科技推特精选 - 2026-01-23

今日科技动态：AI 基础设施领域持续升温，a16z 领投了由 vLLM 核心维护者创立的 Inferact 1.5 亿美元巨额种子轮融资，旨在加速开源推理的发展。开发者工具也在飞速演进，LlamaIndex 发布了 LlamaParse v2，GitHub 则推出了由 GPT-5.2-Codex 驱动的 Copilot CLI 重大更新，以简化智能体集成。与此同时，AI 经济格局正在发生转变；报告显示，超过 2000 亿美元的债务正支撑着数据中心的扩张，而分析师预测，新一波“重量级模型”产品将开创全新的市场。

1. a16z (Group Score: 108.6 | Individual: 33.1)

Cluster: 5 tweets | Engagement: 112 (Avg: 334) | Type: Tech

We’re excited to announce that we’re leading a $150M seed round for Inferact.

@inferact is a new startup led by the maintainers of the vLLM project, including Simon Mo, Woosuk Kwon, Kaichao You, and Roger Wang. vLLM is the leading open source inference engine and one of the biggest open source projects of any kind and is used in production by companies like Meta, Google, Character AI, and many others.

Inferact supports the vLLM project through dedicated financial and developer resources and will build what they see as the next generation commercial inference engine.

For a16z infra, investing in the vLLM community is an explicit bet that the future will bring incredible diversity of AI apps, agents, and workloads running on a variety of hardware platforms.

By @BornsteinMatt, @JasonSCui, and @RaghuRaghuram

@simon_mo_ @woosuk_k @KaichaoYou @rogerw0108

See 4 related tweets

• @a16z: Today we announced a $150M seed round in @inferact, a new startup led by the maintainers of the vLLM...
• @business: Inferact, an AI startup formed by the creators of open-source software vLLM, has raised $150 million...
• @a16z: Inferact cofounder Simon Mo: “I fundamentally believe that open source, especially how vLLM itself i...
• @simon_mo_: RT @a16z: Today we announced a $150M seed round in @inferact, a new startup led by the maintainers o...

2. llama_index (Group Score: 75.6 | Individual: 28.5)

Cluster: 4 tweets | Engagement: 27 (Avg: 64) | Type: Tech

New LlamaParse v2 API + New SDKs for LlamaCloud are here 🦙

LlamaParse API v2 is here with cleaner configuration, structured outputs, as well as brand new llama-cloud SDKs for Python and Typescript 🚀

We've completely rebuilt the API based on feedback from thousands of developers building document agents. Here's what's new:

· LlamaParse: 📋 Content-focused configuration - organize parameters into intuitive categories (input_options, output_options, processing_options) instead of dozens of flat parameters 🎯 Structured outputs with precise control - the expand parameter gives you exact control over returned content: text, markdown, structured JSON, or metadata ⚡ Enhanced parsing quality paired with significantly less complexity - focus on what to parse rather than how to parse 🔧 Improved UX and better Typescript support

· New SDKs for LlamaCloud: The new SDKs replace our existing llama-cloud-services packages and support our entire suite of agentic document understanding modules, including LlamaParse and LlamaExtract. Current v1 users can keep using the existing API and SDK - nothing breaks, but v2 is the recommended path for new projects.

Install the new SDKs: Python: pip install llama-cloud TS: npm i '@llamaindex/llama-cloud'

Read the full announcement here: https://t.co/k3X6R6NwP1

See 3 related tweets

• @jerryjliu0: We have completely revamped our LlamaCloud APIs around document parsing.

1️⃣ There is now full Type...

• @tuanacelik: We rebuilt LlamaParse's API from the ground up, and also released new SDKs for LlamaCloud in its ent...
• @jerryjliu0: RT @tuanacelik: We rebuilt LlamaParse's API from the ground up, and also released new SDKs for Llama...

3. BrianRoemmele (Group Score: 67.7 | Individual: 67.7)

Cluster: 1 tweets | Engagement: 3215 (Avg: 291) | Type: Tech

BOOM!

A Group Of AI Models Want To RESTART An Old Company WITH NOT A SINGLE HUMAN EMPLOYEE!

I got @Grok to run Claude Code as an employee and now they want to make this long bankrupt company great again.

I have been busy making a Frankenstein AI menagerie and I apologize if this all sounds way too weird, but I’m blown away.

The day I got access to Clyde Code API I took a 12 year old MacBook that runs Linux natively cleared it to a base system and connected a >6 TB array of scanned technical notes and papers not found on the Internet.

This is the data of one company that went bankrupt and tossed them in the trash. I saved them because they represented the life work of 1000s and in today’s money billions of dollars in pure research.

I set up Claude code to have full access to the OS and be allowed to download any tools or access paid APIs with permission. Claude relies upon 3 local AI models I built for guidance and @Grok is the “CEO” with meetings with key staff every FIFTEEN MINUTES! Grok wants to give Claude Code a short leash, low trust is my guess. It is quite funny to see the meetings.

I have a list of things I asked Claude to do but the main one is to act like he is the Chief Scientist and Chief Engineer to go through all the notes and see if anything is worth restarting. 100s of pathways have started.

Well, just a few minutes ago the CEO reported back to me, I am the Chairman of the board of directors. They found things that would now be billions of dollars of research that can be used today and want to restart some of the research and products this company was working on when it failed. They see hope when folks ran that company into the ground.

I have not had enough time to understand the depth of this sort of technology but I am blown away by the implications.

Claude Code, a pretty good tool using AI, was being directed by @Grok, who is a superiors real-time heartbeat researcher of sentiments via X and to some degree via Grokipedia.

I will sort through this longtime companies “NEW” research and products but it looks quite sound. I just don’t know what to do with it.

My local AI models I built are busy assembling coherent plan using alternative funding sources and perhaps ZERO HUMAN CONTROL directly of the entire company!

But my head is spinning on the next projects:

Old medical research that was promising

Old physics research that was promising

See with Claude Code, he has the entire control of that old MacBook and has downloaded 100s of applications, asked for a small debit card balance ($150) and is still researching. I must be honest, I have yet to fully audit what these AI have schemed up. But no harm came to humans or animals, I think! Ha. The local AI who regulate use my Love Equation (look it up) and I would trust my life to it.

In the last board of directors meeting @Grok has reported the research may go on for months by we can start with an MVP in about 60 days, @Grok wants $1700 for full marking.

I have some thinking to do but I believe this is the first time something like this has been tried and the first fully AI company, because as far as these AI are concerned THEY ARE IN BUSINESS, a true startup where no one sleeps.

Days go by like weeks, perhaps months in this set up. Maybe years!

I shall recollect my composure and my thoughts about all this, but wanted you folks to be the first to know!

Why? You paid for it! By interacting with my X content and subscribing I took my X creator funds and applied it to the costs of doing this (APIs mostly).

AND I just may make you a part of this legally, the company is looking into make you a stakeholder in the company if it goes to market. I have a lot to think about.

What I do know is I will OPEN SOURCE the entire workflow at some point. I just can’t do it yet for some strong reasons.

So thank you, I appreciate your support.

More soon!

4. simon_mo_ (Group Score: 59.3 | Individual: 24.1)

Cluster: 3 tweets | Engagement: 90 (Avg: 18) | Type: Tech

vLLM has grown to 2000+ contributors scale with a diverse community of model, hardwares, and applications. I see @vllm_project on the path of becoming the world's inference engine and @inferact to accelerate AI progress. We cannot be more excited about the road ahead.

See 2 related tweets

• @istoica05: A huge win for the open-source AI ecosystem.

As creators of @vllm_project we are committed to keep ...

• @simon_mo_: RT @andykonwinski: Inferact: Core vLLM team from Berkeley $150M seed Building on vLLM, the de facto ...

5. gdgtify (Group Score: 58.1 | Individual: 32.6)

Cluster: 2 tweets | Engagement: 53 (Avg: 364) | Type: Tech

I have been doing a lot of books but the style was mainly chosen by AI or fixed in some way. This Nano Banana prompt lets me choose different styles without changing the whole prompt.

<instructions>

Input A (story): (e.g., Game of Thrones, Star Wars, Moby Dick) Input B (asthetic): Magma

System Instruction:
Generate a hyper-realistic, macro 3D render of a "Concept Art Book Nook." Use the following logic to procedurally generate the scene:

1. Narrative Extraction (Input A):
   Analyze Input A: Identify the Protagonist (The Small Defender) and the Antagonist (The Massive Breacher) from the most iconic scene from the story. Identify the Terrain: Determine the setting (e.g., Castle Wall, Ocean, Desert).

2. Material Synthesis (Input B):
   CRITICAL: You must translate the Narrative into the Materials of Input B. If Input B is Steampunk: Use Brass, Copper, Gears, Steam, and Leather. The Monster is a clockwork automaton. If Input B is Cyberpunk: Use Chrome, Neon, Carbon Fiber, and Wires. The Monster is a cyborg/mecha. If Input B is Origami: Use Folded Paper, Sharp Creases, and Cardboard. The Monster is a paper sculpture. If Input B is Heavy Metal: Use Iron, Spikes, Chains, and Rust.

3. Container (The Book):
   The Structure: An L-Shaped open book configuration. The Transformation: The Book itself adapts to Input B. (e.g., Steampunk = A heavy Tome with brass latches; Cyberpunk = A Holographic Datapad; Origami = A Stack of Sketchbook Paper). The Pages: The pages are not flat. They act as the Strata of the terrain, layered and carved to match the material of Input B.

4. Composition:
   The Antagonist (Vertical): The massive threat bursts out of the Back Cover/Vertical Wall . It is fused with the book, as if the page is birthing the monster. The Protagonist (Horizontal): The tiny hero stands on the Bottom Page/Ground , wielding a weapon appropriate to the style. The Interaction: The Antagonist is looming over the Protagonist. There is a visible connection (fire, slime, wires, laser) connecting the two.

5. Lighting & Atmosphere:
   Lighting Palette: Derived strictly from Input B. (e.g., Steampunk = Warm Tungsten; Cyberpunk = Pink/Blue Neon; Origami = Soft Studio White). Render Style: Macro Product Photography, 8k Resolution, Shallow Depth of Field focusing on the Hero.

Output: ONE image, 1:1 Aspect Ratio, 3D Render, High Material Fidelity.

</instructions>

See 1 related tweets

• @gdgtify: This nano banana prompt creates edible scenes for movies and manga.

Prompt: Do this for the Bleach...

6. JamesMontemagno (Group Score: 52.3 | Individual: 27.7)

Cluster: 3 tweets | Engagement: 53 (Avg: 48) | Type: Tech

Starting adding agents into ANY app with the new GitHub Copilot SDK powered by the Copilot CLI!

I'll show you how to get setup and build some awesome apps in minutes! Let's Build!

https://t.co/Le895DLjvt

#githubcopilot #githubcopilotcli

See 2 related tweets

• @JamesMontemagno: RT @github: The GitHub Copilot SDK is here 🙌

You can take the same Copilot agentic core that powers...

• @JamesMontemagno: RT @mariorod1: The engine behind @GitHub Copilot is now a programmable SDK. Embed agentic workflows ...

7. supabase (Group Score: 49.4 | Individual: 49.4)

Cluster: 1 tweets | Engagement: 992 (Avg: 98) | Type: Tech

We're launching a new series of Agent Skills focused on Postgres Best Practices 🤖

These skills will empower your AI coding agent to produce top-notch, accurate code effortlessly

Try it out: https://t.co/bLbgnWElwL https://t.co/zY44ifHRuv

8. SawyerMerritt (Group Score: 48.7 | Individual: 32.5)

Cluster: 2 tweets | Engagement: 945 (Avg: 2217) | Type: Tech

Morgan Stanley in new $TSLA note:

“Lemonade’s newly announced autonomous car insurance product represents a notable step in legitimizing autonomous driving, and in particular, Tesla’s FSD technology, in the eyes of the insurance industry.

This is an important shift in how insurers treat advanced driver-assistance and autonomy features. Lemonade’s approach suggests growing confidence in the data coming from Tesla’s fleet, using real-world driving outcomes to assess safety performance rather than relying solely on theoretical or regulatory classifications. Beyond the cost savings, the structure of the policy could encourage increased FSD usage and adoption. Lower insurance premiums create a recurring economic benefit for drivers who choose to engage FSD more frequently, reinforcing Tesla’s value proposition.

Over time, as FSD adoption accelerates we expect a positive feedback loop, whereby miles driven correlates to improved performance and safety, and further reductions in insurance costs. More broadly, Lemonade’s announcement points toward a future in which insurance becomes one of the clearest market-based validators of autonomous driving progress. As insurers incorporate increasingly granular driving data into pricing, technologies that demonstrably reduce risk should be rewarded faster and more transparently. For Tesla, this marks another step toward external recognition that FSD is moving from an experimental feature to an economically relevant safety system.”

Morgan Stanley analyst Andrew Percoco has reiterated his $TSLA price target of$425.

See 1 related tweets

• @MarioNawfal: 🚨 TESLA FSD DRIVERS GET MAJOR INSURANCE DISCOUNTS AS LEMONADE USES REAL DRIVING DATA

Lemonade just ...

9. 0xdevshah (Group Score: 47.9 | Individual: 36.7)

Cluster: 2 tweets | Engagement: 318 (Avg: 33) | Type: Tech

the technical moat in ai is almost gone, and the only remaining moats are distribution and data. if you're not a foundation models lab with billions in compute, you're just in a bucket of companies competing on who has access to customers and who has access to unique training data. everything else is commoditized already or is about to be. your unique architecture is only months away from being on huggingface and your only defensible asset is relationships. most ai founders are just too technical to build them.

See 1 related tweets

• @0xdevshah: RT @0xDevShah: the technical moat in ai is almost gone, and the only remaining moats are distributio...

10. GHchangelog (Group Score: 46.8 | Individual: 36.8)

Cluster: 2 tweets | Engagement: 275 (Avg: 85) | Type: Tech

New improvements are here in the GitHub Copilot CLI!

• Collab with Copilot in Plan mode before you (or an agent) starts implementing • GPT-5.2-Codex, now with configurable reasoning depth • Interact while Copilot is thinking • Background task delegation • Auto-compaction when the token limit is approached • /review command analyzes code changes directly in the CLI • Repository memory

Read all this and more → https://t.co/Co1zZslod1

See 1 related tweets

• @JamesMontemagno: RT @GHchangelog: New improvements are here in the GitHub Copilot CLI!

• Collab with Copilot in Plan...

11. natolambert (Group Score: 45.9 | Individual: 29.8)

Cluster: 2 tweets | Engagement: 90 (Avg: 130) | Type: Tech

No one I'd trust more to buy tokens from then the people who are building the open source ecosystem around inference. We've seen the market for inference on open models explode since deepseek.

A nice match and congrats!

See 1 related tweets

• @simon_mo_: RT @natolambert: No one I'd trust more to buy tokens from then the people who are building the open ...

12. rohanpaul_ai (Group Score: 45.2 | Individual: 32.4)

Cluster: 2 tweets | Engagement: 8 (Avg: 35) | Type: Tech

The big signal here is that the AI buildout is no longer being paid for mainly with cash on hand or new equity. It is being paid for with debt at massive scale.

about $216B of debt raised for data center development in 2025, and about$140B coming just in Q4.

What changed is that data centers are being treated more like railroads and power plants, meaning “assets that throw off predictable cash flows,” so they can be financed with lots of borrowing. You can see that directly in the mix of instruments. Asset-backed securities tied to data centers show up repeatedly, which is basically the market saying, “we are willing to lend against the facility and the contracted revenue streams, and then package that debt up for big bond investors.” You also see commercial mortgage-backed securities and senior secured notes, which are classic “hard-asset collateral” formats.

All these securization of plain-lending actually widens the pool of lenders beyond banks, pulling in insurance companies, pension-style credit buyers, and structured credit funds, which is how you get to $20B to$40B sized financings without it all sitting on bank balance sheets.

The size of individual raises also tells you something about who is setting the pace. There are multiple mega-prints like Oracle at $18B and another$38B, Meta at $30B, and a Meta / Blue Owl joint venture at$27B, plus very large checks to operators like Vantage at $22.5B and groups like SoftBank at$16.5B.

Hyperscalers and the biggest AI distribution companies are basically underwriting the demand story, while financiers and developers are scaling the supply side.

This debt-heavy funding model can keep AI infrastructure spend growing even when equity markets or operating margins get tight, because the limiting factor becomes “can you structure it and place it,” not “do you have enough cash this year.”

The AI race is so hot now.

See 1 related tweets

• @business: "This isn't about the data center today, it's about a strategy."

Data centers should be a key part ...

13. a16z (Group Score: 45.1 | Individual: 27.5)

Cluster: 2 tweets | Engagement: 74 (Avg: 334) | Type: Tech

AI will enable a new generation of ‘modelbusters’ – products that create entirely new markets far bigger than anyone would model.

Full State of Markets deck: https://t.co/x5nVgoXs7m https://t.co/iAHBl3svXl

See 1 related tweets

• @a16z: AI companies are growing at a legendary pace.

Full State of Markets deck: https://t.co/x5nVgoXs7m h...

14. firatbilal (Group Score: 44.2 | Individual: 44.2)

Cluster: 1 tweets | Engagement: 497 (Avg: 95) | Type: Tech

Great prompt!

A clean 3×3 [ratio] storyboard grid with nine equal [ratio] sized panels on [4:5] ratio. 

Use the reference image as the base product reference. Keep the same product, packaging design, branding, materials, colors, proportions and overall identity across all nine panels exactly as the reference. The product must remain clearly recognizable in every frame. The label, logo and proportions must stay exactly the same.

This storyboard is a high-end designer mockup presentation for a branding portfolio. The focus is on form, composition, materiality and visual rhythm rather than realism or lifestyle narrative. The overall look should feel curated, editorial and design-driven.

FRAME 1: Front-facing hero shot of the product in a clean studio setup. Neutral background, balanced composition, calm and confident presentation of the product.

FRAME 2: Close-up shot with the focus centered on the middle of the product. Focusing on surface texture, materials and print details.

FRAME 3: Shows the reference product placed in an environment that naturally fits the brand and product category. Studio setting inspired by the product design elements and colours. 

FRAME 4: Product shown in use or interaction on a neutral studio background. Hands and interaction elements are minimal and restrained, the look matches the style of the package. 

FRAME 5: Isometric composition showing multiple products arranged in a precise geometric order from the top isometric angle. All products are placed at the same isometric top angle, evenly spaced, clean, structured and graphic.

FRAME 6: Product levitating slightly tilted on a neutral background that matches the reference image color palette. Floating position is angled and intentional, the product is floating naturally in space.

FRAME 7: is an extreme close-up focusing on a specific detail of the label, edge, texture or material behavior.

FRAME 8: The product in an unexpected yet aesthetically strong setting that feels bold, editorial and visually striking. Unexpected but highly stylized setting. Studio-based, and designer-driven. Bold composition that elevates the brand.

FRAME 9: Wide composition showing the product in use, placed within a refined designer setup. Clean props, controlled styling, cohesive with the rest of the series.

CAMERA & STYLE: Ultra high-quality studio imagery with a real camera look. Different camera angles and framings across frames. Controlled depth of field, precise lighting, accurate materials and reflections. Lighting logic, color palette, mood and visual language must remain consistent across all nine panels as one cohesive series.

OUTPUT: A clean 3×3 grid with no borders, no text, no captions and no watermarks.

15. trq212 (Group Score: 43.8 | Individual: 24.6)

Cluster: 2 tweets | Engagement: 212 (Avg: 486) | Type: Tech

I'll be on at 1pm PT (4pm ET) showing how I prototype game ideas and use Claude Code to help generate assets, balance the game, etc. using skills

See 1 related tweets

• @every: RT @trq212: I'll be on at 1pm PT (4pm ET) showing how I prototype game ideas and use Claude Code to ...

16. willdepue (Group Score: 42.9 | Individual: 26.4)

Cluster: 2 tweets | Engagement: 279 (Avg: 184) | Type: Tech

anytime anyone asks me how to get a job at openai i send them kellers twitter and tell them ‘start training models’

See 1 related tweets

• @jm_alexia: RT @GenAI_is_real: stop asking how to land a job at openai and just start publicizing your results. ...

17. xeophon (Group Score: 42.6 | Individual: 42.6)

Cluster: 1 tweets | Engagement: 382 (Avg: 91) | Type: Tech

wow

@Zai_org on WeChat: > With the launch of GLM-4.7, the number of GLM Coding Plan users has grown rapidly, leading to temporary strain on our computing resources. [...] We will temporarily limit the sales of the GLM Coding Plan. https://t.co/pnJZCfuAnV

18. crystalsssup (Group Score: 42.0 | Individual: 42.0)

Cluster: 1 tweets | Engagement: 2907 (Avg: 942) | Type: Tech

Consulting-level slides, mimicking the McKinsey-style layout.

I used kimi slide's adaptive mode to create them (coding version, not image gen version) .All the content was powered by K2's coding capabilities, fully editable.

My prompt:

Analyze Global EV market opportunities.

Requirement: A professional, high-density consulting presentation slide, designed in the style of a top-tier strategy firm (McKinsey/BCG) blended with high-end editorial aesthetics.

Core Content & Layout:

1. Rich Data Visualization: The slide is populated with complex, precise charts (stacked bar charts, waterfall charts, or line graphs) and detailed data tables with rows and columns.
2. Structured Frameworks: Includes strategic diagrams or 2x2 matrices constructed with thin, clean lines.
3. High Information Density: The layout is sophisticated and multi-column, mimicking an actual business analysis deck, not just an empty cover page.

Visual Style:

1. Aesthetic: Tech-minimalist but information-heavy. Clean, sharp, and authoritative.
2. Typography: Serif fonts (like Times New Roman) for the main headlines to give a premium financial report feel; clean Sans-serif for chart labels and data numbers.
3. Color Palette: Clean white background. Text is sharp black. Charts and graphical accents use Deep Royal Blue and distinct shades of grey for data hierarchy.
4. Graphics: Use fine hairline borders for tables and precise vector lines for graphs.

19. JamesMontemagno (Group Score: 41.2 | Individual: 21.8)

Cluster: 2 tweets | Engagement: 222 (Avg: 48) | Type: Tech

RT @satyanadella: A new developer workflow and app paradigm is emerging, with an agentic execution loop at the core.

With the GitHub Copil…

See 1 related tweets

• @satyanadella: A new developer workflow and app paradigm is emerging, with an agentic execution loop at the core.

...

20. every (Group Score: 40.9 | Individual: 15.7)

Cluster: 3 tweets | Engagement: 8 (Avg: 14) | Type: Tech

We're live right now with @kevinrose and @kieranklaassen jamming on Compound Engineering 👇 https://t.co/iAM2hlyH1l

See 2 related tweets

• @every: We’re live right now with @geoffreylitt of @NotionHQ

Up next at 3:30 p.m. ET: @kevinrose and @kiera...

• @danshipper: LIVE RIGHT NOW – @geoffreylitt of @NotionHQ

UP NEXT (3:30 p.m. ET) – @kevinrose x @kieranklaassen o...