本周我阅读的内容…… 1）@Google 于 2025 年 11 月 18 日正式发布 Gemini 3，称其为迄今为止“最智能”的模型，并在发布当天就全面整合进 Google Search（AI 模式）、Gemini App、AI Studio、Vertex AI，以及一个名为 Google Antigravity 的全新智能代理开发环境。 Gemini 3 在多个榜单上取得了领先成绩，包括： * LMArena 排行榜：以突破性分数超越 Grok 4.1 版本 * Humanity’s Last Exam：达到博士级的推理能力 * GPQA Diamond（科学知识基准测试） 并在更多基准测试中占据领先，显示出大模型的规模定律依然有效。 一个关键洞察是：来自多方来源（包括模型卡与媒体报道）的信息确认：Gemini 3 Pro 的训练完全依赖 Google 自家的 TPU（张量处理器），而非 Nvidia GPU。 这让 Google 在当下处理器资源竞争极度激烈的市场环境中获得了强大的优势。 此外，Google 的 Nano Banana 也升级为 Pro 版本，成为一款先进的图像生成与编辑工具。 它引入了： * 4K 输出 * 实时迭代式优化 * 复杂场景的混合式 JSON Prompting * 性能较上一代提高约 4 倍 这项速度与品质的提升明确归功于在 TPU 上优化后的张量处理能力。 What I read this week... 1) @Google announced Gemini 3 on November 18, 2025, calling it its “most intelligent” model to date and rolling it out day-one across Search (AI Mode), the Gemini app, AI Studio, Vertex AI, and a new agentic dev environment called Google Antigravity. It has topped the charts on: - LMArena Leaderboard with a breakthrough score right above Grok’s 4.1 release - Humanity’s Last Exam with PhD-level reasoning - GPQA Diamond (for scientific knowledge) And leading many more benchmarks, showing that scaling laws are still intact. The key insight is that multiple sources, including model cards and press coverage, confirm that Gemini 3 Pro was trained on Google’s own Tensor Processing Units (TPUs), not Nvidia GPUs. This gives Google a strong advantage over other firms that are in extreme competition for processing units. Nano Banana has also been upgraded to a Pro version and is Google’s advanced image generation and editing tool. It introduces 4K output, real-time iterative refinement, and hybrid JSON prompting for complex scenes, at roughly a 4x speedup over prior Nano Banana generations. The additional speed and quality is attributed explicitly to optimized tensor processing on TPUs.
Chamath PalihapitiyaFounder & CEO, Social Capital
观点原文
本周我阅读的内容……
1）@Google 于 2025 年 11 月 18 日正式发布 Gemini 3，称其为迄今为止“最智能”的模型，并在发布当天就全面整合进 Google Search（AI 模式）、Gemini App、AI Studio、Vertex AI，以及一个名为 Google Antigravity 的全新智能代理开发环境。
Gemini 3 在多个榜单上取得了领先成绩，包括：
- LMArena 排行榜：以突破性分数超越 Grok 4.1 版本
- Humanity’s Last Exam：达到博士级的推理能力
- GPQA Diamond（科学知识基准测试）
并在更多基准测试中占据领先，显示出大模型的规模定律依然有效。
一个关键洞察是：来自多方来源（包括模型卡与媒体报道）的信息确认：Gemini 3 Pro 的训练完全依赖 Google 自家的 TPU（张量处理器），而非 Nvidia GPU。
这让 Google 在当下处理器资源竞争极度激烈的市场环境中获得了强大的优势。
此外，Google 的 Nano Banana 也升级为 Pro 版本，成为一款先进的图像生成与编辑工具。
它引入了：
- 4K 输出
- 实时迭代式优化
- 复杂场景的混合式 JSON Prompting
- 性能较上一代提高约 4 倍
这项速度与品质的提升明确归功于在 TPU 上优化后的张量处理能力。
What I read this week...
- @Google announced Gemini 3 on November 18, 2025, calling it its “most intelligent” model to date and rolling it out day-one across Search (AI Mode), the Gemini app, AI Studio, Vertex AI, and a new agentic dev environment called Google Antigravity.
It has topped the charts on:
- LMArena Leaderboard with a breakthrough score right above Grok’s 4.1 release
- Humanity’s Last Exam with PhD-level reasoning
- GPQA Diamond (for scientific knowledge)
And leading many more benchmarks, showing that scaling laws are still intact.
The key insight is that multiple sources, including model cards and press coverage, confirm that Gemini 3 Pro was trained on Google’s own Tensor Processing Units (TPUs), not Nvidia GPUs. This gives Google a strong advantage over other firms that are in extreme competition for processing units.
Nano Banana has also been upgraded to a Pro version and is Google’s advanced image generation and editing tool. It introduces 4K output, real-time iterative refinement, and hybrid JSON prompting for complex scenes, at roughly a 4x speedup over prior Nano Banana generations. The additional speed and quality is attributed explicitly to optimized tensor processing on TPUs.
NvidiaAIGoogle
相关观点
- 你之所以负担不起医疗，是因为：医疗保险保费太高了。 多向杰夫·贝索斯征税，出于其他原因也许是对的， 但这对解决医疗负担问题毫无帮助。 保费之所以这么高，是因为保险公司在利用《奥巴马医改》（Obamacare）引入的定价机制进行博弈。 《奥巴马医改》将毛利率上限设为 15%。结果是，医疗保险公司如果想赚更多钱，并不需要提高效率或加剧竞争，他们只需要协同提价即可——而他们确实这么做了。 对他们来说，在 10,000 美元的保费上赚 15%，远比在 1,000 美元的保费上赚 15%（甚至更高）、但要面对激烈竞争，更划算。 现在，《奥巴马医改》要求这些多出来的钱必须“用于医疗服务”，但如果你看看医疗体系中行政人员数量的暴涨，以及预期寿命的停滞不前，真正发生的事情是： 这些额外资金只是在系统内部来回流转，最终以各种形式流向了股东，却没有对医疗质量产生任何可衡量的改善——看看预期寿命 vs 医疗行政人员增长，或者医疗股股价 vs 预期寿命，你就明白了。 因此，要让医疗重新变得负担得起，就必须改变毛利率规则，并强制私营保险公司 / 药品福利管理机构（PBMs）/ 医疗服务提供者之间展开真正的竞争。 The reason you can’t afford healthcare: Healthcare premiums are too high. Taxing Jeff Bezos more may be right for other reasons but it will do nothing to fix this. Premiums are too high because insurance companies game the pricing model brought by Obamacare. Obamacare capped gross margins at 15%. So for healthcare companies to make more money, they didn’t have to find efficiency or compete harder, they could coordinate and raise prices. And they did that. Making 15% on $10,000 with little competition is better for them than making 15% or more on $1,000. Now Obamacare says you must spend these extra dollars on services but if you look at the growth of administrative staff and the flatlining of life expectancy, what really happened was that the extra dollars just sloshed around in the system and eventually made its way to stockholders with no measurable impact to quality - look at life expectancy vs admin staff growth in healthcare or stock price of health care stocks vs life expectancy. So to make healthcare affordable again, change the gross margin rules and force competition between the private insurers/PBMs/providers.Chamath Palihapitiya · 2026/01/01
- 这不是“每年 1%，持续 5 年”。 而是一次性对所有资产征收 5% 的税，而且它会彻底扼杀加州的创业精神。 我给你举个例子： John Doe 创办了一家公司。他拿的是象征性的薪水——假设是 15 万美元/年——其余收入以公司股权形式获得。假设他持有 20% 的股份。2026 年，他从风投那里融资，有人向公司投资 1 亿美元，并给公司估值 60 亿美元。这意味着他那 20% 的股份“价值”12 亿美元。 我给“价值”打引号，是因为他根本卖不了。 他只是拥有一个“账面价值”，名义上看起来是个亿万富翁。但他的实际生活收入仍然只有 15 万美元/年。 仅仅因为有人押注这家公司，并不意味着你的银行账户里就凭空多出了 12 亿美元。 但在这个拟议税制下，John Doe 在 2027 年需要向加州缴纳 6000 万美元现金税款。 问题是—— 👉 他从哪儿拿这 6000 万？ 你认识什么买家吗？ 是市场上其他人都不知道的那种？ 会在一个被迫卖资产、急需现金的卖家身上， 按最高估值接盘？ 再想象一下：税刚评估完，2027 年初，公司估值被大幅下调到 2 亿美元。那他的 20% 现在只值 4000 万美元。但他仍然欠加州 6000 万美元税款。 还是那句话：👉 他的股份依然没有买家。 他的年收入，依旧只有 15 万美元。 那你告诉我，👉 这个人该怎么办？ 他现在“身价”4000 万，却欠加州 6000 万。 他是不是应该现在就去申请破产？ 就因为他“蠢到”在加州创业？ 所以你们真的“搞定了亿万富翁”吗？ 没有。 真正的超级富豪要么早就离开了加州， 要么通过税务结构规避了这类税， 要么有能力把官司一路打下去。 但你们确实会把一大批—— 年轻的、有创造力的、想创业、想雇人的人—— 直接拖进破产法庭。 干得漂亮，Ro。你应该为此感到骄傲。 It’s not 1% a year for 5 years. It’s a one time 5% tax on all assets and it will kill entrepreneurship in California. Here is an example: John Doe starts a company. He takes a nominal salary - say $150k for this example - and the rest in equity in the company. Let’s say he owns 20%. He raises VC capital in 2026 from someone that invests $100M into the company and values the company at $6B. This means his 20% is “worth” $1.2B. I put it in quotes because he can’t actually sell. He has a paper value that putatively says he’s a billionaire. But he actually lives on $150k because that is what his income is. Just because someone decides to make a bet on the business does not mean some bank account in your name magically gets created with $1.2B in it. Under the proposed tax, however, John Doe would now owe $60M in cash to California in 2027. How will he pay it? Is there some buyer you know of, that the rest of the market doesn’t, that will do a deal at the max value when there is a distressed seller like John Doe who needs money he doesn’t have to pay taxes on value he also doesn’t have! Now imagine that after the tax is assessed, in early 2027, the company takes a write down to $200M. Now his share is $40M. But he still owes $60M. Again, there are no buyers for his shares per se. He still only makes $150k/yr. What is this person supposed to do? He now has a “worth” of $40M but owes California $60M. Should he declare bankruptcy now because he tried to start a business but was retarded enough to do it in California? So did you really get the billionaires?? No. Because the mega billionaires have already left or are tax structured to minimize the tax or will fight it. You will, however, drag a bunch of young, energetic folks who want to make things and hire people into bankruptcy court. Awesome work, Ro. You should be proud.Chamath Palihapitiya · 2025/12/28
- 以“精准”构建 vs 以“速度”构建。 前者是在一开始放慢节奏，把目标和细节想清楚、做对，这样就能在中期和后期实现真正的快速推进。 后者则是一上来就追求速度，结果很快发现自己做出了大量次优决策，代码质量糟糕、结构混乱，项目迅速变得难以维护、充满“垃圾代码”。 2026 年将是一个重要的年份，越来越多的公司会意识到这一点，并开始要求一种协作式、有明确立场、结构化且纪律严明的软件开发方式。 这正是我们在 @8090solutions 打造 Software Factory 时所创建的东西：一种全新的、协作式的软件开发理念——自律、有主张、系统化、方法论清晰、注重细节。 自今年推出以来，我们已经在多个要求极高的行业中成功赢得客户，包括：医疗健康、生命科学、制造业、保险、金融服务和能源等。这些行业的公司无法承受“凭感觉写代码（vibe coding）”的风险。他们编写的代码必须足够精确、按预期运行、易于维护，并具备极高的可靠性和可用性。 目前，Software Factory 已被广泛用于这些行业中，以系统化、可控的方式，用成本仅为原来一小部分的定制化新软件，逐步替换“软件工业复合体”时代遗留下来的旧系统。 如果你想了解更多，欢迎联系我们。 Building with precision vs building with "speed". The former is about going slow up front so you get the intention and details right. This then allows you to go fast in the middle and the end. The latter is about going fast up front only to realize you made many sub optimal decisions, the code is janky and it quickly becomes unwieldy and slop. 2026 will be an important year when more and more companies come to this realization and demand a collaborative, opinionated way to build software in a structured and disciplined way. This is what we created when we built Software Factory at @8090solutions. A collaborative new way to imagine software development: disciplined, opinionated, methodical, detailed... Since launching this this year, we have successfully won customers in many demanding industries: healthcare, life sciences, manufacturing, insurance, financial services and energy among others. The companies in these sectors can't afford vibe coding. The code they write needs to be precise, work as advertised, be easy to maintain and have high reliability/uptime. Software Factory is being used in all of these industries to methodically replace legacy software from the Software Industrial Complex with new, purpose-built custom software that is a fraction of the cost. Ping us if you want to learn more...Chamath Palihapitiya · 2025/12/16
- 2）在过去的数个月里，Tether——长期以来以发行最大稳定币 USDT 而闻名——悄然成为全球最大的私人黄金持有者之一。 Jefferies 的股票分析师 Fahad Tariq 和 Andrew Moss 最近发布了一份关于 Tether 在黄金市场影响力不断上升的报告。 Jefferies 的数据估算，Tether 当前持有 116 吨黄金（价值约 140 亿美元），其购金量约占： * 全球黄金需求的 ~2% * 全球各国央行购金量的 ~12% 这意味着，Tether 的黄金头寸规模已经与一些中小型央行（例如韩国、匈牙利或希腊）相当，这些国家的官方黄金储备均在 100–115 吨之间。 在这些持有量中，约 12 吨用于支撑 Tether 的黄金背书代币 Tether Gold（XAUt）。 Jefferies 指出，Tether 将金条视为一种长期的结构性资产： * 通过利润与稳定币业务收入持续增持黄金 * 将黄金作为长期价值储存手段 甚至正在探索对“黄金特许权（royalty）与金矿流媒体公司（streaming companies）”的投资，从而在黄金供应链的多个环节建立敞口，而不仅仅是购买放在金库中的金条。 2) In the past several months, Tether, long known as the issuer of the leading stablecoin USDT, has quietly become one of the largest private gold holders in the world. Fahad Tariq and Andrew Moss, equity analysts at Jefferies, recently published a report on Tether’s growing influence in the gold market. Jefferies quantifies Tether’s holdings at 116 tonnes (valued at roughly US$14 billion) and states its purchases comprised ~2% of global gold demand and ~12% of central bank buying. This corresponds to a gold position that is in the order of magnitude of the holdings of smaller central banks such as South Korea, Hungary or Greece (each holding around 100-115 tonnes of gold). A small portion of these holdings, about 12 tons, directly underpins the gold-backed token Tether Gold (XAUt). Jefferies suggests that Tether sees bullion as a structural asset: using profits and stable-coin revenue to build a long-term store of value, and even exploring investments in gold royalty and streaming companies, thus gaining exposure across the gold supply chain, not just to bars in vaults.Chamath Palihapitiya · 2025/11/30