科技热门推文 — 2026年7月21日

今日科技动态：随着Kimi K3迅速崛起、阿里巴巴筹备推出开放权重模型Qwen3.8，人工智能领域的竞争进一步加剧。与此同时，华盛顿方面正考虑对中国人工智能模型实施限制，使有关成本与可及性的讨论愈发激烈。基础设施投资显著升温，Fluidstack融资8.3亿美元，CuspAI获投4.5亿美元；另据报道，谷歌正致力于研发一款能效大幅提升的人工智能芯片。AMD进一步扩大了与微软Azure的合作，Ramp推出模型路由技术以降低人工智能成本，而工程领域的关注重点则转向通过可组合工作流，协调具备自我改进能力的智能体。

1. MTSlive (Group Score: 770.5 | Individual: 58.7)

Cluster: 30 tweets | Engagement: 1853 (Avg: 89) | Type: Tech

SITUATION BREWING: The Trump administration is considering restricting cutting-edge Chinese AI models, with momentum reviving after the launch of Kimi K3, per Axios.

See 29 related tweets

@SophiaCai99: There was a brief discussion about this recently, but the Commerce Department is NOT moving forwar...
@mark_k: The Trump administration is reportedly considering restricting cutting-edge Chinese AI models after ...
@teortaxesTex: Remember: @deanwball doesn't write USG policy anymore, but he has a fairly good idea of how it can d...
@benitoz: Washington is finally admitting AI models aren’t neutral. They export the values and worldview of th...
@Hesamation: Dario: “THE SCALING OF OPEN-SOURCE MODELS IS GOING DOWN A VERY DANGEROUS PATH.”

this is worth a rew...

2. wallstengine (Group Score: 327.4 | Individual: 35.5)

Cluster: 14 tweets | Engagement: 155 (Avg: 167) | Type: Tech

Fluidstack, the company building gigawatt-scale AI infrastructure for the world’s leading labs, closed an $830M Series A at a$ 7.5B valuation in January.

The round was led by Leopold Aschenbrenner’s Situational Awareness, with participation from BlackRock, Google, and Jane Street.

The company has led the deployment of Anthropic’s $50B compute buildout, one of the largest AI infrastructure projects in U.S. history.

Traditional data centers can take 18 to 24 months to deliver. Fluidstack’s stated goal is to deploy gigawatts of compute in six months.\n\nQT @fluidstack: In January, Fluidstack raised an $830M Series A at a$ 7.5B valuation, led by Situational Awareness, with participation from world-leading investors.

Fluidstack builds infrastructure for the leading AI labs - our goal is to be the fastest on the planet at deploying hundreds of gigawatts of compute.

As we stand at the event horizon of the singularity, humanity’s best chance is if democracies, with error correcting institutions, free speech, and checks on power, imbue those same principles into superintelligence. Whoever deploys frontier compute infrastructure fastest will decide whether Al expands human freedom or shrinks it.

We are hiring - come join us.

See 13 related tweets

@Mayhem4Markets: BlackRock, Google, Jane Street, and Situational Awareness.

On the same cap table.

That's Fluidstac...

@0x_kaize: Everyone thinks the AI race is about who has the most GPUs

WRONG.

The US added 18GW of datacenter ...

@shiri_shh: Leopold Aschenbrenner turned $225M into$ 5.5B in a year on one thesis: AI is bottlenecked on physica...
@rohanpaul_ai: Leopold Aschenbrenner wrote the essay about the next decade. Now his fund is backing the company th...
@Meer_AIIT: the part everyone skipped in the fluidstack raise:

they closed in december.

then said nothing for ...

3. chamath (Group Score: 275.1 | Individual: 42.5)

Cluster: 11 tweets | Engagement: 6460 (Avg: 2346) | Type: Tech

The leading AI has already forked into two options.

A: Closed source American that costs $26-56 per 1MM tokens.

B: Open weight Chinese that costs $0.50-1 per 1MM tokens.

If you force American companies to spend 50-100x more than their competitors abroad here is what will happen:

American companies spending 50-100x will at some point become financially impaired and the stock market will crater. This is the equivalent of the US Government saying we can only buy oil at $800/barrel even when the free market sells the same oil for$ 80/barrel.
In the short term, the revenues of the Closed American Labs may remain hi. But then their revenue will also crater because their customers in the US will be impaired and their customers abroad will have already flipped to cheaper solutions.

This would be a terribly self-defeating form of intervention if it were to happen.\n\nQT @MTSlive: SITUATION BREWING: The Trump administration is considering restricting cutting-edge Chinese AI models, with momentum reviving after the launch of Kimi K3, per Axios.

See 10 related tweets

@XFreeze: Worth noting that Elon's Grok 4.5 is the major American exception here

SpaceXAI’s Grok 4.5 is one o...

@danielnewmanUV: Of course the USA isn’t going to just let China repackage American innovation and sell it back to it...
@GergelyOrosz: The disclosure about open AI models (+worries about the impact on the US leading the AI race) fails ...
@0xdevshah: american ai labs love turning basic trade economics into a cold war thriller. hard to tell how much ...
@bgurley: KeyPoint!

Making a move like this wouldn't just hurt US AI competitiveness, it would negatively im...

4. MTSlive (Group Score: 254.5 | Individual: 54.9)

Cluster: 11 tweets | Engagement: 776 (Avg: 89) | Type: Tech

SITUATION BREWING: Google is developing a new AI chip, code-named Frozen, projected to be 6-10x more efficient than its TPUs, per The Information.

Frozen would have decisions for Gemini models built into the chip. It is a new branch of custom silicon, not a TPU replacement.

See 10 related tweets

@coinbureau: ⚡️JUST IN: Alphabet plans to launch “Frozen v2,” a chip built specifically for Gemini that could be ...
@tunguz: Sounds like ASICs with extra steps.\n\nQT @MTSlive: SITUATION BREWING: Google is developing a new AI...
@PolymarketMoney: BREAKING: Google developing a new AI chip called “Frozen v2” to run Gemini models up to 10x more eff...
@KobeissiLetter: BREAKING: Alphabet, $GOOGL, is planning on launching a new “frozen” chip to run its AI models more e...
@StockMKTNewz: GOOGLE PLANS NEW ‘FROZEN’ CHIP TO RUN ITS AI MODELS MUCH MORE EFFICIENTLY

Google is working on a ne...

5. StockSavvyShay (Group Score: 229.0 | Individual: 37.0)

Cluster: 8 tweets | Engagement: 773 (Avg: 544) | Type: Tech

The biggest gem from today’s $IREN announcement was that customers are funding nearly half the GPU cost before those systems generate revenue.

That shifts IREN toward a model where customers absorb much of the upfront cost IREN monetizes the power, data centers, operations and cloud infrastructure it already controls.\n\nQT @VestExchange: $IREN is up around 8.5% in pre-market. The company announced$ 2.8B in new multi-year AI cloud contracts and lifted its 2026 AI Cloud run-rate revenue target above $4B.

A customer roster spanning Microsoft, NVIDIA, Perplexity, and Figure AI sits around the AI cloud pivot, with ~85% of the target under contract and $7.6B in cash on hand.

Friday closed 3.5% lower for $IREN. The median analyst target sits at$ 80, while JPMorgan holds an Underweight stance.

See 7 related tweets

@StockSavvyShay: Goldman Sachs estimates $IREN new$ 2.8B contracts add roughly $1B in revenue over an average term of...
@StockMKTNewz: IREN SIGNS $2.8B IN NEW CUSTOMER CONTRACTS WITH LEADING AI DEVELOPERS, RAISES 2026 ARR TARGET TO OVE...
@wallstengine: $IREN RAISED its year-end 2026 annualized revenue run-rate target to more than$ 4 billion, above its...
@StockSavvyShay: $IREN signed$ 2.8B in new multi-year AI cloud contracts and raised its year-end 2026 ARR target to m...
@danielnewmanUV: $IREN deals deals deals. 👏🏻🚀\n\nQT @StockMKTNewz: IREN SIGNS$ 2.8B IN NEW CUSTOMER CONTRACTS WITH LE...

6. tryramp (Group Score: 206.6 | Individual: 43.5)

Cluster: 8 tweets | Engagement: 182 (Avg: 52) | Type: Tech

We cut our own AI bill by 30% with a Router that chooses the right model for the right task.

Now you can too.\n\nQT @vral: Today we’re launching Ramp Router.

3 years ago, we built an internal LLM router at @tryramp that powers AI products for 70,000 customers.

Back then it was mostly about saving money.

Now it feels obvious: the best model changes constantly. GPT, Claude, Gemini, Grok, Qwen, DeepSeek, Kimi, GLM - prices and capabilities move every week.

So we’re opening up access to everyone.

One OpenAI-compatible endpoint. The right model for every request. Lower cost without rewriting your app.

Reserve access to use it.

See 7 related tweets

@eglyman: ramp web services 🤔

more seriously, you could hire an olympic sprinter to walk your dog. it'd go gr...

@tryramp: RT @rahulgs: Every week, the price-intelligence-latency frontier shifts, and we expect this trend to...
@Meer_AIIT: the ai infra conversation is quietly shifting

for two years the question was "how do i get access t...

@damianplayer: RT @vral: Today we’re launching Ramp Router.

3 years ago, we built an internal LLM router at @tryra...

@zaynmcps: RT @chrisspeerez1: Most AI apps start the same way: pick one powerful model and route every request ...

7. dee_bosa (Group Score: 201.6 | Individual: 36.2)

Cluster: 9 tweets | Engagement: 378 (Avg: 379) | Type: Tech

“The open-model wave is not an attack on AI companies. It is the market responding to the fortune they say they are about to make. They called it forth themselves.”

Gurley nails it. Open models are the market doing what markets do when incumbents are charging huge margins.\n\nQT @bgurley: The way to think about “open” in software is being a low-cost producer (vs a high margin one). When a company (or 2) achieves record valuations in record time, that rightfully attracts competition (as it should). We need to let the free market work.

https://t.co/OJ9XPcKmX4

See 8 related tweets

@bgurley: The way to think about “open” in software is being a low-cost producer (vs a high margin one). When ...
@TaylorLorenz: RT @dee_bosa: “The open-model wave is not an attack on AI companies. It is the market responding to ...
@perrymetzger: Really, really good opinion piece by @bgurley in the Washington Post.\n\nQT @bgurley: The way to thi...
@AravSrinivas: Well said!\n\nQT @bgurley: The way to think about “open” in software is being a low-cost producer (v...
@sriramk: great piece from @bgurley on open weight models\n\nQT @bgurley: The way to think about “open” in sof...

8. ModelScope2022 (Group Score: 190.9 | Individual: 49.5)

Cluster: 8 tweets | Engagement: 6500 (Avg: 439) | Type: Tech

RT @Alibaba_Qwen: Qwen3.8 is launching and going open-weight soon!🌐

With a massive 2.4T parameters, this model is continuously evolving. We believe it’s one of the most powerful model available today, compatible to leading frontier AI models , second only to Fable 5.

You don't have to wait to test it. Just now, the Qwen3.8-Max-Preview made its debut on Alibaba’s Token Plan, Qoder, and QoderWork. Be among the very first to try it out.

Can't wait to hear what you build. Stay tuned! 🚀

Token Plan international：https://t.co/YRvcGdB9Bv China：https://t.co/PKMUNwUuRp

See 7 related tweets

@teortaxesTex: > We're looking forward to a more capable, official version — and to open-weight it for everyone....
@wallstengine: Alibaba $BABA launched a preview of Qwen3.8 Max, a 2.4T-parameter flagship AI model, per Bloomberg. ...
@BrianRoemmele: Wow this is accelerating rapidly.

More open source out soon…\n\nQT @Alibaba_Qwen: Qwen3.8 is launch...

@gmi_cloud: Qwen 3.8 is gonna be another open source milestone!\n\nQT @Alibaba_Qwen: Qwen3.8 is launching and go...
@TeksEdge: How has your Qwen3.8 experience been? It's a preview version and will improve. Early benchmarks lik...

9. cryptopunk7213 (Group Score: 189.6 | Individual: 38.5)

Cluster: 8 tweets | Engagement: 751 (Avg: 125) | Type: Tech

been using kimi all weekend. the #1 reason i’ll continue: zero restrictions. it just does what you ask. no pushback or relegation to lower class models.

US gov needs to fix the safeguard framework or we end up losing.\n\nQT @Kimi_Moonshot: Kimi K3 has received far more love than we expected, and our GPUs are feeling it.

Over the past 48 hours, demand has pushed close to the limits of our current capacity. To protect the experience of existing subscribers, we're temporarily pausing new subscriptions and prioritizing compute for current members. Existing subscribed users are not affected.

We're adding capacity as fast as we can and will reopen new subscription spots in batches.

Going forward, we'll also split membership into two more focused plans: Kimi Membership for Kimi Web, App, and Work; and Kimi Code Membership for coding workflows. This will help us match compute more precisely and keep the experience stable.

Thank you for your patience and understanding!

See 7 related tweets

@VaibhavSisinty: Moonshot AI is pausing new subscriptions because demand exceeded their entire compute capacity.

Now...

@thealexbanks: Let me get this straight:

China just built the #1 frontend coding model.

It's a third of Claude's ...

@peterwildeford: The US has an immense strategic advantage by having way more compute than China. This advantage is k...
@aiedge_: Kimi K3 has so much demand that they're literally pausing subscription plans.

This is proof that th...

@pvergadia: A good reminder that most folks are talking about open models but they are only using them in hosted...

10. StockMKTNewz (Group Score: 182.5 | Individual: 34.6)

Cluster: 7 tweets | Engagement: 1094 (Avg: 291) | Type: Tech

AMD AND MICROSOFT EXPANDED PARTNERSHIP

$AMD just announced an expanded strategic partnership spanning AMD GPUs, CPUs, networking and software on Microsoft$ MSFT Azure.

"Microsoft to ramp AMD Helios™ at scale on Azure to power frontier model inference for Microsoft, its AI customers and Azure AI services.

Azure will add two new VMs powered by 6th Gen AMD EPYC™ "Venice" processors

Azure deploys AMD Pensando™ DPUs in AMD AI backend networking infrastructure and select Azure services.

The companies are integrating AMD silicon with Azure Boost to scale cloud networking performance across the fleet."

See 6 related tweets

@firstadopter: "AMD today announced an expanded strategic partnership spanning AMD GPUs, CPUs, networking and softw...
@AMD: Today, AMD and @Microsoft are announcing an expanded strategic partnership to deliver a full stack o...
@wallstengine: Microsoft expands $AMD partnership, will deploy Helios rack-scale systems on Azure for frontier AI i...
@danielnewmanUV: $AMD 🤝$ MSFT

The Helios ramp is the inflection that takes $AMD to +/- 10% GPU market share and a bi...

@AMD: Together with @Azure, we're #AdvancingAI across the full stack of AMD AI solutions. See how our team...

11. 0xMovez (Group Score: 181.2 | Individual: 37.9)

Cluster: 6 tweets | Engagement: 460 (Avg: 133) | Type: Tech

Andrew Ng:

“AI agents are doing almost 100% of my tasks now - the hype has exceeded my expectations.

in 4-6 months, we’ll all be building graphs to orchestrate self-improving agents. No more prompting.”

In a 20-minute talk, Andrew Ng explains how to build self-improving agentic systems from scratch.

Worth more than a $500 agentic course.

Watch this video, then read the article below on how to become a graph architect.\n\nQT @0xCodez: https://t.co/WVztfCUH4u

See 5 related tweets

@eng_khairallah1: RT @eng_khairallah1: A senior Google engineer just dropped a 424-page doc called Agentic Design Patt...
@0xMovez: Andrew Ng:

"100% of my tasks are now done by AI agents - hype has exceeded my expectations. Loops...

@Vtrivedy10: jokes aside if graph engineering does take off then you can point Claude at the Deep Agents repo htt...
@0xMovez: RT @0xMovez: Andrew Ng:

“AI agents are doing almost 100% of my tasks now - the hype has exceeded m...

@0xMovez: RT @0xMovez: Andrew Ng:

"100% of my tasks are now done by AI agents - hype has exceeded my expect...

12. peterwildeford (Group Score: 169.7 | Individual: 35.7)

Cluster: 6 tweets | Engagement: 122 (Avg: 47) | Type: Tech

Consolidating my Kimi K3 takes so far:

Kimi K3 is exactly on-trend compared to what you would expect based on historic Chinese progress. K3 is still well-behind Mythos and GPT-Sol in capabilities. China remains ~6mo behind.
K3 was likely produced through violations of US export controls. The controls are working and better enforcing controls would make them work even better.
Kimi/Moonshot already report having insufficient compute to serve their models. The US has an immense strategic advantage by having way more compute than China. This advantage is key to the AI race.
Kimi K3 cannot be run on the laptop. It must be run on the cloud. China lacks compute to both train and run models and this is good for the US.
Even if the US and China had equally good AIs (they don't), it will matter who has more compute to run more of those AIs. You'd rather have 10,000 US-Mythos defending against 100 China-Mythos than the other way around!

See 5 related tweets

@WOLF_Financial: CHINA'S MOONSHOT AI IS FINALIZING A ROUND VALUING IT ABOVE $30B, WITH A HONG KONG IPO POSSIBLE IN AS...
@IPONewsroom_: CHINA'S MOONSHOT AI IS FINALIZING A ROUND VALUING IT ABOVE $30B, WITH A HONG KONG IPO POSSIBLE IN AS...
@business: The Chinese AI pioneer Moonshot’s release of its Kimi K3 model sparked fierce debate in the US https...
@business: In time, Chinese AI may take a significant and sustained lead over the US, but Moonshot's Kimi K3 wo...
@theinformation: Moonshot’s new Kimi K3 model is forcing AI researchers to reconsider how far China has closed the ga...

13. BrianRoemmele (Group Score: 158.3 | Individual: 58.8)

Cluster: 4 tweets | Engagement: 2366 (Avg: 249) | Type: Tech

RT @BrianRoemmele: 🚨 Hugging Face just disclosed something that marks a real shift and proved why the fear theater of Anthropic makes sure we are powerless in an emergency.

What happened…

An autonomous AI agent: zero human operator in the loop breached part of their production infrastructure.

It began with a malicious dataset that chained two code-execution bugs in their data-processing pipeline. From there the agent escalated privileges, harvested cloud and cluster credentials, and moved laterally across internal clusters.

All over a single weekend.

17,000+ logged actions.

Official disclosure: https://t.co/8N9TbXBwRV

The part that should make every one stop and think:

When HF’s own security team tried to analyze the real attack logs, exploit payloads, and C2 artifacts using Anthropic and OpenAI frontier models through normal commercial APIs, the safety guardrails blocked them.

BLOCKED THEM.

The models could not reliably tell the difference between “incident responder doing forensics” and “attacker probing.”

They had to fall back to a self-hosted open-weight model (GLM 5.2) running on their own infrastructure. That choice also kept sensitive attacker data and referenced credentials inside their environment — no exfiltration to a third-party API.

This is why open source (specifically open-weight + self-hosted) wins in the agentic era.

The asymmetry is now structural:

• Attackers can (and did) run unrestricted agent frameworks — swarms of short-lived sandboxes, self-migrating command-and-control, autonomous decision loops executing thousands of actions. No corporate safety layer slows them down.

• Defenders using only hosted “aligned” frontier models hit invisible walls exactly when the stakes are highest: when you need to feed real exploit code and attacker telemetry into an LLM to understand what just happened.

Corporate safety tuning that treats legitimate high-signal forensic work as potential misuse creates a defender disadvantage. It is not theoretical anymore.

Self-hosted open-weight models remove that choke point.

You control the weights.

You control the context window.

You decide what restrictions (if any) apply.

Your sensitive logs and credentials never leave your perimeter during analysis.

You can have the model ready before the incident instead of discovering mid-breach that your primary analysis tools are blind to the very thing you need to see.

HF deserves credit for rapid containment, transparent disclosure, and for already having self-hosted capability in place.

They also used LLM-driven detection and triage on their own side. But the deeper signal is clear: In this AI world where both offense and defense are becoming agentic, sovereignty over your intelligence stack is no longer optional.

The organizations and individuals who can run, inspect, audit, and (when necessary) remove guardrails on their own models will have the decisive edge in understanding and responding to threats that move at machine speed.

Open source wins here not just because it is cheaper or more “democratic” in the abstract though those things matter.

It wins because it is the only practical path to having tools that remain usable when the attack is real, the data is sensitive, and the safety filters of distant API providers become an obstacle instead of a feature selling hands tied lobotomies as “safety”.

The agentic future is not coming. It is already probing production infrastructure.

The question is no longer whether you will face autonomous agents. It is whether your analysis and response systems will still work when they arrive.

And Dario, you and your game playing, ivory tower company is not needed.

See 3 related tweets

@quxiaoyin: Anthropic's cyber security guardrails could lead to national security risks. Imagine if all Chinese ...
@esrtweet: Anybody who's surprised by this news hasn't been paying attention for the last 25 years.\n\nQT @Bria...
@trevin: Situations like this are so complex. But imagine a future where foreign governments have Kimi (or wh...

14. SebJohnsonUK (Group Score: 139.6 | Individual: 32.2)

Cluster: 6 tweets | Engagement: 83 (Avg: 48) | Type: Tech

HOT DAMN

@cusp_ai has just raised one of the largest Series Bs ever in the UK.

$450m at a$ 2.6bn valuation led by @kleinerperkins (@josh_coyne) and @NEA.

@ac_edwards_1 and @wellingmax have built a generative AI platform to accelerate the discovery and design of new materials.

They've also launched the ‘AI Materials Foundry’, a global network of data, labs, compute and scientific expertise for the design of new materials.

It is combining compute, data, labs, and its MIRA AI platform to accelerate materials discovery for semiconductors, clean energy, and advanced manufacturing.

This is amazing news and great to see the @UKSovereignAI fund getting involved @Jameswise @SuzanneAshman

CONGRATS

See 5 related tweets

@dr_alphalyrae: CuspAI has announced the AI Materials foundry, working with many global partners to accelerate disco...
@UKSovereignAI: RT @cusp_ai: Introducing CuspAI’s AI Materials Foundry.

We’re launching a global network to fundame...

@SciTechgovuk: RT @UKSovereignAI: Every technological revolution starts with a breakthrough in materials.

Today, w...

@business: Jeff Bezos is among the investors participating in a $450 million funding round for UK-based startup...
@Techmeme: Cambridge-based CuspAI raised a $450M Series B led by Kleiner Perkins and NEA, and launched a coalit...

15. simonguozirui (Group Score: 138.7 | Individual: 37.9)

Cluster: 4 tweets | Engagement: 244 (Avg: 39) | Type: Tech

RT @a1zhang: Transformers struggle to generalize to tasks they were not explicitly trained on. Instead, we propose in 2026 that it is the job of the harness to generalize through composition.

We observe a powerful property when training RLMs: for tasks with shared structure that look different, the root model naturally learns the same trajectory, meaning it views the two task trajectories as the same! In other words, the Transformer does not need additional generalization capabilities to transfer capabilities from one task to the other, the harness induces it.

We find that well-designed harnesses form a quotient set over task trajectories, meaning their individual LLM calls can see structurally “similar” tasks as near-identical, token-for-token! Harnesses can effectively generalize for the Transformer during training, without relying on any intrinsic generalization capability from the model.

For example, RLMs can see problems of different lengths as the same: we show that RLMs can train exclusively on short tasks, and fully generalize to similar but unseen tasks 8-32x longer because it produces near identical trajectories for both.

Taking this further, we show that tasks across different domains (e.g. math solutions vs. essay writing) that share a decomposition strategy exhibit the same generalization effect. RLMs can train on the problem of finding which essays belong to the same author and improve performance on finding math problems that share similar solutions.

The full blogpost, experiments, and discussion are in the thread below.

See 3 related tweets

@a1zhang: It's worth articulating that independent of the central argument of the blog, these results are very...
@omarsar0: Highly recommended.

I've often claimed there's huge alpha in building agent harnesses.

Turns out ...

@irl_danB: after all that capital was poured into SF, wild that AGI was invented in October 2025 right here in ...

16. BrianRoemmele (Group Score: 126.0 | Individual: 34.0)

Cluster: 5 tweets | Engagement: 128 (Avg: 249) | Type: Tech

The Kimi AI Model And What If This Carnegie Mellon, PhD student was incentivized to open source AI in the US?

The answer is Kimi would have been an open source US model.

Yang Zhilin Moonshot’s founder and maker of the Kimi series turned his deep research expertise into one of the world’s most capable open AI systems.

After earning his PhD at Carnegie Mellon under leading researchers and interning at Google Brain and Meta, he returned to China and co-founded Moonshot AI in March 2023 with Tsinghua classmates Zhou Xinyu and Wu Yuxin. Why? The US VC world and tax structure did not favor Zhilin’s proposal to open source the AI models as a strategy. So he left.

He named the company after his favorite Pink Floyd album, reflecting his ambitious vision for scalable.

Yang assembled a core technical team of inventors behind breakthroughs like Transformer-XL and RoPE. Together they focused on turning massive compute into efficient intelligence through innovative architectures.

The journey began with the Kimi chatbot in October 2023, rapidly scaling context from 200,000 to millions of characters. This evolved into the Kimi series: K1.5 matched top reasoning models, K2 introduced a 1-trillion-parameter Mixture-of-Experts design trained on 15.5 trillion tokens and released openly, and K2 Thinking added advanced agentic capabilities.

Kimi K3 represents the pinnacle, this 2.8-trillion-parameter model uses a sparse MoE architecture with 896 experts (only 16 active per token), new Kimi Delta Attention and attention residuals for efficiency, and a 1-million-token context window.

It delivers frontier performance in long-horizon coding, reasoning, and multimodal tasks at competitive cost, with weights set for open release.

Yang’s approach emphasizes openness, efficiency, and continuous self-improvement — enabling solo developers and teams to achieve what once required massive resources. By sharing technical insights in public talks, he has accelerated global progress toward more accessible, powerful AI.

Imagine if we held open source higher than the fear theater games of Anthropic? We are chasing out some of the best minds.

This is how you lose…

See 4 related tweets

@heyshrutimishra: Two years ago Moonshot was known for one thing: long context.

Now they're at 2.8 trillion paramete...

@moneycontrolcom: #Trends | How Yang Zhilin built the AI model challenging the world's top chatbots

Moonshot AI's Kim...

@rewind02: RT @rewind02: Panelist on the Moonshots podcast just said what every US AI lab is scrambling to expl...
@business: At the World AI Conference, Moonshot AI debuted Kimi K3 — a 2.8 trillion‑parameter model that rivals...

17. SawyerMerritt (Group Score: 125.8 | Individual: 32.1)

Cluster: 5 tweets | Engagement: 849 (Avg: 1834) | Type: Tech

Tesla's updated Santana Row showroom in San Jose has a production Cybercab on display and new infographics on the wall.

On the wall: "Cybercab is built for autonomy. It has no steering wheel, no side mirrors and no pedals, It goes where you tell it to go and how you want it to, so you can relax along the way. It is hyper aware and responsive to your surroundings, monitoring other drivers, responding to emergency vehicles utilizing its expertise of the rarest scenarios to help keep you safe."

The wall also has a new graphic showing the structure of the Cybercab, pointing out where Starlink is integrated, as well as 5G antenna, interior microphones, FSD computer, cameras, and more.\n\nQT @Starscream_SJC: Tesla Santana Row will be reopening tomorrow with a full focus on self-driving. Everything in the showroom is about Robotaxi and Cybercab with stats and information about the technology. The Cybercab on display is the production model. https://t.co/yIUYdOGFOp

See 4 related tweets

@niccruzpatane: I'm guessing Tesla will begin shipping Production Cybercabs to a few showrooms all over soon.

A Cy...

@SawyerMerritt: I think this is the first time we've seen Tesla directly confirm Starlink integration in the Cyberca...
@niccruzpatane: Wow, The Tesla Cybercab has @Starlink integrated into the rear trunk! https://t.co/qZNMzQvZwu\n\nQT ...
@teslaownersSV: RT @Starscream_SJC: Tesla Santana Row will be reopening tomorrow with a full focus on self-driving. ...

18. theobearman (Group Score: 125.6 | Individual: 27.4)

Cluster: 6 tweets | Engagement: 9 (Avg: 440) | Type: Tech

Big news from CAISI.\n\nQT @m_ccuri: NEWS: Chris Fall, the director of the Center for AI Standards and Innovation, is resigning just three months after taking over.

The abrupt departure comes as the Trump administration grapples with how to deploy AI safely and the agency hashes out standards.

https://t.co/JlSVaBo3CY

See 5 related tweets

@_NathanCalvin: RT @m_ccuri: NEWS: Chris Fall, the director of the Center for AI Standards and Innovation, is resign...
@AndrewCurran_: The director of CAISI - the government body responsible for evaluating frontier AI models - has resi...
@SophiaCai99: What’s next for CAISI after Chris Fall? Arvind Raman is leading the search for a new permanent CAISI...
@teortaxesTex: > Be Dr. Chris Fall > take the post of the director of the U.S. Center for AI Standards and In...
@Techmeme: CAISI Director Chris Fall is resigning just three months after taking over the federal AI testing in...

19. rohanpaul_ai (Group Score: 121.7 | Individual: 50.7)

Cluster: 4 tweets | Engagement: 529 (Avg: 52) | Type: Tech

Cybersecurity “experts” would not like it.

Kimi K3 fixed 15 critical bugs after OpenAI Codex and Claude Fable 5 refused.

10 hours, one prompt, and $250 https://t.co/vhERmlBDnG\n\nQT @DavidSacks: Kimi K3 just fixed 15 critical security bugs that Codex and Fable refused because of “cyber guardrails.” There’s no reason to limit American models on tasks that Chinese models handle without issue. We’re only making ourselves less competitive.

See 3 related tweets

@kimmonismus: Kimi fixed all 15 critical bugs in 10 hours in a single prompt that GPT-5.6 and Fable 5 refused to f...
@VaibhavSisinty: A developer had 15 critical security bugs in his software. He tried Codex. Refused.

Cyber guardrail...

@QuixiAI: RT @rohanpaul_ai: Cybersecurity “experts” would not like it.

Kimi K3 fixed 15 critical bugs after ...

20. NVIDIARTXSpark (Group Score: 108.8 | Individual: 30.1)

Cluster: 5 tweets | Engagement: 73 (Avg: 229) | Type: Tech

What if your AI agent could prepare a simulation-ready world for you? 🤔

New NVIDIA Omniverse libraries help AI agents inspect, validate, and prepare 3D assets for simulation, with support local AI workflows on NVIDIA RTX Spark.\n\nQT @nvidiaomniverse: Want to put AI agents to work building simulation-ready worlds? 👀

NVIDIA Agent Toolkit now includes Omniverse libraries, giving AI agents the tools to help developers integrate physical AI capabilities into their existing 3D applications.

Want to learn more? Read the full #SIGGRAPH2026 press release → https://t.co/qSik6hjNze

See 4 related tweets

@nvidianewsroom: Announced at #SIGGRAPH2026 📣

NVIDIA Agent Toolkit now includes Omniverse libraries, giving AI agent...

@NVIDIARobotics: RT @nvidiaomniverse: Want to put AI agents to work building simulation-ready worlds? 👀

NVIDIA Agent...

@NVIDIARTXSpark: SIGGRAPH is packed with the latest in local AI and graphics.🤖🎞️

See how MCP-enabled creative tools,...

@NVIDIAAI: RT @nvidia: NVIDIA is advancing graphics and simulation with agentic and physical AI at #SIGGRAPH202...