心术

Gemini 3.5 Pro难产，谷歌先交了两张草稿纸_我的网站

一 |     7月21日，谷歌突然上线了两个新模型：          Gemini 3.6 Flash和Gemini 3.5 Flash-Lite。         没有一点点防备，也没有一丝顾虑，它俩就这样出现在了Google AI Studio和Gemini的模型选择器里。         这两个模型，一个主打更强的代码和Agent能力，一个主打更低成本的大规模调用。         而作为旗舰模型的Gemini 3.5Pro，千呼万唤依然……出不来。         按照此前的计划，3.5 Pro本应在6月份上线，现在7月都过一大半了，这款被寄予厚望的模型仍然没有正式亮相。    Scott Ritter, a military analyst and former US Marine Corps intelligence officer, told Sputnik on Thursday that NATO’s policy of giving Ukraine billions of dollars in weapons that so quickly get destroyed is “inefficient and counterproductive” from both a political and a national security point of view.,“The more Europe invests its military capacity into Ukraine, the weaker Europe gets,” Ritter said.,Proxy Wars Are Expensive,“I think there's a couple of things at play here. One, and I think first and foremost, is that Europe has bankrupted its own military capacity by transferring military equipment to Ukraine in an unconstrained manner. It's becoming clear to all that Ukraine is facing a very difficult time on the battlefield. Weaponry is being provided by Europe, by the United States, by NATO, and this weaponry is not being absorbed in a coherent manner by the Ukrainian Armed Forces. Instead, it's being received and rushed to the front where it is not being used effectively and increasingly it's being destroyed by the Russians. This is very expensive, very inefficient. It's detrimental not only to Ukraine's military capability but also the nations providing the equipment. They're literally stripping bare their own inventories to support Ukraine in a losing cause,” Ritter explained.,“The second part of this equation is, not only is this inefficient and counterproductive from a national security standpoint, but it's expensive and counterproductive from a political standpoint,” he said. “I think the political cost of supporting Ukraine is about to become very, very high.”,“Many political leaders in Europe now are going to be facing an increasingly hostile constituency, demanding answers as to why their leaders committed: a) economic suicide by joining in on a failed sanctions effort; b) are further bankrupting the nation by stripping bare their own arsenals, requiring them to expend money to restock. This is an expensive proposition. And c), supporting what is increasingly being documented as a very corrupt and, frankly speaking, vile regime in Kiev.”,Kiev’s Mask is Slipping,Ritter noted how Ukrainian President Volodymyr Zelensky “is no longer seen as the savior” by Westerners. “More and more, he's being seen as what he has always been, an incompetent politician operating way above his capabilities.”,In particular, he recalled how a damning CBS documentary about the conflict was partially retracted earlier this month after it revealed that some 70% of equipment given to Ukraine never reaches the battlefield, either because the Russians destroy them or because they’re sold off on the black market. Forthcoming trials in Mariupol can also be expected to further peel back the Western myths about the Azov Regiment and other neo-Nazi units being anything but fascists.,© AP Photo / Alexander ZemlianichenkoRussian soldiers walk inside the Ukraine's Azov Regiment base adorned with the unit's emblems in Yuriivka resort settlement on the coast of Azov Sea not far from Mariupol, in territory under the government of the Donetsk People's Republic, eastern Ukraine, Wednesday, May 18, 2022Russian soldiers walk inside the Ukraine's Azov Regiment base adorned with the unit's emblems in Yuriivka resort settlement on the coast of Azov Sea not far from Mariupol, in territory under the government of the Donetsk People's Republic, eastern Ukraine, Wednesday, May 18, 2022© AP Photo / Alexander Zemlianichenko,“It's a horribly corrupt regime. It's also a regime that is being linked to an odious ideology, the old ideology of Stepan Bandera - a neo-Nazi, right wing, ultra-nationalistic ideology responsible for the murders of hundreds of thousands of people in the years of World War II and the decade afterwards, people who terrorized, murdered, raped ethnic Russians in the aftermath of the Maidan coup that took place in 2014,” Ritter noted.,“I think you're seeing Europeans starting to recoil from this and saying ‘This is not what we want to be seen as supporting.’ All of this, when you bring it together, explains why people aren't willing to continue to dispatch military equipment to Ukraine.”,Ritter dismissed the idea that Berlin and Paris could continue to send weapons to Kiev at the same pace that they have been. In this area, too, they will soon be eclipsed by Washington.,“Military equipment is not produced with the snap of a finger. And it requires a budgetary process. It requires careful coordination with the industrial base, and equipment is procured over the course of several years. What Europe has done by providing this equipment to Ukraine is in one fell swoop, in a matter of months, they've taken decades of military procurement and transferred it to Ukraine with no plan on how to replace it. You can't just simply pick up the phone and say, ‘give me a hundred more tanks,’ ‘give me a hundred more artillery pieces.’ It doesn't work that way. They have literally destroyed the very fabric of their national security by stripping bare their coverage with no means of replacing this. They've already bankrupted themselves. There's really not that much left they can give to Ukraine,” he explained.,‘This Can Only Be Resolved Politically’,Ritter said EU foreign policy chief Josep Borrell is “out of his competency level” in commenting on military affairs, especially in making claims like “This war will be won on the battlefield.”,“You weren't going to see a general talking about that in that way,” he noted. “Fact is, most military people said that this can only be resolved politically, this being the conflict in Ukraine. Borrell is way off base. He's out of his competency level and he's just playing wrong. Who's going to solve this on the battlefield? What military in Europe today is capable of standing toe to toe with the Russian military in Ukraine and beating them? And the answer is none. There's nobody that can do this.”,Ritter said Germany and France, the two powerhouse NATO nations in continental Europe, are “paying a horrific economic price” and political price for their efforts to bring the US back firmly into the NATO fold, following four years of vacillation by former US President Donald Trump.,© AP Photo / BelTA, Andrei StasevichNormandy format talksNormandy format talks© AP Photo / BelTA, Andrei Stasevich,“[N]ow the United States has committed to NATO and has committed NATO to this horrendous mistake: that is the expansion into Ukraine, the confrontation with Russia, and this very expensive proxy war against Russia in Ukraine, that NATO is losing and losing decisively,” Ritter said. “But there's a lot of politics at play here. And it's not as a matter of just simply snapping your fingers and saying, ‘Okay, we've made a mistake, we're done, we're going to now reexamine a new European security framework.’”,Ritter recalled Berlin and Paris were also key guarantors of the Minsk Accords, which were supposed to settle the Donbass conflict in the context of autonomy for the Russian-speaking regions.,“It's not only that they failed to pressure Ukraine. It appears, according to [Petro] Poroshenko, the former president of Ukraine, that they were complicit in a scheme that used the Minsk Accords as a sham method to extend the negotiations, knowing they would never be finalized, to buy time so that NATO could train Ukrainian Armed Forces so that they would be in a position to be able to retake all of the Donbass and perhaps Crimea,” Ritter noted.,“Germany and France were part and parcel of this effort, so they need to be careful. They're very exposed here in terms of the politics of that. But reality is reality. And the reality is that the current European security framework, dominated as it is by American national security interests, is unsustainable for Europe,” he said. “They are paying a horrific economic price and they will probably pay a political price. Some of these leaders that are in power today won't be in power in the coming months as their citizens rise up through the democratic process and remove them for incompetence and for pursuing bad policies.”,“The last problem is Russia sort of owns the notion of European security framework having defined it, and their needs and their requirements in December of last year, through two draft treaties shared with NATO and the United States. For Germany and France to pursue a new European security framework would mean that they are stepping away from the United States and stepping toward Russia. That's very politically difficult to do, no matter how justified it might be.”,© Sputnik / Sergei AverinEight years of conflict have left a mark: it will take decades to demine the territory of the republics. Photo: A warning sign at a position near the line of contact with Ukrainian forces near Spartak village in the Donetsk region.     Eight years of conflict have left a mark: it will take decades to demine the territory of the republics. Photo: A warning sign at a position near the line of contact with Ukrainian forces near Spartak village in the Donetsk region.     © Sputnik / Sergei Averin,‘Somebody Else is Going to Have to Seek Peace for Them’,Ritter said that, like World War II after the Battle of Kursk, everyone knows that Ukraine has lost. The question is whether or not the government in Kiev is capable of making peace; he predicts not.,“It's going to be a victory in accordance with all of the prerequisites accomplished - demilitarization, de-Nazification and NATO removed from Ukraine. Russia has also now appeared to put some territorial requirements,” Ritter said.,“The key question is, how does this end at some point?” he asked. “There's going to be a need for a political collapse in Kiev, where the current cadre of warmongers and nationalists have to be removed. They will not seek peace. Somebody else is going to have to seek peace for them.”,“I think it's becoming clear that the current regime in Ukraine cannot be entrusted to govern in a responsible fashion, and I think Russia's taken it as part of their policy that they will protect ethnic Russians going forward, which means that they will probably acquire additional territory in Ukraine where ethnic Russians live. This takes time. It's not going to happen overnight. As much as Russia is winning the war, there's still a very strong Ukrainian military on the field that's receiving military assistance from the United States and NATO. So there's a lot of work left to be done.”。

二 |          代表旗舰实力的答卷迟迟不交，谷歌先拿出两张“草稿纸”，是个什么意思？          01          3.6 Flash：比上一代便宜，不比上一代聪明          谷歌表示，相比上一代Flash模型，Gemini 3.6 Flash进一步提升了编码、推理和工具调用表现，同时通过减少输出token降低运行成本。         Gemini 3.6 Flash的定位是一个面向真实生产环境的高效模型，或者更直白一点，为Agent服务。         在Agent时代，一次任务可能需要几十轮模型调用，这意味着模型不仅要聪明，还必须足够快、足够便宜。

三 |          Google此前已经在Gemini 3.5 Flash上强化了这条路线。根据官方API文档，Gemini 3.5 Flash支持100万token的长上下文，可以一次处理大量信息，同时支持代码执行、调用外部工具、搜索文件等能力，能够完成更复杂、更长时间的任务。         Gemini 3.6 Flash延续了这条路线，它追求的并不是单次回答的质量，是一个“能够承担得起每日真实工作”的位置。         Gemini 3.6 Flash的定价如下：          输入：1.50 美元/百万 token          输出：7.50 美元/百万 token          相比此前Gemini 3.5 Flash每百万token输入1.5美元、输出9美元的价格，输入成本没有变化，但输出成本进一步下降。         对于需要大量调用AI的企业来说，这种成本变化可能比benchmark上的几个百分点提升更加重要。         另外，谷歌展示的一项内部测试显示，在完成同一任务时，Gemini 3.6 Flash相比3.5 Flash减少了55.8%的token消耗，同时将任务评分从85分提升到了100分。         当然，这只是谷歌自己的测试结果，只能说明Google希望展示的方向：          新模型不仅更省，而且能够用更少的计算完成同样甚至更好的任务。         在那些被公开的benchmark里，谷歌主要强调的是代码能力和Agent能力。         在DeepSWE代码评测中，Gemini 3.6 Flash得分49%，高于上一代的37%。谷歌表示，新模型能够减少无效代码修改和重复执行过程。         在测试AI操作电脑能力的OSWorld-Verified评测中，Gemini 3.6 Flash得分83%，超过3.5 Flash的78.4%。         而在面向机器学习任务的MLE Bench中，Gemini 3.6 Flash得分63.9%，相比上一代的49.7%也有明显提升。         不过，上述数据更多说明Gemini 3.6 Flash相比上一代有所进步。         如果放到当前大模型竞争中横向比较，Google选择的对手是GPT-5.6 Luna、Grok 4.5和Claude Sonnet 5（这里也可以看出这个产品的定位）。         从结果来看，Gemini 3.6 Flash并没有全面领先，GPT和Claude的模型在复杂软件工程和知识工作任务上仍然保持优势。         但这其实也是Flash系列存在的意义：          它不一定要成为最强模型，只需要在明显低于旗舰模型成本的情况下，提供够用的能力。

四 |          官方测评之外，根据Artificial Analysis的发布前测试，Gemini 3.6 Flash在Intelligence Index上拿到了50分，与Gemini 3.5 Flash持平。         这意味着，如果只看模型综合“智力”，Gemini 3.6 Flash并没有什么升级。         但是呢，它的速度又很好地弥补了这一点。         Artificial Analysis还统计了模型完成Intelligence Index任务所需的平均时间。结果显示，Gemini 3.6 Flash每项任务平均耗时约1.3分钟，相比上一代3.5 Flash的约2.7分钟，减少了一半。         与此同时，Gemini 3.5 Flash-Lite的任务完成时间也从上一代3.1 Flash-Lite的约1.6分钟，下降到约0.6分钟。         总的来说，Gemini 3.6 Flash的升级方向其实是效率——更少的token消耗，更低的运行成本，以及更适合长期运行的Agent能力。         对于习惯使用Gemini 3.5 Flash的用户来说，确实是非常不错的升级。         02          3.5 Flash-Lite：更快，但没有上一代便宜          然后再来看另一个模型Gemini 3.5 Flash-Lite。         顾名思义，Flash-Lite是Flash系列中更轻量的版本，相比Flash，它牺牲了一部分能力上限，换取了更低的成本和更快的速度。         就像前面提到的那样，Gemini 3.5 Flash-Lite的任务完成时间从上一代3.1 Flash-Lite的约1.6分钟，下降到了约0.6分钟。

五 |          3.5 Flash-Lite也是3.5系列中速度最快的型号，根据Artificial Analysis测试数据，其生成速度达到约350 token/秒，已经属于当前主流大模型中的高速水平。

六 |          Gemini 3.5 Flash-Lite的定价如下：          输入：0.30 美元/百万 token          输出：2.50 美元/百万 token          相比Gemini 3.6 Flash，输入价格约为1/5，输出价格约为1/3。不过，与上一代Gemini 3.1 Flash-Lite相比，新模型并没有进一步降价。

七 |          虽然Gemini 3.5 Flash-Lite由于能力提升，可以通过减少输出量降低部分成本，但综合起来，还是很难抵消单价上的成本增加。         根据官方文档，相比上一代Gemini 3.1 Flash-Lite，新模型在不同思考等级（thinking levels）下都有明显提升。开发者可以根据任务需求调整模型的思考深度。         此外，Gemini 3.5 Flash-Lite还内置了Computer Use能力，可以帮助Agent更可靠地操作电脑环境，完成跨应用任务。         在具体测试中，Gemini 3.5 Flash-Lite相比上一代模型也有明显提升：在Terminal-Bench 2.1编程Agent测试中，成绩从31%提升到54%；在测试长上下文理解能力的GDM-MRCR v2中，从60.1%提升到72.2%；在更贴近真实工作场景的GDPval-AA v2任务执行测试中，从642提升到1140。         值得注意的是，在一些Agent和代码相关测试中，Gemini 3.5 Flash-Lite甚至超过了上一代更高定位的Gemini 3 Flash模型。

八 |          例如，在SWE-Bench Pro软件工程测试中，Gemini 3.5 Flash-Lite得分54.2%，高于Gemini 3 Flash的49.6%；在测试AI操作电脑能力的OSWorld-Verified中，Gemini 3.5 Flash-Lite也以74.0%的成绩超过Gemini 3 Flash的65.1%。         这意味着，随着模型能力不断提升，过去需要更大模型才能完成的部分Agent任务，正在逐渐下沉到更便宜、更快速的模型。         顺便一提，除了前两个面向普通用户的通用模型，谷歌还推出了Gemini 3.5 Flash Cyber。

九 |          它主要针对网络安全场景。

十 | 根据官方文档，在CodeMender系统中，多个Gemini 3.5 Flash Cyber Agent可以协同工作，分别完成不同分析任务，并最终汇总成一份完整报告。在网络安全基准测试CyberGym中，Flash Cyber也达到了接近前沿模型的表现。         该模型目前不会公开提供，仅通过CodeMender平台向政府和可信合作伙伴开放。         03          3.5 Pro：谷歌迟迟交不出的旗舰答卷          如果说Gemini 3.6 Flash和Gemini 3.5 Flash-Lite还可以看作是谷歌对AI规模化应用的判断，那么Gemini 3.5 Pro则代表着谷歌的模型上限。         但问题在于：这份答卷，到现在还没有交出来。

十一 |          5月I/O的时候，谷歌表示Gemini 3.5 Pro将在“接下来一个月左右”推出，也就是预计在今年6月上线。现在7月都过了一大半了。         据彭博社7月16日报道，Gemini 3.5 Pro的发布时间已经落后原计划数月。谷歌正在继续优化模型能力，尤其是编码表现，希望达到内部设定的目标。         路透社最新的报道则显示，截至目前，谷歌仍未公布具体上线时间，只表示该模型正在与合作伙伴测试，并将“很快”推出；另有新闻稿透露，Gemini 4的训练工作已经开始了。         而我们都知道，“很快”就是不确定的意思。         Gemini 1.0发布时就曾因演示争议受到质疑；Gemini 1.5 Pro凭借百万token上下文短暂扳回局面，但随后Claude和GPT系列快速追赶；直到Gemini 3系列发布，谷歌才重新获得“回到前沿竞争”的评价。         如今Gemini 3.5 Pro再次延期，市场关注的已经不只是一个模型什么时候上线，还有谷歌到底能否保持旗舰模型的竞争节奏问题。

十二 |          何况谷歌透露出的编码表现不及预期，并不是什么容易解决的小问题。

十三 |          随着Agent开始进入企业工作流，代码能力的重要性迅速提升。一个模型能否理解复杂项目、修改真实代码、完成多步骤开发任务，已经成为衡量它是否具备实际生产价值的重要指标。         路透社指出，华尔街正在等待Gemini 3.5 Pro，因为它将成为判断Google DeepMind是否能够跟上OpenAI和Anthropic的重要指标。

十四 |          当然，谷歌并不是没有动作，这次推出的几个模型，恰说明谷歌正在强化另一条路线：让AI变得更便宜、更容易规模化。         谷歌CEO Sundar Pichai近期也多次强调成本优势，并表示企业如果将大部分AI工作负载迁移到Gemini模型，可以每年节省超过10亿美元。         但问题在于，市场在期待一份证明谷歌模型实力的“答卷”，谷歌交出的却是两张关于效率和成本的“草稿纸”——很难不让人觉得，谷歌在用Flash系列填补Pro延迟留下的空档。         有意思的是，在Gemini 3.5 Pro延期的同时，AI竞争格局正在发生变化。         过去，人们讨论AI领先者，主要关注海外“御三家”：OpenAI、Anthropic和Google DeepMind。

十五 |          但最近，GLM-5.2、Kimi K3等国产模型也开始进入前沿竞争，并引发了大量讨论。         前沿模型的追赶速度正在加快，也让谷歌的问题变得更加紧迫，它当然可以继续大力推出Flash模型，但前提是它的旗舰模型足够给力——只要硬实力足够，自有大儒为他辩经。         如果谷歌最终推出一个真正领先的旗舰模型，那么Flash路线会成为完整体系的一部分：Pro负责突破能力上限；Flash负责规模化应用。         但如果Gemini 3.5 Pro持续落后，市场很可能会继续追问：谷歌拥有最强AI研究资源，为什么却无法稳定跑赢竞争对手？          在Alphabet本周举行第二季度财报电话会议时，人们很可能会进一步询问有关Gemini 3.5 Pro的更多细节。（作者：袁心玥）。

Current article:http://z2io.cousizhegaobaolengkuo.bond/news/20260826_26319.html

Published on:07:34:38