笑看风云

DeepSeek又发新模型，小而美玩出新高度_我的网站

A |     British Prime Minister Boris Johnson is fighting for his political life but his opponents have few options for replacing him — so say two senior academics.,Calls for Johnson to go have reached fever pitch since Chris Pincher, the Conservative MP he appointed as deputy chief whip in February, was sacked for drunkenly groping two men at a posh private members' club in London.,This week it emerged that Johnson had been informed of earlier sexual assault claims against Pincher before he promoted him — leaving cabinet colleagues were embarrassed at having to defend the ever-changing official line.,Laura McAllister, professor of public policy at the Wales Governance Centre, told Sputnik nobody should take Johnson's political demise for granted.,"He survived so much in a way that few other politicians have ever managed to, particularly those in senior public office and prime minister," McAllister says. "He operates by a completely different set of rules and principles and ethics, which we would assume would be his undoing, but actually in reality he seems to have been able to ride the storm".,The academic points to Johnson's survival of the scandals around his advisor Dominic Cummings, 'Partygate', his handling of the COVID-19 pandemic before the Pincher affair. But she conceded that the spate of ministerial resignations led by Sajid Javid as health secretary and Rishi Sunak as chancellor of the exchequer were "hugely significant.","Two Cabinet members who sat in Cabinet yesterday morning had resigned without warning by the evening," the professor stresses "I think it's also fair to say that both of those individuals are thinking about their own potential bids to stand as the next leader of the Conservative Party and therefore timing is very much predicated on that.",McAllister says things have changed since the days of Margaret Thatcher, who resigned in 1990 after failing to win a leadership challenge by an overwhelming margin.,"[T]here was a much stronger Cabinet code, a ministerial code that was upheld broadly, or if it wasn't upheld, then there were consequences," she says "That's not the case now, and it seems very likely that Johnson will not resign of his own accord.","He still has a base of support, of course, even though that’s dwindling away now," she adds.,The academic also doubts that a general election would break the political deadlock.,"Whilst Labour are a few percentage points ahead, they clearly aren’t making sufficient headway to be confident of winning a majority," McAllister says. "The best they could hope for in the current polling would be a minority government and potentially having to seek coalition partners".,Those could include the Liberal Democrats and the Scottish National Party (SNP), who would demand a repeat of the 2014 independence referendum as the price of their support. Ironically, independence for Scotland and its 59 parliamentary seats would shore up the Conservative majority in Westminster and make Labour nigh on unelectable.,"[Labour leader] Keir Starmer has already ruled out any deals with the SNP for obvious constitutional reasons, and therefore with Plaid Cymru in Wales," she points out, "So it seems unlikely that that route to power is open".,McAllister believes a new Tory leader would at least enjoy a "honeymoon period" bounce in the polls.,"And who knows, were Johnson to be defenestrated, it might be that a new leader might choose to capitalize on the honeymoon period and call an earlier election," she speculates.,Beleaguered BoJo Battles With Opposition at Weekly PM's Questions12:52 GMT,No 'Coup', But Then Who?,Mark Garnett, a politics professor at Lancaster University and author of the books "The British Prime Minister in an Age of Upheaval" and "From Anger to Apathy: The British Experience" thinks Johnson has no hope of survival. But he does not see how the party can push the PM out if he refuses to jump.,He says Javid and Sunak's resignations are "serious blows" to Johnson, but they are only "symptoms of a more general crisis of leadership".,"Johnson has been involved in too many incidents which raise questions about his personal integrity," Garnett says, with ministers told to defend the PM against allegations that turned out to be true. "In effect, they have been asked to lie in public on his behalf. Their patience has finally snapped.",But the professor does not see the near simultaneous resignations as a "coup", stressing: "The most surprising thing is that ministers have tolerated Johnson for so long".,While Sunak was seen as the favorite to succeed Johnson just six months ago "he is being held at least partly responsible for the country's economic problems and his chances of becoming prime minister are now very remote," Garnett says. "Sajid Javid's chances are greater, but Conservative members (who make the final decision) seem more attracted to the Defence Secretary, Ben Wallace, who has not resigned.",The academic believes MPs are worried that Johnson is now an "electoral liability" who could cost them their seats at the next general election. "The party has just lost two by-elections in crucial seats, and the main reason given by those who voted against them was Boris Johnson's leadership," he stresses.,But the problem of how to remove Johnson from office if he will not quit remains.,"Currently the party's rules make it impossible to hold a vote of confidence for another year," Garnett points out. "It is very likely that those rules will be changed within the next ten days". But he believes if a vote could be held now, "Johnson would certainly lose and be forced to resign".。

B |     就在刚刚，DeepSeek开源了一个3B模型DeepSeek-OCR。虽然体量不大，但模型思路创新的力度着实不小。

C |                    众所周知，当前所有LLM处理长文本时都面临一个绕不开的困境：计算复杂度是平方级增长的。序列越长，算力烧得越狠。                             于是，DeepSeek团队想到了一个好办法。既然一张图能包含大量文字信息，而且用的Token还少，那不如直接把文本转成图像？这就是所谓的“光学压缩”——用视觉模态来给文本信息“瘦身”。

D |                    而OCR正好天然适合验证这个思路，因为它本身就是在做“视觉→文本”的转换，而且效果还能量化评估。                             论文显示，DeepSeek-OCR的压缩率能达到10倍，OCR准确率还能保持在97%以上。                   啥意思呢？就是说，原本需要1000个文本Token才能表达的内容，现在只用100个视觉Token就搞定了。即使压缩率拉到20倍，准确率也还有60%左右，整体效果相当能打。                   OmniDocBench基准测试结果显示：                    只用100个视觉Token，就超过了GOT-OCR2.0（每页256个Token）的表现；          用不到800个视觉Token，干翻了MinerU2.0（平均每页超过6000个Token）。                        在实际生产中，一块A100-40G显卡就能每天生成超过20万页的LLM/VLM训练数据。20个节点（160块A100）直接飙到每天3300万页。

E |                              DeepSeek-OCR由两个核心组件组成：                    DeepEncoder（编码器）：负责图像特征提取和压缩；          DeepSeek3B-MoE（解码器）：负责从压缩后的视觉Token中重建文本。                        让我们来重点说说DeepEncoder这个引擎。

F |                    它的架构很巧妙，通过把SAM-base（8000万参数）和CLIP-large（3亿参数）串联起来，前者负责“窗口注意力”提取视觉特征，后者负责“全局注意力”理解整体信息。                   中间还加了个16×卷积压缩器，在进入全局注意力层之前把Token数量大幅砍掉。                   举例而言，一张1024×1024的图像，会被切成4096个patch token。

G | 但经过压缩器处理后，进入全局注意力层的Token数量会大幅减少。                   这样的好处是，既保证了处理高分辨率输入的能力，又控制住了激活内存的开销。                   而且DeepEncoder还支持多分辨率输入，从512×512的Tiny模式（64个Token）到1280×1280的Large模式（400个Token），一个模型全搞定。                   目前开源版本支持的模式包括原生分辨率的Tiny、Small、Base、Large四档，还有动态分辨率的Gundam模式，灵活性拉满。                             解码器用的是DeepSeek-3B-MoE架构。                   别看只有3B参数，但采用了MoE（混合专家）设计——64个专家中激活6个，再加2个共享专家，实际激活参数约5.7亿。这也让模型既有30亿参数模型的表达能力，又保持了5亿参数模型的推理效率。

H |                    解码器的任务就是从压缩后的视觉Token中重建出原始文本，这个过程可以通过OCR风格的训练被紧凑型语言模型有效学习。                   数据方面，DeepSeek团队也是下了血本。                   从互联网收集了3000万页多语言PDF数据，涵盖约100种语言，其中中英文占2500万页。                   数据分两类：粗标注直接用fitz从PDF提取，主要训练少数语言的识别能力；精标注用PP-DocLayout、MinerU、GOT-OCR2.0等模型生成，包含检测与识别交织的高质量数据。                   对于少数语言，团队还搞了个“模型飞轮”机制——先用有跨语言泛化能力的版面分析模型做检测，再用fitz生成的数据训练GOT-OCR2.0，然后用训练好的模型反过来标注更多数据，循环往复最终生成了60万条样本。                   此外还有300万条Word文档数据，主要提升公式识别和HTML表格解析能力。                   场景OCR方面，从LAION和Wukong数据集收集图像，用PaddleOCR标注，中英文各1000万条样本。

I |                              DeepSeek-OCR不仅能识别文字，还具备“深度解析”能力，只需一个统一的提示词，就能对各种复杂图像进行结构化提取：                    图表：金融研究报告中的图表可以直接提取为结构化数据；          化学结构式：识别并转换为SMILES格式；          几何图形：对平面几何图形进行复制和结构化解析；          自然图像：生成密集描述（dense captions）。                        这在STEM领域的应用潜力巨大，尤其是化学、物理、数学等需要处理大量符号和图形的场景。                   第一作者Haoran Wei此前曾供职于阶跃星辰，期间发布并开源了GOT-OCR2.0系统                    值得注意的是，DeepSeek团队在论文里还提出了一个脑洞大开的想法——用光学压缩模拟人类的遗忘机制。                   人类的记忆会随时间衰退，越久远的事情记得越模糊。DeepSeek团队想，那能不能让AI也这样？于是，他们的方案是：                    1. 把超过第k轮的历史对话内容渲染成图像；          2. 初步压缩，实现约10倍的Token减少；          3. 对于更久远的上下文，继续缩小图像尺寸；          4. 随着图像越来越小，内容也越来越模糊，最终达到“文本遗忘”的效果。                   这就很像人类记忆的衰退曲线，近期信息保持高保真度，久远记忆自然淡化。                   虽然这还是个早期研究方向，但如果真能实现，对于处理超长上下文将是个巨大突破——近期上下文保持高分辨率，历史上下文占用更少计算资源，理论上可以支撑“无限上下文”。

J |                              简言之，DeepSeek-OCR表面上是个OCR模型，但实际上是在探索一个更宏大的命题：能否用视觉模态作为LLM文本信息处理的高效压缩媒介？                    初步答案是肯定的，7—20倍的Token压缩能力已经展现出来了。                   当然，团队也承认这只是个开始。单纯的OCR还不足以完全验证“上下文光学压缩”，后续还计划开展数字–光学文本交替预训练、“大海捞针”式测试，以及其他系统性评估。                   不过不管怎么说，这在VLM和LLM的进化路上，又多了一条新赛道。

K |                    去年这个时候，大家还在卷怎么让模型“记得更多”。

L | 今年DeepSeek直接反其道行之，不如让模型学会“忘掉一些”。                   确然，AI的进化，有时候不是做加法，而是做减法。小而美，也能玩出大花样，DeepSeek-OCR这个3B小模型就是最好的证明。                   GitHub主页：http://github.com/deepseek-ai/DeepSeek-OCR          论文地址：https://github.com/deepseek-ai/DeepSeek-OCR/blob/main/DeepSeek\_OCR\_paper.pdf          模型下载：https://huggingface.co/deepseek-ai/DeepSeek-OCR                    本文来自微信公众号：APPSO （ID：appsolution）。

Current article:http://eez2ngj.senbatuqiuzhuanlanrilao.sbs/alupdk8/7wczlgc.html

Published on:12:36:00