科技推特精选 - 2026年3月28日

2026年3月28日科技每日简报

Today's top tech conversations are led by @rohanpaul_ai, whose post about 'Fortune just reported about th...' garnered the highest engagement. Key themes trending across the top stories include anthropic, mythos, https, launch, turboquant. The community is actively discussing recent developments in AI, engineering practices, and startup strategies.

1. rohanpaul_ai (Group Score: 246.6 | Individual: 35.0)

Cluster: 11 tweets | Engagement: 67 (Avg: 81) | Type: Tech

Fortune just reported about this leak:

Anthropic's new generation of super-strong models, "Claude Mythos"

and Anthropic believes it "poses unprecedented cybersecurity risks."

fortune. com/2026/03/27/anthropic-data-leak-reveals-powerful-secret-mythos-ai-model/ https://t.co/g4h9rJav9h\n\nQT @rohanpaul_ai: Rumor has it Anthropic is getting ready to launch new models, Mythos and Capybara, with Mythos being a brand-new tier that is larger than Opus.

Mythos is built as a step above Opus 4.6 rather than a routine upgrade.
The leak also says Mythos is extremely expensive to run, which usually means a larger system with heavier computation, longer reasoning, or both, so cost is part of the product story, not just an engineering detail.
Anthropic appears to be slowing release, limiting access, and starting with a small set of security-focused users instead of pushing a broad launch.

IMO, Anthropic may be trying to define a new premium tier where raw capability matters more than wide availability, with cyber defense as the safest first market.

See 10 related tweets

@MarioNawfal: 🚨LEAKED: ANTHROPIC BUILT AN AI SO GOOD AT HACKING THEY'RE AFRAID TO RELEASE IT...

A data leak just ...

@cryptopunk7213: holy shit Anthropic accidentally leaked a new AI model that’s “by far the most powerful AI model we’...
@cryptopunk7213: this Anthropic leak is definitely Claude 5 and it’s so good that it’s a major security risk and can’...
@Cointelegraph: 🚨 NOW: Anthropic confirms it is testing a new AI model called Claude Mythos, described as its most c...
@minchoi: 🚨BREAKING: Fortune says leaked Anthropic docs show "Claude Mythos" is already in testing...

And Ant...

2. rohanpaul_ai (Group Score: 119.0 | Individual: 31.6)

Cluster: 6 tweets | Engagement: 76 (Avg: 81) | Type: Tech

OpenAI just pushed Codex past the editor by rolling out plugins that connect it to tools like Slack, Figma, Notion, Gmail, and Google Drive, so the agent can work across the mess around code instead of only the code itself.

bundles apps, skills, and optional MCP servers into one installable package, which means Codex gets tool access, reusable workflows, and outside context as a single unit rather than a pile of manual setup.

OpenAI also made plugins portable across the app, CLI, and IDE extensions, and added @ plugin-creator so teams can scaffold local plugins and share the same setup across projects.\n\nQT @OpenAIDevs: We're rolling out plugins in Codex.

Codex now works seamlessly out of the box with the most important tools builders already use, like @SlackHQ, @Figma, @NotionHQ, @gmail, and more.

https://t.co/PQDsLqHGA6 https://t.co/TIbsIUAf6S

See 5 related tweets

@testingcatalog: OpenAI released Plugins for Codex! Plugins are combinations of Apps and Skills to enable Codex to wo...
@WesRoth: OpenAI announced a major expansion to its Codex coding platform, rolling out plugins that connect th...
@kimmonismus: Competition between OpenAI and Anthropic is heating up!\n\nQT @thsottiaux: Hello. We have reset Code...
@Cointelegraph: ⚡️ UPDATE: OpenAI is rolling out plugins for Codex, adding native support for tools like Slack, Figm...
@arstechnica: OpenAI brings plugins to Codex, closing some of the gap with Claude Code https://t.co/WSz8AwAoVI...

3. JaredSleeper (Group Score: 112.7 | Individual: 37.1)

Cluster: 5 tweets | Engagement: 277 (Avg: 47) | Type: Tech

Wait, cybersecurity stocks are falling because of reports that Anthropic's new model is going to be so excellent on the attack side that Anthropic is briefing enterprises + giving them time to prepare?

That's a cybersecurity tailwind, no?

See 4 related tweets

@Mayhem4Markets: It's actually kind of funny that this Anthropic headline is causing cybersecurity stocks to fall aga...
@business: Cybersecurity stocks slumped on Friday after a Fortune report raised concern that an Anthropic AI mo...
@CNBC: Cybersecurity stocks fall on report Anthropic is testing a powerful new model https://t.co/gcdKUWLMd...
@zerohedge: AI Disruption Returns: Cybersecurity Stocks Tumble On Report Of New Anthropic "Step Change" AI Model...

4. FirstSquawk (Group Score: 107.4 | Individual: 42.6)

Cluster: 5 tweets | Engagement: 1488 (Avg: 132) | Type: Tech

GOOGLE IS CLOSE TO STRIKING A DEAL TO FUND ANTHROPIC'S DATA CENTER, ACCORDING TO FT.

See 4 related tweets

@BrianRoemmele: Google will likely own Anthropic at some point.

It leaves us with SpaceX and Google in the end.\n...

@Cointelegraph: 🚨 JUST IN: Google is nearing a deal to help finance a $5B+ Texas data center leased to Anthropic, ex...
@FT: Google nears deal to help finance multibillion-dollar data centre leased to Anthropic https://t.co/a...
@FT: FT Exclusive: Google plans to throw its financial support behind a data centre project in Texas, lea...

5. petergostev (Group Score: 103.1 | Individual: 36.4)

Cluster: 3 tweets | Engagement: 102 (Avg: 79) | Type: Tech

Open AI's initial ad revenue in context. Note that this is a log scale, otherwise it would look ridiculous. To be fair, 1/6th of New York Times in a matter of months is not bad at all https://t.co/Ey1VF3d8yT\n\nQT @steph_palazzolo: New: OpenAI has surpassed $100m in ARR from its ads pilot, which launched 6 weeks ago. It's expanded to 600+ advertisers and plans to launch self-serve advertiser access in April.

https://t.co/CihJpMpEZm

See 2 related tweets

@ns123abc: And here’s OAI showing off about $100M “annualized” https://t.co/RmozFXJGwr\n\nQT @steph_palazzolo: ...
@jasonlk: This might just mean they sold $8m in ads

Hard to know really\n\nQT @Techmeme: OpenAI has surpassed...

6. teortaxesTex (Group Score: 101.2 | Individual: 35.3)

Cluster: 3 tweets | Engagement: 9 (Avg: 65) | Type: Tech

Singaporeans did TurboQuant first?\n\nQT @gaoj0017: We need to publicly clarify serious issues in Google’s ICLR 2026 paper TurboQuant.

TurboQuant misrepresents RaBitQ in three ways:

Avoids acknowledging key methodological similarity (JL transform)
Calls our theory “suboptimal” with no evidence
Reports results under unfair experimental settings

We have expressed our concerns to the authors before their submission, but they chose not to fix them in their paper submission. The paper was accepted at ICLR 2026 and heavily promoted by Google (tens of millions of views). At that scale, uncorrected claims quickly become “consensus.”

Facts:

RaBitQ already proves asymptotic optimality (FOCS’17 bound)
TurboQuant uses the same random rotation step but misses stating the connection
Their experiments used single-core CPU for RaBitQ vs A100 GPU for TurboQuant

None of these is properly disclosed.

We’ve filed a formal complaint and posted on OpenReview (https://t.co/nDVjmNhATM). We’ll release a detailed technical report on arXiv.

Our goal is simple: keep the academic record accurate.

Would appreciate people taking a look and sharing.

See 2 related tweets

@yoavgo: i knew the method in that hypey google paper sounded familiar!\n\nQT @gaoj0017: We need to publicly ...
@stanfordnlp: RT @gaoj0017: We need to publicly clarify serious issues in Google’s ICLR 2026 paper TurboQuant.

Tu...

7. rohanpaul_ai (Group Score: 97.7 | Individual: 49.8)

Cluster: 3 tweets | Engagement: 345 (Avg: 81) | Type: Tech

Rumor has it Anthropic is getting ready to launch new models, Mythos and Capybara, with Mythos being a brand-new tier that is larger than Opus.

Mythos is built as a step above Opus 4.6 rather than a routine upgrade.
The leak also says Mythos is extremely expensive to run, which usually means a larger system with heavier computation, longer reasoning, or both, so cost is part of the product story, not just an engineering detail.
Anthropic appears to be slowing release, limiting access, and starting with a small set of security-focused users instead of pushing a broad launch.

IMO, Anthropic may be trying to define a new premium tier where raw capability matters more than wide availability, with cyber defense as the safest first market.\n\nQT @M1Astra: Claude Mythos Blog Post

Saved before it was taken down.

https://t.co/6XIw1LKnkA

See 2 related tweets

@testingcatalog: RT @testingcatalog: BREAKING 🚨: Anthropic is preparing to release new models, Mythos and Capybara, w...
@startupideaspod: Anthropic is already testing their newest model. Mythos

“Compared to Claude Opus 4.6, Mythos achiev...

8. MLStreetTalk (Group Score: 96.6 | Individual: 36.1)

Cluster: 3 tweets | Engagement: 2276 (Avg: 654) | Type: Tech

RT @AIatMeta: Today we're introducing TRIBE v2 (Trimodal Brain Encoder), a foundation model trained to predict how the human brain responds to almost any sight or sound.

Building on our Algonauts 2025 award-winning architecture, TRIBE v2 draws on 500+ hours of fMRI recordings from 700+ people to create a digital twin of neural activity and enable zero-shot predictions for new subjects, languages, and tasks.

Try the demo and learn more here: https://t.co/VkMd1YpQWI

See 2 related tweets

@WesRoth: Meta’s Fundamental AI Research (FAIR) team released TRIBE v2 (Trimodal Brain Encoder), a groundbreak...
@Nick_Davidov: Everyone: let’s make AI cure cancer or enable anyone to build software Meta: let’s make AI that pre...

9. ivanfioravanti (Group Score: 94.8 | Individual: 27.1)

Cluster: 5 tweets | Engagement: 259 (Avg: 78) | Type: Tech

GLM-5.1! And it’s Open!\n\nQT @Zai_org: GLM-5.1 is available to ALL GLM Coding Plan users! https://t.co/Nk8Y98HNhU

See 4 related tweets

@testingcatalog: Z AI released GLM-5.1 and made it available to all GLM Coding Plan users.

It is also expected to b...

@kimmonismus: GLM-5.1 released! Chinese models are coming closer, the gap is shrinking\n\nQT @Zai_org: GLM-5.1 is ...
@gdgtify: GLM 5.1? Did I read that right?\n\nQT @Zai_org: GLM-5.1 is available to ALL GLM Coding Plan users! h...
@ZixuanLi_: RT @Zai_org: GLM-5.1 is available to ALL GLM Coding Plan users! https://t.co/Nk8Y98HNhU...

10. kimmonismus (Group Score: 85.0 | Individual: 35.3)

Cluster: 4 tweets | Engagement: 159 (Avg: 371) | Type: Tech

Seriously, this story illustrates why AI will change all of our lives and why I'm writing against all the doomers.

Three years ago, Paul Conyngham noticed lumps on his dog Rosie. Vets dismissed it for nearly a year. By the time cancer was finally diagnosed, it was terminal. Paul refused to accept it.

What followed was one of the most remarkable stories at the intersection of AI and medicine: a non-biologist, running an AI consulting business on the side, used ChatGPT, Gemini, and Grok to design a custom mRNA cancer vaccine for his dog from scratch.

The pipeline was real science. Full genome sequencing at the Garvan Institute. AlphaFold 2 to model Rosie's mutated c-KIT protein. A million-candidate ligand screen. When those paths hit dead ends, patent walls, regulatory timelines, Paul pivoted to vaccines. AI helped him architect a 7-epitope mRNA construct, write 100 pages of ethics approval documents, and design a multimodal treatment protocol combining the vaccine with a tyrosine kinase inhibitor and a PD-1 checkpoint inhibitor.

Six weeks into treatment: tumors shrinking. Two months in: near-normal. Rosie is improving. Paul is now convinced this can scale, and he's building what comes next.

if this doesnt excite you, I don't know what else will.

This is the story of the year.\n\nQT @paul_conyngham: https://t.co/bpa3HHt8Mg

See 3 related tweets

@minchoi: RT @minchoi: This story is wild…

Paul used AI (ChatGPT, Gemini, and Grok) to help create a personal...

@Cointelegraph: 🇦🇺 NEW: Australian Entrepreneur Paul Conyngham uses AI to design personalized mRNA cancer vaccine fo...
@aakashgupta: This guy literally explains how he cured his dog’s cancer with AI.\n\nQT @paul_conyngham: https://t....

11. bran_don_gell (Group Score: 82.1 | Individual: 33.4)

Cluster: 3 tweets | Engagement: 91 (Avg: 17) | Type: Tech

The new design system (claude code to html to claude code to figma to claude code):

Claude Code design exploration in html (create multiple versions and iterate until you get somewhere you like) → create full design system in html (all buttons, alerts, inputs, spacing, etc) → transfer to figma using use_figma and create all your variable and components → combine your atomic unit components to make bigger components like your nav → tell Claude Code the vision for how your app works/functions. Ask it to explore your code base and make a plan → edit the plan and send it back to claude to build in figma → tell it to critique itself after it builds and to do another pass after that → go to sleep and see how it looks in the morning 💤\n\nQT @figma: Now you can use AI agents to design directly on the Figma canvas, with our new use_figma MCP tool and skills to teach them. Open beta starts today. https://t.co/AQZsFWvvXQ

See 2 related tweets

@bran_don_gell: went to sleep having reviewed the HTML design system. Woke up with a fully designed app and all the ...
@danshipper: RT @bran_don_gell: The new design system (claude code to html to claude code to figma to claude code...

12. WesRoth (Group Score: 80.7 | Individual: 32.0)

Cluster: 3 tweets | Engagement: 44 (Avg: 31) | Type: Tech

Cline introduced Cline Kanban, a new standalone application that orchestrate multiple AI coding agents simultaneously.

Released as a free, open-source tool, it aims to solve the cognitive overload of managing several autonomous agents in separate terminal windows.\n\nQT @cline: Introducing Cline Kanban: A standalone app for CLI-agnostic multi-agent orchestration. Claude and Codex compatible.

npm i -g cline

Tasks run in worktrees, click to review diffs, & link cards together to create dependency chains that complete large amounts of work autonomously. https://t.co/4HjvwSu4Mo

See 2 related tweets

@addyosmani: RT @cline: Introducing Cline Kanban: A standalone app for CLI-agnostic multi-agent orchestration. Cl...
@svpino: RT @svpino: A single board to orchestrate all your coding agents.

The future of software developmen...

13. BrianRoemmele (Group Score: 79.4 | Individual: 49.4)

Cluster: 2 tweets | Engagement: 3565 (Avg: 369) | Type: Tech

This is absolutely astounding!

A printer that prints circuit board traces.

I am going all in on this.

We will have a 100x increase in production and testing! https://t.co/f82BRCjAsc

See 1 related tweets

@BrianRoemmele: About $3,500 and you have a prototype circuit board, conductive trace printer!

No loss of IP while...

14. zerohedge (Group Score: 78.2 | Individual: 28.1)

Cluster: 4 tweets | Engagement: 4765 (Avg: 1105) | Type: Tech

A federal judge has halted the Pentagon’s punishment of Anthropic, temporarily preventing the government from treating the AI start-up as a threat to national security.

Activist judges now dictating US national security

See 3 related tweets

@FT: The Pentagon has been blocked by a US court from punishing Anthropic over its refusal to allow unres...
@Reuters: A US judge temporarily blocked the Pentagon's blacklisting of Anthropic, the latest turn in the Clau...
@ReutersBiz: WATCH: A US judge temporarily blocked the Pentagon's blacklisting of Anthropic, the latest turn in t...

15. jason (Group Score: 74.5 | Individual: 52.6)

Cluster: 2 tweets | Engagement: 5459 (Avg: 519) | Type: Tech

Founders: take my advice... do not talk to the press, go direct and do long-form podcasts.

Wired and the NYT are as biased as Fox News and MSNOW these days

This is a function of their need to pander to one side to survive, be it through $ 3-a-month subs or rage-baiting ad-based stories.

Attacking tech gets views (see Karen Swisher)... and views get advertisers (paradoxically, tech advertisers support the folks trashing tech! let that sink in!)

Founders: If you talk to the NYT or WIRED, they will trash and misrepresent you 95% of the time in order to get more subscribers and page views

It is what it is...\n\nQT @tab_delete: .@Jason says this but Wired added 200k new paid subscriptions last year and the NYT is healthier than ever.

“Don’t engage with the media” is an inherently silly take. Journalism is an essential public interest.

Founders evading scrutiny is how fraud pops up again and again…

See 1 related tweets

@elonmusk: Absolutely\n\nQT @Jason: Founders: take my advice... do not talk to the press, go direct and do long...

16. edzitron (Group Score: 69.9 | Individual: 30.3)

Cluster: 3 tweets | Engagement: 2706 (Avg: 567) | Type: Tech

It begins…\n\nQT @trq212: To manage growing demand for Claude we're adjusting our 5 hour session limits for free/Pro/Max subs during peak hours. Your weekly limits remain unchanged.

During weekdays between 5am–11am PT / 1pm–7pm GMT, you'll move through your 5-hour session limits faster than before.

See 2 related tweets

@Grady_Booch: The enshittification of @claudeai begins.\n\nQT @trq212: To manage growing demand for Claude we're a...
@thdxr: this is probably going to have a positive impact on software\n\nQT @trq212: To manage growing demand...

17. aakashgupta (Group Score: 66.8 | Individual: 37.4)

Cluster: 2 tweets | Engagement: 357 (Avg: 638) | Type: Tech

Meta has lost $73 billion on Reality Labs since 2020. Wall Street calls it the most expensive money pit in tech history.

Then today, quietly, the FAIR team in Paris releases a model that predicts how your brain responds to anything you see, hear, or read. 70x higher resolution than v1. Zero-shot predictions for people it has never scanned.

The training data: 700+ volunteers watched movies and listened to podcasts inside fMRI machines for 1,115 total hours. The model learned how visual cortex, auditory cortex, and language centers fire simultaneously, then built a single architecture that maps all of it.

The competition results tell you how far ahead they are. TRIBE v1 already won first place in Algonauts 2025, beating 262 other teams. V2 is a 2-3x improvement on top of that, with 70x the spatial resolution.

Here's what nobody is connecting. Meta also builds Ray-Ban smart glasses with cameras and microphones. They're developing a neural interface wristband that reads EMG signals from your arm. They run the largest advertising platform on earth, one that generated $200 billion in revenue last year by predicting which content keeps you engaged.

TRIBE v2 tells them exactly which brain regions activate when you watch a 15-second Reel. Which neurons fire when an ad plays in your peripheral vision. How language processing changes when you're listening versus reading.

They open-sourced the model. That's the part that should make you pay closer attention. Meta open-sources things when the research advantage is already captured and the ecosystem benefit of external researchers improving the model exceeds the competitive risk. They did it with LLaMA. They're doing it again.

A company spending $135 billion in capex this year did not build a digital twin of the human brain for academic citations. They built the prediction layer for every piece of hardware and every ad impression they'll sell for the next decade.

The $73 billion was never about the metaverse. It was about understanding the 20-watt computer that decides what every human pays attention to.\n\nQT @AIatMeta: Today we're introducing TRIBE v2 (Trimodal Brain Encoder), a foundation model trained to predict how the human brain responds to almost any sight or sound.

Try the demo and learn more here: https://t.co/VkMd1YpQWI

See 1 related tweets

@ianmiles: Meta's new AI model can predict brain activity across 70,000 regions and handle vision, sound, and l...

18. omarsar0 (Group Score: 66.7 | Individual: 35.2)

Cluster: 2 tweets | Engagement: 162 (Avg: 114) | Type: Tech

NEW research from IBM: Workflow Optimization for LLM Agents.

LLM agent workflows involve interleaving model calls, retrieval, tool use, code execution, memory updates, and verification.

How you wire these together matters more than most teams realize.

This new survey maps the full landscape.

It categorizes approaches along three dimensions: when structure is determined (static templates vs. dynamic runtime graphs), which components get optimized, and what signals guide the optimization (task metrics, verifier feedback, preferences, or trace-derived insights).

It proposes structure-aware evaluation incorporating graph properties, execution cost, robustness, and structural variation.

Most teams either hardcode their agent workflows or let them be fully dynamic with no principled middle ground. This survey provides a unified vocabulary and framework for deciding where your system should sit on the static-to-dynamic spectrum.

Paper: https://t.co/qF8kTaNPYo

Learn to build effective AI agents in our academy: https://t.co/1e8RZKs4uX

See 1 related tweets

@dair_ai: Great survey on Workflow Optimization for LLM Agents.\n\nQT @omarsar0: NEW research from IBM: Workfl...

19. huggingface (Group Score: 66.6 | Individual: 35.1)

Cluster: 2 tweets | Engagement: 220 (Avg: 103) | Type: Tech

Model weights are here: https://t.co/rQlfP51Db7!\n\nQT @trychroma: Introducing Chroma Context-1, a 20B parameter search agent.

> pushes the pareto frontier of agentic search

> order of magnitude faster

> order of magnitude cheaper

> Apache 2.0, open-source https://t.co/bhAkULyBBn

See 1 related tweets

@WesRoth: Chroma announced Context-1, a highly specialized 20-billion parameter model designed specifically to...

20. badlogicgames (Group Score: 65.0 | Individual: 38.6)

Cluster: 2 tweets | Engagement: 338 (Avg: 95) | Type: Tech

chat, how do we feel about this for my AI Engineer London keynote?

(still can't believe anyone would make me a keynoter). https://t.co/oNtnhVpqzz

See 1 related tweets

@aiDotEngineer: chat get in the comments\n\nQT @badlogicgames: chat, how do we feel about this for my AI Engineer Lo...