科技热门推文——2026年8月15日

今日科技动态：人工智能在编程、网络安全和多模态智能体等领域加速发展，GLM-5.3、开放权重的 Qwen3.8 系列模型、Gemini 3.7 Flash，以及小红书的长上下文模型 dots3-note 均引发关注。SpaceX 收购 Cursor，凸显市场对智能体工程工具日益增长的需求；与此同时，据报道，苹果正与阿里巴巴合作开发面向中国市场的模型。此外，人工智能驱动的机器人与视频编辑技术取得新进展，MongoDB 加大对初创企业开发者的推广力度；围绕 OpenAI 营收、实时 API 定价，以及日益普及的人工智能生成广告，相关讨论也愈发激烈。

1. FireworksAI_HQ (Group Score: 1533.4 | Individual: 58.1)

Cluster: 51 tweets | Engagement: 4368 (Avg: 252) | Type: Tech

RT @Zai_org: Introducing GLM-5.3: Built to Code. Ready for Cyber Defense.

Top-tier coding and agentic capabilities, achieved through post-training on the 743B base model
A major leap in cybersecurity, setting a new standard among open models

Tech Blog: https://t.co/ekQkO83jCv https://t.co/y3Y2AB0wxr

See 50 related tweets

@EMostaque: Something has happened with post-training as shown by DeepSeek flash & GLM-5.3 updates.

Same ba...

@kimmonismus: WHAT: Zai just launched GLM-5.3, and its biggest leap may be in cybersecurity.

The 743B base model ...

@teortaxesTex: oh brother https://t.co/cZL4YRnCTU\n\nQT @Zai_org: Introducing GLM-5.3: Built to Code. Ready for Cyb...
@rileybrown: GLM 5.3 was just released!

We have an action packed @agentnative_ weekly update planned tomorrow. ...

@haider1: GLM-5.3 is here

i had early access to it, and the first thing that stood out was how much better it...

2. a16z (Group Score: 964.3 | Individual: 60.8)

Cluster: 30 tweets | Engagement: 6654 (Avg: 581) | Type: Tech

RT @cursor_ai: Cursor is now part of @SpaceX.

Today, we have officially closed our acquisition. We will join the @SpaceXAI team to help make Grok the world's most useful AI and improve Grok Build, Grok Bot, Grok API, Cursor, and more.

SpaceX has built some of the most inspiring and impressive technology in the world, and we’re grateful for the opportunity to become part of such a special company. Onwards.

See 29 related tweets

@Rasmic: The Cursor badge is gone.

Again, what an insane run. Congrats to the entire cursor team\n\nQT @cur...

@SpaceXAI: Welcome to the SpaceXAI team!

Cursor will help us move faster on the path to the world's most usefu...

@rohanpaul_ai: SpaceX Completes $60 B acquisition of Cursor.

So now SpaceX owns the entire stack, model, compute, ...

@kimmonismus: SpaceX has officially completed its $60 billion acquisition of Cursor. The Cursor team is joining Sp...
@leerob: Big day! Cursor has officially joined SpaceX.

I’m proud of our team and what we’ve accomplished so ...

3. victormustar (Group Score: 738.6 | Individual: 60.1)

Cluster: 22 tweets | Engagement: 2954 (Avg: 275) | Type: Tech

RT @Alibaba_Qwen: We promised open weights for Qwen3.8. Now, time to meet them! 🎉

⚡ Qwen3.8-27B:

A native multimodal dense model. With just 27B parameters, it outperforms Qwen3.7-Plus overall and shines in real-world coding & office workflows.
262K native context, easily extendable to 1M tokens via YaRN.
Built for builders. Highly efficient, high-quality, and licensed under Apache 2.0.

🚀 The open weights for Qwen3.8-2.4T-A95B (Max-level) have also been released recently.

Whether you're shipping lightweight applications with Qwen3.8-27B locally or building agents with Qwen3.8-2.4T-A95B, they're yours now!

Download, deploy, and build something we haven't imagined yet. 👀👇

Hugging Face： https://t.co/4kaAcqYEVj
ModelScope： https://t.co/eRIMZCGkhC

See 21 related tweets

@vllm_project: 🎉 Qwen3.8-27B is here from @Alibaba_Qwen, and the whole thing fits on a single GPU. Same hybrid back...
@Prince_Canuma: Qwen3.8-27B comes with day-0 support on MLX-VLM and Nativ

Model weights uploading 🚀\n\nQT @Alibaba...

@Mayhem4Markets: A 27-billion-parameter open-weight model just beat the proprietary frontier on computer use, softwar...
@TeksEdge: 🔥 THIS is the Qwen3.8 model Local AI users have been waiting for! Questions still remain!

The 2.4 T...

@alex: Required contract for inference providers >50m in revenue - seems fair\n\nQT @Alibaba_Qwen: We pr...

4. GeminiApp (Group Score: 352.9 | Individual: 42.0)

Cluster: 15 tweets | Engagement: 1696 (Avg: 384) | Type: Tech

Gemini 3.7 Flash is now available to all Pro and Ultra users in Gemini chat.

This model update delivers improved reasoning and accuracy for multi-step tasks like intelligently connecting the dots across dozens of files and emails into one master document.

Give it a try today on the web or in the app!\n\nQT @GeminiApp: Gemini Spark now runs on Gemini 3.7 Flash.⚡️

Whether you’re using Spark to compile vendors into Sheets or draft negotiation emails, 3.7 Flash makes your personal AI agent more precise and accurate with improved tool use for @GoogleWorkspace apps to help turn ideas into action.

See 14 related tweets

@genspark_ai: Gemini 3.7 Flash just landed on Genspark.

Use it in AI Chat, Code agent, and Genspark Claw starting...

@WesRoth: Google just gave its 24/7 AI agent a major intelligence upgrade.

Gemini Spark is now powered by Gem...

@GoogleAI: It’s (finally) Friday 🎉 Here’s our end-of-week recap:

— This year’s @madebygoogle lineup (Pixel 11 ...

@WesRoth: Google shipped Gemini 3.6 Flash.

Three weeks later: Gemini 3.7 Flash.

43.6% vs 34.4% on FrontierCo...

@osanseviero: Gemini Flash 3.7 was released yesterday and I'm excited to see what folks create with it!

Show us w...

5. vllm_project (Group Score: 288.3 | Individual: 34.8)

Cluster: 9 tweets | Engagement: 134 (Avg: 103) | Type: Tech

Good to see @RedNote back with a new open model, and a much bigger one built for long-horizon agent work. dots3-note preview is a 280B MoE with 16B active at 512K context, it reads images, audio, and video, and it is trained to explore unfamiliar environments and update its own memory as it goes.

Out from @dotsstudioai under Apache 2.0, with day-0 support in vLLM. 🚀

Thanks to the @dotsstudioai team for the collaboration. 🙌 🔗 https://t.co/3ST9wsZdpU\n\nQT @dotsstudioai: Introducing dots3-note preview — a small but mighty step toward long-horizon agency in real life. 🔹 280B MoE with 16B active parameters, a 512K context window, and multimodal understanding across text, vision, and audio 🔹 Introduces TEMPO, a new RL approach for long-horizon agent training through self-critiquing and test-time-scaled value estimation 🔹 Built to reason, explore unfamiliar environments, update memory over time, and combine multimodal perception with coding and tool use to solve complex tasks 🔹 Open weights on Hugging Face, alongside two open benchmarks for real-life agents: VibeSearchBench and VibeLifeBench Competitive with much larger models across reasoning, agentic, and multimodal evaluations. 🔗 Tech blog: https://t.co/kMJtfvB1s7 🔗 Model weights: https://t.co/Lxwbgjc8wk 🔗 Github: https://t.co/QAAeYBTULP

See 8 related tweets

@BanghuaZ: The first language model from Rednote dots studio lab! (Yes, rednote from China is also starting to ...
@TeksEdge: 🤯 A 🆕🆕 16B-active open model scored 81.4 on ARC-AGI-2 beating Opus 4.8’s reported 72.1. 👀

The catch...

@FellMentKE: Real life doesn't happen in a single chat session. Things change, flights get canceled, constraints ...
@teortaxesTex: Guys, look here please. RedNote's AI lab says they have a new RL regimen (TEMPO; beta version) that ...
@kimmonismus: Nice: RedNote just open-sourced dots3-note Preview, a 280b multimodal MoE built for agents that oper...

6. MarioNawfal (Group Score: 227.3 | Individual: 32.1)

Cluster: 12 tweets | Engagement: 156 (Avg: 336) | Type: Tech

🇺🇸🇨🇳 Apple built its own LLM for China with Alibaba's technical backing, ditching its earlier plan to simply license Baidu's existing model.

This gives Apple direct control over its AI stack in its second-largest market, where regulatory approval from Beijing is non-negotiable for any AI feature to ship on iPhones.

Tim Cook just made Alibaba the gatekeeper to 1 billion potential Apple Intelligence users.

Source: Reuters / Writer: Oliver

See 11 related tweets

@MarioNawfal: 🇺🇸🇨🇳 Apple has trained its own AI model specifically for China, with help from Chinese tech giant Al...
@Reuters: EXCLUSIVE: Apple trains its own AI model for China market with Alibaba's support, sources say https:...
@Techmeme: Sources: Apple trained a China-specific LLM with Alibaba's support, which would make Apple the first...
@moneycontrolcom: #Business | Apple trains its own AI model for China market with Alibaba's support: Report
Apple t...
@MacRumors: Apple Trained Its own AI model for China Market With Help From Alibaba https://t.co/hJOLbehSEo https...

7. GavinSBaker (Group Score: 196.9 | Individual: 39.6)

Cluster: 6 tweets | Engagement: 869 (Avg: 745) | Type: Tech

Really does work as advertised.

“Hey Matic go clean the kitchen.”

First 10/10 product rating from Wired in a decade. And kids love it for some strange reason.

Happy that @Atreidesmgmt is an investor.\n\nQT @maticrobots: Introducing Cues: Voice & Gesture control for Matic

We raised $115M and spent 9 years to make the world’s first intuitive home robot

Say you spilled coffee: Point to the spill and say, "Hey Matic, clean this" and it will hear, see, locate in 3D, and go clean on its own.

Matic comes with a lot of features:

Say "Hey Matic" - it locates your voice, turns, and looks at you
Say "Hey Matic, follow me" and start walking. Matic will follow behind
Say "Hey Matic, go clean the living room". Since it knows your house map, it navigates and just does it

It's so easy, a 5 year old and an 80 year old can use it and it understands 75 different languages.

Matic has 8x the airflow, specialised cleaning algorithms for rugs, corners, toekicks, mopping, etc and cleans better than any other robot vacuum. Also keeps improving with software updates.

13,000 families use and love Matic. WIRED magazine gave it a 10/10 (the only hardware to receive this rating in a decade)

Buy yours at https://t.co/nLfY5pchxB and if you don't love it after 6 months, we'll give you a full refund.

To celebrate our launch, we're cleaning 300 homes with Matic in San Francisco and New York City.

Comment "Matic" below, we'll send you the link to sign up and come to your doorstep to clean your home with Matic.

See 5 related tweets

@Scobleizer: Matic Robots was launched in my house in 2024.

So I get why so many people are praising it for a s...

@zaynmcps: MATIC JUST KILLED THE APP FOR ROBOT VACUUMS

→ say “hey matic clean this” and point, that’s it

→ ev...

@pingToven: so i saw this, went to set it up, panicked when the onboarding wanted me to say the name of a room t...
@mattshumer_: Matic generously sent me a robot.

I was super skeptical… last year I bought a Roomba and hated it s...

@shiri_shh: WIRED magazine just gave a 10/10 rating to a piece of hardware for the first time in a decade.

It i...

8. cryptopunk7213 (Group Score: 193.0 | Individual: 31.1)

Cluster: 8 tweets | Engagement: 47 (Avg: 119) | Type: Tech

new deepseek feature allows you to retrofit any ai model (inc claude, got etc) with a custom harness engineered for coding, reasoning, superior memory management.

there’s many available. open source too.

deepseek, zhipu and china in general are excellent at the post-training part of models.\n\nQT @deepseek_ai: 🧩 DeepSeek Harness v0.1 is now available in Developer Preview!

🔹 We’re opening it up to developers building agent harnesses worldwide and open-sourcing the codebase in MIT license. 🔹 Powered by the Cordis meta-framework, DeepSeek Harness is an agent harness built around one core idea: Everything is a plugin. Models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration, and UI are ALL implemented as plugins, and can be mixed, matched, replaced, and extended.

Try it now! https://t.co/2YWSvJHhKA

See 7 related tweets

@tianyi: Thank you! Pi is many DeepSeek researchers and developers favorite daily drive.

DSH reused Pi's LLM...

@badlogicgames: 🤝\n\nQT @tianyi: Thank you! Pi is many DeepSeek researchers and developers favorite daily drive.

DS...

@RoundtableSpace: DEEPSEEK JUST OPEN-SOURCED A FULLY MODULAR AI AGENT HARNESS

DeepSeek Harness makes the model, tools...

@teortaxesTex: RT @tianyi: Thank you! Pi is many DeepSeek researchers and developers favorite daily drive.

DSH reu...

@RoundtableSpace: DEEPSEEK HARNESS JUST BECAME THE CODING AGENT TO WATCH

• Crossed 36K GitHub stars on launch day • M...

9. MTSlive (Group Score: 188.5 | Individual: 32.4)

Cluster: 6 tweets | Engagement: 23 (Avg: 194) | Type: Tech

MTS Host @sophiadew visits @MongoDB's Build Fest to learn how MongoDB is fueling the builder ecosystem.

@jadefurubayashi shares how MongoDB for Startups fuels early-stage founders, @frankzliu unpacks Voyage AI's embedding models and the future of retrieval.

Apoorva Joshi dives into vector search and MongoDB's new Atlas MCP server, and @BazeleyMikiko reveals auto-embedding and MongoDB's role as the memory layer for AI agents.

1:42 Jade Furubayashi gives an overview of the MongoDB for Startups program 3:14 How AI has changed the builder ecosystem: the need for speed and scale 8:10 Frank Liu explains how MongoDB's integrated platform simplifies building AI apps 13:11 Hot take: what is overhyped and what is underhyped? 17:38 Apoorva Joshi introduces the MongoDB Atlas managed MCP server 18:35 Why MongoDB is more than just a database 21:33 Mikiko Bazeley: how auto-embedding removes the need for manual embedding pipelines 22:19 MongoDB as the memory and context layer for AI agents

See 5 related tweets

@MTSlive: MTS Host @sophiadew talks with @BazeleyMikiko, Staff Developer Advocate at @MongoDB, at Build Fest. ...
@MTSlive: MTS Host @sophiadew asks @frankzliu, Voyage AI Research Manager at @MongoDB, for his hot takes at Bu...
@MTSlive: MTS Host @sophiadew talks with Apoorva Joshi, Developer Advocate at @MongoDB, at Build Fest.

Apoorv...

@MTSlive: MTS Host @sophiadew talks with @jadefurubayashi, MongoDB's Director of Startups and Venture, at Buil...
@MongoDB: AI development is moving fast. Your data platform should, too.

Today at #MongoDBlocal Build Fest, ...

10. firstadopter (Group Score: 186.9 | Individual: 34.2)

Cluster: 8 tweets | Engagement: 162 (Avg: 146) | Type: Tech

Bloomberg's report that OpenAI has a $40 billion ARR WAY understates the company's current ARR, according to three sources familiar\n\nQT @business: OpenAI is on track to generate annualized revenue of more than$ 40 billion based on its current performance, according to sources, roughly doubling its run rate from the end of 2025 and bolstering the company’s plans for a Wall Street debut https://t.co/Ize8BUOBNA

See 7 related tweets

@business: OpenAI is on track to generate annualized revenue of more than $40 billion based on its current perf...
@pstAsiatech: OpenAI is on track to generate annualized revenue of more than $40 billion based on its current perf...
@kimmonismus: OpenAI is on track to clear $40 billion in annualized revenue, roughly double its run rate at the en...
@justjoshinyou13: We're one year past GPT-5 and the big wave of OpenAI compute deals, when many were skeptical OAI cou...
@Cointelegraph: 🔥 NEW: OpenAI is on track to top $40 billion in annualized revenue, roughly doubling since the end o...

11. cryptopunk7213 (Group Score: 186.7 | Individual: 35.2)

Cluster: 8 tweets | Engagement: 165 (Avg: 119) | Type: Tech

cursor is the reason $SPCX added$ 100Bs to their market cap this week.

the market underestimated how good the team is at building ai models. they undervalued the importance of an agent harness (cursors is #1)

with anthropic rumored to ipo at a $2 trillion valuation the spaceX re-pricing will be brutal

one asset giving you exposure to: ai, chips, neocloud, mobile connectivity, space rockets and the worlds best social platform.\n\nQT @cryptopunk7213: cursor was always a frontier ai lab disguised as a wrapper. elon saw this. he knew all he had to give them was compute (lots of it) and they’d create a killer model.

hilarious part is the models aren’t even their best product - grok bot, cloud agents and anything that includes the secret cursor harness sauce will be an instant hit

best acquisition of 2026

See 7 related tweets

@XFreeze: SpaceX has officially completed its $60 billion acquisition of Cursor

The merger became effective o...

@StockMKTNewz: SpaceX $SPCX has completed its$ 60 billion acquisition of AI coding startup Cursor\n\nQT @IPONewsroo...
@IPONewsroom_: SpaceX $SPCX has completed its$ 60 billion acquisition of AI coding startup Cursor

The deal became ...

@PolymarketMoney: JUST IN: SpaceX has completed its $60,000,000,000.00 acquisition of AI coding startup Cursor....
@BullTheoryio: BREAKING: SpaceX has completed its $60 billion acquisition of Cursor.

The all-stock deal is the lar...

12. daleverett (Group Score: 161.1 | Individual: 33.8)

Cluster: 5 tweets | Engagement: 11 (Avg: 12) | Type: Tech

Watch me sing dumb ways to die: ai edition

Animations by @daltonmeon\n\nQT @daleverett: Introducing: 🎵Dumb Ways to AI🎵

a 3-minute song-cumentary about your $40M AI startup:

you have 3 databases and no idea what’s happening.

Stop. Use @polygres.

It’s free right now. Use it → get credits at launch.

This song will now follow you into your dreams.

@daltonmeon @damienhe @fdotinc

See 4 related tweets

@daleverett: If u liked dumb ways to die? Meet dumb ways to AI\n\nQT @daleverett: Introducing: 🎵Dumb Ways to AI🎵 ...
@damienhe: this is funny af LOL\n\nQT @daleverett: Introducing: 🎵Dumb Ways to AI🎵

a 3-minute song-cumentary ab...

@polygres: I love this song\n\nQT @daleverett: Introducing: 🎵Dumb Ways to AI🎵

a 3-minute song-cumentary about ...

@daltonmeon: RT @daleverett: Introducing: 🎵Dumb Ways to AI🎵

a 3-minute song-cumentary about your $40M AI startup...

13. VadimStrizheus (Group Score: 154.8 | Individual: 34.0)

Cluster: 5 tweets | Engagement: 16 (Avg: 88) | Type: Tech

nah bro how is this even possible?! 😭

Aracds' Actor+ is INSANE for ugc ads

you can’t even tell the difference between AI UGC and real UGC now

every founder can market their app now without a excuse

making money online in 2026, has never been easier. https://t.co/C8oRV8XDel\n\nQT @arcads_ai: Introducing Actor+

The most realistic AI actor library ever built.

As of today, the "is AI UGC good enough?" debate is over.

Now live on Arcads. https://t.co/GCOoOQVnZN

See 4 related tweets

@Parul_Gautam7: ↳ Pick the right actor ↳ Build the scene ↳Add your own creative touch ↳Make it feel real

What I lik...

@rohanpaul_ai: Arcads' Actors+ is a good example of why AI video quality is becoming a perceptual problem now.

You...

@rom1trs: The "is AI UGC good enough?" debate is officially over.

Excited to launch Actor+

Most realistic AI...

@alexcooldev: The gap between AI UGC and real UGC is getting ridiculously small.

The skin, lighting, phone camer...

14. tszzl (Group Score: 153.1 | Individual: 33.1)

Cluster: 5 tweets | Engagement: 1394 (Avg: 1025) | Type: Tech

the major ai companies should commit to a real time priced API product. ai demand varies wildly over a day/night curve and the industry is broadly extremely capacity crunched. meanwhile agents make dealing with variable pricing, batching, and projecting total costs very easy

See 4 related tweets

@swyx: damn deepseek moves fast on roon tweets https://t.co/PSvo3MaDPJ\n\nQT @tszzl: the major ai companies...
@pingToven: already on openrouter, get the best prices by default https://t.co/XPT4SmhgVP\n\nQT @tszzl: the majo...
@ShanuMathew93: peak pricing incoming\n\nQT @tszzl: the major ai companies should commit to a real time priced API p...
@tugot17: tokens are THE product for dynamic pricing, your draft acceptance rate, your cache hit (which memory...

15. ethancole_ai (Group Score: 152.9 | Individual: 42.3)

Cluster: 6 tweets | Engagement: 270 (Avg: 64) | Type: Tech

A useful AI video workflow shouldn’t make you restart every time one detail is wrong.

MiniMax H3 supports instruction-based editing and motion transfer, so you can keep working from footage you already have.

Now available on Magnific with 50% OFF in 2K.

https://t.co/mSEGBx45s4 https://t.co/bYCfVmKuwA

See 5 related tweets

@TheoBuildsAI: Another excuse for me to generate way too many AI videos this week 😂

MiniMax H3 in 2K at half price...

@alexaiworks: AI video is becoming less about writing the perfect prompt and more about giving the model the right...
@alexaiworks: RT @NovaIAHQ: The more AI video improves, the more I care about control.

Give the model your produc...

@alexaiworks: RT @SynapseOpsAI: This is the kind of AI video workflow I actually want.

Start with a real product ...

@alexaiworks: RT @EmmaUsesAi: AI video gets far more practical when you can control the camera, character and voic...

16. guohao_li (Group Score: 139.9 | Individual: 34.1)

Cluster: 5 tweets | Engagement: 60 (Avg: 34) | Type: Tech

Chinese instagram (@xiaohongshu) dropped a 280B model and a new RL training algorithm for long-horizon rollouts based on test-time-scaled value estimation with macro-step policy optimization - TEMPO\n\nQT @ChaoQiao42: WE SHIPPED IT! 🚀

Introducing dots3-note Preview: our open-source 280B/16B-active multimodal model, built for complex reasoning and long-horizon agents.

Beyond the model, we’re sharing TEMPO (Test-time-scaled Value Estimation with Macro-step Policy Optimization)—our answer to an increasingly important question for the field: how can RL train agents when a single rollout takes tens of hours? TEMPO turns intermediate self-critiquing into learning signals before the task is complete.

We’re also open-sourcing VibeSearchBench & VibeLifeBench.

A small but mighty step toward long-horizon agency in real life—and an incredibly emotional day for me and my team.

https://t.co/OOeoS7KMB8

See 4 related tweets

@teortaxesTex: RT @ChaoQiao42: WE SHIPPED IT! 🚀

Introducing dots3-note Preview: our open-source 280B/16B-active mu...

@RoundtableSpace: THE OPEN SOURCE MODEL THAT TRAINS ITSELF WHILE IT'S STILL WORKING - Dots 3-note.

280B parameters. ...

@no_stp_on_snek: The models dropped this week just don’t stop! Love it.\n\nQT @ChaoQiao42: WE SHIPPED IT! 🚀

Introduc...

@teortaxesTex: RT @ChaoQiao42: How do you train an agent when a single rollout takes tens of hours?

Alongside the ...

17. akshay_pachaar (Group Score: 137.2 | Individual: 29.3)

Cluster: 7 tweets | Engagement: 92 (Avg: 574) | Type: Tech

Gemini 3.7 Flash matches Sonnet 5 at half the price.

On Cognition's FrontierCode 1.1 inside Devin, this performance was measured on tightly scoped refactors with minimal diffs that match repo conventions.

Available at an extra 50% off through Aug 27. https://t.co/fkY8ei2Pm2\n\nQT @cognition: Gemini 3.7 Flash is now available in Devin.

On FrontierCode 1.1, it reaches Claude Sonnet 5-level performance at less than half the cost, with the low latency the Flash series is known for. https://t.co/ATpmjX2qAJ

See 6 related tweets

@rohanpaul_ai: Low latency matters more inside an agent because one task can stack many sequential model calls. Gem...
@arsh_goyal: RT @arsh_goyal: Gemini 3.7 Flash is now inside Devin, and the pitch is cost and latency.

Cognition...

@astropol0: Anthropic has no choice but to drop Sonnet 5.5 soon.....

gemini 3.7 Flash is putting real pressure ...

@sairahul1: sir...google just dropped a new model with 5x speed and 5x cheaper than Sonnet 5 https://t.co/K3fnSm...
@sundarpichai: RT @cognition: Within Devin, Gemini 3.7 Flash performs particularly well on tightly scoped refactors...

18. MollySOShea (Group Score: 137.1 | Individual: 37.3)

Cluster: 4 tweets | Engagement: 80 (Avg: 54) | Type: Tech

The most efficient company in public markets is an ad company

Revenue per employee:

› AppLovin: ~ $17M (~400 people, approaching$ 7B in annualized EBITDA) › Nvidia: ~ $5M (3x less) › Netflix: ~$ 4.2M (4x less) › Apple: ~ $2.4M (7x less) › Meta: ~$ 2.2M (8x less) › Google: ~ $1.9M (9x less) › Microsoft: ~$ 1.1M (15x less)

Engineering hasn't grown in 3 years while cash flow scaled 20x. 4 execs total. No one-on-ones, no PMs.

In 2024 AppLovin cut ~40% of its team while its core ads business was growing nearly 100% YoY.

Key point: AppLovin cut during hypergrowth to deliberately increase efficiency. Unlike most companies who add more headcount.

Full breakdown with CEO Adam Foroughi and CTO Gio Ge below:\n\nQT @MollySOShea: BREAKING: Inside AppLovin's $100B+ Ad Engine

Co-Founder & CEO Adam Foroughi CTO Giovanni Ge

FULL INTERVIEW

AppLovin is one of the most efficient companies in public markets at ~400 people. Approaching $7B in EBITDA. That's ~$ 17M per employee, roughly 3x the revenue per head of Nvidia & 7x Apple, Meta, or Google. Engineering hasn't grown in 3 years (100 people) while cash flow scaled 20x.

The most extensive breakdown yet of how they built the ad model that grew @AppLovin ( $APP ) from a -92% post-IPO decline to$ 100B+ company, with billions in profit

Gio joined AppLovin in late 2022, when the company was worth ~$5.5 billion to help architect Axon 2, the recommendation engine that became central to AppLovin’s transformation. The new architecture improved predictions & advertiser returns, helping drive the company’s expansion beyond mobile gaming into e-commerce ads & a full recovery

"Technology moves really fast, & you gotta be humble about what you have. If technology goes faster than what you've developed, you gotta throw away what you have & rebuild it. And it's probably gonna happen every couple years, or even faster in the future." - Adam Foroughi

We go inside the machine: › Rebuilding Axon from scratch with 100 engineers › No one-on-ones, no training manuals, almost no PMs › New hires ship production code in week 1 › Engineers on top of AI, not next to it › The e-commerce product designed on a napkin › Dropping 92%, then becoming their own best investor › The math to $1 trillion

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒 (00:00) Adam Foroughi, Co-Founder & CEO & Giovanni Ge, CTO at AppLovin (00:52) Gio's first impression of AppLovin (03:38) Joining AppLovin after a 30% stock drop (09:59) How Axon 2 actually got built (20:23) Why AppLovin never needed a sales team (21:59) From gaming ads to e-commerce (24:47) From a napkin drawing to a real product (30:54) The onboarding mistake most companies make (34:37) What "AI-Native" actually means (36:37) Should new engineers still learn to code? (41:41) UGC vs. traditional ads: what's winning? (47:56) Adapting to AI without rebuilding from scratch (54:50) Gio's daily tech stack (1:02:17) AppLovin's path to a trillion-dollar valuation (1:06:19) Recovering from a 90% stock drop (1:09:47) What's next for AppLovin? (1:13:44) What happens when AI takes over? (1:15:48) Adam and Gio's hottest takes (1:19:50) How to actually stay locked in (1:24:47) Building something that lasts

See 3 related tweets

@MollySOShea: AppLovin likely has the highest concentration of super individual contributors

Aka "Super ICs"

it'...

@MollySOShea: Sector Head at @Coatuemgmt Michael Barton (@lordbartonjr) says AppLovin CEO Adam Foroughi is the "mo...
@MollySOShea: RT @DimitryNakhla: Adam Foroughi was just asked how AppLovin becomes a $1T company.

Here’s the clea...

19. SebJohnsonUK (Group Score: 129.8 | Individual: 30.7)

Cluster: 5 tweets | Engagement: 106 (Avg: 78) | Type: Tech

A London AI Lab has just released an agent that outperforms Claude Opus 4.8 and GPT-5.5 on the task of replicating research.

NICE

The company, founded by DeepMinders, just emerged from stealth with a $50m seed led by Index.

The lab is called Inherent, and it’s just released Faraday - a 27B-parameter “AI Scientist” agent that outperforms Claude Opus 4.8 and GPT-5.5 on the task of replicating research.

It’s been trained via long-horizon RL with coding agents as a tool to learn the skills of a scientist.

This is VERY COOL. London AI is flying right now.

See 4 related tweets

@KanishkaNarayan: Congrats to the @inherent_labs team on their first research release: an AI scientist, mixing intuiti...
@SebJohnsonUK: London is not only dominating in many AI applications (@synthesiaIO & @ElevenLabs ) our labs are now...
@bakkermichiel: Super cool progress from @tantumscollins @edwardfhughes and the team at @inherent_labs towards opti...
@matthewclifford: I’m very excited about the Inherent team - this paper gives a glimpse of what they’ve been working o...

20. mark_k (Group Score: 127.8 | Individual: 34.0)

Cluster: 5 tweets | Engagement: 144 (Avg: 216) | Type: Tech

Currently, competitors of @SpaceXAI such as OpenAI are still downplaying the need for orbital compute (AI satellites). In fact, Sam Altman has repeatedly ridiculed the idea!

But this is obviously going to change quite soon, once resistance against datacenters intensifies. SpaceX will be the only provider able to scale further. Nobody else has the capacity to build and launch a meaningful number of AI satellites.

That alone is a strong indication why SpaceX might win the AI race, and I think the competitors are starting to realize this. The ridicule will turn into bitter tears soon enough.\n\nQT @elonmusk: Orbital compute will be the only way to scale AI probably sometime in 2029 due to power availability& permitting problems on land

See 4 related tweets

@niccruzpatane: RT @elonmusk: Orbital compute will be the only way to scale AI probably sometime in 2029 due to powe...
@elonmusk: Orbital compute will be the only way to scale AI probably sometime in 2029 due to power availability...
@Cointelegraph: ⚡️ INSIGHT: Elon Musk says orbital computing may be the only way to scale AI by 2029 due to power av...
@MTSlive: SITUATION DETECTED: Elon Musk says orbital compute is the only viable path to scaling AI, as by arou...