科技推特精选 - 2026年1月27日

今日科技要闻：AI 基础设施领域迎来爆发式增长，CoreWeave 扩大了与 NVIDIA 的合作伙伴关系，计划到 2030 年建成 5GW 的算力容量。产品创新方面，NVIDIA 发布了 Earth-2 开源气象模型，OpenAI 则悄然升级 ChatGPT 的代码解释器以支持外部包安装。初创生态领域，MiniMax 推出全新的智能体工作流工具，Locally Translate 发布了针对 Apple Silicon 的端侧翻译功能。与此同时，监管压力日益增大，欧盟正就 Grok 聊天机器人涉及的深度伪造风险对 X 平台展开调查。

1. CoreWeave (Group Score: 82.0 | Individual: 27.8)

Cluster: 4 tweets | Engagement: 71 (Avg: 46) | Type: Tech

We’re expanding our partnership with NVIDIA to build 5GW+ of AI compute capacity by 2030. By combining world-class hardware with CoreWeave Mission Control and SUNK, we’re building the foundation for the next wave of global innovation.

2. business (Group Score: 81.0 | Individual: 18.2)

Cluster: 5 tweets | Engagement: 78 (Avg: 60) | Type: Tech

Elon Musk’s X has been hit by a EU probe over concerns it failed to prevent its Grok AI chatbot from spewing out deepfake images “that may amount to child sexual abuse material” https://t.co/4wjzgGxZEN

See 4 related tweets

@Reuters: The European Commission launched an investigation into Elon Musk's AI chatbot Grok over concerns it ...
@business: The EU launched a probe into Elon Musk’s X over concerns it failed to prevent its Grok AI chatbot fr...
@WSJ: The EU launched an investigation into Elon Musk’s X following a public outcry over sexualized deepfa...
@BusinessInsider: The European Union is launching an investigation into Grok as Elon Musk's AI chatbot continues to fa...

3. MiniMax_AI (Group Score: 67.9 | Individual: 43.5)

Cluster: 2 tweets | Engagement: 1278 (Avg: 223) | Type: Tech

MiniMax Agent = Claude cowork + Clawdbot

Lives on your computer AND in the cloud
Agentic + batch workflows
Full browser control
Agent skills library
Expert agents: customized & shared
Runs on macOS & Windows

Upgrade your tools and mindset. Everyone is an Agent Designer. https://t.co/XzaTmAos0V

See 1 related tweets

@omarsar0: I've been playing around with MiniMax Agent.

It's really good.

I've been very impressed with the M...

4. gdgtify (Group Score: 67.0 | Individual: 33.7)

Cluster: 2 tweets | Engagement: 62 (Avg: 138) | Type: Tech

Another fun Nano Banana prompt for scientists.

Prompt:

Nikola Tesla: Input Variable: [INSERT FAMOUS FIGURE]

System Instruction:
Generate a hyper-realistic, 2x2 Grid of Macro Desktop Dioramas . Core idea: "Living Documents." The 3D scenes must emerge physically from 2D paper artifacts lying on a vintage mahogany desk.

4 Panels
Analyze the Input Variable to generate 4 distinct visual chapters based on their history:
Panel 1 Theme: The Rivalry / The Battle of Ideas. Ground: An open Technical Blueprint or City Map lying flat. Action: Two figures (The Protagonist vs. The Rival) stand on opposite sides of the paper. Connection: A visual representation of their conflict connects them (e.g., Electrical Arcs, a Tug-of-War rope, intersecting flight paths). Labels: Blue plastic nameplates stand in front of each figure (e.g., "AC POWER" vs "DC POWER").
Panel 2 Theme: The Magnum Opus / The Vertical Invention. Ground: A Wooden Radio Cabinet or Museum Display Case standing vertically on the desk. Action: The subject's most famous tall invention (e.g., A Tower, A Rocket, A Statue) is built inside the cabinet, glowing with internal power. Observer: A tiny figure of the subject stands outside the box, reading a scroll/clipboard, admiring the work.
Panel 3
Theme: The Myth / The Destructive Experiment. Ground: A thick, Red/Brown Hardcover Book lying flat. Action: A miniature Workbench sits on top of the book. The subject sits at the desk operating a device. Chaos: Next to the desk, a miniature building or object is Crumbling/Exploding (e.g., Bricks falling, smoke rising) due to the device's power. Prop: A large real-world Magnifying Glass rests on the book edge for scale.
Panel 4 Theme: The Weapon / The Future Application. Ground: An open Scientific Journal or Patent Text . Action: A weaponized version of their invention is mounted on the page. Target: The weapon is shooting a Beam or Projectile at a floating aerial target (e.g., A Blimp, A Drone, A Bird) hovering above the book.
2D to 3D :
Ink-to-Matter: The diagrams drawn on the paper floor must physically rise up to become the 3D metal/wood of the machines. Debris: Dust, sparks, and brick rubble must litter the paper surfaces to sell the realism.
Scale Indicators:
Clutter: Scattered around the edges of the diorama "zones" are real-world desk objects to prove the miniature scale: Rolled-up Blueprint Tubes. Fountain Pens. Vintage Newspapers with readable headlines about the event.
Lighting & Atmosphere:
Setting: A dark, rich Mahogany Library Desk . Lighting: Warm, cozy "Study Lamp" lighting. Highlights on the brass instruments and leather book bindings. Atmosphere: Tiny wisps of smoke or electrical ozone haze in the air.

Output: 2x2 Grid, 1:1 Aspect Ratio, Octane Render, "Tilt-Shift" Macro Photography, 8k Resolution.

See 1 related tweets

@gdgtify: Modified my prompt for food and other products. Nano Banana Pro vs. Nano Banana

Prompt: Input Var...

5. code (Group Score: 62.6 | Individual: 26.3)

Cluster: 3 tweets | Engagement: 98 (Avg: 145) | Type: Tech

🔥 We're going live in 60 minutes! Join @JamesMontemagno @digitarald and Den Delimarsky as they demo MCP Apps, the first official MCP extension in @code ▶️https://t.co/swkGJHeAYz #AnthropicAI #mcp https://t.co/qg5wCt8OB9

See 2 related tweets

@burkeholland: RT @code: 🔥 We're going live in 60 minutes! Join @JamesMontemagno @digitarald and Den Delimarsky as ...
@JamesMontemagno: MCP Apps! UI from your MCP servers are here! And inside @code!

https://t.co/UtSv3cPPkh

Come hang ...

6. aakashgupta (Group Score: 53.5 | Individual: 53.5)

Cluster: 1 tweets | Engagement: 584 (Avg: 90) | Type: Tech

Anthropic employs world-class engineers who could build an HR system in weeks. They use Workday anyway. The reason tells you exactly where enterprise SaaS is headed.

Building HR software requires knowing labor law across 50 states and 100+ countries. Payroll tax compliance changes quarterly. Healthcare benefit structures shift annually. One classification error creates seven-figure liability.

No engineering team wants to own that surface area. The maintenance burden compounds forever while delivering zero competitive advantage.

This is why enterprise SaaS moats actually strengthen with AI. The value was never “we built software you couldn’t.” The value was always “we absorb compliance risk and regulatory complexity you don’t want.”

AI makes custom software cheaper to build. It doesn’t make compliance cheaper to own. Workday’s real product is liability absorption, and that product just got more valuable as build-vs-buy calculations everywhere else shift.

The companies getting disrupted are the ones selling capability. The ones selling risk transfer are about to have their best decade.

7. MarioNawfal (Group Score: 53.5 | Individual: 34.2)

Cluster: 2 tweets | Engagement: 847 (Avg: 931) | Type: Tech

🇪🇺 OPINION: EU DITCHING U.S. TECH IS A NIGHTMARE OF PROTECTIONIST STUPIDITY

Oh boy, the EU's at it again... cooking up resolutions to ditch American tech giants like AWS, Azure, and Google Cloud for some homegrown "Eurostack."

The only issue? Nobody's heard of that sh*t, but it's all in the name of "digital sovereignty."

It's one of these new technocratic jingles that Lula and Maduro were already using to silence dissent in South America.

This is peak bureaucratic idiocy: governments forcing markets to "buy local" when the free choice has already spoken.

U.S. firms dominate because they're better, cheaper, and innovate without a nanny state holding their hand.

Think about it, instead of letting entrepreneurs compete on merit, the EU's basically admitting their local options can't hack it without mandates and subsidies.

Sorry, they have none! What I mean in this case is the hypothetical idea of global companies that could hold against the U.S.'s digital dominance.

They're trying to spank companies out of a sheet of paper like it's alchemy

That's cronyism on steroids, propping up inferior products with taxpayer cash while stifling real progress.

Innovation thrives in voluntary exchanges, not Brussels' red tape labyrinth.

This'll just drive up costs, slow down EU tech, and leave users stuck with second-rate services (and incidentally, make VPN companies a lot of money).

The ugly dinosaur of statism reared its ugly head: Wake up, Europe, free markets or bust!

Source: EU Parliament Resolution, @GlobeObserver

See 1 related tweets

@rohanpaul_ai: 🚨🇪🇺 The EU is taking steps to replace U.S. tech companies with regional options.

Europe is starting...

8. nvidianewsroom (Group Score: 52.2 | Individual: 36.5)

Cluster: 2 tweets | Engagement: 611 (Avg: 176) | Type: Tech

🌍 Weather forecasting has always relied on powerful supercomputers running physics-based models.

We are proud to announce the NVIDIA Earth-2 family of open models — the world’s first fully open, accelerated AI weather stack — saving computational time and costs, and enabling more nations, enterprises, and businesses to run application-specific forecasting systems.

Weather AI is now accessible worldwide at every stage. #AMS2026

9. Miles_Brundage (Group Score: 52.0 | Individual: 52.0)

Cluster: 1 tweets | Engagement: 46008 (Avg: 1272) | Type: Tech

RT @ExtremeBlitz__: in a world full of AI slop edits, we need to start appreciating hard working editors that take the time to make banger…

10. simonw (Group Score: 51.6 | Individual: 29.0)

Cluster: 2 tweets | Engagement: 676 (Avg: 541) | Type: Tech

OpenAI shipped a HUGE upgrade to ChatGPT Code Interpreter and failed to document it (even in the release notes) - but ChatGPT can now pip/npm install packages and run code in Python, Node.js, Bash, Ruby, Perl, PHP, Go, Java, Swift, Kotlin, C and C++! https://t.co/El5osALZGx

See 1 related tweets

@xeophon: RT @simonw: OpenAI shipped a HUGE upgrade to ChatGPT Code Interpreter and failed to document it (eve...

11. rohanpaul_ai (Group Score: 48.3 | Individual: 29.0)

Cluster: 2 tweets | Engagement: 51 (Avg: 95) | Type: Tech

Usual “AI detector” tools are easier to fool than people think.

People use detectors to catch fake reviews, fake school work, or fake news, but those detectors often learn simple clues like a certain stiff writing style, and then they break when the AI writes in a more human way.

The authors of this paper build a new benchmark called MAGA-Bench, which is basically a big set of human text plus AI text that has been polished to look human, across many types of writing like forums, reviews, questions and answers, and news.

To make the AI text more human-like, they do 4 things: persona prompting, automatic prompt improvement, self critique and rewrite, and detector feedback as a training score so the AI learns what gets flagged and avoids it.

When they test existing detectors on this harder data, detector quality drops by about 8.13% on the main separation score, meaning the detectors miss more AI text.

Also they found training a detector on this harder data can make it better on new, real world style data, with about 4.60% average gain on that same separation score.

Paper Link – arxiv. org/abs/2601.04633

Paper Title: "MAGA-Bench: Machine-Augment-Generated Text via Alignment Detection Benchmark"

See 1 related tweets

@rohanpaul_ai: Another paper showing just how fragile AI detectors are.

They report 92% Attack Success Rate, meani...

12. MiniMax_AI (Group Score: 48.0 | Individual: 31.1)

Cluster: 2 tweets | Engagement: 1193 (Avg: 223) | Type: Tech

You don’t need a mac mini to escape the permanent underclass.

$10 MiniMax coding plan.$ 5 Railway hobby plan.

@clawdbot + @MiniMax_AI = 24/7 personal intelligence. https://t.co/LHqllPJLDl

See 1 related tweets

@KarthiDreamr: RT @MiniMax_AI: You don’t need a mac mini to escape the permanent underclass.

$10 MiniMax coding pl...

13. LocallyAIApp (Group Score: 47.7 | Individual: 30.0)

Cluster: 2 tweets | Engagement: 349 (Avg: 175) | Type: Tech

Introducing our new app: Locally Translate — instant, on-device translation in 50+ languages.

Available now on the App Store for iPhone 15 Pro and newer. https://t.co/uRfoKu9TsZ

See 1 related tweets

@awnihannun: RT @LocallyAIApp: Introducing our new app: Locally Translate — instant, on-device translation in 50+...

14. HiTw93 (Group Score: 47.5 | Individual: 34.8)

Cluster: 2 tweets | Engagement: 917 (Avg: 144) | Type: Tech

我最近很喜欢使用 NotebookLM 来帮助学习理解新东西，但只能导出 PDF，有时候对于中文的渲染也不是很好。一直想找一个好用的能把 PDF 转成可编辑 PPT 的工具，找一圈下来发现 Codia AI NoteSlide 做得最好用，相比传统的 pdf2ppt 简洁易用很多，可快速转化成可编辑的内容，效果很美观精致无差错，很好用，有相同痛点的小伙伴可以去试试看。 https://t.co/FhXZjcWblv

See 1 related tweets

@HiTw93: RT @HiTw93: 我最近很喜欢使用 NotebookLM 来帮助学习理解新东西，但只能导出 PDF，有时候对于中文的渲染也不是很好。一直想找一个好用的能把 PDF 转成可编辑 PPT 的工具，找...

15. aakashgupta (Group Score: 45.8 | Individual: 39.7)

Cluster: 2 tweets | Engagement: 229 (Avg: 90) | Type: Tech

The story of Synthesia is insane.

2017: Two 25-year-olds pitch AI video generation to VCs. Everyone thinks they’re crazy.

100 European VCs turn them down. The technology barely works. One co-founder had been running sand mines and chicken farms in Africa before this. The other didn’t have a PhD or Google on his resume.

They’re about to give up.

One night, Steffen Tjerrild finds Mark Cuban’s email address in the leaked Sony hack files. Sends a cold email at 11pm UK time.

Cuban responds in 5 minutes.

14 hours of back-and-forth emails later. No calls. Just emails. At 4am UK time, Cuban wires $1 million.

That $1M bought them time to stop building what they thought they should build (AI dubbing) and start building what enterprises actually wanted (AI training videos).

The next 7 years:

$3.1M seed (2019)$ 12.5M Series A (2021) $50M Series B (2021)$ 90M Series C at $1B (2023)$ 180M Series D at $2.1B (Jan 2025)$ 200M Series E at $4B (today)

Total raised: $536M

Adobe invested, then tried to buy them for $3B. Synthesia said no. Meta tried too. Also no.

60,000+ customers. 70% of the Fortune 100. $150M ARR. On track for$ 200M this year.

The company that 100 investors thought was “completely crazy” just became Europe’s most valuable AI media startup.

See 1 related tweets

@synthesiaIO: RT @thealexbanks: Synthesia is now one of the most valuable AI companies in Britain.

This transform...

16. 0xdevshah (Group Score: 42.9 | Individual: 31.8)

Cluster: 2 tweets | Engagement: 36 (Avg: 21) | Type: Tech

the narrative that we don't have enough energy on earth to run agi-scale systems continuously is probably true for the current architectures. but what is also true is that the neocortex does mass general reasoning, planning and world modelling at just 20 watts/day. either there is a 10000x efficiency gain hiding somewhere in the algorithms, or biological compute is exploiting physics in ways we don't know yet. k0 civilization might not solve this, but k1 definitely will. the plausibility is that this bitter lesson will have a sweet corollary.

See 1 related tweets

@0xdevshah: RT @0xDevShah: the narrative that we don't have enough energy on earth to run agi-scale systems cont...

17. badlogicgames (Group Score: 41.7 | Individual: 41.7)

Cluster: 1 tweets | Engagement: 2168 (Avg: 60) | Type: Tech

RT @karpathy: A few random notes from claude coding quite a bit last few weeks.

Coding workflow. Given the latest lift in LLM coding capab…

18. business (Group Score: 40.2 | Individual: 20.6)

Cluster: 2 tweets | Engagement: 56 (Avg: 60) | Type: Tech

A year ago, the Chinese startup DeepSeek freaked out the stock market with the idea that developing AI was much easier and cheaper than everyone imagined. But that’s turned out to be largely a mirage. https://t.co/1BDW4jNKPr

See 1 related tweets

@TheEconomist: In the year since DeepSeek shocked the world with a whizzy new AI model, China’s clout in the tech h...

19. rohanpaul_ai (Group Score: 37.6 | Individual: 37.6)

Cluster: 1 tweets | Engagement: 146 (Avg: 95) | Type: Tech

The paper shows why LLM agents fail, then uses that failure map to guide strict answer verification.

Deep research agents, meaning LLMs that browse and use tools, often fail because 1 early mistake can snowball across many steps.

The authors use 1 idea: checking an answer is often easier than producing it, so the agent should spend extra effort verifying.

DeepVerifier turns that into a workflow that flags likely failure types, asks up to 3 targeted follow up questions, then gives written feedback for a retry.

They first study common agent mistakes to build the checklist, then test the verifier and retry loop on GAIA and other hard benchmarks.

DeepVerifier is a stronger checker than common LLM judges, and its feedback loop raises agent accuracy by about 8% to 11% on hard tasks.

Because it works at inference time, meaning while the agent is answering, it can improve without extra training, and DeepVerifier-4K with 4,646 examples trains open models to self check.

Paper Link – arxiv. org/abs/2601.15808

Paper Title: "Inference-Time Scaling of Verification: Self-Evolving Deep Research Agents via Test-Time Rubric-Guided Verification"

20. BrianRoemmele (Group Score: 37.5 | Individual: 37.5)

Cluster: 1 tweets | Engagement: 189 (Avg: 295) | Type: Tech

To the $ZHC community,

First off, I want to absolutely thank you all for this support. I had no idea what this was, and still learning.

You Bagged me.

I believe we are at a precipice of a massive change within our society. I’ve been writing about this for a long time and the Zero-Human Company experiment. Is just the first of what will be millions of AI and robots that will operate with no need of a key human employee.

It is no longer science fiction: the world just woke up, well the 1/2 of 1% like you folks that are on the very frontier, they woke up to Cawdbot as a manifestation of an employee. I have written how YOU can have your own employee(s). DO IT.

This is only one part of my experiments. I do a lot of things and can’t drop them either. This project is new and takes a bit more of my energies. Today it is a process to understand where this all goes. But the goal is to continue the research and understand the best paths. The company is iterating on its own rapidly and 100s of paths are being tried. The basis for this today is the old data from a company that went bankrupt and finding out if we can go to market with something, but there are no guarantees on anything here.

This is a project today, it may always be, but it may become something absolutely amazing. We are making history here folks. You are here seeing it. And I want you to build too. So you make history. I want to support you. This is YOUR time, this is YOUR moment to build.

I SUPPORT YOU AND WHAT YOU WILL BUILD.

I think https://t.co/WK9MzIYHjq has the prospects to do great good for everyone involved. It will need to cool down the anxiousness of some. I get it, but understand many have not asked to be Bagged, and frankly have absolutely no idea what it all means. It can and will scare away some, honestly it could scare away me. There needs to be clear education to all involved. Remember I and many did not ask for this but will do their best. But the anxiousness can be a big detriment to this being a major transformative force… far too many as you know have been burned and have burned themselves on many new ideas. Don’t make https://t.co/WK9MzIYHjq be one of them. And demands may not be possible for some no matter how simple they may seem to outsiders.

Time in research, is not minutes or hours, but maybe today days or weeks. It could be months. But it takes time and maybe out of garages comes Apple or a lawnmower.

So this is the way of it. Here is what I do know. I love you my sisters and brothers who want to see my work proceed, no matter what the reason is.

Thank you.

Deep gratitude.

∞Brian