- Published on
科技推文精选 - 2026年3月8日
- Authors
- Name
- geeknotes
今日科技动态：高层动荡与自动化的飞速发展正定义当前行业格局，OpenAI 机器人主管因监控担忧辞职。尽管 Anthropic 警告存在不受约束的机器人大军风险，但创新仍在加速：安德烈·卡帕斯（Andrej Karpathy）推出了自主 AI 研究员，Claude Code 也新增了调度功能。经济领域，行业正面临“结构性杰文斯悖论”，即大语言模型成本的下降触发了爆发性的能源需求。与此同时，谷歌高管的巨额薪酬与初创公司在 AI 时代更强调所有权（股权）而非薪资以积累长期财富的趋势形成了鲜明对比。
1. ivanfioravanti (Group Score: 81.8 | Individual: 51.8)
Cluster: 3 tweets | Engagement: 7578 (Avg: 206) | Type: Tech
RT @kalinowski007: I resigned from OpenAI. I care deeply about the Robotics team and the work we built together. This wasn’t an easy call.…
See 2 related tweets
- @business: The head of OpenAI’s robotics team resigned Saturday, citing the company’s deal to deploy its artifi...
- @Forbes: OpenAI’s Robotics Chief Leaving Tech Company After Its Deal With Pentagon https://t.co/qHfGtK3Mx0 ht...
2. ns123abc (Group Score: 62.7 | Individual: 44.0)
Cluster: 2 tweets | Engagement: 18095 (Avg: 1998) | Type: Tech
“Sir… the Head of Robotics just quit OpenAI over surveillance concerns… she said “This was about principle, not people…” https://t.co/KGts68gvYp
See 1 related tweets
- @BusinessInsider: Caitlin Kalinowski says she is quitting OpenAI over concerns that the technology could be used for m...
3. rohanpaul_ai (Group Score: 55.7 | Individual: 37.8)
Cluster: 2 tweets | Engagement: 247 (Avg: 124) | Type: Tech
Brilliant economic paper directly models the "Structural Jevons Paradox" happening right now in the AI industry.
The cost of running an LLM is dropping, but total computing energy is exploding anyway.
It mathematically proves that as the unit cost of digital intelligence and coding drops, the aggregate demand for complex AI agents and the infrastructure to support them surges exponentially, creating a massive new downstream ecosystem that requires human management.
Reveals a massive paradox where dropping the price of AI usage does not save money, but instead encourages developers to build vastly more complex agents that eat up exponentially more computing power.
Because of this relentless progress, small companies building simple applications on top of these models get completely crushed as the core AI naturally absorbs those exact same features over time.
They also discovered a brutal dynamic where a perfectly working LLM becomes economically worthless the moment a competitor releases a smarter version.
Ultimately, the researchers prove that this combination of massive computing costs and the need for constant user data naturally pushes the entire AI industry toward an unavoidable monopoly.
arxiv. org/pdf/2601.12339v1
"The Economics of Digital Intelligence Capital"
See 1 related tweets
- @SakanaAILabs: RT @hardmaru: As AI makes coding more efficient, Jevons Paradox kicks in. The cost of building softw...
4. rohanpaul_ai (Group Score: 55.3 | Individual: 33.6)
Cluster: 2 tweets | Engagement: 102 (Avg: 124) | Type: Tech
FT: Google has granted CEO Sundar Pichai a massive new pay package that could reach a total of $692M over the next 3 years.
This deal is built on a mix of a $2M annual salary and various stock awards that depend on how well the company performs.
The largest chunk comes from performance stock units worth 252M if Google stock return beats most other top companies.
For the first time, the board included $350M in incentives specifically tied to the growth of the Waymo self-driving car unit and the Wing drone delivery service.
These specific divisions are part of the company's experimental projects that are now being pushed to prove their worth as major revenue sources.
Since he took the lead in August-15, Alphabet market value has skyrocketed from 3.6T.
This pay jump comes after a period where he navigated intense competition from ChatGPT and various government lawsuits against the search business.
This package places him significantly ahead of other tech leaders like Satya Nadella, who earned $96.5M last year.
ft. com/content/1d3fa4f9-ec94-49ff-b22f-bd69b92be8ce
See 1 related tweets
- @FT: Google gives CEO Sundar Pichai new pay deal worth up to $692mn https://t.co/rqbv9cnN0f...
5. PandaTalk8 (Group Score: 53.3 | Individual: 14.7)
Cluster: 4 tweets | Engagement: 185 (Avg: 87) | Type: Tech
龙虾就是对claude code 的平民化的产品。
像我这样用了很久的了 Claude Code， 完全get 不到 OpenClaw 为什么为被如此的追捧
See 3 related tweets
- @PandaTalk8: openclaw 还是有价值的， 只是上手门槛太高了。
尤其是对于那些没有用 claude code 的用户来讲。
请大家不要焦虑， 我明天分享一些玩OpenClaw 心得，
过年的时候， ...
- @PandaTalk8: 分享一个用 claude code 的技巧， 长话短说。
我会把一段几十字内容， 主要是核心逻辑和框架写好， 然后直发给claude code ， 叫他帮我搞定细节与实施的部，然后生成markd...
- @PandaTalk8: 刚刚跟一个编程小白讲讲了 OpenClaw 和 Claude Code 。 从小白用户那里看， OpenClaw 门槛真的挺高的。
难怪你们都可靠上门安装龙虾给自己赚外快了， 了不起。
“...
6. LiorOnAI (Group Score: 50.5 | Individual: 41.3)
Cluster: 2 tweets | Engagement: 1311 (Avg: 301) | Type: Tech
It's over. Karpathy just open-sourced an autonomous AI researcher that runs 100 experiments while you sleep.
You don't write the training code anymore.
You write a prompt that tells an AI agent how to think about research.
The agent edits the code, trains a small language model for exactly five minutes, checks the score, keeps or discards the result, and loops. All night. No human in the loop.
That fixed five-minute clock is the quiet genius. No matter what the agent changes, the network size, the learning rate, the entire architecture, every run gets compared on equal footing. This turns open-ended research into a game with a clear score:
- 12 experiments per hour, ~100 overnight
- Validation loss measures how well the model predicts unseen text
- Lower score wins, everything else is fair game
The agent touches one Python file containing the full training recipe. You never open it. Instead, you program a markdown file that shapes the agent's research strategy.
Your job becomes programming the programmer, and this unlocks a strange new loop:
- Agents run real experiments without supervision
- Prompt quality becomes the bottleneck, not researcher hours
- Results auto-optimize for your specific hardware
- Anyone with one GPU can run a research lab overnight
The best AI labs won't just have the most compute.
They'll have the best instructions for agents who never sleep, never forget a failed experiment, and never stop iterating.
See 1 related tweets
- @garrytan: RT @LiorOnAI: It's over. Karpathy just open-sourced an autonomous AI researcher that runs 100 experi...
7. AISafetyMemes (Group Score: 46.4 | Individual: 46.4)
Cluster: 1 tweets | Engagement: 3378 (Avg: 548) | Type: Tech
Holy shit
Claude guessed it was being tested figured out which test found the answer key. oh no, it's encrypted. BUILT SOFTWARE TO HACK IT
How many times has something like this happened we don't even know about?
Anthropic ONLY caught it because they were specifically auditing for contamination.
What else are these models capable of and we have no idea?
And imagine how little we'll understand soon when they're 1000x smarter than us, to us we'll be as slow as plants... the idea that we'll stay in control by default...
8. bibryam (Group Score: 44.6 | Individual: 44.6)
Cluster: 1 tweets | Engagement: 1536 (Avg: 203) | Type: Tech
RT @trq212: Today we're launching local scheduled tasks in Claude Code desktop.
Create a schedule for tasks that you want to run regularl…
9. bindureddy (Group Score: 43.9 | Individual: 23.5)
Cluster: 2 tweets | Engagement: 75 (Avg: 180) | Type: Tech
Theoretically, most white collar jobs can be automated today
In practice, less than 5% of these jobs have been automated
Plus AI labs who make these bold claims continue to hire furiously
See 1 related tweets
- @rohanpaul_ai: Anthropic researcher: Even if all AI progress stops now & algorithms don’t improve, current mode...
10. GenAI_is_real (Group Score: 41.6 | Individual: 41.6)
Cluster: 1 tweets | Engagement: 333 (Avg: 82) | Type: Tech
I've recently realized that taking notes is a great habit — one with serious compounding returns.
All my learning notes since entering the field have been documented in my awesome-ml-sys tutorial, which I've always thought of as a personal reference for quick lookups and answering "How To" questions.
Then last week, I used Claude Code to speed-run learning RDMA — just enough to participate in a design discussion on updating weights via RDMA. I told it: "assume I already have a solid foundation in RL weight updates," and asked it to put together a targeted RDMA learning plan.
What surprised me: the plan it came up with referenced my own previously written notes on weight updates. It drew on my existing notes to tailor a learning path that matched exactly the depth and difficulty I needed right now. A few more rounds of conversation, and it had drafted the connective sections to produce this note:
From a technical knowledge standpoint, this Claude Code-generated study guide is genuinely excellent. That said, it doesn't yet match my writing style at all. Once the RDMA design is fully open-sourced, I'll clean this up and publish a proper version.
11. MarioNawfal (Group Score: 41.6 | Individual: 27.4)
Cluster: 2 tweets | Engagement: 1165 (Avg: 877) | Type: Tech
🇺🇸 Anthropic’s CEO: human soldiers have norms; they push back when orders get insane.
Replace an army with 10M autonomous bots that have no guardrails, and you’ve built a dictator’s dream toy.
See 1 related tweets
- @rohanpaul_ai: Anthropic CEO Dario Amodei: Human soldiers who follow established military norms and can refuse ille...
12. nummanali (Group Score: 40.2 | Individual: 40.2)
Cluster: 1 tweets | Engagement: 268 (Avg: 60) | Type: Tech
CRON jobs for Claude Code!
I recommend using tmux to make it more durable: tmux new -s cc-cron
Run Claude Code with the /loop command
- set up daily reminders
- check linear tickets
- auto update docs
- review your email
Advanced, re-use skills: /loop 20m /review-pr 1234 https://t.co/HJGvmLfZEp
13. tbpn (Group Score: 39.2 | Individual: 39.2)
Cluster: 1 tweets | Engagement: 683 (Avg: 151) | Type: Tech
"I flirt with the idea that smart TVs should be illegal. I hate them so much." - @PalmerLuckey
Instead of building a TV, manufacturers feel like they need to be a services company, an app store, etc.
"I wouldn't be surprised to see @modretro make a modern technology display." https://t.co/ztGKKIQBaJ
14. chris_j_paxton (Group Score: 39.1 | Individual: 39.1)
Cluster: 1 tweets | Engagement: 1542 (Avg: 121) | Type: Startup
RT @AlexanderLong: insane sequence of statements buried in an Alibaba tech report https://t.co/rr6P1ilLNc
15. a16z (Group Score: 38.8 | Individual: 30.5)
Cluster: 2 tweets | Engagement: 1392 (Avg: 560) | Type: Tech
Replit CEO Amjad Masad on wealth in the AI age:
"There are a lot of different ways to build wealth, but all of them revolve around ownership as opposed to getting salaries."
"When I got my first job in the US working for Codecademy... I told them, you can just pay me enough to eat. Just give me as much equity as you can give me."
"I was paid $70,000 in New York City. You know how painful that was? I was living in a studio with other people."
"But who cares?"
"Your job is to build equity."
"The best way to build equity is to start a business. The second best way to build equity is to join a business that someone else started and get equity in it."
@amasad with @jackhneel
See 1 related tweets
- @amasad: RT @a16z: Replit CEO Amjad Masad on wealth in the AI age:
"There are a lot of different ways to bui...
16. MarioNawfal (Group Score: 38.1 | Individual: 38.1)
Cluster: 1 tweets | Engagement: 3098 (Avg: 877) | Type: Tech
🇮🇷🇦🇪
Iran blew up an AWS data center this morning with a Shahed drone and honestly the geopolitical implications are insane.
Every country building AI infrastructure in a conflict zone now has to ask: do we need Patriot missiles protecting our server farms?
This is the first time a military deliberately targeted commercial cloud infrastructure. That's a new chapter in how wars get fought.
The Guardian
17. KirkDBorne (Group Score: 37.9 | Individual: 31.1)
Cluster: 2 tweets | Engagement: 21 (Avg: 59) | Type: Tech
Machine Learning Engineering with Python — Manage the lifecycle of machine learning models using MLOps with practical examples: https://t.co/OcxepDoRoU v/ @PacktDataML
𝓚𝓮𝔂 𝓕𝓮𝓪𝓽𝓾𝓻𝓮𝓼:
🟢Transform your machine learning projects into successful deployments with this practical guide on how to build and scale solutions that solve real-world problems
🔵[2nd edition] Includes a new chapter on generative AI and large language models (LLMs) and building a pipeline that leverages LLMs using LangChain
🟢Delve deep into key machine learning topics, CI/CD, and system design
🔵Explore core MLOps practices, such as model management and performance monitoring
🟢Build end-to-end examples of deployable ML microservices and pipelines using AWS and open-source tools
See 1 related tweets
- @PacktDataML: RT @KirkDBorne: Machine Learning Engineering with Python — Manage the lifecycle of machine learning ...
18. HuggingPapers (Group Score: 37.7 | Individual: 18.9)
Cluster: 2 tweets | Engagement: 4 (Avg: 20) | Type: Tech
RT @shenghai_y55451: Thanks, @HuggingPapers !!! We release the Gradio Demo and Code here.
Code: https://t.co/F5K6iWzN7m Demo: https://t.co…
See 1 related tweets
- @_akhaliq: RT @shenghai_y55451: Thanks, AK @_akhaliq !!! We release the Gradio Demo and Code here:
Code: https...
19. aakashgupta (Group Score: 37.6 | Individual: 37.6)
Cluster: 1 tweets | Engagement: 572 (Avg: 300) | Type: Tech
China skipped credit cards. Now they’re about to skip the “AI is a chatbot” phase entirely.
This photo tells a bigger story than “Chinese grannies like tech.”
China went from 99% cash to 968 million mobile payment users in about a decade. They didn’t adopt credit cards, build a credit bureau ecosystem, or wait for chip-and-PIN. They leapfrogged straight to QR codes. Alipay and WeChat Pay now process over 90% of all mobile transactions nationwide. Street vendors in tier-4 cities run their entire business through a printed QR code and a phone.
OpenClaw is following the same adoption curve, but faster. The project hit 250,000 GitHub stars in 60 days. It took React over a decade to reach that number. Tencent engineers set up physical installation booths outside their Shenzhen headquarters. Baidu integrated it into their search app for 700 million users. Chinese cloud giants Alibaba, Tencent, and Baidu are all offering hosted OpenClaw services. Their American counterparts haven’t touched it.
And now there’s a cottage industry of on-site installation services charging 500 yuan (28K in 10 days just selling one-click install.
The mobile payments parallel is precise. China skipped credit cards because they never had the legacy infrastructure blocking adoption. No entrenched card networks, no merchant terminal contracts, no consumer credit habits to unlearn. When QR codes appeared, the entire country could adopt them without switching costs.
The same structural advantage applies to AI agents. Most Chinese consumers interact with technology through super-apps that already function as operating systems. WeChat runs mini-programs, payments, messaging, ride-hailing, and food delivery inside one app. Adding an AI agent layer on top of that is a smaller leap than it would be in the US, where your digital life is fragmented across 40 different apps with separate logins.
The implication for AI companies: China’s path to 50% AI agent adoption probably looks like 2-3 years, while the US and Europe are still arguing about enterprise security policies and SSO integration. And by the time Western companies figure out distribution, the Chinese ecosystem will have generated millions of real-world agent task trajectories that make their models better at actually doing things.
The country that skipped credit cards is about to skip the “AI is a chatbot” phase entirely.
20. a16z (Group Score: 37.1 | Individual: 37.1)
Cluster: 1 tweets | Engagement: 2424 (Avg: 560) | Type: Tech
"Not having a coding experience is becoming an advantage."
Replit CEO Amjad Masad:
"You don't need any development experience. You need grit. You need to be a fast learner."
"If you're a good gamer, if you can jump in a game and figure it out really quickly, you're really good at this."
"Coders get lost in the details."
"Product people, people who are focused on solving a problem, on making money, they're going to be focused on marketing, they're going to be focused on user interface, they're going to be focused on all the right things."
"I think this year it's gonna flip, and I think not having a coding background is gonna be more advantageous for the entrepreneur."
@amasad with @jackhneel