SpaceHole

Share industry insights / Co-invest and Follow investments

[Oklo @oklo]

We were proud to welcome @ENERGY @SecretaryWright to @INL, where significant progress is underway on our first commercial Aurora powerhouse.

Together with the strides we're making toward criticality at our Groves Isotope Test Reactor in Texas, Oklo’s work on full civil nuclear facilities is moving forward on multiple fronts. Construction, fuel, licensing, and engineering are advancing in parallel, compressing timelines and redefining the pace of advanced nuclear deployment.

We're grateful for Secretary Wright's leadership and the continued partnership of DOE and INL as we work together to turn advanced nuclear from ambition into deployment.

[Elon Musk @elonmusk]

Just follow SpaceX if you want news about our company

[Jason Guo @StarSpaceVC]

★ Y Combinator (YC) 2026 Summer Closed-Door Memo Leaked — Declaring "SaaS Is Dead, Physical AI Reigns Supreme" .

☆ The most circulated document in Silicon Valley's venture capital circle today is the summer venture trend memo just sent by YC's core partners to their internal alumni. This memo has completely stripped away the last veil of the past decade's "software tax (SaaS)."

☆ Core Insight: YC clearly stated that the SaaS venture model of the past—wrapping a large language model, writing a few lines of code, and building an attractive frontend UI to charge enterprises subscription fees—has seen its commercial survival rate drop below 3% in 2026, as the cost of large model inference approaches zero.

★ Where the growth is: Venture capital is fleeing software and pure cloud-based AI, pouring into "Physical AI" in full force.

☆ Includes industrial robots and offline edge privacy computing hardware.

★ [StarSpace This Week's Reflection]:

☆ YC's report indicates that Silicon Valley's top venture capital trends have completely abandoned "inflated soft AI" and are now fully shifting toward physical infrastructure and edge hardware.

☆ However, physical AI cannot achieve mass production or scale up in the short term (within six months to a year), and thus cannot deliver performance results.

★ Then why do top-tier investors like a16z and YC keep pouring money into it so aggressively?

★ How should we defend and position ourselves when investing in both primary and secondary markets?

[Reuters @Reuters]

Nvidia clinches deals with South Korean giants including SK Group to advance AI boom https://t.co/yO402affTB https://t.co/yO402affTB

[How To Prompt @HowToPrompt__]

Google has published a paper that might end the transformer era.

For the last 7 years, every major AI, ChatGPT, Claude, Gemini, has been built on the exact same architecture: The Transformer.

But Transformers have a fatal flaw.

To remember context, they have to process every single word against every other word. It’s called quadratic complexity. As your prompt gets longer, the compute cost explodes.

The alternative is the old-school RNN (Recurrent Neural Network). RNNs are incredibly cheap and fast, but they have a fixed memory size. If you give them a long document, they get amnesia.

Until today.

Google researchers published Memory Caching: RNNs with Growing Memory.

And it fixes the biggest bottleneck in AI.

Instead of an RNN having a fixed, rigid memory that constantly overwrites itself, Google gave it a "save" button.

The technique allows the RNN to cache checkpoints of its hidden states as it reads.

The memory capacity of the RNN can now dynamically grow as the sequence gets longer.

They built four different variants, including sparse selective mechanisms where the AI actively chooses exactly which checkpoints matter most.

The results rewrite the rules of efficiency.

On long-context understanding and recall-intensive tasks, these new Memory-Cached RNNs closed the gap with Transformers.

They achieved competitive accuracy without the explosive, quadratic compute cost. It perfectly bridges the gap between the cheap efficiency of an RNN and the massive capability of a Transformer.

We have spent billions scaling Transformers because we thought they were the only way an AI could remember a long conversation.

But Google just proved we don't need to process the whole history every single time.

We just needed a smarter cache.

[Jason Guo @StarSpaceVC]

【StarSpace 星空前哨 一周思考】

★ 观察：在几场闭门半导体沙龙（如 HTEC 2026 趋势研判）中，顶尖架构师提出了一个 “金字塔倒悬” 的怪象：

☆ 云端AI的“金字塔倒悬”危机 vs 边缘端的“物理防线”；

☆ 最上方是数百家前沿芯片设计商、数千亿美元资本，疯狂砸向 AI 硬件创新；

☆ 而最下方的支撑点，却是极度稀薄、尚未实现大规模闭环盈利的真实企业级落地场景。

★ StarSpace 解析：

1、依靠巨额资本开支（CapEx）堆砌的“中心化算力奇点” 正在撞墙；

2、真正的 AI 增量在：物理AI（Physical AI）和 分布式边缘（Edge AI）；

3、智能不能只停留在云端，必须沉淀到物理实体（设备、传感器、人形/工业机器人）。

《StarSpace 非共识观点》：

1、未来的胜负手不再是谁能买到更多 GPU；

2、而是谁能将大模型进行极致的模型蒸馏与量化（Model Distillation & Quantization）；

【未来趋势判断】：

★ 算力正在走向分布式与本地化；

★ 协议正走向机器原生化；

★ 硬件正在寻找物理能效比的极限。

[Claude @claudeai]

Introducing Claude Opus 4.8: it builds on Opus 4.7 with sharper judgment, more honesty about its own progress, and the ability to work independently for longer than its predecessors.

Available today at the same price.

[David Sacks @DavidSacks]

Q: How are job postings for software engineers rising rapidly despite AI agents automating coding?

A: Because there’s far more code to manage than ever before. We’re already seeing a 14x YoY increase in GitHub commits, and it’s accelerating.

AI has dramatically lowered the cost of writing code, so it’s now being used across far more businesses, applications, and use cases.

We’re at the beginning of a massive productivity boom driven by the proliferation of bespoke software throughout the entire economy.

Coding has been AI’s breakout use case this year. The fact that it’s increased demand for software engineers — rather than decreased it — should call into question the entire “AI will cause mass job loss” narrative.

[Oklo @oklo]

In less than two months, Oklo’s isotope test reactor will go critical.

We’re building, iterating, and executing in real time across power, fuel, and isotopes.

Follow our journey: https://t.co/28lyGSS380

[Jason Guo @StarSpaceVC]

■ 以下几个长期被主流叙事掩盖、但极具破坏力、非共识/全新前沿科技赛道：

★ 物理层安全与逆全球化供应链的“非显性替代”

☆ 通常，大家只盯着看最顶尖的先进制程，但现在真正有大量隐性资金涌入的，是支撑分布式、高隐私需求的边缘侧主权硬件生态。

1、例如：防溯源的底层加密芯片，
2、完全脱离中心化、云控的、本地、物理网关，
3、在极其严苛、断网环境下运行的自组织网络（Mesh Network）硬件。

☆ 这不只是工程问题，而是地缘政治和隐私刚需，催生出的全新百亿级、非共识、公开市场。

★ 生物算力与非硅基计算 (Non-Silicon Computing)

☆ 硅基半导体的物理极限、和能耗墙大家都清楚。

☆ 现在顶尖的Frontier Tech VC，正在秘密押注合成生物学与计算的交叉点

☆ 比如：DNA数据存储和神经元芯片（Biocomputing）。

☆ 利用生物体天然的低能耗、超高密度特性，来做边缘侧的认知和计算。

☆ 这可能会彻底颠覆我们对“智能载体”的定义。

★ AI Agent 经济学：从“工具”到“原生经济实体”。

☆ 除了技术本身的迭代，更性感的是机 M2M 原生金融基础设施。

☆ 当 AI Agent 开始拥有独立的钱包、能够自主进行微型支付、竞价买卖算力和数据资产时，

☆ 整个互联网的流量分配逻辑、品牌营销规则（比如从给人类看变成给Agent看）都会重写。

☆ 这背后的清结算网络和协议，是传统金融根本无法承载的。

★ 真正意义上的工业/制造端“具身认知”

☆ 不是那些在发布会上走两步、翻个跟头的营销型人形机器人。

☆ 而是深嵌在极端复杂供应链、精密制造和高危环境中的 “隐形具身智能”。

1、可能是一条自我优化的、多维的机械臂/灵巧手；
2、可能是一个动态调配整个车间物理流向的边缘大脑；
3、可能是某敷膜/涂层新型材料，或者元器件；
4、可能是某种通讯协议/路径。

☆ 这里的壁垒，不是算法有多炫，而是对物理世界的真实洞察、和工业know-how的结合。

～～～End~~~

[Jason Guo @StarSpaceVC]

■ 从一名实战派天使投资人视角，看：人类程序员 vs AI 编程，消失的“直觉”与级别的鸿沟。

本文是我通过这两天亲自开发 “星空前哨投资”App，悟道了现在硅谷创投圈最核心的非共识的投资密码：“AI 编程” 和 “人类编程” 之间，隔着一条太平洋。

文章有点长，喜欢看短文的，只需看本文的最后一句话，就可以离开了。（这也就是“知道了” 和“懂了”的差距）

本文最后的一句结论就是：

“ AI ，是人类的 “黄金右手”；但脑子，依然是人类的”。

事情起因背景 1：

我背景是长期做一级市场天使投资的，现在又同时延伸到二级市场投资。两边业绩，我自己还很满意的，哈哈。

基于我这15+年硬科技赛道早期投资经验，我发现一个机会：现在没有一款专业的投资 APP，是同时适合一、二级市场投资人实用的。

虽然一级市场和二级市场，现在分别都有很多专业的 APP。

背景 2：我是完全不懂一行代码，杠杆的 coding 小白。

于是，我花了两天时间，在 Gemini 帮助下，成功完成了 APP 的开发，还在 AI 的建议下，取名“星空前哨 Starspace Sentinal ” APP。

在开发过程中，并不是像网上、自媒体、硅谷 KOL 们说的那样，交给 AI 编程，完全由 AI 帮你搞定。这些都是忽悠没有真正做过的人听的。

其实，你的系统、软硬件环境、各种软件版本的不同、系统硬件防火墙、哪怕程序代码中的标点符号、特别是软件不同版本的默认设置。。。都会影响到你的 APP，是否能真的能够跑起来。

真的，别太相信网上的 PR 忽悠，都有各自的目的，或者盲区。

这些 coding 技术上的事，并不是我本文想说的重点。

我本文想分享的是我过程后的感悟：

★ 现在的 AI 编程，无论是通过 VS Code 的 Cursor、GitHub Copilot 还是直接用大模型抠代码，其底层逻辑都是大规模概率预测。

低级别 AI 模型 vs 高级别 AI 模型：差别非常大。

特别是在分析 coding 的 bugs 能力上，关注点不一样，甚至有可能误导你，进入了死循环。

我一段时间就是这样被 AI 误导，进入了死循环，一天没有吃饭的。

平心而论，AI 初心并不是想忽悠你，后来，我换了更高级别模型后，我才知道 ，是 AI 认知问题。

所以花钱买高级别的模型，是值得的。

可能这就是未来的认知差，重点是，你当时还不知道是死局。

低级别的模型，只能做“单点代码补全”，或回答简单的语法问题。

而高级别模型，具备复杂推理能力的最新一代 Gemini 或 Claude 3.5/Opus，具备了上下文长文本（Long Context）的全局掌控力。

它能看懂你整个 GitHub 仓库的架构，帮你做模块设计。

这，就是 高 vs 低级别算法和推理模型的差异，也是套壳和蒸馏模型的差异，也是你掏更高价钱，买好模型的理由。

★ 人类与 AI 的本质难度差异： 

AI 的短板：AI 缺乏“架构直觉” 和 “边界长周期考虑”。

AI 写一个具体的 Python 爬虫函数极快，

但 AI 不知道，当 OpenD 的底层 API 发生网络抖动、频次限制（Rate Limit）或者数据格式微调时，系统应该如何处理。

AI 不知道：数据源的残缺、物理延时、漏斗效应。

AI 不知道，人类平台为了保护自己的商业护城河和防止被白嫖，对 OpenD 做了大量的频次限制（Throttling）、数据字段脱敏和非核心资讯截断。

您用 Python 程序去抓，抓到的本就是平台“挑剩下”或“故意延迟”的二手、三手数据。

这需要人类程序员十几年试错积累出来的“防御性编程”经验，更关键的是，对人类天性的理解。

★ 资本市场吹的泡沫：

现在一级市场高喊“AI 彻底消灭程序员”，那是写给不懂技术的 LP 看的。

真相是：AI 极大地消灭了“低级码农”（只会写增删改查、调包的中间层），但它极大地提高了对“系统架构师（Prompt / Product Architect）”的要求。

你不需要懂底层语法，但您必须拥有清晰的系统逻辑、模块解耦思维和对边界条件的判断力。

而这恰恰是有实战经验的人类，作为产品专家和投资人的核心优势。

AI 变成了您的“黄金右手”，但脑子依然是人类的。


[Elara Grace @ElaraGrace_AI]

🚨 Anthropic just gave Claude the ability to DREAM.

While you're asleep, Claude reviews everything it did that day, finds patterns in its own mistakes, and rewrites its own memory.

Yes — they literally call the feature "dreaming."

The co-founder thinks there's a 60% chance these models will train their own successors by 2028.

We are not ready for what's coming. 🧵👇

[Jason Guo @StarSpaceVC]

■ The full-scale commercialization of HBM4 and the battle for production capacity

★ Based on the current intent orders from cloud giants for 2027, the global supply and demand gap for HBM in 2027 is expected to further widen compared to 2026.

■ Storage factories strongly promote the "quarterly fixed price system":

★ Due to the introduction of advanced logic base die in HBM4 and the extremely high "wafer penalty" (the wafer area required to produce the same capacity is 3 times of traditional DRAM), the entire production capacity has been sold out.

★ Samsung and SK Hynix are currently fully rejecting the cloud giants' long-term fixed-price contracts of 2 to 3 years and strongly promoting short-term contracts with quarterly price hikes, with the sellers' pricing power reaching an all-time high.

♥︎♥︎♥︎【Starspace Perspective】:

♥︎ From the perspective of the "physical layer supply chain bottleneck" mainly composed of HBM4 and silicon photonics in the upstream, HBM4/LITE/MRVL will remain highly popular in 2027.

♥︎ NVIDIA has joined forces with Nous Research to launch the AI agent Hermes and deeply integrate it with the new hardware DGX Spark.

♥︎ Trend: "Industrial application opportunities" dominated by all-weather agentic independent
hardware in the downstream.

[a16z @a16z]

Mounting evidence of productivity lift across high-AI industries

Charts of the Week: https://t.co/nVVf3gIw3U

[Oklo @oklo]

At @INL, Oklo has moved from site mobilization into deep foundation excavation for our first commercial Aurora powerhouse.

What makes this moment significant is not just the physical progress at the site, but the coordinated progress taking shape across the entire project.

Construction, fuel, and licensing are all advancing at a pace that once seemed inconceivable. Oklo is demonstrating what is possible when the critical pieces move forward in parallel.

[Louisa Nicola @louisanicola_]

You can feel when your cognitive precision changes before anyone else notices it.

That is not intuition. That is your brain detecting instability in its own neural dynamics.

This new whole-brain model used empirical EEG and fMRI data to simulate how excitability and inhibition shape large-scale brain function. When that balance shifts, the systems responsible for executive control and cognitive flexibility shift with it.

Your thinking is biological infrastructure. Not abstraction.

[Jason Guo @StarSpaceVC]

■ Storage and Architecture:

★ HBM4 Dominance Established: SK Hynix is joining the "trillion-dollar club".

☆ Due to the strong demand for HBM4 on the Vera Rubin platform, its entire production capacity for 2026 has been sold out.

☆ This is not only an upgrade of memory, but also the beginning of large-scale commercialization of "integrated computing and storage".

☆ SK Hynix's leading position in the HBM4 field (accounting for approximately two-thirds of Nvidia's demand) makes it perform more like an "AI infrastructure company" during the semiconductor cycle.

☆ Note: I hold this stock.

★ CPO (Co-Packaged Optics): The Next Gen trend.

☆ As the bandwidth of a single node approaches 1.6T, traditional copper wire interconnection has reached its physical limit.

☆ Co-packaged optics (CPO) is moving from the laboratory to data centers, integrating optical engines directly on the chip side, which can reduce power consumption by 30% to 50%.

☆ This is not only a technological update but also a reshuffling of the equipment manufacturers' landscape.

[Jason Guo @StarSpaceVC]

■ Now that the tech giants have challenged "Moore's Law", will the next target be the "Von Neumann" architecture？

★ The biggest efficiency killer of silicon-based chips is the "von Neumann bottleneck".

☆ Data is constantly shuttled between the processor and memory, and 80% of the energy efficiency is wasted on this "highway".

☆ Traditional chips are "clock-driven", consuming energy even when there are no tasks.

☆ The biological architecture is "event-driven", and it only discharges when the input signal (Spike) reaches the threshold.

■ [Starspace Investment Observation]:

★ Focus on Brain-like Intelligence

★ Focus on Compute-in-Memory Technology

[Jason Guo @StarSpaceVC]

★【Starspace Predictive AI Agent RoadMap】:

★ Today, by the second half of 2026,
☆ Enterprise-level AI will shift comprehensively from "content generation" to "autonomous decision-making",
☆ AI-native organizations will start to emerge on a large scale.

★ Tomorrow, E-commerce | Business Iteration of AI Agents:
☆ The A2A payment protocol triangle battle:
☆ Google AP2, OpenAI ACP, and Coinbase x402 have begun their exclusive period in the ecosystem.

☆ Stripe has completed the integration of x402 payments on the Base chain.

★ AI algorithmic negotiation:
☆ AI shopping robots are beginning to have the "dynamic pricing confrontation" function, and online retail is evolving into real-time price negotiations between algorithms.

★ Seamless checkout penetration:
☆ Mastercard and Microsoft have collaborated to embed Copilot Checkout into retailers' private domain websites, allowing AI agents to directly complete mixed payments of fiat currency and stablecoins within the conversation flow.

[Jason Guo @StarSpaceVC]

■ The Undercurrents of Agent Payment Protocols (AP2 and x402)

★ The current bottleneck of the Agentic Economy (Agent-based Economy) is not that AI cannot do the work, but that AI lacks a "legally recognized payment identity".

★ Unification of the protocol layer: Recently, Google has taken the lead in formulating the AP2 (Agent Payments Protocol) protocol.

★ This agreement introduces "Mandates" with encrypted signatures, allowing AI to complete payment actions within a preset range without real-time human confirmation.

★ Native crypto challenger: Coinbase has launched x402 (A2A extension), attempting to establish a decentralized AI payment standard to compete with Google's AP2 for the underlying discourse power of "machine payment".

★ This means that in the future, M2M (machine-to-machine) transactions will no longer rely on traditional bank interfaces but will be completed through this native AI protocol.