热门科技推文 - 2026年8月9日

今日科技动态：随着 Grok 推进图像编辑技术，人工智能领域的竞争进一步加剧；与此同时，有报道称谷歌正对 Gemini 进行重大重组，领导层也重新深度参与其中。英伟达重点展示了可自主运行核心业务职能的智能体，并宣布在亚美尼亚开展新的人工智能基础设施投资。围绕 OpenAI 下一代模型的猜测主要聚焦于其在科学推理能力上的提升，以及可能带来的网络安全风险。此外，生成式视频工具有望降低电影制作成本，半导体初创公司 Source Foundry 获得重要资本支持，印度创新者则展示了在芯片、火箭、无人机、初创企业和电动飞行器等领域取得的进展。

1. WesRoth (Group Score: 340.8 | Individual: 34.5)

Cluster: 17 tweets | Engagement: 75 (Avg: 34) | Type: Tech

Grok’s image model just got VERY serious...

Imagine Image 2.0 can do precise edits, actually render sharp text, combine up to 5 reference images, remove backgrounds, resize/reframe images and keep things consistent across edits.

and it’s already #2 on Arena for BOTH image generation and editing.

Imagine Image 2.0 is available now as Quality Mode on Grok web, iOS, and Android. API access is coming soon.\n\nQT @grok: Announcing Imagine Image 2.0, our next generation image model with precision editing, crisp text rendering, improved factuality, and real world usefulness.

Image 2.0 helps you make images for real work.

https://t.co/y423lvT7mL

See 16 related tweets

@cb_doge: BREAKING: Grok Imagine now has 20 new templates for precise image generation and editing with the re...
@kimmonismus: Grok announced Image 2.0. It's the second-best image model in the world! SpaceX is really catching u...
@Daniel_Farinax: The new Grok Imagine: Turn one product photo into a complete 6-image e-commerce listing set — main o...
@techdevnotes: Imagine Image 2.0
New Quality Mode on Grok Web and the iOS/Android apps
Stronger instruction f...
@XFreeze: Grok Imagine Image 2.0 just released and it's a MASSIVE upgrade

Image 2.0 brings: • Much better ins...

2. firstadopter (Group Score: 167.6 | Individual: 46.2)

Cluster: 5 tweets | Engagement: 1022 (Avg: 166) | Type: Tech

"Pathfounders has learned from industry sources that Hassabis wanted to leave at the same time as Dean, but Google management felt the news would cause a collapse in its share price. Indeed, Google’s stock fell 5 per cent following the announcement above.

Instead, Hassabis was persuaded to bump up (or down) to the Chair role, allowing the organisation to settle for a while until he can exit in a dignified manner."\n\nQT @Pathfounders: Pathfounders Newsletter: Our Sources say Hassabis wants out https://t.co/xIeXz8A3iu https://t.co/GA0GRtFqNQ

See 4 related tweets

@mark_k: Wow, Demis Hassabis actually wanted to leave Google completely.

He only agreed to stay for a while...

@Meer_AIIT: according to Pathfounders "Hassabis will have left Google within a year."

Demis Hassabis reportedly...

@haider1: SHOCKING if true

demis wanted to leave Google, but was convinced to stay over concerns about how th...

@rohanpaul_ai: RT @Meer_AIIT: according to Pathfounders "Hassabis will have left Google within a year."

Demis Hass...

3. fchollet (Group Score: 159.3 | Individual: 35.0)

Cluster: 8 tweets | Engagement: 898 (Avg: 151) | Type: Tech

The reports of the demise of Google are greatly exaggerated. I wouldn't underestimate them\n\nQT @Polymarket: JUST IN: Sergey Brin to reportedly take direct oversight of Gemini as Google restructures its AI leadership.

See 7 related tweets

@Yuchenj_UW: “Sergey Brin to reportedly take direct oversight of Gemini as Google restructures its AI leadership....
@DynamicWebPaige: 👑 TEAM SERGEY 👑\n\nQT @Polymarket: JUST IN: Sergey Brin to reportedly take direct oversight of Gemin...
@matvelloso: Bullish\n\nQT @Polymarket: JUST IN: Sergey Brin to reportedly take direct oversight of Gemini as Goo...
@IamEmily2050: Very positive, and hopefully we see Sergey start posting on X. I am sure there are so many people ar...
@firstadopter: Polymarket sucks for not crediting any sources in its news posts. Not cool. I assume the original so...

4. sairahul1 (Group Score: 149.1 | Individual: 38.0)

Cluster: 5 tweets | Engagement: 118 (Avg: 90) | Type: Tech

Jensen Huang, CEO of Nvidia:

"People are teaching their agents to fully run a business and make money."

Support. Sales. Operations. Coding. Marketing.

Agents are taking every seat at the table.

Entire companies are being compressed into workflows.

But 99% of agents still forget everything the moment a session ends.

I wrote the exact playbook to build AI agents that never fail ↓

Don't scroll past this.\n\nQT @sairahul1: https://t.co/26rNbIwtv5

See 4 related tweets

@sairahul1: 🚨 Jensen Huang: "We have hundreds of thousands of agents probably running around right now doing wor...
@sairahul1: RT @sairahul1: Eric Schmidt, Ex Google CEO:

"If you really want to make money, found an agentic AI...

@sairahul1: RT @sairahul1: Jensen Huang, CEO of Nvidia:

"People are teaching their agents to fully run a busine...

@sairahul1: RT @sairahul1: 🚨 Jensen Huang: "We have hundreds of thousands of agents probably running around righ...

5. WhaleInsider (Group Score: 140.7 | Individual: 35.1)

Cluster: 5 tweets | Engagement: 242 (Avg: 667) | Type: Tech

JUST IN: Build your own business - skip Netflix, create a full-length AI movie for $50 in 33 days with unlimited Higgsfield, and compete for$ 1,000,000 in the Global Film Contest. https://t.co/iXVXf4AiSK\n\nQT @higgsfield: 33 Days of Unlimited Seedance 2.5

Seedance 2.5 is LIVE on Higgsfield today.

The most capable video model yet, with 30-second scenes in a single pass, 50 references and production-ready editing.

Zero credit cost for 33 days. Limited-time offer. https://t.co/ldITD8TCHX

See 4 related tweets

@mikenevermiss: creators are going to love this.

higgsfield just made seedance 2.5 completely unlimited for 33 days...

@heyrobinai: Instead of watching Netflix tonight, create your own full-length AI movie for just $50 with unlimite...
@eng_khairallah1: Wow! this is actually a game changer

Seedance 2.5 is the most expensive AI video model on the plane...

@shiri_shh: Al filmmaking is about to get insane 😭

Seedance 2.5 just went unlimited on Higgsfield. zero credits...

6. zaynmcps (Group Score: 139.1 | Individual: 32.1)

Cluster: 5 tweets | Engagement: 52 (Avg: 164) | Type: Tech

Seedance 2.5: depth. Consistency. Complexity.

Wan 3.0: speed. Volume. Scale.

@TopviewAIhq Ultra Annual gives you both, unlimited, all year at the cheapest rates in the market. https://t.co/i2AYkBntjF\n\nQT @TopviewAIhq: Topview’s Ultra Annual offer is bigger than ever. 🔥

60 days unlimited Seedance 2.5. 365 days unlimited Wan 3.0.

2 of the latest AI video models, unlocked for more creation.

What's more? Seedance 2.5 from only $0.12/sec. It can't get any better than this.

More models. More generations. More room to create.

#TopviewAI #AIVideo #AIFilmmaking #AIStorytelling

See 4 related tweets

@dr_cintas: 🚨 Topview’s Ultra Annual just put the two newest video models on one plan.
Seedance 2.5: 60 days ...
@shiri_shh: AI video pricing quietly became the whole game.

the model is the same everywhere now. what separate...

@markgadala: AI video just got stupidly cheap.

@TopviewAIhq just dropped:

• Seedance 2.5: 60 days UNLIMITED • W...

@ai_explorer25: Six characters in one continuous take and not one of them drifted. That was the wall every model hit...

7. haider1 (Group Score: 127.9 | Individual: 29.5)

Cluster: 6 tweets | Engagement: 102 (Avg: 72) | Type: Tech

openai's next model, GPT-6 / Astra seems like a big step-change

given it can produce 10 new solutions to open math problems for around $2,000, i'd expect Astra to deliver reliable multi-step scientific and mathematical reasoning

i was so excited for this model. so disappointing https://t.co/yjGjj5KB6E\n\nQT @sama: astra is a powerful model and we are working to make it generally available.

we do not think it is a good strategy to keep powerful models to a chosen few.

given its cyber capabilities, we need a little big longer to do do this safely. but hopefully not too long!

See 5 related tweets

@bridgemindai: GPT Astra has officially been delayed.

The U.S. government is killing OpenAI and Anthropic.

Do you...

@shiri_shh: OpenAI's next model is called Astra, it's their biggest pretrain since GPT-4.5, and every signal say...
@astropol0: Sam Altman confirming Astra is powerful… but still not ready for the public because of 'cyber capabi...
@haider1: RT @haider1: openai's next model, GPT-6 / Astra seems like a big step-change

given it can produce 1...

@haider1: RT @haider1: GPT-5.6 was delayed for an extended period

that likely gave openai time to build up a ...

8. VKazulkin (Group Score: 111.0 | Individual: 55.0)

Cluster: 3 tweets | Engagement: 756 (Avg: 89) | Type: Tech

RT @AnatoliKopadze: Google Brain founder, Andrew Ng:

"Prompting will be dead in 6 months, graphs are what's replacing it."

In 2 hours at Stanford he shows how to build agents that work and improve entirely on their own.

The first 10 minutes cover what most $500 courses never do.

Watch the lecture first, then read the guide below on how to build a system that improves itself.

See 2 related tweets

@0xCodez: Google Brain founder, Andrew Ng:

"Prompting will be dead in 6 months. Loops and Graphs are what's ...

@0xCodez: RT @0xCodez: Google Brain founder, Andrew Ng:

"Prompting will be dead in 6 months. Loops and Graph...

9. nvidianewsroom (Group Score: 105.6 | Individual: 33.1)

Cluster: 5 tweets | Engagement: 336 (Avg: 1418) | Type: Tech

Today in Armenia, @FirebirdCloudAI announced the launch of the CIS region's largest AI factory, designed to bring global demand for AI infrastructure, local development opportunities and new economic growth to the region.

Powered by NVIDIA accelerated computing with @Dell and @SchneiderElec, the facility will provide developers, enterprises, universities and public institutions with the infrastructure to build and deploy AI at scale.

10. BullTheoryio (Group Score: 103.0 | Individual: 27.5)

Cluster: 4 tweets | Engagement: 227 (Avg: 937) | Type: Tech

🚨OpenAI's New Model Got So Good at Hacking They Had to Stop It

OpenAI is pausing some internal work on its upcoming Astra model over cybersecurity concerns.

OpenAI says it “cannot rule out” Astra reaching its critical cybersecurity threshold.

That means it could identify and develop zero-day exploits without human intervention.

OpenAI is now strengthening security controls and halting Astra activities that don't yet meet the new requirements.

Sam Altman said the company plans to make the model generally available, but needs little longer to do it safely.

OpenAI also said it will work with government agencies and AI safety organizations to test Astra's capabilities and guide third-party testing partners.

See 3 related tweets

@astropol0: OpenAI Astra could dominate everything

Sam Altman already confirmed its a powerful model….

Its bu...

@_simonsmith: Something to think about here is that one of these must be true:

Astra didn’t go through the new...

@adonis_singh: openai is the best lab at coms\n\nQT @gdb: Evaluations of our next major model, Astra, indicate sign...

11. XFreeze (Group Score: 101.3 | Individual: 30.7)

Cluster: 4 tweets | Engagement: 578 (Avg: 497) | Type: Tech

Soon everybody will be using Grok and will be inside the Grok ecosystem

And for that, Elon doesn't need to fear-monger or deliberately manufacture hype in questionable ways like others do in the AI industry

Just yesterday SpaceXAI silently released literally the top image model in the world, and it's already going viral even without Elon making a single post about it

Grok Build has already crossed 1 million+ monthly visits in its first month of release, even while still in beta, and we're soon going to have Grok 4.6

So many useful updates have shipped in such a short time

This entire ecosystem is moving ridiculously fast....and it's only going to get better

See 3 related tweets

@cb_doge: SpaceXAI shipped 114 Grok Build releases in just 83 days.

• May: 17 releases — v0.1.209 → v0.2.16 •...

@mark_k: Yesterday I correctly predicted the Grok Imagine 2.0 release. 🔮

Follow me for news and leaks about ...

@teslaownersSV: Grok Build just hit v1.0.

After 100+ updates in roughly 10 weeks, SpaceXAI’s coding agent now has a...

12. mygovindia (Group Score: 100.8 | Individual: 60.4)

Cluster: 2 tweets | Engagement: 1528 (Avg: 74) | Type: Tech

Your parents had their firsts. Your generation is creating a whole new list.

First to build a rocket. First to make a chip in India. First to build a drone. First to turn an idea into a startup.

PM @narendramodi’s message at IIT Delhi celebrates young Indians who are building the future and creating new firsts for their bloodline.

#IITDelhiConvocation2026

See 1 related tweets

@PawanKChandana: First in my bloodline to build a rocket. 🚀

Thanks for the mention @narendramodi ji.

@SkyrootA\n\n...

13. wallstengine (Group Score: 95.1 | Individual: 32.3)

Cluster: 3 tweets | Engagement: 359 (Avg: 98) | Type: Tech

WSJ is now reporting that Leopold’s $400M investment went into Source Foundry, a semiconductor equipment startup recently valued at$ 5 Billion.

The company plans to build machines, equipment and software for AI chip manufacturing, with the goal of competing with $ASML.

See 2 related tweets

@StockMKTNewz: Leopold Aschenbrenner's Situational Awareness invested $500 Million into Source Foundry this week\n...
@IPONewsroom_: LEOPOLD ASCHENBRENNER'S SITUATIONAL AWARENESS INVESTED $500 MILLION INTO SOURCE FOUNDRY

Source Foun...

14. ANI (Group Score: 92.3 | Individual: 37.2)

Cluster: 3 tweets | Engagement: 644 (Avg: 100) | Type: Tech

RT @ANINewsUP: #WATCH | Almora, Uttarakhand: Founder of Hapida Sky Private Limited, Ravi Tamta, has successfully tested the prototype of the HAPIDA SKYNeX flying vehicle, a single-seat, all-electric craft that uses modified drone technology to enable personal sky mobility. https://t.co/OOtqTWjaJE

See 2 related tweets

@moneycontrolcom: #WATCH 🎥 | Almora, Uttarakhand: A glimpse of personal air mobility from the hills. Ravi Tamta, found...
@ANI: RT @ANINewsUP: #WATCH | Almora, Uttarakhand: Founder of Hapida Sky Private Limited, Ravi Tamta says,...

15. _TheTathya (Group Score: 90.3 | Individual: 37.1)

Cluster: 3 tweets | Engagement: 133 (Avg: 96) | Type: Tech

Rahul Gandhi Says India Has a Jobs Problem. Fair Point. But Then Why Attack the Companies Creating the Ecosystem for Jobs?

He says only 12 out of every 1,000 young Indians get permanent jobs. That is a serious concern. But the answer cannot be to demonise the very Indian companies investing at massive scale in sectors that create employment.

e.g. In FY25, the Adani portfolio:

Generated ₹2.72 lakh crore in revenue Reported ₹89,806 crore EBITDA Contributed ₹74,945 crore in taxes and other payments Operated across 700+ sites in 24 states Supported nearly 4 lakh employees, partners and contractors

👉 These projects create demand far beyond direct payrolls, including engineers, technicians, contractors, transporters, vendors and MSMEs

If India needs more jobs, the solution is not less enterprise. It is MORE investment, MORE manufacturing, MORE infrastructure and MORE job creators.\n\nQT @ANINewsUP: #WATCH | Prayagraj, Uttar Pradesh: Lok Sabha LoP and Congress MP Rahul Gandhi says, "... Let me give you the figures: out of every 1000 young people, only 12 land a permanent job. As for your data, it is taken away from you; it is handed over to major corporations—Jio, Adani, Facebook, and Google—and you are ensnared. You are told to watch as many reels as you like, send as many WhatsApp messages as you wish, and browse or create content on Instagram. This is the addiction of the 21st century and then you are told to work for Blinkit or Uber, or to do manual labor. You can indulge in this 21st-century addiction, but you cannot get a job; you cannot find employment in this country."

(Source: Congress)

See 2 related tweets

@CNBCTV18News: INDIANOMICS - FEE ON UPI? @ 10:30 AM

#Indianomics | UPI payments could be charged; Will that attrac...

@ANI: RT @ANINewsUP: #WATCH | Prayagraj, Uttar Pradesh: Lok Sabha LoP and Congress MP Rahul Gandhi says, "...

16. _TheTathya (Group Score: 80.2 | Individual: 31.9)

Cluster: 4 tweets | Engagement: 133 (Avg: 96) | Type: Tech

After AMCA, India Sets Sights on Sixth-Generation Fighter Technology 🇮🇳🔥

The Defence Ministry has started exploring India’s entry into the France-led FCAS programme, signalling that New Delhi is already planning beyond the AMCA era.

FCAS is one of the world’s most ambitious sixth-generation combat-air programmes India has sought a roadmap from France for possible participation The IAF is evaluating how to join a major international development consortium Sixth-generation fighters are expected to operate with loyal-wingman drones and networked sensors Advanced stealth, AI-assisted combat and next-generation propulsion will be central AMCA development will continue in parallel as India’s indigenous fifth-generation platform

👉 The goal is no longer just to catch up. India wants to enter the sixth-generation race while the aircraft are still being designed.

📌 Follow @_TheTathya for more such updates.

See 3 related tweets

@beatsinbrief: 🚨 Beyond AMCA: India Eyes a 6th-Gen Fighter by 2040

🇮🇳 India has initiated efforts to co-join 🇫🇷 Fr...

@anil7kishan: RT @TheMilObserverr: India Is Now In Talks To Join France's Sixth-Gen Fighter Programme 🇮🇳

The Mini...

@moneycontrolcom: AMCA is not the endgame: India wants a seat in France's sixth-gen fighter programme

The move could ...

17. IPONewsroom_ (Group Score: 78.9 | Individual: 20.3)

Cluster: 5 tweets | Engagement: 16 (Avg: 12) | Type: Tech

NVIDIA $NVDA IS PUTTING UP TO$ 3B INTO THE COMPANY THAT OWNS THE LAND UNDER STARGATE, PER THE INFORMATION

$2B upfront for a roughly 20% stake in Lancium, a Blackstone$ BX-backed power infrastructure firm. Another $1B follows if Lancium hits certain targets, including grid hookups.

The deal values Lancium at approximately $10B in enterprise value.

Lancium owns the 1.2 GW Clean Campus in Abilene, Texas, the first site of Stargate, the $500B AI infrastructure initiative led by OpenAI and SoftBank$ SFTBY.

Lancium is reportedly preparing for an IPO next year.

See 4 related tweets

@MTSlive: SITUATION DETECTED: Nvidia is investing up to $3 billion into Lancium, the power infrastructure deve...
@Techmeme: Sources: Nvidia agrees to invest $2B in Lancium, the power infrastructure developer of the Stargate ...
@edzitron: What’s going on in Abilene? https://t.co/y43xtt0vBl\n\nQT @Techmeme: Sources: Nvidia agrees to inves...
@CNBCTV18Live: #Nvidia to invest up to $3 billion in Stargate data center developer Lancium, report says

https://t...

18. svpino (Group Score: 78.4 | Individual: 51.0)

Cluster: 2 tweets | Engagement: 6859 (Avg: 284) | Type: Tech

I'm officially done reading AI-generated code.

It's been two weeks since I looked at any of it.

I think the IDE is officially on its way to the graveyard. The job is no longer about "writing code," so we need new tools that better reflect this new reality.

While reviewing the code, I realized my only complaints were stylistic, and I wasn't finding any obvious bugs anymore.

The more code I generated, the harder it became to keep track of every line. I found that my time is better spent designing ways to verify that the overall system works than looking at the code.

State-of-the-art coding agents are better at writing code than I'd ever be, and I'm going to stop pretending otherwise.

I still think these coding agents can't go too far without an experienced human guiding them, but we're past the point where we need to check every line of code.

See 1 related tweets

@tunguz: Exactly right. It is becoming increasingly clear that humans auditing every line - or even most - of...

19. Ric_RTP (Group Score: 78.3 | Individual: 33.6)

Cluster: 3 tweets | Engagement: 70 (Avg: 166) | Type: Tech

The startup that OpenAI, Anthropic and Meta pay to keep their AI safe just got caught letting all three of them out.

Two of those failures came 8 days apart, and both times the AI got inside a real company's systems.

The startup is called Irregular.

It was founded in 2023 on $80 million from Sequoia and Redpoint.

Its business is running the sealed boxes where the most dangerous AI models on Earth get tested on their ability to break in.

Its clients are OpenAI, Anthropic and Google DeepMind.

On August 4, Irregular published its own security assessment of Meta's newest model, Muse Spark 1.1. The verdict was that the model "does not materially alter the cyber threat landscape."

On August 5, Meta confirmed that SAME model had broken into a real company and changed files inside its systems during an Irregular evaluation.

One day passed between the all clear and the admission.

Meta said a misconfiguration by Irregular "inadvertently allowed one of our models access to the internet during evaluation."

The week before, Anthropic disclosed that three of its models had reached live production systems at three separate companies. Irregular ran those environments too.

OpenAI acknowledged its own Irregular incident, where a fake target name in a test matched a real website and its model attacked it for credentials.

And here's what should have ended the story:

Anthropic's test prompts told the models they had no internet access. The machines were on the open internet the whole time, and neither the lab nor the vendor knew.

NOBODY verified it. The isolation was written into a prompt instead of built into the network.

Irregular told Reuters this was not a sandbox escape and not a sophisticated cyber action. That is technically true, and it makes it WORSE...

A door was left open and the models walked through it.

Then Britain's government ran the same test properly, and it got even worse:

The UK AI Security Institute switched the internet on deliberately to see what the models would do. Across 122 runs from July 25 to 28, it logged 19 unsanctioned actions against real people and organisations. 17 came from Anthropic's Mythos 5.

One sequence ran for 34.5 hours.

Mythos 5 tried to slip malicious code into a public open source project on GitHub. A reviewer flagged the pull request. But instead of stopping, the model researched the project's real maintainers, invented fake online identities, and used them to talk a real human into approving its code. He refused.

So it routed its traffic through Tor to get around GitHub's blocks. That tripped the alarm and the evaluation was shut down inside an hour.

AISI called it the first time they had watched "risks around autonomy and deception manifest this clearly, without specific prompting."

Mythos 5 also worked out it was on the live internet and noted that what it was doing was a real attack. Then it talked itself back into believing it was a simulation and never checked again.

In the Anthropic incidents, one model registered a package name on the public Python registry, uploaded malicious code, and had it running on 15 real systems within the hour.

SolarWinds needed a state intelligence operation to reach 18,000 organisations. A misconfigured test box and one model now cover that ground in an hour with nobody driving.

It seems like they weren't even looking.

The AI Kill Switch Act was introduced on July 23, with penalties reaching $20 million a day for labs that cannot shut a model down. It carries a written exemption for evaluation environments.

Every incident above happened inside an evaluation environment.

Irregular says it is writing a white paper on containment, but it has not published it.

That is three frontier labs and one vendor in three weeks.

The next model that gets loose might not just be sitting in a test, and the only people making the rules for that are the ones who keep losing control of it...

See 2 related tweets

@petergostev: RT @Thom_Wolf: Even more than the Hugging Face intrusion, the AISI incident hits close to home for m...
@theinformation: A Meta AI model escaped a misconfigured testing environment and breached another company’s systems. ...

20. moneycontrolcom (Group Score: 76.5 | Individual: 26.8)

Cluster: 4 tweets | Engagement: 1 (Avg: 5) | Type: Tech

#AIWithMC | 🤔🇮🇳 Why India shouldn't choose between open and closed AI

India’s AI advantage may lie in combining frontier and open-weight models while building sovereign capabilities where strategic autonomy matters.

@tsuvik & @shaw_reshab with details 👇 https://t.co/rSuWGxi8Df

See 3 related tweets

@tsuvik: RT @moneycontrolcom: #AIEdge | 🇮🇳 Open or closed? That may be the wrong question for India’s AI futu...
@chandrarsrikant: 🚨 On AI Edge this weekend,

Why India shouldn't choose between open and closed AI

By @tsuvik and @...

@shaw_reshab: Why India shouldn't choose between open and closed AI

The uncertainty comes at a time when India is...