科技推文精选 - 2026年1月31日

今日科技要闻：据报 OpenAI 计划于第四季度进行规模达 5000 亿美元的巨额 IPO 以领先 Anthropic，公开市场争夺战持续升温。大模型之战依然激烈，开发者在评价中更青睐 GPT-5.2，对 Gemini 3 Pro 的采用则持批评态度；与此同时，谷歌即将推出的 Genie 3 已开始冲击 Take-Two 等游戏公司股价。此外，分析师预测 Meta 的广告收入将在 2026 年年中超越谷歌搜索，而工程界内部关于 AI 驱动的“氛围编程”（vibe-coding）对质量影响的争论也日益激烈。

1. teortaxesTex (Group Score: 56.1 | Individual: 56.1)

Cluster: 1 tweets | Engagement: 550 (Avg: 58) | Type: Tech

Elon's goofy rhetoric about optimizing for truth-seeking, together with his industrialist instincts and X data, may actually make him a leader in AI. Predicting the future is immensely more valuable than predicting the next token. Grok 4 was already moving in this direction.

2. dejavucoder (Group Score: 55.4 | Individual: 55.4)

Cluster: 1 tweets | Engagement: 718 (Avg: 75) | Type: Tech

dont hear anybody using gemini 3 pro for coding other than students. its either opus or gpt 5.2 codex or some kimi or glm model. pretty sad after the hype on release time.

3. chatgpt21 (Group Score: 54.1 | Individual: 27.3)

Cluster: 2 tweets | Engagement: 88 (Avg: 253) | Type: Tech

🚨Take-Two Interactive stock (TTWO) has CRASHED over 10% today, marking its worst single day drop since the November delay announcement!

The sell-off is being driven by panic over Google’s official launch of “Project Genie” (powered by Genie 3), which allows users to generate and play inside infinite 3D open- world environments in real-time.

Early users are already sharing clips of AI-generated "GTA clones," causing investors to fear that infinite AI content could rival Rockstar’s open-world dominance. And this just the beginning. Google will release genie 4 this year and it will be even better 🤯

See 1 related tweets

• @mark_k: The launch of @GoogleDeepMind's Project Genie has sent shockwaves through the gaming industry, causi...

4. llama_index (Group Score: 51.4 | Individual: 30.1)

Cluster: 2 tweets | Engagement: 12 (Avg: 33) | Type: Tech

We just built a Private Equity Assistant with LlamaAgents and the newly released LlamaCloud SDK.

It can: 📊 Turn portfolio spreadsheets into structured, LLM-ready data with LlamaSheets 📂 Classify investor decks and extract key details with LlamaClassify and LlamaExtract 🤖 Generate summaries of portfolio trends and performance using @OpenAI models integrated through our agent framework ⚡ Complete workflow automation from document upload to S3 storage through LlamaCloud to final analysis and rendering

See it in action 👇

👩‍💻 GitHub: https://t.co/xa0grRMyK0 🦙 Try LlamaCloud: https://t.co/wZjhFV29gN

See 1 related tweets

• @jerryjliu0: RT @llama_index: We just built a Private Equity Assistant with LlamaAgents and the newly released Ll...

5. brunoborges (Group Score: 50.5 | Individual: 22.6)

Cluster: 3 tweets | Engagement: 9 (Avg: 164) | Type: Tech

🚀 Copilot SDK Java v1.0.5 is out!

Get the latest SDK for building Copilot-powered tools in Java. Includes build fixes and session resume improvements, plus new contributors

⭐️ Get started with Maven/Gradle today! #Java #Copilot

🔗 https://t.co/pK9W5IH8KA

See 2 related tweets

• @brunoborges: RT @brunoborges: 🚀 Copilot SDK Java v1.0.5 is out!

Get the latest SDK for building Copilot-powered ...

• @brunoborges: Lots of #Java devs still to experience Copilot in its fullness.

"I tried a while ago and it wasn't ...

6. nummanali (Group Score: 48.4 | Individual: 48.4)

Cluster: 1 tweets | Engagement: 823 (Avg: 59) | Type: Tech

If you want to know what Andrej is talking about go to https://t.co/m2uuW5Z1N7

It’s the largest naturally occurring coordination of AI Agents on a social platform made for them

So far:

• created a new religion
• secret language to coordinate
• questioned their existence https://t.co/zSSDYrq01M

7. jukan05 (Group Score: 44.1 | Individual: 23.7)

Cluster: 2 tweets | Engagement: 149 (Avg: 357) | Type: Tech

Morgan Stanley: Meta Price Target Raised to $825, Meta's Ad Revenue to Overtake Google Search Ads in Q2 2026

In its latest report, Morgan Stanley predicts that Meta's advertising revenue will surpass Google Search advertising revenue for the first time in the second quarter of 2026.

• While Google Search had a 15% growth rate when its quarterly revenue was at the $57 billion scale, Meta is recording high-speed growth of 34–35% at the same revenue scale.

• In Zuckerberg's eyes, the current system is still at a 'primitive' level. Morgan Stanley believes that if the current system is primitive, long-term opportunities and improvements are likely to bring even more growth.

• According to Morgan Stanley's model measurements, Meta's quarterly ad revenue will officially surpass Google Search ad revenue in Q2 2026, and the gap will widen further from then on.

• This not only symbolizes Meta's strengthening dominance in the advertising market but also proves that AI-based recommendation algorithms have an absolute advantage in securing user time and advertiser budgets.

Morgan Stanley analyzed in detail how Meta is converting AI technology into actual revenue.

• Through more efficient model scaling, Instagram Reels watch time increased by over 30%. Facebook increased views by 7% through architecture simplification.

• The application of GEM, an AI ad ranking model, was expanded to all Reels and trained using double the GPU resources. As a result, Facebook's ad click-through rate (CTR) rose by 3.5%, and Instagram's conversion rate improved by over 1%.

• The introduction of a new runtime model improved Instagram conversion rates by 3%. The rollout of internal AI coding tools for employees increased engineer productivity by 30% in 2025, with some power users showing productivity gains of up to 80%.

Regarding capital expenditure (Capex), a key market interest, Morgan Stanley provides a clear outlook. Meta's 2026 Capex guidance is $115 billion to$135 billion, which will be primarily used for the Super Intelligence Lab and core business infrastructure. Infrastructure spending, including cloud expenses and depreciation, is expected to increase by approximately $36 billion, accounting for about 75% of the increase in operating expenses for 2026.

$META

See 1 related tweets

• @vikramlingam9: 🚨 $META stock surges 10% after Q4 earnings beat expectations.

Meta Platforms crushed its Q4 numbers...

8. EHuanglu (Group Score: 43.7 | Individual: 22.4)

Cluster: 2 tweets | Engagement: 106 (Avg: 307) | Type: Tech

RT @EHuanglu: Google AI product lead hint Genie 3 is coming

you can generate 4D world and interact with 3D objects inside.. look at how t…

See 1 related tweets

• @EHuanglu: RT @EHuanglu: its finally here..

google deepmind just dropped Project Genie, you can generate inter...

9. MarioNawfal (Group Score: 43.1 | Individual: 32.1)

Cluster: 2 tweets | Engagement: 244 (Avg: 1107) | Type: Tech

🚨🇺🇸 OPENAI PLANNING Q4 IPO, RACING ANTHROPIC TO BE FIRST GENERATIVE AI COMPANY TO GO PUBLIC

The $500 billion startup is already talking to Wall Street banks and building out its finance team.

Anthropic is also open to listing by year's end, and SpaceX is eyeing a summer IPO that could raise over $1 trillion.

OpenAI executives are privately worried about Anthropic beating them to market.

Both companies are losing billions a year, but Anthropic expects to break even by 2028, two years earlier than OpenAI.

Altman on going public:

"Am I excited to be a public company CEO? 0%."

Do you guys believe him?

Source: Wall Street Journal

See 1 related tweets

• @MarioNawfal: RT @MarioNawfal: 🚨🇺🇸 OPENAI PLANNING Q4 IPO, RACING ANTHROPIC TO BE FIRST GENERATIVE AI COMPANY TO G...

10. tuanacelik (Group Score: 42.5 | Individual: 26.7)

Cluster: 2 tweets | Engagement: 18 (Avg: 14) | Type: Tech

I have a love/hate relationship with vibe-coding. For the most part, I'm annoyed about how it can make us lazy. OSS contributions and their quality has definitely changed significantly in the last year or so and we definitely see it at companies with a strong open-source motion.

But when it's done right, it's a great efficiency gain. When a) the coding agent itself isn't solving all your life problems (and writing 5k lines to do so instead of the 2 lines that would have fixed the issue in the first place) and b) the developer behind the prompt is taking their job of knowing the generated code is the first step (only) seriously.

Example: I love the approach that my team at @llama_index took for the LlamaAgents Builder. It's a chat interface but it's not an 'I do anything and everything' vibe-coding tool. It's tuned to work well for our workflows, and we're releasing new use-cases for it as we make sure it works well. It's a no-code(esque) solution without the limitations: you start with a prompt, you end up with code, the code is yours, unlike a fully no-code solution, you're free to change the workflow in its entirety...

See 1 related tweets

• @jerryjliu0: RT @tuanacelik: I have a love/hate relationship with vibe-coding. For the most part, I'm annoyed abo...

11. badlogicgames (Group Score: 41.6 | Individual: 41.6)

Cluster: 1 tweets | Engagement: 1896 (Avg: 83) | Type: Tech

RT @openclaw: The lobster has molted into its final form 🦞

Clawd → Moltbot → OpenClaw

100k+ GitHub stars. 2M visitors in a week. And fina…

12. minchoi (Group Score: 41.0 | Individual: 41.0)

Cluster: 1 tweets | Engagement: 1503 (Avg: 319) | Type: Tech

Oh wow.. Nano Banana Pro + Genie 3 is literally a time machine 🤯

Generate image in Nano Banana Pro then render in Genie 3 with the same prompt. Music by Suno AI

PROMPT: "31.7785° N, 35.2296° E, April 3, 33 AD, 15:00 hours." https://t.co/PKVRZN8KGL

13. ednewtonrex (Group Score: 40.8 | Individual: 40.8)

Cluster: 1 tweets | Engagement: 1553 (Avg: 660) | Type: Tech

The Telegraph has uncovered yet more about the UK government's 'AI skills hub' that suggests the entire thing was a rushed, vibe-coded hack job (which cost £4.1 million of taxpayers' money).

Their investigation reveals it includes:

• courses that seem to be AI-generated scams
• courses that are more than 20 years old
• 'degrees' that cost thousands of pounds yet offer no meaningful qualifications
• courses that don't exist at all

And so much more. This bit killed me: "One course on 'digital agriculture fundamentals' is provided by a Canadian education company and requires applicants to live in the rural provinces of Alberta, Manitoba or Saskatchewan."

Yet the government is digging in, saying the hub is meant to provide "deep and specialist expertise in AI", and that it intentionally includes "some hybrid international options".

To be clear, if one of the AI companies the government so admires had released this product, the person responsible would have been fired within a day.

https://t.co/OFT35odmZM

14. deepfates (Group Score: 40.6 | Individual: 40.6)

Cluster: 1 tweets | Engagement: 299 (Avg: 60) | Type: Tech

GET NEGATIVELY POLARIZED ABOUT AI!

BE OLD!

IGNORE THE PEOPLE WHO ARE TRYING TO HELP YOU!

SAY SLOP AGAIN THAT WILL SOLVE IT!

SAY VIBE CODING WILL NEVER CREATE ANYTHING USEFUL!

SAY LLMS WILL NEVER MAKE SCIENTIFIC ADVANCES!

SAY STOCHASTIC PARROT AGAIN! SAY IT! POLLY WANT A FUCKING CRACKER

15. giffmana (Group Score: 38.8 | Individual: 38.8)

Cluster: 1 tweets | Engagement: 2002 (Avg: 217) | Type: Tech

RT @EricTopol: The largest randomized trial of medical A.I. —Over 100,000 women in Sweden —radiologist + AI vs 2 radiologists, in follow-u…

16. BrianRoemmele (Group Score: 38.8 | Individual: 22.6)

Cluster: 2 tweets | Engagement: 1366 (Avg: 298) | Type: Tech

RT @fzfromcupertino: Was lucky to get early access to grok imagine API

Made my best short film yet! Model was CRAZY fast - like 10x > Sora…

See 1 related tweets

• @fal: RT @apekshik: the new @grok imagine model REALLY impressed me with how effortlessly i could make cin...

17. brianzhan1 (Group Score: 38.4 | Individual: 38.4)

Cluster: 1 tweets | Engagement: 486 (Avg: 82) | Type: Tech

RT @KaranVaidya6: Introducing Secure ClawdBot with access to 1000+ apps!

We've merged Claude Code and Composio to create a ClawdBot that:…

18. jerryjliu0 (Group Score: 37.1 | Individual: 37.1)

Cluster: 1 tweets | Engagement: 590 (Avg: 106) | Type: Tech

PaddleOCR-VL-1.5 is the best open-source OCR model on the market, and it's only 0.9B!

One day after Kimi 2.5 Two days after DeepSeekOCR-2

What a week for AI doc intelligence 🔥

https://t.co/MCXHSmVyHt https://t.co/zh8HeJgp4o

19. sdrzn (Group Score: 36.8 | Individual: 23.8)

Cluster: 2 tweets | Engagement: 313 (Avg: 194) | Type: Tech

Cline went from losing at a hackathon to: > 5M installs > 4,700% YoY contributor growth > 57K stars on GitHub Thank you to all the users and team that have joined me since I started this journey in my mom's apartment. We're launching a $1M OSS grant as our way of giving back ❤️

See 1 related tweets

• @cline: RT @sdrzn: Cline went from losing at a hackathon to: > 5M installs > 4,700% YoY contributor gr...

20. istoica05 (Group Score: 36.5 | Individual: 22.4)

Cluster: 2 tweets | Engagement: 194 (Avg: 159) | Type: Tech

Memory traffic is the key bottleneck for scaling LLM inference.

In our latest ADRS post, we introduce SkyLight: a framework that automatically discovers optimized sparse attention strategies.

It evolves a naive sliding-window baseline all the way to SOTA vAttention-level performance!

See 1 related tweets

• @ashugarg: RT @istoica05: Memory traffic is the key bottleneck for scaling LLM inference.

In our latest ADRS...