Tweets

Is there an AI bubble? With the massive number of dollars going into AI infrastructure such as OpenAI’s $1.4 trillion plan and Nvidia briefly reaching a $5 trillion market cap, many have asked if speculation and hype have driven the values of AI investments above sustainable values. However, AI isn’t monolithic, and different areas look bubbly to different degrees.
- AI application layer: There is underinvestment. The potential is still much greater than most realize.
- AI infrastructure for inference: This still needs significant investment.
- AI infrastructure for model training: I’m still cautiously optimistic about this sector, but there could also be a bubble.

Caveat: I am absolutely not giving investment advice!

AI application layer. There are many applications yet to be built over the coming decade using new AI technology. Almost by definition, applications that are built on top of AI infrastructure/technology (such as LLM APIs) have to be more valuable than the infrastructure, since we need them to be able to pay the infrastructure and technology providers.

I am seeing many green shoots across many businesses that are applying agentic workflows, and am confident this will grow. I have also spoken with many Venture Capital investors who hesitate to invest in AI applications because they feel they don’t know how to pick winners, whereas the recipe for deploying $1B to build AI infrastructure is better understood. Some have also bought into the hype that almost all AI applications will be wiped out merely by frontier LLM companies improving their foundation models. Overall, I believe there is significant underinvestment in AI applications. This area remains a huge focus for my venture studio, AI Fund.

AI infrastructure for inference. Despite AI’s low penetration today, infrastructure providers are already struggling to fulfill demand for processing power to generate tokens. Several of my teams are worried about whether we can get enough inference capacity, and both cost and inference throughput are limiting our ability to use even more. It is a good problem to have that businesses are supply-constrained rather than demand-constrained. The latter is a much more common problem, when not enough people want your product. But insufficient supply is nonetheless a problem, which is why I am glad our industry is investing significantly in scaling up inference capacity.

As one concrete example of high demand for token generation, highly agentic coders are progressing rapidly. I’ve long been a fan of Claude Code; OpenAI Codex also improved dramatically with the release of GPT-5; and Gemini 3 has made Google CLI very competitive. As these tools improve, their adoption will grow. At the same time, overall market penetration is still low, and many developers are still using older generations of coding tools (and some aren’t even using any agentic coding tools). As market penetration grows —  I’m confident it will, given how useful these tools are — aggregate demand for token generation will grow.

I predicted early last year that we’d need more inference capacity, partly because of agentic workflows. Since then, the need has become more acute. As a society, we need more capacity for AI inference. 

Having said that, I’m not saying it’s impossible to lose money investing in this sector. If we end up overbuilding — and I don’t currently know if we will — then providers may end up having to sell capacity at a loss or at low returns. I hope investors in this space do well financially. The good news, however, is that even if we overbuild, this capacity will get used, and it will be good for application builders!

AI infrastructure for model training. I am happy to see the investments going into training bigger models. But, of the three buckets of investments, this seems the riskiest. If open-source/open-weight models continue to grow in market share, then some companies that are pouring billions into training models might not see an attractive financial return on their investment.

Additionally, algorithmic and hardware improvements are making it cheaper each year to train models of a given level of capability, so the “technology moat” for training frontier models is weak. (That said, ChatGPT has become a strong consumer brand, and so it enjoys a strong brand moat, while Gemini, assisted by Google's massive distribution advantage, is also making a strong showing.)

I remain bullish about AI investments broadly. But what is the downside scenario — that is, is there a bubble that will pop? One scenario that worries me: If part of the AI stack (perhaps in training infra) suffers from overinvestment and collapses, it could lead to negative market sentiment around AI more broadly and an irrational outflow of interest away from investing in AI, despite the field overall having strong fundamentals. I don’t think this will happen, but if it does, it would be unfortunate since there’s still a lot of work in AI that I consider highly deserving of much more investment.

Warren Buffett popularized Benjamin Graham’s quote, “In the short run, the market is a voting machine, but in the long run, it is a weighing machine.” He meant that in the short term, stock prices are driven by investor sentiment and speculation; but in the long term, they are driven by fundamental, intrinsic value. I find it hard to forecast sentiment and speculation, but am very confident about the long-term health of AI’s fundamentals. So my plan is just to keep building!

[Original text: deeplearning.ai/the-batch/issu… ]

NeurIPS received 21,575 paper submissions this year. Our Agentic Reviewer, released last week, just surpassed this in number of papers submitted and reviewed. It's clear agentic paper reviewing is here to stay and will be impactful!

🛠 Open Source Release

📦 DeepSeek-V3.2 Modelhuggingface.co/deepseek-ai/De…uX
📦 DeepSeek-V3.2-Speciale Modehuggingface.co/deepseek-ai/De…0AJ
📄 Tech repohuggingface.co/deepseek-ai/De…lMpX

5/n

💻 API Update

🔹 V3.2: Same usage pattern as V3.2-Exp.
🔹 V3.2-Speciale: Served via a temporary endpoint: base_urapi.deepseek.com/v3.2_speciale_…Lbm". Same pricing as V3.2, no tool calls, available until Dec 15th, 2025, 15:59 (UTC Time).
💡 V3.2 now supports Thinking in Tool-Use — detaiapi-docs.deepseek.com/guides/thinkin…DI8T

4/n

cool new capabilities in langchain 1.1 - ability to inspect LangChain LLM integrations to detect their capabilities (input tokens, multi-modality, etc)

allows you to compact message history differently for different llms depending on context window, etc

My biggest learnings from Jeanne DeWitt Grosser (ex-Chief Business Officer at @Stripe, now @Vercel COO):

1. What failed seven years ago now works with AI. In 2017, Jeanne tried to build a system at Stripe that would automatically personalize outbound emails based on company data. Despite working with world-class data scientists, it failed due to too many errors. Today, that exact same approach works. This shows how AI has made previously impossible ideas suddenly viable.

2. A single GTM engineer at Vercel reduced a 10-person sales team to 1 (in just 6 weeks). Jeanne’s team at Vercel had an engineer build an AI agent that handles inbound lead qualification, outbound prospecting, and deal loss evaluation. The agent costs $1,000 per year to run versus over $1 million in salaries for the sales team. The nine displaced team members moved to higher-value work rather than being laid off, and the remaining salesperson is 10 times more efficient.

3. Their AI deal-loss bot has become better at understanding what went wrong than humans. When Jeanne analyzed her biggest loss of the quarter, the salesperson blamed pricing. But an AI agent reviewed every email, call transcript, and Slack message and discovered the real reason: they never spoke to the person who controls the budget, and when ROI came up, the customer clearly didn’t believe the value claims. They are now using AI to analyze sales calls in real time and send alerts like “You’re halfway through the sales process and haven’t talked to a budget decision-maker yet.”

4. Wait until $1 million in revenue before hiring your first salesperson. Founders should continue selling themselves until they reach around $1 million in annual revenue with a repeatable process. The key is having a defined ideal customer profile—customers who look alike.

5. Segment customers on what drives their buying decisions, not just company size. OpenAI has roughly 3,000 employees, which would typically put them in the “mid-market” category. But they’re a top-25 website globally by traffic, so Vercel treats them as enterprise customers requiring complex sales. Effective segmentation combines company size with growth rate, web traffic, workload type, and industry—because selling to e-commerce companies requires completely different language than selling to crypto companies.

6. Most customers buy to avoid risk, not to gain opportunity. About 80% of customers purchase to reduce pain or avoid problems, while only 20% buy to increase upside. This means you should focus your sales messaging on what could go wrong without your product—like falling behind competitors or damaging their reputation—rather than just talking about exciting features. This is especially true when selling to larger companies, where individual careers are on the line.

7. Sales teams should be indistinguishable from product managers—for a bit. Jeanne hires salespeople who have such deep product knowledge that if you put one in front of a group of engineers, it should take 10 minutes to realize they’re not a product manager. This credibility allows sales teams to serve as an extension of research and development—a 20-person sales team talks to hundreds of customers weekly and can translate those conversations into product insights at scale.

8. Building your own AI sales tools may beat buying off-the-shelf software. Because AI is so new and every company’s sales process is unique, Jeanne finds that building custom internal agents often delivers more value than buying vendor solutions. A single go-to-market engineer built their deal analysis bot in just two days, perfectly tailored to their specific workflow. These engineers shadow top salespeople to understand their workflows, then build automation that would have taken months or been impossible just a few years ago.

9. Make every sales interaction great, whether customers buy or not. Jeanne replaced boring discovery calls at Stripe with collaborative whiteboarding sessions where customers drew their payment architecture. Many customers had never visualized their own systems before. They left with a useful asset and a feeling of collaboration, regardless of whether they bought. Many returned years later to purchase. Think about your go-to-market process like a product, not just a sales function.

10. Product-led growth has a ceiling—no $100 billion company runs on it alone. While product-led growth (where users can sign up and start using a product without talking to sales) works well for early growth, customers generally won’t spend a million dollars through a self-service flow. Every major technology company eventually builds a sales team for larger deals. The mistake is waiting too long, since building a predictable sales process takes time.

Great Engineers are Also Artists.

“I characterize art as something that is done for its own sake, and done well, and often creates a sense of beauty or some strong emotion.

And a lot of engineers are introverts.

As an aside, I hate the term “incel.” It’s just a way of putting introverts down. It’s the new “nerd,” if you will. If someone says that somebody is an incel, I’m more likely to want to interview them. So let’s move away from the slurs.

But introverts tend to want to express themselves through other things rather than going out and expressing themselves directly. So what are they going to do? They’re going to express themselves through their craft. They’re going to create art.

In my current company, at least half the engineers have serious artwork they’ve done on the side. World-class artwork—everything from elegant mathematical proofs to beautiful computer art, to literally sculpting things with clay, designing clothing, designing doorknobs, water bottles. There’s one who’s done incredible music videos, really good stuff. And I see a lot of the better engineers tinker with the AI art products, much more so than even so-called artists do. I think a lot of artists are scared by AI art products saying, “This is going to replace me.” Whereas someone who doesn’t have that identity of an artist and doesn’t feel threatened by it—it’s just a tool and they try it out to see what it can create.

Anything done for its own sake and done as well as one possibly can is art. And great engineers are also artists. They’re capable of anything. It’s just they’ve chosen to be engineers and focused on building things because engineering is the ability to turn your ideas and your art into things that actually work, that do something useful, that embody some knowledge in a way that it can be repeated and people can get utility out of it. But that doesn’t mean that it can’t be beautiful.”

🏆 我们非常荣幸地宣布，我们的论文《用于大型语言模型的门控注意力：非线性、稀疏性和无注意力陷阱》荣获 NeurIPS 2025 最佳论文奖！

对我们敬业的研究团队表示热烈祝贺，感谢他们为 AI 发展做出的卓越贡献。

阅读更多：blog.neurips.cc/2025/11/26/ann…

I talked to a startup that's not a software company but uses AI quite a lot. They currently have 6 employees. I asked how many more people they'd need to hire if they didn't use AI, and they said about 10. So AI is increasing their productivity about 2.7x.

🍌Nano Banana Pro PPT 生成提示词

整了一套非常漂亮的渐变拟物玻璃卡片风格 PPT 提示词，可以在 NotebookLM、Youmind、Listenhub、Lovart等支持 Nano Banana Pro 生成 PPT 的位置使用

------

你是一位专家级UI UX演示设计师，请生成高保真、未来科技感的16比9演示文稿幻灯片。请根据视觉平衡美学，自动在封面、网格布局或数据可视化中选择一种最完美的构图。

全局视觉语言方面，风格要无缝融合Apple Keynote的极简主义、现代SaaS产品设计和玻璃拟态风格。整体氛围需要高端、沉浸、洁净且有呼吸感。光照采用电影级体积光、柔和的光线追踪反射和环境光遮蔽。配色方案选择深邃的虚空黑或纯净的陶瓷白作为基底，并以流动的极光渐变色即霓虹紫、电光蓝、柔和珊瑚橙、青色作为背景和UI高光点缀。

关于画面内容模块，请智能整合以下元素：

1. 排版引擎采用Bento便当盒网格系统，将内容组织在模块化的圆角矩形容器中。容器材质必须是带有模糊效果的磨砂玻璃，具有精致的白色边缘和柔和的投影，并强制保留巨大的内部留白，避免拥挤。

2. 插入礼物质感的3D物体，渲染独特的高端抽象3D制品作为视觉锚点。它们的外观应像实体的昂贵礼物或收藏品，材质为抛光金属、幻彩亚克力、透明玻璃或软硅胶，形状可是悬浮胶囊、球体、盾牌、莫比乌斯环或流体波浪。

3. 字体与数据方面，使用干净的无衬线字体，建立高对比度。如果有图表，请使用发光的3D甜甜圈图、胶囊状进度条或悬浮数字，图表应看起来像发光的霓虹灯玩具。

构图逻辑参考： 如果生成封面，请在中心放置一个巨大的复杂3D玻璃物体，并覆盖粗体大字，背景有延伸的极光波浪。 如果生成内容页，请使用Bento网格布局，将3D图标放在小卡片中，文本放在大卡片中。 如果生成数据页，请使用分屏设计，左侧排版文字，右侧悬浮巨大的发光3D数据可视化图表。

渲染质量要求：虚幻引擎5渲染，8k分辨率，超细节纹理，UI设计感，UX界面，Dribbble热门趋势，设计奖获奖作品。