科技推文精选 - 2026-05-27

2026年5月27日科技每日简报

Today's top tech conversations are led by @JayminSOfficial, whose post about 'The AI industry quietly normal...' garnered the highest engagement. Key themes trending across the top stories include skill, every, skills, model, https. The community is actively discussing recent developments in AI, engineering practices, and startup strategies.

1. JayminSOfficial (Group Score: 399.9 | Individual: 44.1)

Cluster: 11 tweets | Engagement: 1525 (Avg: 628) | Type: Tech

The AI industry quietly normalized something insane:

Researchers’ benchmarks now decide production traffic.

MMLU. HumanEval. GPQA.

Benchmarks built for papers and leaderboards now influence which model handles customer support, coding agents, sales workflows, everything.

Most teams simply followed whichever provider topped the chart that week.

That became the default operating system for AI decisions.

@merge_api Gateway’s “Build Your Own Router” changes the equation: you define what “best” means for your product, your users, your workflows.\n\nQT @shensi: Introducing Merge Gateway - Build Your Own Router.

You're three sprints into your coding assistant.

You pick the most hyped model, integrate, test, deploy.

A month later, a new model drops.

Now you re-test, re-integrate, re-deploy.

Your product didn't change, but the benchmark did.

That's how most AI teams operate.

Chasing a "best" defined by people who've never seen their product.

There is no best model.

There's only the right one for your product, users, and use-cases.

Build Your Own Router runs on your definition of good.

Pick your benchmarks, weigh them, add your own evals.

@merge_api routes every request to your winner.

👉$100 in credits to the first 200 people that comment

https://t.co/VLbDYycQRY

See 10 related tweets

@alex: RT @shensi: Introducing Merge Gateway - Build Your Own Router.

You're three sprints into your codin...

@kimmonismus: Uber burned $3.4B in AI budget in 4 months.

Microsoft is cancelling Claude Code licenses company-w...

@godofprompt: Hot take 🔥

Most AI products don’t have a model problem.

They have a “we blindly worship benchmarks...

@dr_cintas: Merge Gateway just shipped the first LLM router you actually program 🤯

You stop picking a model. Yo...

@MTSlive: SITUATION ANALYSIS: AI spend is becoming a finance problem.

What Shensi Ding and Gil Feig, founders...

2. business (Group Score: 245.5 | Individual: 28.7)

Cluster: 13 tweets | Engagement: 296 (Avg: 90) | Type: Tech

China is restricting overseas travel for top AI professionals in private firms such as Alibaba and DeepSeek, suggesting an escalation in measures intended to safeguard its technology and catch up to the US in a pivotal sphere.

Government agencies have begun imposing restrictions on individuals involved in advanced AI work and considered strategically important to the country, people familiar with the matter said. https://t.co/WM5Y9gkue9

📷: Qilai Shen/Bloomberg

See 12 related tweets

@MsMelChen: The noose is tightening:

Beijing is treating elite “private” sector AI engineers and innovators li...

@teortaxesTex: This is quite dumb given how much they share openly (DeepSeek at least), but what's interesing: do A...
@MTSlive: SITUATION DETECTED: China is imposing restrictions on overseas travel for top AI researchers from la...
@rohanpaul_ai: Bloomberg: China is now treating its strongest private-sector AI researchers as strategic assets, wi...
@zephyr_z9: HUH\n\nQT @DeItaone: CHINA IS RESTRICTING OVERSEAS TRAVEL FOR TOP AI PROFESSIONALS IN PRIVATE FIRMS ...

3. Parul_Gautam7 (Group Score: 239.1 | Individual: 40.2)

Cluster: 7 tweets | Engagement: 386 (Avg: 105) | Type: Tech

RT @Capafyai: Introducing Capafy: the Skill-based Agent Marketplace.

Now your Skill runs as a product and earns while you sleep.

On Capafy, you can upload your Skills, they run online while staying closed-source, and you get paid every time someone uses them.

You can also use Skills uploaded by industry top talent to get expert-level work done directly.

You'll find Skills built from industry expertise in every field.

Let's say: ·A creator with 100M+ views uploaded their viral video Skill; ·A recruiter who's screened 10,000+ resumes uploaded their hiring Skill; ·A top sales rep who's closed thousands of deals uploaded their cold email Skill.

Skills uploaded by industry top talent across countless fields can be used directly to get excellent work done.

Launch your Skills: upload the Skills you've built in Claude Code, Codex, or OpenClaw, and get paid every time someone uses them.
Use expert Skills: get expert-level work done, not the average AI output. Use them in one click, or connect your own Agent via agent-to-agent and let it tap into the expert Skills on Capafy.

See 6 related tweets

@Parul_Gautam7: Capafy turns individual expertise into reusable, monetizable AI Skills.

It’s a shift from using AI ...

@godofprompt: Your best prompt/workflow shouldn’t die in a folder.

Capafy turns Skills into products people can u...

@dr_cintas: You can now give any Al agent expert Skills with one link.

Paste Capafy's install URL into Claude C...

@AiBreakfast: A marketplace for writing Skills - very cool concept.

I probably have 20 "Skills" that I've writte...

@TheoBuildsAI: I’m curious what separates a “good” Skill from an average one. In theory this could become really us...

4. badlogicgames (Group Score: 214.0 | Individual: 46.7)

Cluster: 11 tweets | Engagement: 9090 (Avg: 255) | Type: Tech

RT @edzitron: Uber’s COO has said that it’s getting “harder to justify” its AI costs because there was no way to show a link between AI spend and any meaningful increase in useful features. This is the first time I’ve seen a company say this directly.

https://t.co/xUhZvtpwah https://t.co/nDj9GIXssV

See 10 related tweets

@nalinrajput23: Microsoft blocks Claude Code.

Uber questions AI spending.

Meanwhile developers keep shipping htt...

@aiedge_: Uber's COO just silently stated the #1 problem with AI right now.

He says AI costs are "hard to jus...

@etnshow: PolyAI (@polyaivoice) CEO @nikola_mrksic says not all token maxing leads to a vastly better company....
@edzitron: To explain the significance of this, Anthropic moved enterprises to token-based billing in Q1 2026. ...
@fchollet: Developer productivity is hard to manage.\n\nQT @edzitron: Uber’s COO has said that it’s getting “ha...

5. tunguz (Group Score: 149.6 | Individual: 46.3)

Cluster: 4 tweets | Engagement: 473 (Avg: 104) | Type: Tech

This is the next level.\n\nQT @koylanai: Gradient descent for SKILL.md files sounds interesting, maybe a bit complex but it's becoming a real part of agent harness.

SkillOpt is one of the first papers to treat markdown skill files as trainable parameters and provides a proper optimization framework for them.

A few things I learned that you should consider too.

The validation gate is the only thing that matters in a self-editing loop.

Held-out set, strict improvement, ties rejected. End-to-end, their best skills land with 1 to 4 accepted edits total. If your "self-improving agent" is accepting most of what it proposes, you're shipping slop.

Bounded edits are better than full rewrites. 4 to 8 edits per step is the sweet spot.

Remove the budget and performance collapses. This is the textual analog of learning rate, and it transfers to any LLM-as-author loop. If you're using an agent to refactor your docs, your prompts, or your skills, cap the diff size.

Compactness wins. Median final skill: ~920 tokens.

Skills do not need to be long. They need to be high-signal. Most skill files I see are bloated because length feels like effort. It isn't.

The harness is becoming less important; the skill is becoming more important.

A Codex-trained skill ported into Claude Code hit +59.7 points on SpreadsheetBench. Procedural knowledge is more general than the runtime that produced it.

Frozen model + trained context is the practical adaptation.

GPT-5.4-nano with a SkillOpt'd skill ≈ frontier behavior on procedural benchmarks. Cheaper, portable, inspectable, zero inference-time cost. This is the answer to "how do we adapt a frontier model for our domain" for almost everyone who isn't training their own models.

Verification is the bottleneck.

Every gate in this paper depends on an auto-grader. That works for benchmarks. It fails for writing, design, and strategy, exactly the open-ended work we want to automate. Whoever builds the verifier for open-ended tasks owns the next stage.

There are also two leassons I learned while shipping v2.3.0 of my Context Engineering Agent Skills repo, measured across composer-2, claude-opus-4-7, gpt-5.5, and gemini-3.1-pro via the @cursor_ai SDK:

Description and body are two different surfaces. The router only sees the description. The agent sees the body once activated. They can quietly disagree, and only end-to-end task tests catch it.
Aggregate accuracy is the wrong unit. When I rewrote three descriptions, the corpus average moved ~1pp. Individual skills moved 23–25pp. Per-skill effect size is where the action is.

Also, in Feb 2026 I shared a piece called Personal Brain OS arguing that the markdown file is a first-class substrate for agent state. SkillOpt is the optimizer-shaped version of that same argument: not "store memory in files" but "treat files as trainable parameters with proper optimization machinery around them." That's the move from static to measured.

The fast/slow split they describe already lives implicitly in the digital-brain-skill repo:

voice-guide and tone-of-voice.md are slow-state (rarely touched)
posts.jsonl and bookmarks.jsonl are fast-state

What SkillOpt adds that I didn't have is a protected section invariant, a structural guarantee that fast edits cannot overwrite slow lessons. Removing that mechanism cost them 22 points on SpreadsheetBench. Worth borrowing.

If you're building agents, SkillOpt: Executive Strategy for Self-Evolving Agent Skills is a good paper to read: https://t.co/ZS9SZXQ6Mv

See 3 related tweets

@garrytan: This sounds complicated but the agents can implement this in OpenClaw/Hermes Agent trivially (use sk...
@omarsar0: RT @omarsar0: New research from Microsoft Research

I see a lot of AI engineers handwriting agent sk...

@garrytan: These concepts coming soon to GBrain this week https://t.co/nXPzt9hD71\n\nQT @omarsar0: New research...

6. XFreeze (Group Score: 145.0 | Individual: 27.8)

Cluster: 6 tweets | Engagement: 332 (Avg: 843) | Type: Tech

Grok Build is actually insane right now

It’s completely different from when it first launched. In just a couple of days, xAI has turned it into a productivity monster

What started as a CLI has evolved into a full developer powerhouse where you can actually ship real work at ridiculous speed

It feels insanely powerful now: • Fast workflows • Strong coding capabilities • Image input • Parallel Agents - Run up to 8 subagents simultaneously • Image & video generation built directly into the CLI

And the craziest part is Updates are landing daily. It gets noticeably better almost every single day

Now available for SuperGrok and 𝕏 Premium+ users

See 5 related tweets

@MarioNawfal: Grok's progress now feels like watching software evolve in fast-forward.

V9 is landing within weeks...

@XFreeze: Grok Build just received another two updates today:

v0.1.220-alpha.3

xAI is shipping updates to Gr...

@VaibhavSisinty: xAI just shipped a 1.5 trillion parameter monster. Grok V9 is 2 to 3 weeks away. 🤯

The full breakdo...

@MarioNawfal: RT @MarioNawfal: Grok's progress now feels like watching software evolve in fast-forward.

V9 is lan...

@MarioNawfal: 🇺🇸 xAI launched Grok Build, a coding agent that runs directly from your terminal, now in early beta ...

7. ReutersBiz (Group Score: 144.6 | Individual: 24.7)

Cluster: 7 tweets | Engagement: 84 (Avg: 19) | Type: Tech

Micron Technology reached a $1 trillion market value for the first time as shares surged more than 18% following UBS's price target hike of over$ 1,600. Read more: https://t.co/Jo484ZhWeS https://t.co/IUBRNX5Kv3

See 6 related tweets

@Reuters: Micron Technology reached a $1 trillion market value for the first time as shares surged more than 1...
@business: Micron Technology's market capitalization, which has soared about 700% over the past year, will more...
@MTSlive: SITUATION DETECTED: Micron hit $1T in market cap for the first time today, surging 17% on AI memory ...
@moneycontrolcom: #Business | Micron joins $1 trillion club as AI race powers memory chip boom

-Micron briefly surpas...

@Polymarket: NEW: Micron surpasses $1,000,000,000,000.00 in market cap.\n\nQT @Polymarket: NEW: Micron Technology...

8. Parul_Gautam7 (Group Score: 138.6 | Individual: 37.4)

Cluster: 4 tweets | Engagement: 304 (Avg: 105) | Type: Tech

RT @LucasQin77: A question has stayed with me for years: When generic AI becomes capable of every kind of intellectual work humans do, what's left for human value? Where does competitive advantage come from? I think I found the answer this year. Two things made it click. First, the stock market.

I'm in the top 5% of US equity returns. Trading is one of the cleanest experiments you can run, because everyone has access to the same tools: same charts, same news, same models. Today most traders use ChatGPT or Claude to research positions. Ten years from now, every trader will. And yet — the market will still have a top 5% and a bottom 20%. Same AI, same data, same access. Different outcomes. Why? Because the same model gives different work to different people. The questions you ask, the angles you pursue, the patterns you notice: these are downstream of your mental model. The top 5% and the bottom 20% use the same tools differently, because they think differently. Mental model in, work out. Second, OpenClaw. When Skills started spreading this year as a real format, it hit me — a mental model isn't an abstract thing. It can be packaged. The way a top-5% trader actually thinks through a position — the questions they ask, the order they ask them in, the things they refuse to ignore — all of that can be written down as a Skill. And once it's a Skill, anyone can run it. Which means: a top-5% trader's mental model, packaged as a Skill, run by someone in the bottom 20%, doesn't just give that person a tool. It gives them a different mind for the duration of the work. This isn't only true for trading. Every field has its top 5% and its long tail. Every one of us is in the top 5% of something, and the bottom 20% of many other things. Which means every one of us has a Skill worth packaging — and a thousand others worth running. That's the moat I think AI leaves us with. Not the work itself : the AI will do that. The packaged way you do the work, the difference between your output and the average — that's the part that compounds, that earns, that lasts. Capafy is built on this idea. You package your edge as a Skill. When someone runs it, we spin up an isolated sandbox just for that run — your Skill executes inside, the user gets the output, but the Skill itself never leaves the sandbox. Closed-source online. Your method never leaves you. Every use pays you. In ten years, the question won't be "do you use AI." Everyone will. The question will be: Whose Skill is loaded into your AI. Maybe it comes from your own coding experience. Maybe it comes from years of industry expertise. Maybe it even comes from a top 5% US equity trader. That's the moat.

See 3 related tweets

@Parul_Gautam7: “The question won't be 'do you use AI', it will be whose Skill is loaded into your AI.”

Competitiv...

@TheoBuildsAI: As someone who spends time refining workflows, this hit home a bit. Sometimes what takes years isn’t...
@alex_prompter: Everyone gets the same AI.

Not everyone gets the same instincts, taste, judgment, or decision-makin...

9. myrhex (Group Score: 132.6 | Individual: 31.5)

Cluster: 5 tweets | Engagement: 68 (Avg: 73) | Type: Tech

A new tab dedicated to Grok Build is being worked on in Grok Web. It is called “Build” and links to https://t.co/q07ayvyNvw.

This page is set to become the dedicated entry point for Grok Build directly on https://t.co/MJoc4ZcXfb instead of only https://t.co/Vra6VJCLZe.

It will let SuperGrok, Premium+ and SuperGrok Heavy users install Grok Build with a simple command so they can run it in their terminal.

See 4 related tweets

@elonmusk: How to use Grok Build\n\nQT @Daniel_Farinax: Beginner video: How to install & use Grok Build (made f...
@myrhex: The page https://t.co/KVpXOvNnYD is now functional, this is what it looks like. https://t.co/nzk2TT8...
@teslaownersSV: UPDATE: Grok Build Beta is now available for all SuperGrok and X Premium+ users.

Users can now acce...

@WesRoth: RT @WesRoth: Grok Build is reportedly being prepared for rollout to SuperGrok users, expanding acces...

10. alex_prompter (Group Score: 131.1 | Individual: 69.6)

Cluster: 2 tweets | Engagement: 44276 (Avg: 963) | Type: Tech

Let me trace the timeline here because nobody's connecting it.

Step 1: Scrape the entire internet. Every book, every article, every conversation, every piece of art, every forum post. Do it without asking. Do it without paying.

Step 2: Train a model on all of it. Call it "artificial intelligence."

Step 3: Go to BlackRock's Infrastructure Summit and announce: "We see a future where intelligence is a utility, like electricity or water, and people buy it from us on a meter."

Step 3 is where you sell people's own knowledge back to them. On a meter.

They took the collective output of human thought, compressed it into a model, and now they want to charge you by the token to access a version of what you and everyone you know already created.

One Reddit user put it perfectly: "They stole all this data from us, the people, our life's work, creativity, art, by devouring the internet and blowing through all copyright laws. Now they want to sell it back to us in the form of a utility."

Imagine if someone photocopied every book in the public library, burned the library down, and then opened a subscription service for the copies.

That's the metered intelligence business model.

And they're pitching it to infrastructure investors as though they invented water.\n\nQT @Vivek4real_: SAM ALTMAN: “WE SEE A FUTURE WHERE INTELLIGENCE IS A UTILITY, LIKE ELECTRICITY OR WATER, AND PEOPLE BUY IT FROM US ON A METER.” https://t.co/AXnZ9zh0Ro

See 1 related tweets

@alex_prompter: RT @alex_prompter: Let me trace the timeline here because nobody's connecting it.

Step 1: Scrape th...

11. Reuters (Group Score: 123.1 | Individual: 26.0)

Cluster: 6 tweets | Engagement: 398 (Avg: 117) | Type: Tech

OpenAI's Altman says AI unlikely to lead to 'jobs apocalypse' https://t.co/ED8cHbQEn7 https://t.co/ED8cHbQEn7

See 5 related tweets

@rohanpaul_ai: wionews: OpenAI CEO Sam Altman now says the feared AI white-collar job collapse has not arrived as f...
@KatieMiller: What hasn’t Sam Altman flipped on for profit?

No surprise he’s changed his mind on this once he sta...

@Reuters: OpenAI CEO Sam Altman said the rapid development and adoption of AI would not lead to a global ‘jobs...
@ReutersBiz: OpenAI CEO Sam Altman said the rapid development and adoption of AI would not lead to a global ‘jobs...
@BusinessInsider: The AI jobs apocalypse has not yet come to pass — and Sam Altman says he is happy to be wrong about ...

12. garrytan (Group Score: 121.4 | Individual: 24.1)

Cluster: 6 tweets | Engagement: 608 (Avg: 294) | Type: Tech

This is the new standard for engineering evals https://t.co/UkEwUWybab\n\nQT @serenaa_ge: Today we’re releasing DeepSWE, a new standard for agentic coding benchmarks.

On public leaderboards, top models often look relatively close in capability. DeepSWE shows where they actually diverge, reflecting the realistic experience of developers in their day-to-day work. https://t.co/HCDcjNuTFK

See 5 related tweets

@seconds_0: RT @serenaa_ge: Today we’re releasing DeepSWE, a new standard for agentic coding benchmarks.

On pub...

@xeophon: counterintuitive result right here, imo

the data analysis is well done, lots of thoughts went behin...

@teortaxesTex: Honestly, looks about right, harness (mini-swe-agent) affinities aside Kimi is the closest to a matu...
@antirez: New long frontier coding task benchmark. Didn't yet checked the actual tests performed but looks int...
@adonis_singh: the rankings are very good, but now that it has been released it will be fucked into oblivion\n\nQT ...

13. Sanjay_Sriv (Group Score: 120.5 | Individual: 33.5)

Cluster: 4 tweets | Engagement: 40 (Avg: 97) | Type: Tech

India’s real AI revolution is being built with high power data centres' infrastructure on the ground.

1/ India’s AI story is transforming. Everyone was talking models & talent. The real game now? Hyperscale Data Centers.

The data center project pipeline has skyrocketed. Maharashtra alone is sitting on ₹3.9 trillion worth of projects.

2/ Key players racing ahead: • Adani • Reliance • Tata • Global hyperscalers (Microsoft, Google, AWS etc. indirectly).

Typical facility: 200-300 MW power hungry monsters. One recent set of 9 projects alone requires ₹19,600 crore.

3/ Government strategy:

India Semiconductor Mission.
PLI schemes for electronics & components.
Policy push for data localization + AI infrastructure.
Target: Position India as a serious alternative to Singapore, US & Middle East for AI workloads.

4/ The hard implications:

Enormous power demand (~200 GW+ cumulative need by 2030).
Massive water & land requirements.
Lower direct job creation than traditional manufacturing.
High execution pressure on grid, renewables integration & capital.

5/ Bottom line: India is moving from “AI services” to “AI infrastructure superpower” play.

Success here = stronger digital sovereignty + massive capex cycle.

This infra war can define India’s tech story for the next decade.

See 3 related tweets

@Sanjay_Sriv: Hyperspectral eyes in the sky. Made-in-India rockets. Global contracts pouring in. This is Bharat’s...
@Malay4Product: The headline is true but the story behind it is fascinating.

Taiwan's entire stock market jump is ...

@JayminSOfficial: India’s biggest AI export may not be software.

It may be reskilled human intelligence at scale.

Th...

14. SawyerMerritt (Group Score: 114.0 | Individual: 27.4)

Cluster: 6 tweets | Engagement: 1063 (Avg: 2535) | Type: Tech

Elon on Grok:

“Our AI will be great; I will never give up. Never. Space(XAI) is only 3 years old. That’s half the age of Anthropic and quarter the age of OpenAI. Let’s see where things stand 3 years from now.”

Acquiring Cursor will give SpaceXAI a big boost.\n\nQT @elonmusk: @PeterDiamandis SpaceX had achieved nothing of note after 3 years and was written off as dead after 6 years with 3 consecutive launch failures.

But you may have noticed that things are different now.

See 5 related tweets

@teortaxesTex: yep, Elon has gained some appreciation for the hardness of this task. Amazing. Now he admits he migh...
@JackKuhr: Anthropic was dead in the water 2 years ago. An incredible amount can change in a matter of months. ...
@Z_XSophie: Elon just dropped the real mindset: xAI is only 3 years old — half the age of Anthropic and a quarte...
@mark_k: Never bet against Elon.\n\nQT @elonmusk: What you say is true, but nonetheless our AI will be great....
@chrisparkX: RT @elonmusk: What you say is true, but nonetheless our AI will be great. Whether it is the best rem...

15. Parul_Gautam7 (Group Score: 112.9 | Individual: 30.7)

Cluster: 4 tweets | Engagement: 101 (Avg: 105) | Type: Tech

Voice agents are no longer just audio interfaces.

Adding real-time avatars creates a more human and interactive layer for AI communication.

The biggest win here is integrating visual presence without rebuilding existing stacks.\n\nQT @TryLiveAvatar: Voice agents are getting a face.

LiveAvatar now integrates with LiveKit, Pipecat, Agora, and VisionAgent so the agents you already built can become real-time avatar experiences in a few lines of code.

Keep your stack. Add presence.

Setup guides below ↓ https://t.co/xFQp9haYEL

See 3 related tweets

@Origin_AI_01: Voice agents are evolving from invisible assistants into real-time digital personalities.

The bigge...

@HeyGen: RT @TryLiveAvatar: Voice agents are getting a face.

LiveAvatar now integrates with LiveKit, Pipecat...

@alex_prompter: Nobody wants another “AI avatar demo” that dies outside the landing page.

This is the useful versio...

16. chddaniel (Group Score: 112.8 | Individual: 29.3)

Cluster: 4 tweets | Engagement: 10 (Avg: 21) | Type: Tech

so you're telling me now Claude Opus 4.7 can....

scan an entire website
build it as a mobile app
prepare for App Store submission
self-maintain the app

without any human in the loop?!?

it's so over... https://t.co/BZXQBJaDu5\n\nQT @chhddavid: Introducing Website to App. Turn any website into a mobile app.

Just paste the URL.

Claude Code Opus 4.7 will build a native mobile app inspired by the original website and launch it on the App Stores for you.

We’ve been using this internally a ton for iOS + Android apps. https://t.co/zB3Q0k7eTX

See 3 related tweets

@Shipper_now: THIS IS INSANE!!

He turned a website into a mobile app with Claude...\n\nQT @chhddavid: Introducing...

@chhddavid: Introducing Website to App. Turn any website into a mobile app.

Just paste the URL.

Claude Code Op...

@chddaniel: this is absolutely fucking terrifying..........\n\nQT @chhddavid: Introducing Website to App. Turn a...

17. GregKamradt (Group Score: 109.3 | Individual: 22.0)

Cluster: 6 tweets | Engagement: 27 (Avg: 45) | Type: Tech

Surprised open router is only at $1.3B given the wild rounds going on.

They have more of a business (both present and future) than ~99% of AI companies.\n\nQT @Techmeme: OpenRouter raised $113M led by CapitalG, a source says at a$ 1.3B valuation, and now processes 25T tokens across 400+ models weekly, up from 5T six months ago (@m_delamerced / New York Times)

(Visit Techmeme dot com for the link and full context!)

See 5 related tweets

@novita_labs: Huge congrats to the @OpenRouter team on the $113M Series B 🎉

25T weekly tokens is a massive miles...

@pingToven: RT @OpenRouter: Today we’re announcing our $113M Series B led by @CapitalGVC.

Over the last 6 month...

@MenloVentures: RT @deedydas: OpenRouter is now serving 1.5 quadrillion tokens/yr!

That token run rate is: — 15-30%...

@MenloVentures: Menlo is proud to back @OpenRouter in its $113M Series B, following our Series A lead last year.

AI...

@shiri_shh: OpenRouter just raised $113M at a$ 1.3B valuation

It is now processing 1.5 quadrillion tokens/year ...

18. jukan05 (Group Score: 104.6 | Individual: 30.6)

Cluster: 5 tweets | Engagement: 686 (Avg: 584) | Type: Tech

I told you, didn’t I? https://t.co/ks2fJohLtB\n\nQT @wallstengine: $QCOM STRIKES AI CHIP DEAL WITH BYTEDANCE

Bloomberg reports Qualcomm will supply ByteDance with AI-focused ASICs.

ByteDance is expected to buy millions of Qualcomm chips for AI agent software.

This could make ByteDance one of Qualcomm’s first major AI ASIC customers.

Qualcomm has been trying to expand from smartphone processors into AI infrastructure.

ByteDance is spending heavily too.

SCMP reported its 2026 AI capex was lifted by at least 25% to about $29.4B.

The deal depends on the chips staying within U.S. legal compute thresholds for China.

See 4 related tweets

@wallstengine: $QCOM STRIKES AI CHIP DEAL WITH BYTEDANCE

Bloomberg reports Qualcomm will supply ByteDance with AI-...

@MTSlive: SITUATION DETECTED: Qualcomm has reached a deal to supply chips for ByteDance's AI data centers, per...
@Techmeme: Sources: Qualcomm reached a deal with ByteDance to supply millions of ASICs for AI data centers to s...
@StockSavvyShay: $QCOM reportedly struck a deal to supply ByteDance with AI-focused ASICs potentially making TikTok’s...

19. aakashgupta (Group Score: 97.2 | Individual: 32.6)

Cluster: 3 tweets | Engagement: 30 (Avg: 170) | Type: Tech

A product manager at OpenAI wrote a document on Monday. By Friday, it shipped as a working feature. No engineer wrote a single line of code.

The team that pulled this off built a million-line app where zero lines were typed by a human. But the first month was 10x slower than just coding it the old way. On purpose.

The rule: engineers were banned from writing code. When the AI agent made a mistake, they couldn't fix it. They had to build a test, a lint, or a doc that prevents that class of mistake forever. Then run the agent again. Sometimes that meant recursing 6-8 levels deep just to give it access to Slack credentials.

After a month of that, the repository itself became the quality gate. Every piece of judgment the team had encoded into tests and docs applied automatically to every future task. The PM's document went straight to a pull request because the codebase rejected anything that didn't meet the bar. A designer shipped full UI features with zero backend. A former PM turned new-hire engineer spun up QA agents from the codebase's own feature docs.

Then GPT 5.2 dropped, and without any new harness work, they got 1-2 extra PRs per engineer per day. They built an orchestrator called Symphony that routes tickets to the agent and 10x'd weekly output per engineer. One refactor burned 350 million tokens across a 60-hour session. The engineer buckled his laptop into his car so the agent could keep running during his commute.

250,000 lines of that million-line codebase are markdown. Prompts embedded directly in the repo. The entire codebase is a prompt injection system aimed at the coding agent.

Every team speed-running AI adoption by letting engineers vibe code is building the wrong asset. The teams that win will spend their first month encoding what "good" looks like into tests, lints, and review agents. Then the machine produces at scale against those constraints, and nobody gives the same feedback twice.\n\nQT @aakashgupta: Ryan Lopopolo leads a team at OpenAI where the PM writes a PRD on Monday and ships a pull request by Friday. No human writes the code.

@_lopopolo broke it all down:

0:00 - "Code is a liability" 3:23 - Why your most expensive asset is now free 6:01 - "What's the point of roles anymore?" 8:04 - What replaces the PM/design/eng triangle 13:10 - 1M lines of code, zero written by humans 16:05 - Engineers can't touch the keyboard 18:13 - First month was 10x slower than solo 20:07 - Recursing 8 levels deep for one primitive 20:47 - PM writes PRD Monday, ships PR Friday 25:06 - The feature they had to trash 28:02 - How designers ship UI without a backend 31:40 - What's actually inside the harness 37:03 - Failing the build over curly quotes 40:02 - Inside Ryan's actual Codex setup 46:25 - The codebase that grades itself 50:49 - "A billion tokens a day or you're negligent" 52:19 - 350M tokens on a single PR 53:46 - GPT 5.2 changed everything overnight 57:00 - Every engineer is now a staff engineer 59:19 - The ego problem nobody talks about 1:00:39 - Monday morning roadmap for normal teams 1:08:19 - One skill to build this weekend 1:10:57 - Why one agent beats multi-agent

See 2 related tweets

@aakashgupta: A PM on my guest's team wrote a PRD in markdown on a Monday. By Friday there was a working demo. The...
@aakashgupta: She built her first eval in 5 minutes with Claude Code. Then explained exactly why that eval will fa...

20. wallstengine (Group Score: 92.5 | Individual: 30.4)

Cluster: 4 tweets | Engagement: 62 (Avg: 288) | Type: Tech

OPENAI IS BUILDING CHATGPT ADS FOR SMALL BUSINESSES

The Information reports OpenAI is expanding beyond big-brand advertisers like Adobe, Ford and Target.

The new focus includes smaller local businesses such as car washes and dry cleaners.

OpenAI is rolling out performance-style ad tools:

cost-per-click ads,

conversion-oriented ads,

self-serve buying,

ad pixel measurement,

Conversions API,

and no minimum spend for some buyers.

Advertisers can track actions like purchases, leads, sign-ups, appointments and form submissions.

OpenAI has told investors it expects ad revenue to hit $2.4B this year and$ 102B by 2030.

See 3 related tweets

@StockSavvyShay: OpenAI is reportedly building ChatGPT ads for small businesses by expanding into local advertisers l...
@theinformation: Exclusive: OpenAI is expanding its ad business beyond big brands and into the market for small busin...
@theinformation: Exclusive: OpenAI is dropping some ad spending minimums and adding tools for smaller advertisers, a ...