Back to News
SourceMarsBit

从GPT-2到Kimi K3:大模型七年架构演进,本质是一场记忆争夺战

据 动察 Beating 监测,Baseten 工程师 Ali 发布长文,从 GPT-2 一路拆解到 Kimi K3。按参数量计算,一个 K3 相当于约 22580 个 GPT-2。但他认为,过去七年的真正变化并非单纯扩大模型,而是不断改造模型保存、更新和找回信息的方式。 GPT-2 使用全注意力。每个 Token 都留下自己的 Key 和 Value,后续可以随时回看。信息保存最完整,但上下文越长,KV Cache 和计算成本越高。 线性注意力把整段历史压进固定大小的状态。成本不再随上下文持续膨胀,代价是不同信息会挤在同一块空间里。内容越来越多后,旧记忆开始互相干扰。 此后的多项架构创新,都在给这块有限记忆增加管理规则。DeltaNet 允许模型先读取旧内容,再用新旧差异进行覆盖。Gated DeltaNet 加入整体遗忘,让旧信息能够随场景变化衰减。 KDA 进一步把遗忘控制细化到每个通道。不同信息可以拥有不同保存时间:长期事实保留更久,临时指令和局部变量更快衰减。Ali 认为,这才是 Kimi Linear 最关键的进步。 但固定状态不可能保留所有细节。K3 因此没有只用 KDA,而是把 23 组四层结构排列成「三层 KDA 加一层 MLA」,最后再补一层 MLA。KDA 负责低成本维护长期记忆,MLA 定期回看完整上下文,找回数字、代码和早期指令等精确信息。 K3 还引入 Attention Residuals。普通模型会把每一层结果不断相加,早期提取出的信息容易在 93 层计算中被后续结果稀释。K3 把网络按 12 层分块,让后面的层按需调用较早阶段形成的中间表示,不必每次都从原文重新推导。 因此,K3 的关键并不是某一项突然出现的黑科技。它真正特别的地方,是把循环记忆、全局检索、层间回看和专家路由组合成一套分层信息系统。 过去七年的模型架构演进,本质上是在教模型如何选择什么值得留下,什么应该覆盖,什么时候必须翻回原文。
Disclaimer: The views above are the author's only and do not represent 711BTC. Nothing here constitutes investment advice.

Related

07-08 09:00Important

Robinhood's Meme token CASHCAT has surpassed a market capitalization of $45 million, surging over 370% in 24 hours.

According to BlockBeats, on July 8th, GMGN monitoring showed that the market capitalization of Robinhood's on-chain meme token CASHCAT surpassed $45 million, a 24-hour increase of over 370%. The early mascot of Robinhood's US stock app was CASHCAT, later renamed Robinhood. On July 1st of this year, Robinhood launched its own L2 public chain, Robinhood Chain, focusing on on-chain finance and RWA. BlockBeats reminds users that meme tokens often have little intrinsic value and their prices are highly volatile; trading should be approached with caution.

07-05 09:39

Opinion: The next stage of tokenization will be "personalized portfolios," rather than simply improving settlement efficiency.

According to BlockBeats, on July 5th, Thomas Sy, Head of Multi-Asset Solutions at NYLIM (New York Life Investment Management), stated that the next core application of tokenization will be enabling "personalized portfolio construction," rather than simply improving settlement efficiency or extending trading hours. NYLIM manages approximately $807 billion, with about $11 billion handled by Sy's team. He pointed out that blockchain technology promises to allow asset management institutions to customize complex portfolio strategies for different investors at a large scale, a capability currently difficult to achieve in the traditional financial system. Sy stated that the core of future asset management will shift towards "high customization," and blockchain is the only technological path capable of achieving this at scale. He believes that tokenization is not just about putting ETFs, bonds, or private credit on the blockchain, but more importantly, about reshaping the way portfolios are constructed. He also pointed out that current portfolios typically involve a mix of ETFs, bonds, and private equity assets, but due to operational complexity, personalized strategies are difficult to scale. Tokenization has the potential to "embed customization logic into the assets themselves," reducing operational costs and improving efficiency. Furthermore, Sy stated that stablecoins have become a key entry point for traditional finance to enter the blockchain space, with the current market capitalization exceeding $300 billion, and are being used for cross-border payments and fund management. He believes this trend will gradually drive institutional demand for on-chain yield-generating assets. Regarding decentralized finance (DeFi), NYLIM is still researching related applications, but Sy emphasized that institutional participation still requires more mature infrastructure, including improved tokenized collateral, liquidation mechanisms, and prime brokerage services.

07-06 18:22

GMX: Since March, a total of 313,650 GMX tokens have been repurchased at an average price of $6.27, for a total value of $1.965 million.

According to Odaily Odaily, GMX announced on the X platform that the GMX DAO repurchased 23,280 GMX tokens between June 24th and 30th at an average price of approximately $5.37, for a total value of approximately $125,000. Since its launch on March 5th, the repurchase program has cumulatively repurchased 313,650 GMX tokens, with a total value of approximately $1,965,000 and an average price of approximately $6.27. In the second quarter of 2026, GMX DAO repurchased a total of 228,030 GMX tokens, with a total value of approximately US$1.41 million and an average price of approximately US$6.18.

07-05 21:27

Myanmar's telecom fraud AI industrialization exposed: Starlink becomes key infrastructure, encrypted payments and OpenAI/Google models are incorporated into the toolchain.

According to a report by Red Star News, citing Mars Finance, an investigative report leaked from a scam industrial park in Myanmar reveals that global telecom fraud is rapidly evolving towards an "AI industrialization + cross-border encrypted payment" system. These fraud networks use cryptocurrencies to transfer funds and employ automated tools based on large models for multilingual script generation, identity spoofing, and emotional manipulation. The investigation shows that these systems heavily utilize OpenAI's ChatGPT and Google's Gemini to support "large-scale social fraud," while funds are rapidly laundered and transferred through on-chain payments and cross-border channels, forming a two-tiered structure of "AI customer acquisition + encrypted settlement," enabling the fraud industry to achieve a high degree of automation and transnational expansion. Furthermore, Elon Musk's Starlink has become the leading network service provider in the Myanmar scam industrial park, with US ISPs handling nearly one-fifth of the park's traffic. In response to the allegations, OpenAI stated that the behavior of fraudsters using ChatGPT is highly similar to that of ordinary users, making identification difficult, but they have already blocked approximately 100,000 suspicious accounts monthly through behavioral pattern recognition and risk control systems. Google stated that its AI models have safety barriers in place and emphasized its commitment to "responsible AI development" to limit the tools from being used for illegal purposes such as fraud.

07-05 08:17

New York Life Asset Management executive: The biggest opportunity for tokenization lies in personalized investing.

PANews reported on July 5th, citing Coindesk, that Thomas Sy, Head of Multi-Asset Solutions at New York Life Investment Management (NYLIM), which manages $800 million in assets, stated in an interview that blockchain could ultimately enable asset management firms to tailor portfolios for individual investors on a large scale—something the current financial system cannot achieve. "We believe the future of asset management lies in personalization. And the only technology that can help us achieve this at scale is blockchain." Personalized investment strategies often combine ETFs, bonds, private credit, and other assets, creating operational complexity that makes personalization difficult to scale. Tokenization can simplify transfer brokerage, settlement, and other back-office processes, reducing costs and ultimately benefiting investors.

07-08 18:38Important

ZachXBT: LAB has plummeted 85% in 24 hours; trading this token is not recommended.

According to ChainCatcher, on-chain investigator ZachXBT stated that the LAB token has plummeted from $14 to less than $2 in the past 24 hours, a drop of approximately 85%, resulting in a significant reduction in its fully diluted valuation (FDV) from approximately $14 billion. ZachXBT expressed disappointment that Binance, Bitget, and Gate failed to take proactive measures to prevent price manipulation, stating that if centralized exchanges truly valued user interests, they should at least return the profits of accounts suspected of market manipulation to users. He further pointed out that LAB investors were originally scheduled to begin unlocking their tokens later this month, but the vesting arrangements have been repeatedly altered. Insiders have long controlled almost the entire circulating supply and have been using market makers to conduct extreme price manipulation on centralized exchanges; therefore, he "does not recommend trading LAB under any circumstances."