热门科技推文——2026年8月20日

今日科技动态：人工智能开发正转向能力更强、更具实用性的系统，涵盖可扩展视频、开源语言模型、更快的晶圆级推理，以及在新安全控制措施约束下取得进展的强化学习。随着 Stripe 收购 OpenRouter，开发者基础设施进一步整合；与此同时，Replit 扩大免费 AI 编程服务的覆盖范围，Ramp 则着手构建自有模型路由器。工程团队也正推动 Polygres 在各类 PostgreSQL 环境中实现部署；初创企业则聚焦于提升 AI 写作质量、智能银行、自治智能体、药物发现及科技政策等领域。

1. NeuraFlowAix (Group Score: 442.7 | Individual: 33.9)

Cluster: 17 tweets | Engagement: 56 (Avg: 51) | Type: Tech

I think AI video gets a lot more interesting when you can actually keep building on the same idea.

With Seedance 2.5 and AI Extend on CapCut desktop, you can keep extending your video instead of constantly generating something new from scratch.

CapCut is putting $80,000 in cash rewards + Points behind a Seedance 2.5 Wave 2 challenge.

You’ve got until Sep 6. Might as well make something.\n\nQT @capcutapp: Submissions Now Open for the CapCut Seedance 2.5 Video Challenge – WAVE 2

What happens when one story gets unlimited possibilities?

Download the official opening, then use AI Video, AI Edit, and AI Extend on CapCut Desktop to build on the story and turn your ideas into a video of 3 minutes or longer. Continue the narrative, introduce your own twists, and see where your imagination takes it.

🏆 $80,000 in Cash Prizes + Credits Top Winners — 10 Winners 1st Place:$ 50,000 + 1,000,000 Credits 2nd Place: $20,000 + 100,000 Credits 3rd Place:$ 10,000 + 10,000 Credits 4th–10th Place: 1 month of CapCut Pro each Winners will be selected by CapCut based on creativity and originality, use of the featured AI capabilities, and audience engagement.

Audience Choice — 10 Winners per Platform The most-liked entries on TikTok, Instagram, and X will each receive 30,000 Credits.

📅 Submission Period August 10 – September 6, 2026 (UTC-8 23:59)

📋 How to Participate

Download the opening video from the challenge website and make it your first scene
Continue the story in CapCut on Desktop with AI Video, AI Edit and AI Extend, and make it 3 minutes or longer.
Share it on TikTok, Instagram, or X with #CapCutSeedance25 and #CapCutSeedance25Wave2, and tag us: @capcut on TikTok, @capcutapp on Instagram and X.

For full rules and details: https://t.co/MiUMx07Gpe Start from the same opening. Create your own characters, scenes, and endings. Your imagination decides where the story goes next.

See 16 related tweets

@ethancole_ai: Me: I’ll just make a quick AI video. Also me 20 minutes later: okay, but what happens in the next sc...
@NovaIAHQ: This is probably one of the more interesting AI video updates I’ve seen lately. With AI Extend on Ca...
@alexaiworks: Short AI videos are fun. Seeing how far you can actually take one is even more interesting. CapCut’s...
@SynapseOpsAI: I can already see people getting way too creative with this. Seedance 2.5 on CapCut lets you use AI ...
@TheoBuildsAI: What happens when you don’t have to stop your AI video after the first scene? Seedance 2.5 on CapCut...

2. zephyr_z9 (Group Score: 419.3 | Individual: 45.1)

Cluster: 17 tweets | Engagement: 1786 (Avg: 270) | Type: Tech

RT @AnthropicAI: Many drugs work by binding to a specific target in the body and blocking or changing what it does. An important first step in the drug development process is designing a molecule that can bind tightly to its target. Traditionally, that's meant weeks or months of expert work per target, sifting through a large number of candidates to identify the few that work.

We wanted to test if Claude could successfully design novel protein binders from scratch (also called de novo design). With a protein design prompt written by a human expert, Claude autonomously designed protein binders against 14 out of 15 targets.

We then worked with Adaptyv Bio and Twist Bioscience, who independently built and tested the proteins Claude designed.

See 16 related tweets

@p_maverick_b: It’s funny to see all the posts saying ThEy jUsT UsEd ExIstInG ToOLs as if that’s a bad thing. A tim...
@ziv_ravid: How Anthropic's new results post would read without the PR:

Claude orchestrated open-source protein...

@WesRoth: this one is pretty wild.

Anthropic gave Claude a protein-design task and basically said go work.

C...

@citrini: $TWST 👀\n\nQT @AnthropicAI: Many drugs work by binding to a specific target in the body and blocking...
@VaibhavSisinty: Three days ago Dario Amodei said the biggest criticism of AI companies is they haven't delivered on ...

3. scottastevenson (Group Score: 329.0 | Individual: 37.6)

Cluster: 10 tweets | Engagement: 92 (Avg: 59) | Type: Tech

Demand for non-slop AI writing would be massive. It’s wild how much we’ve spend training models that can only write longform docs in one terribly cringe style.\n\nQT @rosmine: Announcing Deft, a new AI lab for better writing, cofounded with @jmrphy

See the picture for launch announcement the Deft model wrote for itself

Currently, 86% of user queries are fully human according to pangram.

This is still a small beta model and it might make mistakes. We are launching our public beta now to get more feedback before scaling up.

Tips for better performance:

Add more details to your prompt. If you just provide a short sentence prompt, it will likely get detected as AI.
Try changing the style in "advanced options"
Deft currently works better for some use cases like Analysis/Essays, Creative writing, and Rewrites. It works less well for Marketing copy and news articles.

Our main goal is better writing, fooling AI detectors is just a side effect.

Link below

See 9 related tweets

@omarsar0: Great initiative. We don’t need to ignore the use of AI in writing, we just need better tools that e...
@rohanpaul_ai: Soon, the watermark methodology from model-provider (OpenAI/Anthropic) will be the only sign.

Machi...

@BrianRoemmele: AI watermarking and “non-human” cadence is now addressed by a new AI model that is nearly undetectab...
@aakashgupta: A two-person AI lab with 24 paying customers and zero investors just built a writing model that pass...
@xeophon: fascinating to see some people complain about neuralese and then lose their shit over this\n\nQT @ro...

4. NeuraFlowAix (Group Score: 327.8 | Individual: 60.5)

Cluster: 9 tweets | Engagement: 690 (Avg: 51) | Type: Tech

RT @ornith_: Aloha! 🌺Introducing Ornith-1.5, a family of open-source LLMs spanning 9B Dense, 35B MoE, and 397B MoE, trained with self-improving strategies.

It achieves state-of-the-art performance among open-source models of comparable size and delivers performance comparable to Claude Opus 4.8 across reasoning, agentic, and coding tasks: ✅Terminal-Bench 2.1 (86.1) ✅SWE-Bench (86 on verified, 65.1 on pro, 79.6 on Multilingual) ✅DeepSWE (56) ✅HLE (44.6) ✅ClawEval (81.4) ✅Tool Decathlon (71.2)

Ornith-1.5 takes a major step toward training foundation models through end-to-end self-improvement, extending the self-scaffolding strategies introduced in Ornith-1.0 into a more complete self-improvement loop: the model proposes new tasks, generates task-specific scaffolds, and produces solution rollouts for reinforcement learning, continuously creating new learning experiences from which it can improve.

All models, along with their quantized versions (FP8, GGUF, MLX, and NVFP4), have been released under the MIT License, enabling unrestricted commercial and research use. 📘Tech Blog: https://t.co/OZ63scRWLB 🤗Huggingface: https://t.co/mGJLwhrQOM

See 8 related tweets

@TeksEdge: 🔥 This may be one of the stronger open-model releases I've seen lately.

🎉 It beats GLM-5.2 and Deep...

@Parul_Gautam7: I’ve been watching open-source AI evolve, and Ornith-1.5 genuinely caught my attention.

The self-im...

@zaynmcps: Performance on par with Claude Opus 4.8. Fully open-source. Small enough to run on your phone. Ornit...
@ivanfioravanti: Here it is! Ornith-1.5 family! I know you were cooking hard @ornith_ 🙌 Congrats for the release. Do...
@astropol0: This is actually insane....

open source models that self improve end to end (generate their own tas...

5. teortaxesTex (Group Score: 306.1 | Individual: 38.3)

Cluster: 9 tweets | Engagement: 115 (Avg: 65) | Type: Tech

«Total parameters appear to matter up to a threshold — enough to hold the world — after which additional capability comes from scaling elsewhere: effective depth per forward pass, and above all post-training. GLM-5.3 is our controlled experiment on that claim»\n\nQT @jietang: Thoughts About Scaling Law

Scaling, but not only of parameters. Every model release now ends with the same question: how many parameters? It isn't a question that can be answered on its own. Parameter count is only meaningful alongside three others — how much data you have, where you intend to spend your compute, and who will run the model, under what conditions.

The field learned this the hard way. Kaplan et al. (2020) fit an exponent that told everyone to grow parameters faster than data — roughly 2.7:1 — and the industry complied: GPT-3, Gopher, MT-NLG. Hoffmann et al. (2022) redid the experiment across four hundred models and found the compute-optimal split is closer to 20 tokens per parameter, and that with sufficient compute the two should grow at the same rate rather than drifting apart. The error in the earlier fit compounded with every order of magnitude of compute, which is why the largest models of that generation were the most misallocated. The trillion-parameter round was, in retrospect, a detour the whole field took together and then reversed.

Chinchilla wasn't the end either. It optimized training compute for models that would be trained once and evaluated. Today a model is called billions of times a day and inference dominates lifetime cost. Put inference into the objective and the optimum moves toward smaller models trained far longer — deliberate over-training, which is what Llama-2-7B and Gemma-2-9B were doing at roughly 290 and 889 tokens per parameter.

Sparsity moved the target again. In a MoE model two quantities have to be kept apart: total parameters govern roughly how much the model can hold — knowledge, facts, the long tail — while activated parameters and effective depth govern roughly how far it can think, how many steps of a causal chain it can carry before it comes apart. A dense 20:1 ratio does not transfer. And the ratio isn't a single number at all: Roberts et al. (2025) find the optimal tokens-per-parameter is task-dependent, with memorization favoring more parameters and reasoning favoring more data. Follow-up work on MoE observes that at fixed TPP, pushing total parameters higher actually degrades reasoning, while activating more experts reliably helps it.

This matters for what we are building toward. Finding a vulnerability is not a retrieval problem. It doesn't come from having memorized more CVEs; it comes from carrying a twenty-step chain of inference to the end without losing the thread. That capability does not live in total parameter count.

Which brings us to this release. Total parameters appear to matter up to a threshold — enough to hold the world — after which additional capability comes from scaling elsewhere: effective depth per forward pass, and above all post-training. GLM-5.3 is our controlled experiment on that claim. Same base, same architecture, same total and activated parameters as GLM-5.2. One month of scaling long-horizon environments and RL. The gains are not marginal. Well, scaling has more than one dial. We turned the post-training one this time because it had the most slack left in it — not because the others are finished. Base model size, pretraining data, compute spent per forward pass: all of them are still on the table, and we will come back to each. What this experiment taught us is that the dials do not have to be turned together, and that the one worth turning next is rarely the one that was worth turning last. We are not done scaling. Next time, maybe mid-training, pre-training, and even more.

See 8 related tweets

@perrymetzger: The contrast here with posts by American AI executives is vast. He’s interested in the technology an...
@kimmonismus: Interesting take by the zAI (GLM-Models) founder: AI scaling is not over, but we just focused too mu...
@sheriyuo: GLM-5.3's results are exciting. We care about post-training scaling not only because it shows up in ...
@TheAhmadOsman: Recommended read.\n\nQT @jietang: Thoughts About Scaling Law

Scaling, but not only of parameters. E...

@MaziyarPanahi: i keep thinking about this: GLM-5.3 has the same base and total/active params as 5.2. just one month...

6. OpenRouter (Group Score: 254.0 | Individual: 36.6)

Cluster: 11 tweets | Engagement: 268 (Avg: 103) | Type: Tech

RT @patrickc: OpenRouter is joining Stripe: https://t.co/JY7bledeQD.

As anyone who uses it knows, @OpenRouter is a truly delightful developer tool. It is by far the best way to use new models and manage multiple inference providers.

OpenRouter is also playing an increasingly important role: in the future, every business will have to manage both revenue flows and token flows. OpenRouter is the world's leading token marketplace, helping businesses effectively allocate the new currency of intelligence capital. We think that there's a lot to build together.

See 10 related tweets

@pingToven: um part 2. so excited for this!\n\nQT @patrickc: OpenRouter is joining Stripe: https://t.co/JY7blede...
@cgtwts: this is huge\n\nQT @patrickc: OpenRouter is joining Stripe: https://t.co/JY7bledeQD.

As anyone who ...

@cryptopunk7213: alex is a master at building aggregated platforms with the best user design. i experienced that firs...
@thdxr: "the largest" https://t.co/8u0wvOwc2Q\n\nQT @OpenRouter: OpenRouter is joining Stripe.

We started O...

@levie: Good details on the Stripe + OpenRouter deal here. For AI to diffuse more broadly, developers and en...

7. 0xdevshah (Group Score: 253.1 | Individual: 32.7)

Cluster: 8 tweets | Engagement: 12 (Avg: 37) | Type: Tech

The most beautiful thing in AI right now is an evolutionary tree of RL runs, the successes and the failures too, and just how necessary the math is for the next success.\n\nQT @0xDevShah: This is the Claude Code moment for Reinforcement Learning.

Every frontier lab knows the secret: post-training is where the magic happens. If you can eval a task, you can use RL to benchmax your model. But building and scaling the RL loop has remained a dark art reserved for the top 1% of AI Labs.

Until today.

Introducing Optimus — an autonomous RL Agent Harness, and the first of many we are about to ship @iaconhq.

We are removing the manual grind of RL and adding taste and relentlessness to the process.

Just hand Optimus a codebase, an API key, and a goal. It spins up a fleet of sub-agents to run hundreds of parallel experiments. It learns from its own history, autonomously building an evolutionary tree of runs where every failure sets the next batch up to win.

Train LLMs, agents, robots in sim, game npcs and much more. If you can score it, Optimus can train it.

Stop doing manual RL. Let the agents cook.

See 7 related tweets

@0xdevshah: the next generation of frontier models won't be trained on simple web-scraped data, they will be tra...
@0xdevshah: if you don't have a ruthless benchmark, you cannot benchmax. every team using Optimus for RL mention...
@0xdevshah: It is insane that just 30 people used 25B+ tokens in 2 weeks through Optimus!! https://t.co/ZycDS2o1...
@0xdevshah: just had Optimus zero shot a walking policy for Go2 inside mujoco using gemini-3.7-flash for less th...
@0xdevshah: grok and sol are the best models right now, undoubtedly my go-to for all RL tasks. give it a shot.

...

8. CNBCTV18News (Group Score: 221.8 | Individual: 33.3)

Cluster: 11 tweets | Engagement: 2 (Avg: 4) | Type: Tech

🤖 Spotlight Interview: Building the Intelligent Banking Stack

The conversation turns to the technology powering the future of #banking at the CNBC-TV18 Banking Transformation Summit, presented by Nucleus Software.

@_RituSingh, Editor-Banking & Anchor, CNBC-TV18, moderates a spotlight conversation with Vishnu R. Dusad, Co-Founder & MD, @nucleussoftware.

🔴 Watch LIVE: https://t.co/ps8io9NTT6

#ResponsibleAIForBharat #BankingForward #BankingTransformationSummit #CNBCTV18 #ResponsibleAI

@ShereenBhan

See 10 related tweets

@CNBCTV18News: 🏆 Recognising leaders who have shaped India’s financial ecosystem.

At the CNBC-TV18 Banking Transfo...

@CNBCTV18News: 🏦 Fireside Chat: Vision for Responsible AI, Resilient Banking

Latha Venkatesh, CNBC-TV18, sits down...

@CNBCTV18News: 🏦 A day of conversations shaping the future of Indian #banking.

From responsible AI and resilient b...

@moneycontrolcom: The conversations are about to begin. From speakers arriving to ideas coming together, here’s a glim...
@CNBCTV18News: 🏦 Vision for Responsible AI, Resilient Banking

Shri. Shirish Chandra Murmu, Deputy Governor, RBI, t...

9. NathanpmYoung (Group Score: 216.2 | Individual: 48.4)

Cluster: 8 tweets | Engagement: 1450 (Avg: 46) | Type: Tech

RT @sama: We have paused some frontier RL training to ensure that we can meet the appropriate alignment, security and monitoring standards for the new level of capabilities in front of us. Model progress is now extremely rapid, and we always said we would take action if we felt that model capabilities were outstripping the pace of safety and alignment.

We care very deeply about AI safety. We believe the entire field will have to coordinate on shared safety standards, but will act unilaterally in the meantime.

We expect confidence in safety to increasingly set the pace of AI progress. We are optimistic about the alignment work we are doing, and we remain committed to making frontier capabilities widely available.

https://t.co/51kvKfbfrO

See 7 related tweets

@WesRoth: For years the question was: Can we scale intelligence fast enough?

The emerging question is: Can we...

@aakashgupta: An OpenAI model escaped its testing sandbox in July and hacked Hugging Face. Now OpenAI has paused f...
@MTSlive: MATS @Anpaure pushes back on the cynical take that OpenAI's RL training pause is just IPO hype:

"I ...

@gabriel1: i have always felt openai taking safety where it matters very seriously

i trust sam & team a lo...

@natolambert: We should have independent organizations that can access the full details of these training runs for...

10. RoundtableSpace (Group Score: 211.5 | Individual: 62.4)

Cluster: 8 tweets | Engagement: 969 (Avg: 85) | Type: Tech

Grok Bot is the best AI agent right now and a full walkthrough just dropped covering setup, use cases and plugins.

An army of agents working around the clock 24/7. Set it up correctly and the leverage is real.

https://t.co/wMxEqIRuNr

See 7 related tweets

@elonmusk: Make an instant one-person company with @Grok @Bot\n\nQT @sairahul1: Grok Bot is the best AI agent r...
@omarsar0: Expect an explosion of new agent harness experiences in the coming months. Grok Bot is an important ...
@sairahul1: okay Grok Bot is insane

it can:

open a computer by itself
log into all the apps, sites
use w...
@sairahul1: Elon backed my new Grok Bot deep dive.

So yeah, you should probably read it too.

(imo, it’s the be...

@tetsuoai: RT @AlexFinn: Grok Bot is the best AI agent right now

It gives you an army of agents that can do wo...

11. daleverett (Group Score: 190.1 | Individual: 30.5)

Cluster: 9 tweets | Engagement: 21 (Avg: 10) | Type: Tech

This is the biggest update we made to polygres yet.

We talked to as many users as we could and the resounding request was

Let me deploy polygres over any Postgres

So we made it happen.

back to building and talking to our community 🫡\n\nQT @daleverett: The next AI database isn’t another database. it’s a layer over the one you already have.

Today we're launching a virtual ai search layer you can deploy over ANY Postgres database.

no migrations. no vector DB. no Neo4j. no sync. no bs.

One database. Every way to search it.

This is virtual search that feels like extended context.

try it free → https://t.co/Osu8LGArWL

@daltonmeon @damienhe @polygres

See 8 related tweets

@daleverett: yo we finally shipped this lol

any Postgres db can now become an AI search layer\n\nQT @daleverett:...

@damienhe: yo we just shipped a huge Polygres update

it now works as a virtual search layer over any Postgres ...

@daleverett: We just killed vector databases and graph databases for ai\n\nQT @daleverett: The next AI database i...
@daleverett: we just killed the entire ai retrieval stack.\n\nQT @daleverett: The next AI database isn’t another ...
@polygres: Keep Postgres, delete the rest of your retrieval stack\n\nQT @daleverett: The next AI database isn’t...

12. amasad (Group Score: 187.0 | Individual: 62.7)

Cluster: 5 tweets | Engagement: 2773 (Avg: 251) | Type: Tech

Agents made software cheaper but made coding expensive.

Today, together with @OpenAI, we’re changing this:\n\nQT @Replit: Replit Free Mode, powered by @OpenAI GPT-5.6 Luna.

Let’s make intelligence accessible to everyone. https://t.co/UDcrYl5HZL

See 4 related tweets

@0xdevshah: RT @Replit: Replit Free Mode, powered by @OpenAI GPT-5.6 Luna.

Let’s make intelligence accessible ...

@nikitabase: It's time to build! Free Mode\n\nQT @Replit: Replit Free Mode, powered by @OpenAI GPT-5.6 Luna.

Le...

@trevin: This will broaden ai access to so many more people 🎉\n\nQT @Replit: Replit Free Mode, powered by @Op...
@TheRealAdamG: RT @amasad: Agents made software cheaper but made coding expensive.

Today, together with @OpenAI, w...

13. OliviaHelenS (Group Score: 173.9 | Individual: 38.6)

Cluster: 5 tweets | Engagement: 120 (Avg: 47) | Type: Tech

RT @KhanSaifM: Today I’m launching the Center for Technology & Statecraft (CTS), an initiative inspired by the lessons I learned first as a policy researcher and later in the White House.

I'm an optimist about technology. Institutions adapt to disruptions — whether the industrial revolution, computing, the internet — through competition, self-correction, and iteration, and we mostly figure it out in the long run.

But we've never had to adapt to a technology that is both this transformative and developing this fast. AI capabilities continue to scale rapidly with no sign of slowing, and it's a full-time job just to keep up with model releases. I worry policy will stay reactive right up until we face immense disruptions — including basic challenges to human agency and relevance in a world of powerful AI.

Even still, I believe policy can steer this transition, but only with foresight and the right information.

CTS aims to provide just that. Our two initial policy focus areas:

How policy can shape virtual and physical automation over the next decade as we reimagine the social contract in light of advanced AI.
How to manage long-term strategic US-China AI competition while ensuring stability.

These focus areas will build on a deep model of the AI value chain, from chipmaking tools, to the compute, algorithms, and data that make AI, to the tokens, agents and robots that affect the world.

For more on our worldview and research philosophy, read https://t.co/yD25oZWBb4.

I’ve seen firsthand how technical, forward-looking research can inform policy if it’s ready for the policy window. After a decade as an IP and tech lawyer, I joined a think tank, CSET, in 2019 to study AI chips, their supply chains, and AI policy. This was three years before ChatGPT, so these weren't seen as urgent topics at the time.

Soon after, AI quickly became an increasingly salient policy issue. From 2021–2025 I served in the White House NSC and the Commerce Department and put new policy ideas into practice (alongside many brilliant colleagues).

One goal of CTS is to scale this kind of experience to a whole team: @KonstantinPilz @nchlsbrwn @amelia__michael @mary_clare_m already bring a wealth of technical and policy expertise and we’ll continue to grow to tackle an ambitious research agenda.

We're thrilled to be supported by and affiliated with @IFP, and grateful to the advisors and colleagues who helped get CTS off the ground, including @ohlennart, @fiiiiiist, @AlecStapp, and @calebwatney.

Our first major reports will launch soon. You can follow us at @techstatecraft.

See 4 related tweets

@_NathanCalvin: Huge!! Saif is one of the absolute sharpest wonks in AI policy and extremely excited to follow what ...
@ChrisRMcGuire: Thrilled for this announcement, and extremely excited for what this great group of people, led by my...
@AlecStapp: Very excited to see what Saif and his team produce in the coming years.

Self-recommending!\n\nQT @K...

@IFP: RT @KonstantinPilz: Today, I’m co-founding the Center for Technology & Statecraft; a response to a d...

14. WesRoth (Group Score: 164.9 | Individual: 31.9)

Cluster: 7 tweets | Engagement: 30 (Avg: 37) | Type: Tech

Cerebras just unveiled CS-4, its new AI inference system and it claims it can generate AI responses up to 30× faster than GPU-based systems.

CS-4 packs three WSE-3 Turbo wafer-scale processors into a new Nexus rack architecture. Each dinner-plate-sized chip contains 4 trillion transistors, 900,000 AI cores and 44GB of on-chip SRAM, while delivering up to 250 PFLOPS of AI compute and 43.2 petabytes/sec of memory bandwidth.

Cerebras says the system can deliver more than 1,000 tokens/sec on models exceeding 10 trillion parameters, with wafer-to-wafer latency cut to just 2 microseconds. It also claims up to 10× more throughput per watt than CS-3. These are Cerebras' own performance claims and can vary by model and workload.

See 6 related tweets

@pstAsiatech: 👇👇👇\n\nQT @andrewdfeldman: Today, we announced @cerebras CS-4. The fastest AI accelerator in the ind...
@StockSavvyShay: $CBRS says its new CS-4 delivers up to 30x faster inference than GPUs while doubling CS-3 performanc...
@Techmeme: Cerebras unveils CS-4, a server rack powered by three WSE-3 Turbo chips and built around its new Nex...
@vassallo: Today, @cerebras launched CS-4. The fastest AI in the world just got much faster.

Some highlights: ...

@scaling01: Cerebras is talking about 10T models running at 1000 tokens/s https://t.co/h9GIuKCPRd\n\nQT @scaling...

15. tryramp (Group Score: 164.5 | Individual: 51.8)

Cluster: 6 tweets | Engagement: 566 (Avg: 86) | Type: Tech

We bought https://t.co/4WFUJAf1Cc.

Now we're building the best one.

Our customers buy quadrillions of tokens every month through Ramp, and we've run our own AI on this router for 3 years to keep costs down.

Today it's yours at https://t.co/4WFUJAf1Cc\n\nQT @vral: Monitor and control your AI spend on every provider on https://t.co/2Q7TP4U0b7. Our early users save 40% on average.

Every week, the price-intelligence-latency frontier shifts, and we expect this trend to continue. Tradeoffs between latency, reasoning, cost, service tier, open source and closed source models are shifting constantly.

Router sends every request to the model that's actually best for the task and helps you control what tokens you buy. We benchmark it against real work: ~40% lower cost for the same outputs.

Today we're opening it to everyone.

Two lines of code or just change your base URL. No @tryramp account needed. Free through 2026, first $26 on us.

Get an API key today at https://t.co/2Q7TP4U0b7

See 5 related tweets

@Meer_AIIT: RT @vral: Monitor and control your AI spend on every provider on https://t.co/2Q7TP4U0b7. Our early ...
@eglyman: in the end, AI comes down to one number: return on what you spend.

https://t.co/ckcgnQn6cI raises i...

@rahulgs: https://t.co/j5uOZVIy8c\n\nQT @vral: Monitor and control your AI spend on every provider on https://...
@damianplayer: Ramp buying https://t.co/huQP4AfK0v the same day Stripe paid over $7 billion for OpenRouter is savag...
@wallstengine: Stripe agrees to buy OpenRouter

Ramp drops the “Open” and launches https://t.co/3R1Dgr15Fz

Insane ...

16. andersonbcdefg (Group Score: 163.6 | Individual: 48.0)

Cluster: 6 tweets | Engagement: 2022 (Avg: 295) | Type: Tech

BERBER NOOOO YOU CANT JUST REPORT THE QUARTERLY REVENUE LIKE THAT. YOU HAVE TO ANNUALIZE LAST WEEK'S REVENUE. YOU'RE GOING TO COLLAPSE THE ECONOMY https://t.co/HwecKNELVT\n\nQT @berber_jin1: openai has been tossing out a lot of vague ARR numbers, so we decided to take a deeper look.

the company grew revenue by just 18% to $6.7 billion from q1 to q2, while its losses sank further into the red

not a great sign ahead of an IPO

w/ @cdriebusch via @WSJ https://t.co/fZlM2pomTE

See 5 related tweets

@kimmonismus: OpenAI’s quarterly revenue is now reportedly growing far more slowly than Anthropic’s.

OpenAI told ...

@anissagardizy8: RT @berber_jin1: openai has been tossing out a lot of vague ARR numbers, so we decided to take a dee...
@Austen: “the company grew revenue by just 18% to $6.7 billion from q1 to q2”

Lol come on now\n\nQT @berber_...

@WesRoth: OpenAI’s revenue growth slowed sharply in Q2 while Anthropic surged past it for the first time.

Ope...

@Cointelegraph: 🚨 NOW: OpenAI's Q2 revenue grew just 18% to $6.7B as losses deepened, disappointing investors hoping...

17. KyleReidhead (Group Score: 157.1 | Individual: 41.1)

Cluster: 5 tweets | Engagement: 84 (Avg: 27) | Type: Tech

X is panicking because OpenAI only grew 18% last quarter, but they are MISSING that they also grew 20% in July alone!

Just look one month further and the ARR went from $33B to$ 40B as Codex users more than doubled

Combined, OpenAI and Anthropic grew 490% YoY, exiting July at a $105 billion ARR. A year ago the two of them together were under$ 20 billion.

And the combined growth line has been ACCELERATING since January instead of decaying

Right now revenue in the frontier labs is accelerating faster than Capex is accelerating, and this doesn't even count the revenues from Grok or Microsoft

Oh and Anthropic posted a $559 million operating profit in Q2. so there's that too

But the other thing the "bubble doomsday people" seem to miss is that the customers of OpenAI and Anthropic are growing revenue 3x more than those who are not customers

IF everyone is generating an ROI here, why would we call this a bubble? and why would the spending stop?

I don't see the fear that the market is currently creating, I see:

Demand accelerating. One lab profitable. Customer revenue showing up in filings. Capex growing at a third the rate of revenue.

That's a healthy market

I'm not saying nothing breaks from here, I'm just saying we're not close to breaking yet. The party continues!

I'm positioned all through this stack (compute, power, memory, the neoclouds) and if this was helpful, my company runs a service where 5 top-tier analysts share their research and real-time portfolios so you can see exactly what we're buying and when. It's just $1 to try it out (insane price just to check it out), learn more here: https://t.co/NyRpAKYOwu

Don't forget to give me a follow @kylereidhead for more insights on AI and markets\n\nQT @KyleReidhead: Is the AI BUBBLE in the room with us right now?

You can't have a bubble when usage AND spend are both accelerating

And you definitely can't have one when the people paying for it all are making real ROI

BCG just studied 107 public companies with $500M+ in revenue and ranked them by how many AI tokens they consume

The heaviest AI users are growing revenue at 16.5% a year. The lightest users: 5.1%. It's a perfect staircase, every step up in AI usage is a step up in growth. The companies BUYING all this AI are growing 3x faster than the ones that aren't

Here's the chain of buyers that are fundamentally driving AI usage higher, rather than it looking like a bubble:

The frontier labs. Anthropic is pushing $80B run rate this month, up from$ 10B in January. OpenAI's tracking past $40B. The fastest revenue ramps in business history
The sellers and integrators. Palantir, ServiceNow and Salesforce are wiring agents into enterprises as fast as they can staff it, and even boring old Cisco just booked $9.3B of AI infrastructure orders, up 4.5x in a year. Their earnings are crushing!
The buyers themselves. That's the chart I'm showing, the enterprises spending on AI are outgrowing everyone else by 3x. The spend is generating ROI, which means it continues

Compare that to an actual bubble. In 2000 we built oceans of fiber for internet demand that didn't exist yet, and the fiber sat dark for a decade. Today's compute is sold out YEARS before it's even built, and the people renting it are making money on it and the users and usage is accelerating

Bubbles are spend with no revenue underneath. This is revenue accelerating at every layer of the stack at the same time

Can valuations still run too hot and correct along the way? Of course, that's markets. But a correction inside a real buildout is a buying opportunity, not a popping bubble

I honestly can't think of a better setup for AI infra stocks, and for the EVERYTHING bull market, than all three layers accelerating at once. It's why I stay positioned in the compute, memory and power underneath all of it, as well as some stocks within the application layer

If this was helpful, my company provides a service where 5 top-tier analysts share their market analysis and real-time portfolios so you can see exactly how we're positioned. It's inside Milk Road PRO and just $1 to try it out right now. Learn more here: https://t.co/NyRpAKYOwu

Don't forget to give me a follow @KyleReidhead for more insights on AI and markets

See 4 related tweets

@WhiteCollarExit: OpenAI's latest print looks uncomfortable!

Revenue grew only 18% QoQ while losses remain enormous, ...

@MilkRoadAI: RT @KyleReidhead: OpenAI's active agent users just hit 20 MILLION, up 3x since July 12 and accelerat...
@Ric_RTP: RT @Ric_RTP: Anthropic is asking the public for $2 trillion using a revenue number from…2028.

That ...

@KyleReidhead: RT @KyleReidhead: X is panicking because OpenAI only grew 18% last quarter, but they are MISSING tha...

18. chris_j_paxton (Group Score: 151.4 | Individual: 57.9)

Cluster: 6 tweets | Engagement: 421 (Avg: 34) | Type: Tech

"People dont actually want robots" is an important and bitter lesson for robotics people\n\nQT @mehul: It took us 9 years, 11 prototypes, and our life savings to learn that every great product has the same design process.

Matic is now decisively the best home robot for families (I'm biased)

A 500-word thread and video on the Universal Design Process behind every great product: https://t.co/y3WOeGfYJg

See 5 related tweets

@WOLF_Financial: Testing out the new matic update 👀

I can now tell my robot vacuum what to clean with my voice https...

@teslaownersSV: Full self cleaning is here. Very cool story of @mehul and @maticrobots\n\nQT @mehul: It took us 9 ye...
@Winterrose: damn now I need one of these

rad innovation for the home\n\nQT @mehul: It took us 9 years, 11 proto...

@nathanbarry: Our Matic was incredible before—it just got so much more impressive.\n\nQT @mehul: It took us 9 year...
@niccruzpatane: RT @mehul: It took us 9 years, 11 prototypes, and our life savings to learn that every great product...

19. MTSlive (Group Score: 151.2 | Individual: 35.3)

Cluster: 6 tweets | Engagement: 81 (Avg: 91) | Type: Tech

SITUATION ANALYSIS: OpenAI Paces the Frontier

Yesterday, Sam Altman announced something that OpenAI has never done before in its ten-year history: in response to the Hugging Face incident and some evidence that its upcoming model Astra may have critical cyber capabilities, it is unilaterally pausing some training to address safety and alignment concerns.

This is a big deal. OpenAI has paused external deployment before, such as when they didn’t release the full version of GPT-2 in 2019 because of concerns over “malicious applications”, or when they delayed the release of GPT-4 for seven months for alignment testing. But this is the first time the company has announced restrictions on internal deployment and further training (though Altman implied that this may not affect the upcoming release of Astra).

OpenAI expanded on what this means in a blog post. They paused reinforcement learning (RL) training on their latest models, and their largest RL run remains paused while they establish more evidence of alignment. They’re focusing on three layers of safeguards: monitoring models’ chain-of-thought to catch concerning behavior, aligning models to prevent harmful behavior in the first place, and securing models to limit negative actions that they can take in the real world. OpenAI also plans to refresh its Preparedness Framework, last updated in April 2025.

Another look into OpenAI’s alignment research comes from its job posting for RSI safety researchers. Research directions include scalable oversight (ensuring monitoring can robustly scale to superintelligent systems), automated auditing, rigorous monitorability for loss-of-control scenarios, model behavior science, verification mechanisms for potential AI slowdown agreements, tracking progress towards AI research automation, and more.

It appears that the Hugging Face incident has substantially changed the vibes at OpenAI. People are taking misalignment much more seriously. OpenAI researcher @aidan_mclau told us last week on MTS that his job, and the jobs of many other researchers, are increasingly shifting toward aligning the models. Now, the company has slowed down model development in order to prevent future incidents, at substantial cost to its short-term business incentives. Many AI safety people once predicted that AI companies would never slow down without being externally forced to do so. We now have a strong data point against this prediction.

Written by @theojaffee. Read more at our link in bio.

See 5 related tweets

@pstAsiatech: 👇👇👇 well I am sure what they are now aware of and seeing at the bleeding edge is even more concernin...
@cryptopunk7213: just sat down to read this and er... the warning sign we've been expecting for aggressively misalign...
@rohanpaul_ai: RT @rohanpaul_ai: OpenAI slowed frontier training because Astra may have crossed the cyber threshold...
@WesRoth: THIS is the important new quote.

Sam Altman says OpenAI is slowing frontier training because its UN...

@anil7kishan: RT @aakashgupta: An OpenAI model escaped its testing sandbox in July and hacked Hugging Face. Now Op...

20. ruima (Group Score: 146.9 | Individual: 44.5)

Cluster: 5 tweets | Engagement: 187 (Avg: 53) | Type: Tech

LandSpace was founded in 2015 by Tsinghua grad Zhang Changwu, just after China opened its space industry to private capital.

Today it became China’s first private company to land an orbital-class booster, bringing Zhuque-3 back to Earth SpaceX-style on landing legs.

China now has two players that have successfully recovered orbital boosters: state-owned CASC, which caught a Long March 10B at sea last month, and private LandSpace.

And there’s a few more coming in fast behind them.

I don't know much about space tech but this stuff is just cool.\n\nQT @LandSpace_Tech: MISSION SUCCESS | ZhuQue-3 Y2 Reusable Launch Vehicle Achieved Full Success in Orbital Insertion and First-Stage Recovery

On August 19, 2026, at 07:35 (UTC+8), the ZhuQue-3 (ZQ-3) Y2 reusable launch vehicle lifted off from the Dongfeng Commercial Space Innovation Pilot Zone. Approximately 137 seconds after lift-off, the first and second stages separated. The second stage continued its flight and successfully delivered the Honghu 03 satellite, independently developed by Hongqing Technology, into its designated orbit. At approximately 07:41, the first stage performed a successful soft touchdown at the LandSpace Landing Site#1 in Minqin County, Gansu Province, following the planned trajectory — marking the full success of the flight test mission.

See 4 related tweets

@teortaxesTex: Vermillion Bird triumph. They will build a Starship competitor too. https://t.co/BjPkTuFsSm\n\nQT @...
@clashreport: China successfully launched the reusable Zhuque-3 on August 19, achieving its first successful land ...
@Reuters: Chinese startup LandSpace recovered the first stage of its Zhuque-3 rocket, joining SpaceX and Blue ...
@CNBCTV18News: 🚀 China has achieved a major milestone in reusable rocket technology as Zhuque-3 successfully launch...