
#Crypto #Web3
kimi.k3 will be fully open sourced on July 27. Yang Zhilin is awesome, Alibaba is awesome $BABA
Yang Zhilin: Many geniuses call me a genius
半沐
·
--
Achievements of Yang Zhilin, Founder of The Dark Side of the Moon
1. Academic achievements: laying the foundational bedrock for global long-context large language models (the highest-value long-term asset)
1) Transformer-XL (2018, co-first author)
- Core breakthrough: addressed the fatal limitations of the original Transformer—its fixed context window and inability to capture long-range text dependencies—by introducing a segment-level recurrent memory mechanism.
- Industry significance: one of the technical sources behind today’s long-context models. Many long-text optimization approaches in Claude, Kimi, and the GPT series borrow heavily from this idea. Without this work, models with one million-token long context would be next to impossible to even discuss. It also serves as a theoretical foundation for Kimi’s later focus on ultra-long documents.
2) XLNet (2019, first author, co-released with Google Brain)
- Core breakthrough: tackled information bias in BERT-style masked pretraining by adopting permutation language modeling; it surpassed BERT across 20 mainstream NLP benchmarks.
- Historical standing: widely recognized in the industry as one of the most important advances in pretraining models after BERT. Selected for a NeurIPS Oral presentation; the paper has been cited tens of thousands of times, making it one of the most influential foundational NLP works by a Chinese scholar.
3) Other academic contributions
- Collaborated with Turing Award winner Bengio to build the HotpotQA multi-turn reasoning question-answering dataset;
- Completed a PhD in just 4 years (regular program is 6 years), studied under Salakhutdinov, the former AI lead at Apple; interned at Google Brain and Meta FAIR, contributing to early R&D for Gemini and Bard;
- Among NLP papers by scholars under 35 in China, total citations have long remained in the top tier.
1. Academic achievements: laying the foundational bedrock for global long-context large language models (the highest-value long-term asset)
1) Transformer-XL (2018, co-first author)
- Core breakthrough: addressed the fatal limitations of the original Transformer—its fixed context window and inability to capture long-range text dependencies—by introducing a segment-level recurrent memory mechanism.
- Industry significance: one of the technical sources behind today’s long-context models. Many long-text optimization approaches in Claude, Kimi, and the GPT series borrow heavily from this idea. Without this work, models with one million-token long context would be next to impossible to even discuss. It also serves as a theoretical foundation for Kimi’s later focus on ultra-long documents.
2) XLNet (2019, first author, co-released with Google Brain)
- Core breakthrough: tackled information bias in BERT-style masked pretraining by adopting permutation language modeling; it surpassed BERT across 20 mainstream NLP benchmarks.
- Historical standing: widely recognized in the industry as one of the most important advances in pretraining models after BERT. Selected for a NeurIPS Oral presentation; the paper has been cited tens of thousands of times, making it one of the most influential foundational NLP works by a Chinese scholar.
3) Other academic contributions
- Collaborated with Turing Award winner Bengio to build the HotpotQA multi-turn reasoning question-answering dataset;
- Completed a PhD in just 4 years (regular program is 6 years), studied under Salakhutdinov, the former AI lead at Apple; interned at Google Brain and Meta FAIR, contributing to early R&D for Gemini and Bard;
- Among NLP papers by scholars under 35 in China, total citations have long remained in the top tier.
Genius is only a matter of seeing its threshold
半沐
·
--
Achievements of Yang Zhilin, Founder of The Dark Side of the Moon
1. Academic achievements: laying the foundational bedrock for global long-context large language models (the highest-value long-term asset)
1) Transformer-XL (2018, co-first author)
- Core breakthrough: addressed the fatal limitations of the original Transformer—its fixed context window and inability to capture long-range text dependencies—by introducing a segment-level recurrent memory mechanism.
- Industry significance: one of the technical sources behind today’s long-context models. Many long-text optimization approaches in Claude, Kimi, and the GPT series borrow heavily from this idea. Without this work, models with one million-token long context would be next to impossible to even discuss. It also serves as a theoretical foundation for Kimi’s later focus on ultra-long documents.
2) XLNet (2019, first author, co-released with Google Brain)
- Core breakthrough: tackled information bias in BERT-style masked pretraining by adopting permutation language modeling; it surpassed BERT across 20 mainstream NLP benchmarks.
- Historical standing: widely recognized in the industry as one of the most important advances in pretraining models after BERT. Selected for a NeurIPS Oral presentation; the paper has been cited tens of thousands of times, making it one of the most influential foundational NLP works by a Chinese scholar.
3) Other academic contributions
- Collaborated with Turing Award winner Bengio to build the HotpotQA multi-turn reasoning question-answering dataset;
- Completed a PhD in just 4 years (regular program is 6 years), studied under Salakhutdinov, the former AI lead at Apple; interned at Google Brain and Meta FAIR, contributing to early R&D for Gemini and Bard;
- Among NLP papers by scholars under 35 in China, total citations have long remained in the top tier.
1. Academic achievements: laying the foundational bedrock for global long-context large language models (the highest-value long-term asset)
1) Transformer-XL (2018, co-first author)
- Core breakthrough: addressed the fatal limitations of the original Transformer—its fixed context window and inability to capture long-range text dependencies—by introducing a segment-level recurrent memory mechanism.
- Industry significance: one of the technical sources behind today’s long-context models. Many long-text optimization approaches in Claude, Kimi, and the GPT series borrow heavily from this idea. Without this work, models with one million-token long context would be next to impossible to even discuss. It also serves as a theoretical foundation for Kimi’s later focus on ultra-long documents.
2) XLNet (2019, first author, co-released with Google Brain)
- Core breakthrough: tackled information bias in BERT-style masked pretraining by adopting permutation language modeling; it surpassed BERT across 20 mainstream NLP benchmarks.
- Historical standing: widely recognized in the industry as one of the most important advances in pretraining models after BERT. Selected for a NeurIPS Oral presentation; the paper has been cited tens of thousands of times, making it one of the most influential foundational NLP works by a Chinese scholar.
3) Other academic contributions
- Collaborated with Turing Award winner Bengio to build the HotpotQA multi-turn reasoning question-answering dataset;
- Completed a PhD in just 4 years (regular program is 6 years), studied under Salakhutdinov, the former AI lead at Apple; interned at Google Brain and Meta FAIR, contributing to early R&D for Gemini and Bard;
- Among NLP papers by scholars under 35 in China, total citations have long remained in the top tier.
Achievements of Yang Zhilin, Founder of The Dark Side of the Moon
1. Academic achievements: laying the foundational bedrock for global long-context large language models (the highest-value long-term asset)
1) Transformer-XL (2018, co-first author)
- Core breakthrough: addressed the fatal limitations of the original Transformer—its fixed context window and inability to capture long-range text dependencies—by introducing a segment-level recurrent memory mechanism.
- Industry significance: one of the technical sources behind today’s long-context models. Many long-text optimization approaches in Claude, Kimi, and the GPT series borrow heavily from this idea. Without this work, models with one million-token long context would be next to impossible to even discuss. It also serves as a theoretical foundation for Kimi’s later focus on ultra-long documents.
2) XLNet (2019, first author, co-released with Google Brain)
- Core breakthrough: tackled information bias in BERT-style masked pretraining by adopting permutation language modeling; it surpassed BERT across 20 mainstream NLP benchmarks.
- Historical standing: widely recognized in the industry as one of the most important advances in pretraining models after BERT. Selected for a NeurIPS Oral presentation; the paper has been cited tens of thousands of times, making it one of the most influential foundational NLP works by a Chinese scholar.
3) Other academic contributions
- Collaborated with Turing Award winner Bengio to build the HotpotQA multi-turn reasoning question-answering dataset;
- Completed a PhD in just 4 years (regular program is 6 years), studied under Salakhutdinov, the former AI lead at Apple; interned at Google Brain and Meta FAIR, contributing to early R&D for Gemini and Bard;
- Among NLP papers by scholars under 35 in China, total citations have long remained in the top tier.
1. Academic achievements: laying the foundational bedrock for global long-context large language models (the highest-value long-term asset)
1) Transformer-XL (2018, co-first author)
- Core breakthrough: addressed the fatal limitations of the original Transformer—its fixed context window and inability to capture long-range text dependencies—by introducing a segment-level recurrent memory mechanism.
- Industry significance: one of the technical sources behind today’s long-context models. Many long-text optimization approaches in Claude, Kimi, and the GPT series borrow heavily from this idea. Without this work, models with one million-token long context would be next to impossible to even discuss. It also serves as a theoretical foundation for Kimi’s later focus on ultra-long documents.
2) XLNet (2019, first author, co-released with Google Brain)
- Core breakthrough: tackled information bias in BERT-style masked pretraining by adopting permutation language modeling; it surpassed BERT across 20 mainstream NLP benchmarks.
- Historical standing: widely recognized in the industry as one of the most important advances in pretraining models after BERT. Selected for a NeurIPS Oral presentation; the paper has been cited tens of thousands of times, making it one of the most influential foundational NLP works by a Chinese scholar.
3) Other academic contributions
- Collaborated with Turing Award winner Bengio to build the HotpotQA multi-turn reasoning question-answering dataset;
- Completed a PhD in just 4 years (regular program is 6 years), studied under Salakhutdinov, the former AI lead at Apple; interned at Google Brain and Meta FAIR, contributing to early R&D for Gemini and Bard;
- Among NLP papers by scholars under 35 in China, total citations have long remained in the top tier.
半沐
·
--
#kimi k3#
Moon’s Dark Side co-founder, will release open-source Kimi k3. Wow— the world’s number one large language model is going open-source. How does this not kill OpenAI and Claude Code’s closed-source offerings? $BABA
Moon’s Dark Side co-founder, will release open-source Kimi k3. Wow— the world’s number one large language model is going open-source. How does this not kill OpenAI and Claude Code’s closed-source offerings? $BABA
Log in to explore more content