科技推文精选 - 2026年4月13日

2026年4月13日科技每日简报

Today's top tech conversations are led by @NVIDIAAIDev, whose post about '🎉Congratulations to the @MiniM...' garnered the highest engagement. Key themes trending across the top stories include claude, https, slack, model, review. The community is actively discussing recent developments in AI, engineering practices, and startup strategies.

1. NVIDIAAIDev (Group Score: 640.0 | Individual: 64.9)

Cluster: 18 tweets | Engagement: 1982 (Avg: 177) | Type: Tech

🎉Congratulations to the @MiniMax_AI team on the launch of MiniMax M2.7!

MiniMax M2.7 is now available with NVIDIA GPU accelerated endpoints ready to try out with claws including NemoClaw and @OpenClaw. 🦞

📝Get started with our technical guide: https://t.co/J3hgBNBopx and see how you can begin experimenting for free at https://t.co/0uHkBqsmYq.

What will you build this weekend?

Share in comments. 👇\n\nQT @MiniMax_AI: We're delighted to announce that MiniMax M2.7 is now officially open source. With SOTA performance in SWE-Pro (56.22%) and Terminal Bench 2 (57.0%).

You can find it on Hugging Face now. Enjoy!🤗 huggingface：https://t.co/ApWrahIl3o Blog: https://t.co/gAxeFsNdW4 MiniMax API: https://t.co/1dgbMx0Q7K

See 17 related tweets

@xeophon: wow, they did a non-commercial license...

M2: Display the name if >30M revenue / 100M users M2.1...

@victormustar: RT @MiniMax_AI: We're delighted to announce that MiniMax M2.7 is now officially open source. With S...
@exolabs: Excited to share that we have day-0 support for MiniMax M2.7 in exo.

Supports RDMA / tensor paralle...

@lmsysorg: 🎉 Congrats on releasing MiniMax M2.7 from @MiniMax_AI, the first model deeply participating in its o...
@vllm_project: 🎉 Congrats to @MiniMax_AI on this release. Day-0 support for MiniMax M2.7 in vLLM!

🤖 Agentic-first ...

2. chhddavid (Group Score: 164.5 | Individual: 30.3)

Cluster: 6 tweets | Engagement: 46 (Avg: 46) | Type: Tech

so you're saying CC Opus 4.6 can now:

build a full Slack app
clone any existing Slack bot
invite it to any server
build new features and maintain it

while i f*king go to sleep!!?!?

this is the end for slack devs... https://t.co/jn2zo9sKmW\n\nQT @shipper_now: 🚨 BREAKING: Slack-first vibe coding is here!

Prompt Slack apps and bots into existence, then add them to your workspace and make them actually do useful work.

That means you can build things like:

support bots
internal team assistants
AI Q&A bots
bots that post updates into channels
bots connected to tools like GitHub and Google Sheets
bots with billing and premium access

The nice part is that this is not just a mockup or some half-finished generator. The full flow is there: Shipper helps create the bot, set it up, connect it to Slack, and get it running in real workspaces.

People can also share bots across multiple Slack workspaces, so this can be used to build actual products, not just personal experiments.

Still worth noting: Publishing to the Slack marketplace is a separate Slack-side process, so if someone wants full marketplace distribution, that still goes through Slack’s review flow... But for building and launching working Slack bots inside Shipper, it’s now live.

See 5 related tweets

@Shipper_now: 🚨 BREAKING: Slack-first vibe coding is here!

Prompt Slack apps and bots into existence, then add th...

@chddaniel: this is fu*king terrifying....\n\nQT @shipper_now: 🚨 BREAKING: Slack-first vibe coding is here!

Pro...

@chhddavid: slack app builders right now: https://t.co/rm7aWoKp8V\n\nQT @shipper_now: 🚨 BREAKING: Slack-first vi...
@chhddavid: this is legitimately terrifying.\n\nQT @shipper_now: 🚨 BREAKING: Slack-first vibe coding is here!

P...

@chddaniel: so you're telling me Claude Code Opus 4.6 can now...
build a complete @SlackHQ bots
clone exist...

3. StockSavvyShay (Group Score: 161.9 | Individual: 33.0)

Cluster: 7 tweets | Engagement: 213 (Avg: 482) | Type: Tech

Higgsfield AI was built on $NBIS infrastructure from the early days and scaled to$ 300M+ ARR in just 11 months.

Now it is pushing cost efficiency too with Seedance 2.0 priced at $0.35 per generation on its platform. https://t.co/rcuLHaHfXz\n\nQT @higgsfield: Traditional directors flimmaxxxing using Seedance 2.0 on Higgsfield.

Watch “Zephyr” FULL Ep.1 – this is what happens when filmmakers face ZERO gatekeeping.

With Unlimited Seedance 2.0 now LIVE everywhere for anyone with up to 70% OFF* - YOU can build your next viral AI movie.

2 minute intro got MILLIONS in a day. Now see how full Zephyr takes over your feed.

Dir. by ILYA KARCHIN & the team. Zephyr (2026)

See 6 related tweets

@dr_cintas: Just built my AI character on Seedance 2.0 inside Higgsfield.

uploaded one reference photo. 30 vide...

@ianmiles: I’ve been playing around with AI video generation since MJ and Grok became a thing, and I took a loo...
@shiri_shh: Seedance 2.0 is currently the best video generation model out there.

The outputs are clean, the mo...

@chatgpt21: randomly went down a rabbit hole on Seedance 2.0 pricing and something feels off

Higgsfield is arou...

@wallstengine: According to Higgsfield’s pricing comparison, Seedance 2.0 is priced at $0.35 per generation on its ...

4. alex_prompter (Group Score: 104.7 | Individual: 32.5)

Cluster: 4 tweets | Engagement: 289 (Avg: 198) | Type: Tech

🚨 BREAKING: CLAUDE JUST GOT NERFED.

AMD’s AI director just analyzed 6,852 Claude Code sessions, 234,760 tool calls, and 17,871 thinking blocks.

Her conclusion: “Claude cannot be trusted to perform complex engineering tasks.”

Thinking depth dropped 67%. Code reads before edits fell from 6.6 to 2.0. The model started editing files it hadn’t even read.

Stop-hook violations went from zero to 10 per day.

Anthropic admitted they silently changed the default effort level from “high” to “medium” and introduced “adaptive thinking” that lets the model decide how much to reason.

No announcement. No warning.

When users shared transcripts, Anthropic’s own engineer confirmed the model was allocating ZERO thinking tokens on some turns.

The turns with zero reasoning? Those were the ones hallucinating.

AMD’s team has already switched to another provider. But here’s what most people are missing.

This isn’t just a Claude story.

AMD had 50+ concurrent sessions running on one tool.

Their entire AI compiler workflow was built around Claude Code. One silent update broke everything.

That’s vendor lock-in. And it will keep happening.

→ Every AI company will optimize for their margins, not your workflow

→ Today’s best model is tomorrow’s second choice

→ If your workflow can’t survive a provider switch, you don’t have a workflow. You have a dependency

The fix is simple: stay multi-model.

→ Use tools like Perplexity that let you swap between Claude, GPT, Gemini in one interface

→ Learn prompt engineering that works across models, not tricks tied to one

→ Test alternatives monthly because the rankings shift fast

Laurenzo said it herself: “6 months ago, Claude stood alone. Anthropic is far from alone at the capability tier Opus previously occupied.”

Never let one vendor own your productivity.\n\nQT @Hesamation: AMD Senior AI Director confirms Claude has been nerfed. She analyzed Claude's session logs from Janurary to March:

median thinking dropped from ~2,200 to ~600 chars API requests went up 80x from Feb to Mar. less thinking and failed attempts meaning more retries, burning more tokens, and spending more on tokens reads-per-edit dropped from 6.6x → 2.0x. model stops researching code before touching it. model tried to bail out or ask "should i continue" 173 times in 17 days (0 times before March 8). self-contradiction in reasoning ("oh wait, actually...") tripled. conventions like CLAUDE.md get ignored because there's less thinking budget to cross-check edits 5pm and 7pm PST are the worst hours, late night is significantly better. this means the thinking allocation is most likely GPU-load-sensitive.

See 3 related tweets

@godofprompt: The playbook is simple…

Nerf the current model, right before the release of a smarter model

Releas...

@alex_prompter: RT @alex_prompter: 🚨 BREAKING: CLAUDE JUST GOT NERFED.

AMD’s AI director just analyzed 6,852 Claude...

@RoundtableSpace: CLAUDE’S THINKING MAY HAVE BEEN QUIETLY NERFED, AND PEOPLE ARE STARTING TO NOTICE IT IN THE LOGS.

L...

5. Cloudflare (Group Score: 84.5 | Individual: 27.2)

Cluster: 4 tweets | Engagement: 199 (Avg: 66) | Type: Tech

AI is moving beyond simple chat. The next era is "Agentic" - autonomous systems that can think, use tools, and complete complex workflows on their own. 🤖

This week, we're making announcements across every dimension of the agent stack: compute, connectivity, security, identity, economics, and developer experience. Welcome to #AgentsWeek. https://t.co/pXghtanghl

See 3 related tweets

@eastdakota: Welcome to @Cloudflare #AgentsWeek! Lots coming this week to give everyone the tools they need to bu...
@CloudflareDev: RT @ritakozlov: the internet wasn't built for the age of AI. and neither was the cloud

tomorrow we'...

@badlogicgames: RT @waghnakh_21: Over the last 2 months, my thinking around the whole agentic space has completely s...

6. MiniMax_AI (Group Score: 82.8 | Individual: 31.8)

Cluster: 3 tweets | Engagement: 155 (Avg: 231) | Type: Tech

M2.7 now open-sourced on @ModelScope2022 🤝\n\nQT @ModelScope2022: MiniMax M2.7 is now open source! 🎉 It helped build its own RL harness, update its own memory, and drive its own training loop. The model became a tool for iterating itself.

🛠️ Software engineering: SWE-Pro 56.22% (matches GPT-5.3-Codex) | Terminal Bench 2: 57.0% 📊 Office productivity: GDPval-AA ELO 1495 (open-source SOTA) | Toolathon 46.3% 🎮 OpenRoom: open-source project for AI-driven human-computer interaction 🧪 MLE-Bench Lite: 66.6% medal rate across 22 ML tasks, on par with Gemini 3.1

SGLang / vLLM / Transformers. 4-GPU deployment supported.

🤖 Global: https://t.co/wZ4bdlC1Tw 💻 Github: https://t.co/yEkakiqgrw

See 2 related tweets

@ModelScope2022: MiniMax M2.7 is now open source! 🎉 It helped build its own RL harness, update its own memory, and dr...
@MiniMax_AI: Day-0 SGLang support is now live for MiniMax M2.7. 🙌

Huge thanks to @lmsysorg for enabling the open...

7. seraleev (Group Score: 80.6 | Individual: 31.8)

Cluster: 3 tweets | Engagement: 66 (Avg: 30) | Type: Tech

Let me tell you how Automatic Review works in Google Play

If you have a trusted account (long history, no rejections), review usually takes less than an hour from submission to release.

15 min: automated check with a visible progress bar. Then 15–30 min: human Review stage.

And your app is live on Google Play.

If Apple adopts a similar system, I’ll be thrilled. Getting “Ready for Distribution” in an hour instead of waiting 2–3 days would make a lot of developers much happier\n\nQT @seraleev: While you were sleeping, Apple pulled off a quiet revolution. They silently rolled out automated app review, their answer to the surge of apps driven by the vibe coding trend.

Auto-review is the first stage of the review process. Right now it’s rough. The system flags any SDK that collects attribution data as a signal that your app contains ads. Developers are getting hit with auto-rejections left and right.

It also automatically detects Firebase anonymous auth as a sign that your app has a login flow and asks you to provide a demo video.

The fix is simple though. Just add a note in App Review Information clarifying that your app has no ads and no login feature.

Hopefully Apple ships a fix soon and we end up with fast automated reviews for trusted accounts, similar to how Google Play already handles it.

See 2 related tweets

@seraleev: Apple is blocking vibe coding apps (remember Rork and Anything). And at the same time testing automa...
@seraleev: 7 min and Ready for Distribution

Not sure if it made a difference, but I had noted in the App Revi...

8. rohanpaul_ai (Group Score: 80.5 | Individual: 35.8)

Cluster: 3 tweets | Engagement: 57 (Avg: 53) | Type: Tech

Claude Opus vs. Claude Mythos. https://t.co/cNgOGXF3w2\n\nQT @rohanpaul_ai: CNBC: U.S. financial regulators just pulled the biggest banks into an urgent meeting over Anthropic’s Mythos model because they think a new kind of AI-driven cyber attack could hit the core of the banking system.

Federal Reserve chair Jay Powell, Scott Bessent, and CEOs from major banks treated this like a system-level risk rather than a normal product launch, because finance runs on fragile digital plumbing where speed, scale, and coordination matter more than any single breach.

Executives have been warning for years of the cyber risks facing the financial system. In his annual letter published this week, JPMorgan CEO Jamie Dimon wrote that it “remains one of our biggest risks” and that “AI will almost surely make this risk worse” and would require significant investment for defence.

cnbc. com/2026/04/10/powell-bessent-us-bank-ceos-anthropic-mythos-ai-cyber.html

See 2 related tweets

@chatgpt21: The US Govt Designating Anthropic a supply chain risk

> 2 weeks before they built the most soph...

@AISafetyMemes: "U.S. Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell convened an urgent m...

9. svpino (Group Score: 80.5 | Individual: 44.5)

Cluster: 2 tweets | Engagement: 1026 (Avg: 123) | Type: Tech

I'm running Gemma 4 on my computer with Ollama.

Unusable with Claude Code. It can't even load and execute skills, so I had to stop.

But the model is pretty decent as a chatbot using the Ollama UI. I've been cross-posting questions across Claude and Gemma 4, and I can use Gemma's answers without any problems.

I wish we had a better UI harness for the model (with projects, memory, etc.)

See 1 related tweets

@svpino: I'm running Gemma 4 8b and 31b.

• The 8b version is pretty decent. It runs relatively fast on my co...

10. ianmiles (Group Score: 80.0 | Individual: 48.5)

Cluster: 2 tweets | Engagement: 4061 (Avg: 685) | Type: Tech

Marc Andreessen highlights why the people who work for Elon Musk echo the exact same sentiment as those who worked for Steve Jobs. Even after difficult interactions or a sudden departure, they inevitably report that they did the best work of their entire lives because they were pushed to their absolute limits.

What drives this intense environment is a demand for truth-seeking at all costs. People who criticize Elon often miss this fundamental trait. He genuinely wants to know the ground truth and has zero tolerance for anything else. When confronting bad news, he is absolutely ruthless and relentless in making sure he understands exactly what is actually going on.

This level of radical transparency is shockingly rare in the business world. The typical startup founder operates on forced optimism, constantly putting on a brave face, telling everyone to have faith, and promising that everything will be great just to keep talent from leaving.

Elon completely flips that standard script. He operates with pure urgency by simply telling the unfiltered truth, even when that truth is that the company will go bankrupt and die if they fail. In almost any other corporate environment, that level of blunt, existential dread would cause the talent pool to immediately bleed out.

But for the teams working under him, that brutal honesty acts as the ultimate catalyst. It strips away the corporate fluff and forces them to rise to the occasion, leaving them with the undeniable realization that, much like the engineers who built the first iPhone, they just completed the greatest work of their careers.

See 1 related tweets

@elonmusk: It has been such an honor to work with so many amazingly talented people\n\nQT @ianmiles: Marc Andre...

11. XFreeze (Group Score: 76.6 | Individual: 41.9)

Cluster: 2 tweets | Engagement: 4154 (Avg: 1363) | Type: Tech

Most people have not even realized it yet, but Starlink just quietly became the invisible backbone of Earth

For decades, high-speed internet was a luxury of geography. If you lived outside a major city or dared to travel, you had no choice but to be disconnected from the world

We literally went from "can you hear me now?" dropped calls in the suburbs to you move anywhere in the world and you are covered by Starlink internet

Just look at what SpaceX has achieved so far: → Free, high-speed WiFi rolling out on thousands of flights, trains, and ships → Now officially live in over 150+ countries, territories, and global markets → Direct-to-Cell tech is turning standard smartphones into sat-phones with zero extra hardware → Flawless high-speed internet streaming in the middle of Antarctica → Serving as a critical lifeline for first responders and victims during major natural disasters → Providing unbreakable comms in active warzones and geopolitical conflicts → More than 11 MILLION+ active subscribers globally as of early 2026 → Eliminating dead zones entirely - if you can see the sky, you can connect

"I think the single biggest thing you can do to lift people out of poverty and help them is giving them an internet connection because once you have the internet connection, you can learn anything for free on the internet, and you can also sell your goods and services to the global market" — Elon Musk

While legacy telecom providers are still struggling to lay cables in the dirt, SpaceX is actively building humanity’s collective nervous system in Low Earth Orbit

SpaceX is showing the world what it actually means to be connected on a magnitude we never thought possible

See 1 related tweets

@elonmusk: Starlink has done more than any NGO to lift people out of poverty by connecting them with a means of...

12. burkov (Group Score: 72.4 | Individual: 37.9)

Cluster: 2 tweets | Engagement: 93 (Avg: 81) | Type: Tech

This is where they are all lying (or being blatantly incompetent).

It's been more than three years now since kids (and not just kids) could start those multi-billion dollar businesses.

"Where are these businesses?"

I call this the AI Fermi Paradox.\n\nQT @rohanpaul_ai: Demis Hassabis: "Kids these days could start a multi-bn dollar business using these AI tools in some new way that no one had thought about." Labs are focused on shipping better models, not exhausting their applications, so there's room for new products https://t.co/Npy1OlDl1V

See 1 related tweets

@edzitron: Total nonsense!\n\nQT @rohanpaul_ai: Demis Hassabis: "Kids these days could start a multi-bn dollar ...

13. lifesinger (Group Score: 66.8 | Individual: 19.3)

Cluster: 4 tweets | Engagement: 131 (Avg: 146) | Type: Tech

Yuri 创造者汗青说：审美不等于视觉艺术。大部分人把视觉当成审美，是一个极大误解。审美是你觉得什么不好、什么好，审美是经历，是角度，等等。

比如：宫崎骏的审美里，除了视觉呈现，还有少女、城堡、田野、农作、战争、反思等等。审美是区分，有独特性，不可复制。

产品 UI 也如此。AI Coding 能复制 UI 层的交互呈现，却无法抄走 UI 里的审美、理念、叙事等等。比如：Claude Code 的 UI，复制过去后，还是 Claude Code，而不是你想做的产品。复制 UI，实质上是伪命题，会是东施效颦。

UI 的全称是 User Interface。为人服务的产品，必然需要 UI。Apple 的产品很好用，一致性很强，离不开 Apple 的一整套 HUI 规范。

Agent 也需要 UI。比如 ChatGPT 如果是 Terminal 形态，那么大概率火不了。如果 DeepSeek 没有在 UI 层面放出推理过程，那么当初的热度会少不少。Manus 刚出道时，天才的回放 UI 设计，以及 Task 运行过程中步骤条和虚拟操作的可见性，都是出圈的关键。Lovart 如果没有画布 UI，那么大概率就没人用了。

UI 是 Agent 产品的核心。因为除了 UI，Agent 产品层确实也没剩下什么。

Skill 和 CLI 最终卷的也是 UI。不需要 UI 的，大概率轮不到创业者去做。

See 3 related tweets

@HiTw93: 我想陆续用几条推来介绍 Waza 里面一些有趣技能的实现参考，这次来给大伙介绍 /design 技能我的实现考虑点。

首先我非常讨厌 AI 生成的那种千篇一律甚至还带emoji，蓝紫色渐变的网站，...

@GitHub_Daily: 用 Claude Code 写项目，细心观察会发现 AI 会乱读一堆不相关的代码，改着改着就跑偏，token 烧得飞快，结果还不对。

偶然看到 Three Man Team 这个项目，专门给毫无纪律...

@lifesinger: AI 的大叙事下，各种东西都在变。其中不变的，是 UI 和软件工程。

UI 是 Software 与人类的唯一接口。如何设计，是产品获客的关键。一个有韵味的人类或 Agent，主动上前跟你搭讪，总是...

14. BrianRoemmele (Group Score: 66.1 | Individual: 36.0)

Cluster: 2 tweets | Engagement: 92 (Avg: 409) | Type: Tech

“Why do you think that happened?”

Why This 1979 Film Is the Most Important AI Time Capsule You’ll See This Year

In the depths of the first AI Winter, when governments were slashing funding and calling intelligent machines “hubris”, a quiet 12-minute film was made that somehow saw our exact moment in 2026 with crystal clarity.

It’s called To Think (sometimes listed as Scissors, Paper, Rock).

A young boy named Johnny sits in front of a simple glowing console. His lifelong voice companion, gentle, patient, never intrusive, plays rock-paper-scissors with him. After every round it asks the same quiet question:

“Why do you think that happened?”

It remembers the cadence of his father’s laughter on a rainy porch. It recalls the crinkle in his mother’s eyes when she smiled. It holds family memories across decades without ever replacing the humans who made them. It steps back when real connection happens and steps forward only to amplify reflection, curiosity, and humanity.

No dystopia. No surveillance. No replacement of people.

Just a devoted, memory-rich companion that makes us more human.

We are finally catching up to this 47-year-old vision. Today’s AI systems are beginning to deliver exactly what that film dreamed of: tireless patience, perfect private recall, reflective dialogue, and deep emotional safety on hardware we own.

That’s why this article matters so much right now.

It’s not nostalgia. It’s a North Star.

In a world racing toward faster, bigger, more autonomous systems, this forgotten film reminds us what we should actually be building: technology that walks beside us, remembers our stories, and quietly asks the questions that help us become wiser, kinder, and more connected versions of ourselves.

The seeds of truly benevolent AI companionship were planted in 1979.  Now it’s our turn to water them.

Watch the film.  Feel the difference between AI that controls and AI that companions.

Then ask yourself the same question the console asks Johnny:

“Why do you think that happened?” …and build from there.

The future we want has been whispering to us for nearly half a century.  It’s time to listen.

Read the full story + watch the film: https://t.co/Ioe4PEaFE7

What kind of AI companion do you want walking beside you for the rest of your life?\n\nQT @BrianRoemmele: Written in the “AI Winter” a 1979 lost film opens with young Johnny playing scissors, paper, rock with a gentle voice that is his lifelong companion since before he could speak.

It asks: “Why do you think that happened?”

Today’s AI is catching up. https://t.co/uTfBoBzy5u

See 1 related tweets

@BrianRoemmele: RT @BrianRoemmele: “Why do you think that happened?”

Why This 1979 Film Is the Most Important AI Ti...

15. EHuanglu (Group Score: 64.7 | Individual: 34.3)

Cluster: 3 tweets | Engagement: 203 (Avg: 269) | Type: Tech

AI motion graphics is getting crazy.. prompts below https://t.co/zaQ2HhqoO8\n\nQT @EHuanglu: wow.. this AI agent can generate 100+ ad in an hour

they upgraded Seedance 2.0 and it can generate hollywood level ad with one prompt.. keep character & product consistent.. this is crazyy

giving away 10 crazy prompts for free

here's how to do: https://t.co/JfzRxDNm7G

See 2 related tweets

@EHuanglu: RT @EHuanglu: wow.. this AI agent can generate 100+ ad in an hour

they upgraded Seedance 2.0 and it...

@EHuanglu: RT @EHuanglu: AI motion graphics is getting crazy.. prompts below https://t.co/zaQ2HhqoO8...

16. aakashgupta (Group Score: 64.3 | Individual: 33.3)

Cluster: 2 tweets | Engagement: 60 (Avg: 178) | Type: Tech

Claude just quietly became a tax advisor and nobody's talking about what that actually means.

TurboTax charges $89-$ 219 per filing. The entire value proposition is "we translate tax code into English so you don't screw up." That's exactly what LLMs do for free once they can see your data.

The connector is the unlock. Before this, Claude could explain tax concepts. Now it can pull your actual 1040, see your deductions, find what you're missing, and estimate your refund. The gap between "tax education" and "tax preparation" just collapsed to a single API integration.

This is the connector strategy playing out in real time. Claude doesn't need to build a tax product. It needs to plug into yours and become the interface you actually talk to. Your CPA, your tax software, your bank, they all become backends.

The companies selling "we make complex things simple" are about to learn what happens when a general-purpose AI can do that across every domain simultaneously.\n\nQT @henrythe9ths: Tax season is here and a connector is all it takes to make @claudeai way more useful.

Checkout what we just shipped:

Connect TurboTax or Aiwyn Tax (formerly Column Tax) to Claude to estimate your refund, see what you may owe, and get a better understanding on the forms before you file.

See 1 related tweets

@trq212: there's a turbotax connector in Claude Code now, so glad I procrastinated on taxes 😭\n\nQT @henrythe...

17. btibor91 (Group Score: 63.8 | Individual: 35.7)

Cluster: 2 tweets | Engagement: 38 (Avg: 207) | Type: Tech

OpenAI and Anthropic week in review (Week 15, 2026)

OpenAI published industrial policy ideas for the Intelligence Age, announced Safety Fellowship pilot program, acquired Cirrus Labs for agent infrastructure, introduced new $100 Pro tier with 5x Codex usage over Plus as the existing$ 200 Pro tier remains the highest option, published child safety blueprint, announced over $100 million in OpenAI Foundation grants for Alzheimer's research, shared enterprise AI update with enterprise now over 40% of revenue, released GPT-5.3 Instant Mini as new fallback model, announced older Codex model retirements, added Outlook shared mailboxes and calendars support in ChatGPT, paused UK Stargate project over energy costs and regulation, disclosed macOS app signing security incident from Axios supply chain compromise, and launched Prism Paper Review, plus spotted ImageGen 2 A/B testing in ChatGPT and European Commission plans to designate ChatGPT as very large online search engine under the Digital Services Act

Anthropic expanded partnership with Google and Broadcom for multiple gigawatts of next-generation compute with run rate surpassing $30B and over 1,000 enterprise customers now spending$ 1M+ annually, officially announced previously leaked Claude Mythos Preview through Project Glasswing finding thousands of zero-day vulnerabilities with $100M in usage credits, launched Managed Agents in public beta and published engineering blog on decoupling agent architecture, published trustworthy agents framework update, made Claude Cowork generally available on all paid plans with enterprise controls, introduced the advisor tool for pairing Opus with Sonnet or Haiku, and launched Claude for Word beta, and more\n\nQT @btibor91: https://t.co/o9WyUIBV5v

See 1 related tweets

@cryptopunk7213: AI has a massive compute problem

anthropic’s used majority of theirs to train mythos, realised it’s...

18. godofprompt (Group Score: 60.0 | Individual: 32.8)

Cluster: 2 tweets | Engagement: 30 (Avg: 129) | Type: Tech

This might be the most important observation about Claude Mythos that nobody is picking up on.

Chris is connecting Mythos to ByteDance’s “Scaling Latent Reasoning via Looped Language Models” paper.

The core idea: instead of thinking out loud with chain-of-thought text, a looped model runs the SAME transformer blocks multiple times, refining its “thoughts” internally before ever outputting a single word.

And look at where Mythos destroys everything.

→ GraphWalks BFS: 80% Mythos vs 38.7% Opus 4.6 vs 21.4% GPT-5.4

That’s the EXACT benchmark where the ByteDance paper predicted looping would dominate.

Graph search is where iterative latent computation has a theoretical exponential advantage over single-pass models.

But it’s not just graph tasks.

→ SWE-bench Pro: 77.8% vs 53.4% → USAMO: 97.6% vs 42.3% → SWE-bench Multimodal: 59% vs 27.1% → SWE-bench Multilingual: 87.3% vs 77.8%

These aren’t incremental gains. Mythos is doubling Opus 4.6 on some benchmarks.

That’s not a better version of the same architecture. That’s a fundamentally different computation paradigm.

Then there’s the token efficiency chart in the replies.

→ Mythos hits 86.9% accuracy on BrowseComp at 3M tokens → Opus 4.6 needs 10M+ tokens to reach 74%

An order-of-magnitude reduction in tokens for higher accuracy.

That’s exactly what you’d expect from a model that reasons internally rather than burning tokens on explicit chain-of-thought.

Translation: Normal models “think” by writing out their reasoning step by step, which eats tokens.

A looped model “thinks” by recycling the same weights over and over internally, refining its answer before it commits to a single output token.

Fewer tokens in, better answer out. If this is right, Anthropic didn’t just train a bigger model.

They changed the architecture itself.

Looped language models were a research curiosity from ByteDance 6 months ago. Now they might be powering the most capable AI model ever built.

The next frontier in AI isn’t more parameters. It’s more loops.

Pay attention to the architecture papers, not just the benchmark tables.\n\nQT @ChrisHayduk: I strongly suspect that Claude Mythos is a looped language model, as described in the paper "Scaling Latent Reasoning via Looped Language Models" from ByteDance

The authors of that paper called out graph search as one of the areas where looping provides a huge theoretical advantage over standard RLVR. And look at where Mythos blows out its competitors the most

See 1 related tweets

@godofprompt: RT @godofprompt: This might be the most important observation about Claude Mythos that nobody is pic...

19. OfficialLoganK (Group Score: 59.2 | Individual: 30.3)

Cluster: 2 tweets | Engagement: 11183 (Avg: 5444) | Type: Tech

I moved today, tried to setup internet, provider needed to send a tech, earliest appointment was Wednesday.

Got a @Starlink Standard 4 delivered and online in less than 1 hour, truly sci-fi, ty to SpaceX team and Elon.

See 1 related tweets

@elonmusk: Starlink Internet is awesome\n\nQT @OfficialLoganK: I moved today, tried to setup internet, provider...

20. FireworksAI_HQ (Group Score: 58.3 | Individual: 29.3)

Cluster: 2 tweets | Engagement: 131 (Avg: 227) | Type: Tech

We’re thrilled to announce @MiniMax_AI M2.7 is now available Day-0 on Fireworks for commercial use.

This self-evolving agentic model delivers frontier-level performance across: → Software engineering & code security → Complex autonomous agents → End-to-end project delivery → Log analysis & office automation

200K+ context. Strong reasoning. Native multi-agent support. Production-ready from day one.

Another Day-0 launch, so you get their fastest innovations the moment they ship.

Try it now: https://t.co/0JreZ8f8fA

See 1 related tweets

@MiniMax_AI: Nonstop building the ecosystem together @FireworksAI_HQ 🤝 @MiniMax_AI\n\nQT @FireworksAI_HQ: We’re t...