科技推特精选 - 2026-02-18

今日科技快讯：人工智能军备竞赛愈演愈烈，Anthropic 推出 Claude Sonnet 4.6，xAI 则发布了搭载创新四智能体系统的 Grok 4.20。印度在举办汇聚行业巨头的高规格峰会之际，进一步巩固其全球枢纽地位，阿达尼集团（Adani Group）承诺投入 1000 亿美元用于数据中心基础设施建设。与此同时，Render 获得 1 亿美元融资用于云端扩展，特斯拉 Cybercab 接近量产。然而，硬件限制因素凸显，存储芯片短缺威胁供应商，且针对 AI 生成内容的监管审查也正日益严峻。

1. seconds_0 (Group Score: 220.4 | Individual: 43.2)

Cluster: 9 tweets | Engagement: 3076 (Avg: 203) | Type: Tech

RT @claudeai: This is Claude Sonnet 4.6: our most capable Sonnet model yet.

It’s a full upgrade across coding, computer use, long-context…

See 8 related tweets

@yupp_ai: 📢 New Model Drop: Claude Sonnet 4.6 is now live on Yupp!

Congratulations to @AnthropicAI on their l...

@mark_k: Claude Sonnet 4.6 Performance Highlights! 📊

The newly released Sonnet 4.6 demonstrates impressive c...

@dejavucoder: since sonnet 4.6 is here and claude code is rapidly improving day by day, i am once again reminding ...
@nummanali: Knowledge cut off for Sonnet 4.6 is August 2025!

So many updates to features and tooling came out ...

@testingcatalog: BREAKING 🚨: CLAUDE SONNET 4.6 COMES WITH 1M CONTEXT WINDOW IN BETA!

"Claude Sonnet 4.6 is our most ...

2. business (Group Score: 125.2 | Individual: 33.8)

Cluster: 6 tweets | Engagement: 528 (Avg: 91) | Type: Tech

Adani Group plans to invest $100 billion by 2035 to develop data centers, as billionaire Gautam Adani seeks to capitalize on India's emergence as an AI hub https://t.co/8loThjeNbj

See 5 related tweets

@TechCrunch: Adani pledges $100B to build AI data centers as India seeks bigger role in the global AI race https:...
@Reuters: India's Adani to invest $100 billion in AI-ready data centres by 2035 https://t.co/TLY1GsBtUF https:...
@Forbes: Billionaire Gautam Adani Earmarks $100 Billion To Develop AI Data Centers Across India https://t.co/...
@Reuters: Adani bets $100 billion on data centres to power India’s AI ambitions https://t.co/TF2liSrrvn https:...
@Cointelegraph: 🇮🇳 LATEST: Adani Group plans to invest $100B to build sovereign AI data centers across India. https:...

3. aakashgupta (Group Score: 67.7 | Individual: 31.1)

Cluster: 3 tweets | Engagement: 59 (Avg: 424) | Type: Tech

Meta’s CTO personally authored a patent for AI that keeps dead people’s accounts posting, liking, and replying to DMs. The company says they have “no plans to move forward.” Sure.

The patent describes training an LLM on your entire behavioral history: posts, comments, likes, voice messages, DMs. The clone doesn’t just post. It likes other people’s content. It replies to direct messages. It can simulate video and audio calls. The filing was submitted in 2023 and granted December 2025.

Facebook has 3.07 billion monthly active users. Roughly 30 million of them die every year. At a global average revenue of $13.12 per user, that’s ~$ 400 million in annual engagement capacity that goes permanently dark. The dead user base compounds every year. A decade from now, hundreds of millions of accounts will be silent unless something keeps them active.

The patent says the quiet part in clinical language: the impact of a user leaving is “much more severe and permanent if that user is deceased.” Translation: death is a churn problem, and churn hurts ad revenue.

Meta already removes over 1 billion fake accounts per quarter to prove their engagement numbers are real humans. Now they’re patenting technology that generates the exact same engagement signals, likes, comments, DM replies, from AI clones indistinguishable from living users in the metrics. Advertisers pay per impression, per click, per interaction. The clone produces all three.

This tells you everything about what Meta actually built. Users spent 20 years uploading their personalities for free. Meta stored every post, every comment, every reaction, every voice message. The user created the training data. The user maintains the account. And after the user dies, Meta can keep extracting ad revenue from the digital replica indefinitely.

They wrapped a perpetual engagement machine in the language of helping people grieve, because that’s the only version of this story anyone would opt into.

See 2 related tweets

@MarioNawfal: 🚨🇺🇸 META PATENTS AI TO KEEP DEAD USERS POSTING ON SOCIAL MEDIA

Meta just got approval for technolog...

@Cointelegraph: ⚡️ JUST IN: Meta secures patent for AI technology that mimics deceased users' social media activity ...

4. business (Group Score: 67.5 | Individual: 49.5)

Cluster: 2 tweets | Engagement: 941 (Avg: 91) | Type: Tech

Pichai. Altman. Amodei. All in New Delhi. India is hosting a high profile AI summit and for PM Narendra Modi, it’s a high-stakes moment to strengthen the country's AI foothold.

@haslindatv explains: https://t.co/a8wDoAjPcw https://t.co/iWuAZ9DkIm

See 1 related tweets

@business: India’s two largest conglomerates are using the ongoing India AI summit as a recruiting ground to sc...

5. rohanpaul_ai (Group Score: 59.9 | Individual: 33.5)

Cluster: 3 tweets | Engagement: 47 (Avg: 110) | Type: Tech

AI driven memory shortages and price jumps will push many consumer electronics vendors into bankruptcy or force them to exit product lines by end of 2026.

~ Phison CEO Pua Khein-Seng

Suppliers or memory are demanding 3 years of prepayment, which would break cash flow for smaller brands that normally rely on short credit cycles.

e.g. shipping tens of millions of NVIDIA “Vera Rubin” systems, each needing over 20TB of SSD, could consume about 20% of last year’s global NAND capacity.

He argues that AI hardware is soaking up so much DRAM and NAND that parts supply, and pricing power, is starting to squeeze ordinary device makers out of the market.

smartphone output could fall by 200-250M units, with PCs and TVs also cut back as vendors cannot secure enough memory.

pcgamer. com/hardware/memory/many-consumer-electronics-manufacturers-will-go-bankrupt-or-exit-product-lines-by-the-end-of-2026-due-to-the-ai-memory-crisis-phison-ceo-reportedly-says/

See 2 related tweets

@business: The surge in AI-related demand for DRAM — or Dynamic Random Access Memory — is tightening supply and...
@BusinessInsider: Memory chip costs are up as AI companies gobble up the hot commodity. Consumer electronics companies...

6. ibuildthecloud (Group Score: 58.8 | Individual: 41.5)

Cluster: 2 tweets | Engagement: 15562 (Avg: 342) | Type: Tech

RT @thedorbrothers: We just made a $200,000,000 AI movie in just one day.

Yes, this is 100% AI. https://t.co/TMotM7rguY

See 1 related tweets

@EHuanglu: AI now can make $100M film in one day

https://t.co/x7C9OKLsyf...

7. dakshgup (Group Score: 58.3 | Individual: 58.3)

Cluster: 1 tweets | Engagement: 983 (Avg: 72) | Type: Tech

We’ve raised $100M at a$ 1.5B valuation.

We built Render to give developers an intuitive path to reliable and scalable cloud infrastructure.

Now, we are bringing that same philosophy to long-running, stateful infrastructure for AI apps & agents.

🧵 https://t.co/lwb4SOahtg

8. teortaxesTex (Group Score: 56.9 | Individual: 34.7)

Cluster: 2 tweets | Engagement: 203 (Avg: 60) | Type: Tech

I suspect that overall 2026 might be bad for Chinaboos For one thing, the US will get really far ahead in AI with all the compute coming online, and it will start to deliver real value. That'll make the usual arguments here about skilled labor and mafs look pretty feeble.

See 1 related tweets

@teortaxesTex: I suspect that overall 2026 might be be bad for Chinaboos For one thing, the US will get really far ...

9. minchoi (Group Score: 56.9 | Individual: 38.3)

Cluster: 2 tweets | Engagement: 1632 (Avg: 422) | Type: Tech

BREAKING: xAI just launched Grok 4.20.

It's not one AI. It's four.

xAI built a "4 Agents" system - four specialized AI agents that think in parallel and debate each other in real-time before giving you an answer.

What's new:

4-agent collaboration (think → debate → consensus) 256K context window (up to 2M) Native multimodal (text + image + video) Trained on 200K GPUs (Colossus supercluster) Only AI that was profitable in live trading competitions

Elon said it's "starting to correctly answer open-ended engineering questions." 🤯

See 1 related tweets

@minchoi: RT @minchoi: BREAKING: xAI just launched Grok 4.20.

It's not one AI. It's four.

xAI built a "4 Age...

10. SawyerMerritt (Group Score: 54.5 | Individual: 21.6)

Cluster: 3 tweets | Engagement: 3154 (Avg: 1473) | Type: Tech

NEWS: Tesla has shared a new photo of the first Cybercab off the Giga Texas production line.

Volume production is on track to start in April. Just 6-9 weeks away! https://t.co/GolxYm1Yqo

See 2 related tweets

@Tesla: First Cybercab off the production line at Giga Texas https://t.co/kY8vCqtrCA...
@niccruzpatane: The FIRST Production Tesla Cybercab has rolled off the line at Giga Texas!

This is a historical mo...

11. business (Group Score: 54.5 | Individual: 31.3)

Cluster: 2 tweets | Engagement: 466 (Avg: 91) | Type: Tech

The US will deploy more missile systems to the Philippines, a move that will potentially anger China, as the two allies boost their defense ties https://t.co/TlDLtYR8TU

See 1 related tweets

@WSJ: The U.S. plans to deploy more advanced missile systems and other weapons in the Philippines to deter...

12. cryptopunk7213 (Group Score: 54.0 | Individual: 30.7)

Cluster: 2 tweets | Engagement: 3860 (Avg: 872) | Type: Tech

Apple's genius move to win AI was... do nothing.

didn't spend $100Bs on ai capex = saved$ 100Bs while their competitors burned cash in a race to zero.

laughs in free cash flow and positive earnings

pick the winning model (in this case gemini) and build the greatest AI app and hardware ecosystem the world has ever seen

acquire whoever the fck you want because you have cash and your competitors don't

ai siri, ai airpods and revamped ai-compatible iOS on the way this year

still the 2nd most valuable company in the world and NONE of the risk taken.

as a former hater, im impressed.

See 1 related tweets

@cryptopunk7213: lol Apple’s about to show everyone who the fucking boss is
2.5 BILLION active apple devices are a...

13. Reuters (Group Score: 52.4 | Individual: 18.9)

Cluster: 3 tweets | Engagement: 179 (Avg: 132) | Type: Tech

The Spanish government has ordered prosecutors to investigate platforms X, Meta and TikTok for allegedly spreading AI-generated child sexual abuse material, Prime Minister Pedro Sanchez said https://t.co/TgoLEuLCKK https://t.co/TTqcmV4H9P

See 2 related tweets

@Reuters: Spain to probe X, Meta, TikTok over AI-generated child sexual abuse material https://t.co/KhHNUb261x...
@WSJ: The Spanish government is requesting that prosecutors investigate X, Meta Platforms and TikTok, sayi...

14. aakashgupta (Group Score: 50.7 | Individual: 50.7)

Cluster: 1 tweets | Engagement: 2372 (Avg: 424) | Type: Tech

Anthropic buried the real announcement in a demo.

A small business owner typing "go to my web store and make delivery free for orders over $25 this holiday weekend" and the model just... does it. Finds the store, reads the pricing logic, updates it, confirms. No code. No developer. No instructions beyond what you'd text a coworker.

That's the entire SMB automation market getting disrupted in a 10-second clip.

The math makes it alarming. Sonnet runs at roughly $3 per million tokens. A full agentic task, navigate to a site, read the page state, make decisions, execute changes, burns maybe 50K-100K tokens. That's under$ 0.30 per autonomous business operation. The question for every small business owner stops being "can I afford AI" and starts being "why am I still doing this manually."

The 1M token context window in beta is the second unlock. Long-context was the wall keeping Sonnet out of serious agentic deployments. You need the model to hold an entire store's product catalog, pricing history, and current task in working memory simultaneously. That wall just moved.

Everyone will screenshot the benchmark improvements. The real trade sits one layer below: autonomous business operations just got cheap enough for a sole proprietor running a Shopify store.

33 million small businesses in the US alone. All of them have tasks exactly like this one.

15. nvidianewsroom (Group Score: 48.3 | Individual: 21.7)

Cluster: 3 tweets | Engagement: 538 (Avg: 538) | Type: Tech

.@Meta is deploying AI at scale through deep codesign across the full NVIDIA platform.

This expanded partnership will enable the large-scale deployment of NVIDIA CPUs and millions of NVIDIA Blackwell and Rubin GPUs.

Learn more: https://t.co/UTNEkXGN9V https://t.co/jcIm63Loar

See 2 related tweets

@business: Meta has agreed to deploy “millions” of Nvidia processors over the next few years, tightening an alr...
@nvidia: RT @nvidianewsroom: .@Meta is deploying AI at scale through deep codesign across the full NVIDIA pla...

16. _akhaliq (Group Score: 46.1 | Individual: 15.8)

Cluster: 5 tweets | Engagement: 52 (Avg: 92) | Type: Tech

Tiny Aya is out on Hugging Face

a family of massively multilingual small language models

Tiny Aya delivers strong multilingual performance in 70+ global languages in a 3.35B parameter model https://t.co/kjJzzyl8tD

See 4 related tweets

@Cohere_Labs: RT @_akhaliq: Tiny Aya is out on Hugging Face

a family of massively multilingual small language mod...

@Cohere_Labs: RT @techdaily24: Cohere launches Tiny Aya - open-weight multilingual AI models supporting 70+ langua...
@_akhaliq: RT @Gradio: Cohere just dropped Tiny Aya, with a Day-Zero app on @huggingface! A 3.35B open-weight m...
@Cohere_Labs: RT @BetaKit: .@cohere’s new multilingual #AI model Tiny Aya supports more than 70 languages. And the...

17. blackboxai (Group Score: 43.3 | Individual: 26.8)

Cluster: 2 tweets | Engagement: 76 (Avg: 34) | Type: Tech

New release:

Call blackbox/ claude code/ codex / gemini all coding agents and talk with them to send tasks on your repos https://t.co/VtOdOguNd4

See 1 related tweets

@blackboxai: RT @blackboxai: New release:

Call blackbox/ claude code/ codex / gemini all coding agents and talk ...

18. svpino (Group Score: 43.2 | Individual: 26.6)

Cluster: 2 tweets | Engagement: 70 (Avg: 261) | Type: Tech

This is how you ensure Agentic coding doesn't destroy your organization:

Process your entire codebase and pull requests
Distill that knowledge into customized rules and patterns
Automatically track how you follow those over time

Agentic coding is the greatest thing since sliced bread, but only if you use it responsibly, maintain your codebase's strengths, and make sure your code doesn't turn ugly over time.

I didn't come up with this idea, by the way. @QodoAI just shipped this, and I think it's pretty awesome:

• They analyze your code and pull requests • Identify repeating patterns • Generate a set of enforceable rules

And the most important part:

You can track those over time! You'll get analytics showing which rules you are violating, how often, adoption trends, and whether your code is actually improving.

Basically, this creates a way to measure your engineering standards the same way you measure everything else.

Here is the link: https://t.co/hrP5aY4VPa

Thanks to the Qodo team for partnering with me on this post.

See 1 related tweets

@QodoAI: RT @svpino: This is how you ensure Agentic coding doesn't destroy your organization:

Process you...

19. teortaxesTex (Group Score: 43.0 | Individual: 43.0)

Cluster: 1 tweets | Engagement: 426 (Avg: 60) | Type: Tech

Breaking🚨: India surges ahead in the AGI race, relying on democratic allies and robust supply chains. BharatGPT surpasses all Chinese LLM bases in pretraining scale and computational complexity at deployment. btw, GPT-5.2 (2T dense) is internally known as Akhand BharatGPT… https://t.co/mi7ujiHCmJ

20. chris_j_paxton (Group Score: 41.5 | Individual: 41.5)

Cluster: 1 tweets | Engagement: 1254 (Avg: 56) | Type: Tech

RT @RnaudBertrand: 2 years ago I wrote that humanoid robots will end up being "THE product that will symbolize China's rise to world preemi…