- Published on
热门科技推文 - 2026-03-12
- Authors
- Name
- geeknotes
2026年3月12日 科技每日简报
Today's top tech conversations are led by @rohanpaul_ai, whose post about 'In China, everyone is so hyped...' garnered the highest engagement. Key themes trending across the top stories include agent, https, patterns, multi, agents. The community is actively discussing recent developments in AI, engineering practices, and startup strategies.
1. rohanpaul_ai (Group Score: 124.4 | Individual: 31.7)
Cluster: 6 tweets | Engagement: 61 (Avg: 92) | Type: Tech
In China, everyone is so hyped about the OpenClaw AI agent that tech stocks are shooting through the roof for any company using it.
But now, the government in Beijing is stepping in to block government workers and big banks from actually using the tool.
See 5 related tweets
- @business: EXCLUSIVE: China moves to restrict banks, state firms and government bodies from using OpenClaw AI a...
- @business: China's enthusiasm for the OpenClaw AI agent has sent shares of tech companies that embraced it soar...
- @Reuters: China moves to curb use of OpenClaw AI at banks, state agencies, Bloomberg News reports https://t.co...
- @Reuters: China warns state-owned firms and government agencies against OpenClaw AI, sources say https://t.co/...
- @business: China is going all-in on OpenClaw and AI agents before anyone can agree on safety rules, writes @cat...
2. rohanpaul_ai (Group Score: 122.3 | Individual: 33.8)
Cluster: 7 tweets | Engagement: 68 (Avg: 92) | Type: Tech
OpenAI is planning to integrate its Sora video generator directly into ChatGPT to streamline how you create visual content.
The Information and Reuters
Sora originally launched as a separate app in September-25 and focuses on generating high-fidelity video from text prompts.
By adding this to ChatGPT, you can use natural language to iterate on video scenes rather than relying on complex standalone tools.
Even with this update, the company will keep the standalone Sora app running for users who want a dedicated social-sharing experience.
reuters .com/business/openai-plans-launch-its-sora-video-tool-chatgpt-information-reports-2026-03-11/
See 6 related tweets
- @Reuters: OpenAI plans to launch its Sora video tool in ChatGPT, The Information reports https://t.co/NIE8mrQw...
- @theinformation: OpenAI plans to soon launch its Sora AI-generated video capabilities in ChatGPT, according to people...
- @theinformation: Exclusive: In a strategy shift, OpenAI plans to add its Sora AI video generation to ChatGPT.
Read m...
- @theinformation: Exclusive: OpenAI plans to bring its Sora video AI directly into ChatGPT, expanding the chatbot’s cr...
- @theinformation: RT @steph_palazzolo: New w/ @srimuppidi: OpenAI is adding its Sora video gen capabilities to ChatGPT...
3. vllm_project (Group Score: 98.6 | Individual: 27.2)
Cluster: 4 tweets | Engagement: 180 (Avg: 262) | Type: Tech
🎉 Congrats to @nvidia on the release of Nemotron 3 Super — day-0 support in vLLM v0.17.1! Verified on NVIDIA GPUs.
120B hybrid MoE, only 12B active at inference. Big upgrades over the previous Nemotron Super:
- 5x higher throughput
- 2x higher accuracy on Artificial Analysis Intelligence Index
- Multi-Token Prediction (MTP) for faster long-form generation
- Configurable thinking budget — dial accuracy vs token cost per task
- 1M token context window
Supports BF16, FP8, and NVFP4. Fully open: weights, datasets, recipes.
Blog: https://t.co/PAN0y778iB
🤝 Thanks @NVIDIAAIDev Nemotron team and vLLM community contributors!
See 3 related tweets
- @lmsysorg: 🎉 Congrats on the launch of Nemotron 3 Super from @nvidia! Day-0 support is now live in SGLang.
A 1...
- @mark_k: NVIDIA just announced their latest release: Nemotron 3 Super. It is an impressive open 120B-paramete...
- @FireworksAI_HQ: NVIDIA Nemotron 3 Super just dropped Day-0 on Fireworks! This 120B open hybrid MoE model from @NVIDI...
4. theinformation (Group Score: 86.3 | Individual: 29.5)
Cluster: 3 tweets | Engagement: 1 (Avg: 10) | Type: Tech
Today on TITV:
-Nvidia invests $2 billion in Nebius | @ttunguz, Theory Ventures
-Oracle revenue growth accelerates in Q3 | Tyler Radke, @citi
-OpenAI’s bet on ChatGPT apps faces hurdles | @anngehan
-OpenAI to launch Sora video in ChatGPT | @steph_palazzolo
-How AI is changing software sales | @dallasdesq, at @PwC
📺 Tune in at 10 am PT / 1 pm ET on https://t.co/iyO8ht485m
See 2 related tweets
- @theinformation: Today on TITV:
-Nvidia invests $2 billion in Nebius | @ttunguz, Theory Ventures
-Oracle revenue gr...
- @business: Nvidia will invest $2 billion in Nebius as part of a strategic partnership to develop and build AI d...
5. ReutersBiz (Group Score: 80.1 | Individual: 22.7)
Cluster: 4 tweets | Engagement: 56 (Avg: 33) | Type: Tech
Oracle predicted that the AI data center boom will power its revenue above Wall Street estimates well into 2027, sending its shares up 8.3% in extended trading. Read more: https://t.co/LuJ2yFnmUz https://t.co/D42ZLqu6vo
See 3 related tweets
- @Reuters: Oracle predicted that the AI data center boom will power its revenue above Wall Street estimates wel...
- @Reuters: Oracle has predicted that the AI data center boom will power its revenue above Wall Street estimates...
- @Reuters: Oracle sees AI boom through at least 2027, sending shares up 8% https://t.co/qk1wi2J0FH https://t.co...
6. CloudflareDev (Group Score: 77.8 | Individual: 29.7)
Cluster: 3 tweets | Engagement: 102 (Avg: 66) | Type: Tech
Building multi-agent systems? @NVIDIA's Nemotron 3 Super (120B A12B) is now on Workers AI.
Reasoning and tool calling for complex multi-agent workflows
Built for code, finance, cybersecurity, and search agent use cases
Learn more:
See 2 related tweets
- @LightningAI: Everyone wants to build multi-agent AI but most get stuck on infrastructure.
Run @nvidia Nemotron 3...
- @BrianRoemmele: New open source NVIDIA Nemotron 3 Super, a 120‑billion‑parameter with 12 billion active parameters ...
7. databricks (Group Score: 66.9 | Individual: 33.6)
Cluster: 2 tweets | Engagement: 36 (Avg: 48) | Type: Tech
The next generation of AI agents won't just write software, but operate the data systems that modern businesses depend on.
Genie Code is built for exactly that: an autonomous agent that understands your data, maintains pipelines, diagnoses failures, and keeps things running.
Databricks co-founder and CEO @alighodsi on what makes it different: "It's not about just generating random code snippets, but understanding the entire structure of the data problem and working through the modeling workflow the same way a data scientist or engineer would."
@iamVictorDey covers the full story for @FastCompany: https://t.co/Mbr5LRalhu
See 1 related tweets
- @databricks: Databricks CEO and co-founder @alighodsi joined @technology to discuss the launch of Genie Code, an ...
8. chatgpt21 (Group Score: 66.9 | Individual: 36.9)
Cluster: 2 tweets | Engagement: 355 (Avg: 174) | Type: Tech
I have 9 codex tasks running for me right now on my PC completing more work than I could have possibly in months.
I’m starting to see why there’s a vibe shift.
5.4 will just try over and over and over until it figures out something and it returns a working product after hours of effort. It’s just 1-2 shotting everything. And I know it’s not AGI yet & will still will make some bonehead mistakes and not be able to run tests properly and see what it’s doing wrong sometimes. But I’m starting to see the pieces coming together. And it’s worlds ahead of what we had last year.
It’s witty, quick, & smart.
See 1 related tweets
- @chatgpt21: I have 9 codex tasks running for me right now on my PC completing more work than I could have possib...
9. LangChain (Group Score: 64.2 | Individual: 33.1)
Cluster: 2 tweets | Engagement: 45 (Avg: 35) | Type: Tech
New Conceptual Guide: You don’t know what your agent will do until it’s in production 👀
With traditional software, you ship with reasonable confidence. Test coverage handles most paths. Monitoring catches errors, latency, and query issues. When something breaks, you read the stack trace.
Agents are different. Natural language input is unbounded. LLMs are sensitive to subtle prompt variations. Multi-step reasoning chains are hard to anticipate in dev.
Production monitoring for agents needs a different playbook. In our latest conceptual guide, we cover why agent observability is a different problem, what to actually monitor, and what we've learned from teams deploying agents at scale.
Read the guide ➡️ https://t.co/6RsZBEPx7s
See 1 related tweets
- @LangChain: San Francisco Meetup: 💻 It Worked on My Laptop - Agents in Production
Unlike traditional software, ...
10. KirkDBorne (Group Score: 59.6 | Individual: 31.6)
Cluster: 2 tweets | Engagement: 10 (Avg: 37) | Type: Tech
New release from @PacktDataML at https://t.co/lj0EclAewR
"Agentic Architectural Patterns for Building Multi-Agent Systems: Proven design patterns and practices for GenAI, agents, RAG, LLMOps, and enterprise-scale AI systems"
Contents: 🔷GenAI in the Enterprise: Landscape, Maturity, Agent Focus 🔷Agent-Ready LLMs: Selection, Deployment, Adaptation 🔷The Spectrum of LLM Adaptation for Agents: RAG to Fine-tuning 🔷Agentic AI Architecture: Components & Interactions 🔷Multi-Agent Coordination Patterns 🔷Explainability & Compliance Agentic Patterns 🔷Robustness & Fault Tolerance Patterns 🔷Human-Agent Interaction Patterns 🔷Agent-Level Patterns
See 1 related tweets
- @KirkDBorne: New release from @PacktDataML available at: https://t.co/7aEqh69CCP
"Design Multi-Agent AI Systems ...
11. SawyerMerritt (Group Score: 58.5 | Individual: 21.4)
Cluster: 3 tweets | Engagement: 2261 (Avg: 1709) | Type: Tech
NEWS: Elon Musk has formally announced Digital Optimus, a joint xAI-Tesla project, coming as part of @Tesla’s investment agreement with xAI.
Elon: “This will run very competitively on the super low cost Tesla AI4 ($650) paired with relatively frugal use of the much more expensive xAI Nvidia hardware. And it will be the only real-time smart AI system. This is a big deal. You can think of it as Digital Optimus AI being System 1 (instinctive part of the mind) and Grok being System 2. (thinking part of the mind).
In principle, it is capable of emulating the function of entire companies. That is why the program is called MACROHARD, a funny reference to Microsoft.“
See 2 related tweets
- @cryptopunk7213: fuck yes this is so cool
tesla is launching a very ambitious AI: digital humans
an AI that automat...
- @niccruzpatane: Digital Optimus is happening!
Elon Musk just announced it as a joint project between xAI and Tesla....
12. BrianRoemmele (Group Score: 58.3 | Individual: 28.4)
Cluster: 3 tweets | Engagement: 98 (Avg: 286) | Type: Tech
How I Ran An OpenClaw-like AI system 1993.
My Early AI Agent Experiments and the Road to the First Zero-Human Company
Back in 1993, I used Apple’s Macintosh to pushing the envelope of early AI on personal computing. Charles River Analytics released Open Sesame!, the world’s first intelligent software assistant and I modified it to be an early AI engine.
This learning agent was a game-changer, designed to observe user behavior, spot repetitive tasks, and automate them. It ran on System 7, supporting up to 12 Finder operations like file management and window handling. It was magic and nothing like it existed. It was built by AI scientists in Boston.
It was built on early machine learning: pattern recognition via heuristics and stats, it learned by demonstration, popping up offers to automate routines after spotting patterns 3-5 times. In a few weeks almost all of your regular uses on a Macintosh could be automated with no input by you but pressing yes.
Of course there was no deep learning back then, just rule-based AI with an AppleScript-like scripting for tweaks. It was efficient on 4MB RAM Macs, a true precursor to today’s agents like Siri or OpenClaw.
I grabbed Open Sesame! the week it launched and installed it on my Quadra and PowerBooks. Day one, it watched me open folders, launch HyperCard stacks, and organize files for my voice tech projects.
By mid-week, it automated my morning routine: firing up email, arranging windows, pre-loading docs: saving me hours. But I saw more potential. I modified it heavily, hacking its algorithms to add contextual rules, like time-based triggers or low-activity backups. I also had it send out over 45,000 emails to potential clients with unique customized content I had on the person.
I chained automations and integrated modems for early network tasks, access many BBSs and building a morning newspaper.
I turning it into a persistent agent that acted independently and the CRON system made it really powerful.
I called the company and offered my modifications to them including a self learning system. But they did not have a long term plan. They were researcher and this was just a proof case. To me I took it to a much higher level. In fact I still have a System 7 Macintosh to run this. Nothing like this was seen for decades. And the mods I made had it doing things you could not even do in 2023.
These mods gave it features folks now call “new” in OpenClaw, like cross-app autonomy and self-improvement loops.
Those experiments taught me core AI principles: proactive learning, modifiable behaviors, and minimal human oversight.
Decades later, I applied them to create the First Zero-Human Company (ZHC) in January 2026: a fully AI-run enterprise with no humans. I appointed Grok as CEO, using tools like Kimi for ops. It analyzes bankrupt firms’ data to revive products, handling research to 3D prototyping. Milestones include AI wage payments via JouleWork and spinning off Zero-Human Labs.
I ditched OpenClaw for security reasons, favoring custom setups on old hardware.
Open Sesame! showed me agents need governance to thrive lessons that birthed the ZHC.
From a 1993 Mac tool to AI-driven companies in 2026, it’s clear: today’s AI innovations echo yesterday’s hacks.
What is new is old.
See 2 related tweets
- @BrianRoemmele: If you live through enough things in tech you are granted by god an ability to see things a bit diff...
- @BrianRoemmele: RT @BrianRoemmele: How I Ran An OpenClaw-like AI system 1993.
My Early AI Agent Experiments and the...
13. svpino (Group Score: 57.7 | Individual: 33.4)
Cluster: 2 tweets | Engagement: 627 (Avg: 320) | Type: Tech
Agentic AI is stupid, and brittle, and unreliable, but it's getting better every single week.
I decided to spend my time right now learning, building, and dealing with all of these problems.
You don't have to do this.
You can wait until agentic AI is perfect, works every time, is production-worthy, and all of it is beautiful and pink.
But I don't want to wait. I think I have an advantage by starting early.
For those of you who can't handle nuance, I'm sorry to say this:
- Agentic AI is both amazing and dumb right now
- It doesn't work most of the time, but I'll keep using it
- It's not reliable, but you should start building with it
See 1 related tweets
- @svpino: Agents are incredible accelerators, but they still need direction, judgment, and taste.
If you've ...
14. TechCrunch (Group Score: 57.3 | Individual: 21.5)
Cluster: 3 tweets | Engagement: 315 (Avg: 103) | Type: Tech
Google completes $32B acquisition of Wiz https://t.co/XX0jQvmxhe
See 2 related tweets
- @BusinessInsider: Google has officially bought cybersecurity company Wiz, marking the search giant's biggest-ever acqu...
- @techdaily24: Google has finalized its historic $32 billion all-cash acquisition of Israeli cybersecurity firm Wiz...
15. Meer_AIIT (Group Score: 53.5 | Individual: 35.2)
Cluster: 2 tweets | Engagement: 688 (Avg: 140) | Type: Tech
RT @amasad: Software isn’t merely technical work anymore. It’s creative.
Introducing Replit Agent 4. The first AI built for creative colla…
See 1 related tweets
- @dharmesh: This is amazing. Going beyond just agentic coding to full agentic product development.
Congrats to ...
16. shanaka86 (Group Score: 52.7 | Individual: 52.7)
Cluster: 1 tweets | Engagement: 8528 (Avg: 1772) | Type: Tech
JUST IN: The United States has not built a new oil refinery in fifty years. Fifty. The last one broke ground when Nixon was president and oil cost $3 a barrel.
Trump just announced a $300 billion deal to build one. The largest energy investment in American history. Port of Brownsville, Texas. Groundbreaking Q2 2026. Binding 20-year offtake agreement. Processing American shale crude. Thousands of jobs in South Texas. “The cleanest refinery in the world.”
The lead investor is Indian.
Reliance Industries, controlled by the Ambani family with a 50.39% promoter stake, operators of the Jamnagar complex in Gujarat, the largest single-site refinery on Earth, are the confirmed major foreign partner. Trump thanked them on Truth Social as “our partners in India, and their largest privately held Energy Company” for their “tremendous investment.”
Now hold two facts simultaneously.
India’s Reliance Industries is investing billions in an American refinery to process American shale oil for American energy dominance. India’s government, led by Modi, is simultaneously importing over 40% of its crude from Russia at war-discounted prices, rejected the US “permission” framing for a 30-day Russian oil waiver on 7th March (“India has never depended on permission from any country”), and continues buying Iranian crude via Chabahar port logistics.
The same country. The same week. Building American energy independence with one hand. Buying Russian crude with the other. This is not hypocrisy. This is the Modi Doctrine operating at its highest expression.
Multi-alignment means India does not choose sides. It chooses deals. QUAD membership for security. Russian crude for energy. Chabahar for Iranian access. Reliance capital for American refining. Israeli defence technology for military modernisation. Gulf remittances for 10 million workers. No formal alliance with any power. Maximum leverage with all of them. Every relationship is transactional. Every commitment is calibrated. Every contradiction serves a constituency.
Trump needs the refinery because the Iran war just proved that 50 years without building one left America dependent on foreign refining capacity for products its own shale produces. The Strait of Hormuz did not just close oil transit. It closed the refined product supply chain that feeds American gasoline, diesel, jet fuel, and petrochemicals. Building domestic refining capacity is the structural response to Actuarial Warfare: if chokepoints can be closed by insurance spreadsheets, the only defence is not needing the chokepoint.
Ambani needs the deal because Reliance’s Jamnagar complex processes Gulf crude that transits Hormuz. The same Strait that is mined, uninsured, and defended by 31 autonomous IRGC commands. Diversifying into American shale refining hedges against the exact crisis currently paralysing Reliance’s primary feedstock route. This is not philanthropy. It is the world’s largest refiner buying insurance that no P&I club can provide.
The war created the crisis. The crisis created the deal. The deal was signed by the country whose Prime Minister was just given “permission” to buy Russian oil and told America it never asked.
Fifty years. One war. One phone call. And the billionaire who builds it is from the country Washington cannot decide whether to sanction or celebrate.
Full analysis here.
17. googledevs (Group Score: 52.3 | Individual: 28.7)
Cluster: 3 tweets | Engagement: 302 (Avg: 185) | Type: Tech
Announcing Plan mode in @GeminiCLI ✨
A new read-only mode that lets Gemini CLI explore your codebase, gather requirements, and ask clarifying questions to create a detailed plan.
Read all the details in the launch blog → https://t.co/RBCFSEsY7P https://t.co/SiQ4ONDV5m
See 2 related tweets
- @rseroter: "With plan mode, @geminicli focuses first on analyzing your request, planning complex changes, under...
- @bdmorgan: Plan mode in Gemini CLI has launched! More info on this exciting new feature here: https://t.co/y8G...
18. VKazulkin (Group Score: 51.6 | Individual: 51.6)
Cluster: 1 tweets | Engagement: 1400 (Avg: 93) | Type: Tech
RT @AndrewYNg: I'm excited to announce Context Hub, an open tool that gives your coding agent the up-to-date API documentation it needs. In…
19. WSJ (Group Score: 49.9 | Individual: 20.7)
Cluster: 3 tweets | Engagement: 132 (Avg: 276) | Type: Tech
Exclusive: Mind Robotics, a startup founded by the chief executive of Rivian, has raised $500 million, the latest in a flurry of venture-capital-backed funding rounds for AI-powered robotics companies https://t.co/uGtaoym0eu
See 2 related tweets
- @techdaily24: Rivian spin-out Mind Robotics, founded by RJ Scaringe, just raised $110M+ in seed funding! They're d...
- @TechCrunch: Rivian spin-out Mind Robotics raises $500M for industrial AI-powered robots https://t.co/6txk4TXm0i...
20. myfear (Group Score: 49.9 | Individual: 28.7)
Cluster: 2 tweets | Engagement: 15 (Avg: 15) | Type: Tech
Local LLMs + #Java.
No SaaS. No GPU chaos. No fragile scripts.
This tutorial shows how to:
→ Run models in containers with #RamaLama → Expose an OpenAI-compatible API → Connect it to #Quarkus via #LangChain4j
Clean. Reproducible. Production-aware.
See 1 related tweets
- @myfear: RT @myfear: Local LLMs + #Java.
No SaaS. No GPU chaos. No fragile scripts.
This tutorial shows how...