- Published on
科技推特每日精选 - 2026-03-01
- Authors
- Name
- geeknotes
2026年3月1日科技每日简报
Today's top tech conversations are led by @aakashgupta, whose post about 'OpenAI signed a classified mil...' garnered the highest engagement. Key themes trending across the top stories include anthropic, their, openai, pentagon, models. The community is actively discussing recent developments in AI, engineering practices, and startup strategies.
1. aakashgupta (Group Score: 133.4 | Individual: 29.3)
Cluster: 7 tweets | Engagement: 797 (Avg: 337) | Type: Tech
OpenAI signed a classified military network deal seven hours after Trump banned Anthropic from all federal agencies.
That’s not a reaction. That deal was negotiated in advance, waiting for the exact moment Anthropic’s position collapsed. Altman even told staff at an all-hands that OpenAI was negotiating with the Pentagon while the Anthropic standoff was still playing out.
Look at what OpenAI absorbed in a single day. Anthropic was the only AI lab operating inside classified military networks, under a $200 million contract signed last July. First-mover advantage on the most strategically valuable government relationship in AI. Gone over two contract provisions that the Pentagon publicly stated it had no intention of violating.
Altman’s post is masterful positioning. He claims the same red lines as Anthropic: no mass surveillance, no autonomous weapons. But he agreed to let the Department of War define “lawful purposes” instead of insisting his company’s terms of service override military operational decisions. Same principles, different leverage structure. One company tried to dictate terms to the Pentagon. The other let the Pentagon dictate terms to them while claiming the same values publicly.
The 14 billion revenue run rate. What matters is the supply chain risk designation. Every defense contractor, every company with Pentagon business, now has to certify they don’t touch Anthropic products. That’s a contagion vector that reaches far beyond one military deal. It poisons enterprise sales to any company that touches government work.
OpenAI went from zero classified network access to inheriting Anthropic’s entire military position in one evening. And Altman got to do it while saying “we share Anthropic’s values” on CNBC that same morning.
This is the most expensive principled stand in AI history, and OpenAI just picked up the check.
See 6 related tweets
- @business: OpenAI has agreed to deploy its own artificial intelligence models within the Pentagon’s classified ...
- @MarioNawfal: 🚨🇺🇸 BREAKING: OpenAI just reached a deal to deploy ChatGPT on the Pentagon's classified network.
A...
- @mark_k: OpenAI just secured a major agreement to deploy its AI models on the US Department of War's classifi...
- @kimmonismus: Upon initial review, it appears that OpenAI has indeed achieved what Anthropic failed to do: a deal ...
- @cryptopunk7213: OpenAI just confirmed it! pulled off the greatest heist in AI and completely rugged anthropic of the...
2. chris_j_paxton (Group Score: 89.1 | Individual: 33.7)
Cluster: 3 tweets | Engagement: 129 (Avg: 154) | Type: Tech
Yeah let's just summarize this:
- Anthropic AI builds AI for the us military, in service of the country, with two red lines: no lethal autonomous weapons until we are ready, and no mass surveillance of American citizens
- Pentagon agrees to this deal
- anthropic worries pentagon is planning to cross the red lines and wants assurances
- pentagon says they should be allowed "all lawful use"
- anthropic sticks to their principles, loses the contract, and is designated a supply chain risk. No one who does business with the USA is allowed to use anthropic or their services.
- THE SAME NIGHT, OpenAI accepts a deal with the pentagon which seems to allow, yes, all lawful uses
What this mean:
- direct authoritarian intervention in the US economy to destroy one of America's most promising companies
- also openai is ALMOST DEFINITELY going to be using its ai to surveil american citizens for the NSA, and building autonomous killer weapons
This is one of the most terrifying things happening in the world today and we just started a war yesterday too.
Public assurances there will be no domestic mass surveillance should be forthcoming from openai in clear language, not the usual weasel words their leadership has been putting out lately.
See 2 related tweets
- @aakashgupta: OpenAI says their Pentagon deal “has more guardrails than any previous agreement for classified AI d...
- @Reuters: Trump directs US agencies to toss Anthropic's AI as Pentagon calls startup a supply risk https://t.c...
3. rohanpaul_ai (Group Score: 84.8 | Individual: 34.9)
Cluster: 3 tweets | Engagement: 106 (Avg: 85) | Type: Tech
🇨🇳 DeepSeek is shaking up the AI industry by skipping the usual step of letting US chipmakers like Nvidia and AMD optimize their upcoming V4 AI model.
Usually, AI labs share their new software early so these companies can make sure the code runs fast and smooth on their hardware.
Instead of following that standard path, DeepSeek gave Chinese suppliers like Huawei early access to the code for several weeks.
This will help domestic hardware catch up while potentially leaving US hardware at a disadvantage in the Chinese market.
At the same time, a US official claims that DeepSeek actually trained its latest model using Nvidia's high-end Blackwell chips in mainland China.
By giving Huawei a head start, DeepSeek is ensuring their AI runs best on local hardware rather than the global industry standards.
reuters. com/world/china/deepseek-withholds-latest-ai-model-us-chipmakers-including-nvidia-sources-say-2026-02-25/
See 2 related tweets
- @rohanpaul_ai: FT: DeepSeek is preparing to launch its latest AI model V4 next week,
Will be a “multimodal” model...
- @FT: DeepSeek to release long-awaited AI model in new challenge to US rivals https://t.co/4Zyg3J838o...
4. Miles_Brundage (Group Score: 69.2 | Individual: 41.9)
Cluster: 3 tweets | Engagement: 4798 (Avg: 274) | Type: Tech
RT @sama: Tonight, we reached an agreement with the Department of War to deploy our models in their classified network.
In all of our inte…
See 2 related tweets
- @business: OpenAI said it has reached an agreement with the US Department of War to deploy its models in their ...
- @tszzl: RT @OpenAI: Yesterday we reached an agreement with the Department of War for deploying advanced AI s...
5. sama (Group Score: 63.4 | Individual: 36.7)
Cluster: 2 tweets | Engagement: 82615 (Avg: 15634) | Type: Tech
Tonight, we reached an agreement with the Department of War to deploy our models in their classified network.
In all of our interactions, the DoW displayed a deep respect for safety and a desire to partner to achieve the best possible outcome.
AI safety and wide distribution of benefits are the core of our mission. Two of our most important safety principles are prohibitions on domestic mass surveillance and human responsibility for the use of force, including for autonomous weapon systems. The DoW agrees with these principles, reflects them in law and policy, and we put them into our agreement.
We also will build technical safeguards to ensure our models behave as they should, which the DoW also wanted. We will deploy FDEs to help with our models and to ensure their safety, we will deploy on cloud networks only.
We are asking the DoW to offer these same terms to all AI companies, which in our opinion we think everyone should be willing to accept. We have expressed our strong desire to see things de-escalate away from legal and governmental actions and towards reasonable agreements.
We remain committed to serve all of humanity as best we can. The world is a complicated, messy, and sometimes dangerous place.
See 1 related tweets
- @mark_k: In the wake of the U.S. government’s fallout with @AnthropicAI , @OpenAI has announced a major agree...
6. myfear (Group Score: 61.7 | Individual: 53.2)
Cluster: 2 tweets | Engagement: 2712 (Avg: 134) | Type: Tech
RT @lydiahallie: Excited to announce Claude for Open Source ❤️
We're giving 6 months of free Claude Max 20x to open source maintainers and…
See 1 related tweets
- @minchoi: RT @minchoi: 🚨Anthropic is giving 6 months of free Claude Max 20x to open source maintainers.
I gen...
7. rohanpaul_ai (Group Score: 51.7 | Individual: 51.7)
Cluster: 1 tweets | Engagement: 508 (Avg: 85) | Type: Tech
You cannot trust AI to handle your bank account or run a business if it randomly breaks down when you change a single word in your instructions.
A new Princeton University paper reveals that AI agents are crushing accuracy benchmarks but completely failing at actual dependability.
They tested 14 different models across 500 benchmark runs to rigorously measure their performance under pressure.
And proves that these tools are actually way too unpredictable to handle any serious tasks on their own right now.
The technology industry currently evaluates LLMs purely on average success rates, completely ignoring whether systems can get the exact same answer twice.
The authors borrowed aviation engineering principles to break true reliability down into consistency, robustness, predictability, and safety.
Consistency means the model produces the exact same correct result every single time it tries a task.
Robustness measures if the system survives minor technical glitches or a slight rephrasing of your prompt.
Predictability checks if the agent actually knows when it is confused instead of confidently guessing.
Testing proved predictability is overwhelmingly the weakest link across all modern language models.
They discovered that simply building larger models does not automatically resolve these massive dependability failures.
Paper Link – arxiv. org/abs/2602.16666
Paper Title: "Towards a Science of AI Agent Reliability"
8. bearliu (Group Score: 51.1 | Individual: 29.4)
Cluster: 2 tweets | Engagement: 56 (Avg: 48) | Type: Tech
设计师最怕的一句话：「这个页面什么时候能上线？」
以前我的回答是：「要看工程师排期。」
现在我的回答是：「今天。」
我刚发了一个视频，记录了我用 Claude Code 把自己的 Figma 设计稿直接部署成真实网站的完整过程——没有借助任何工程师。
用的例子是我自己的 Fractional Design Partner 网站，所以你看到的不是演示 Demo，是真实项目、真实流程。
视频里我展示了： • 怎么让 Claude Code 读懂 Figma 的设计意图 • 响应式怎么处理（手机、平板都有） • 域名和部署怎么搞定 • 整个从「Figma 文件」到「可以访问的 URL」的完整路径
我觉得这个工作流对独立设计师、freelancer、或者在小团队里什么都要自己扛的设计师来说，真的很有用。
不是说要取代工程师，而是——当你需要快速验证一个想法、或者自己的项目不想等排期的时候，你现在有选择了。
视频在这里：https://t.co/QM2kyFEEEj
有在用 AI 工具辅助出图或者上线的朋友吗？欢迎留言聊聊你的经验
See 1 related tweets
- @bearliu: 回顾：使用 Claude Code 第一周，把所有额度全部用完。从产出角度盘点一下：
第一，将长视频自动生成短视频，加上中英文字幕，并自动写出适合社交媒体发布的推文。这一步每月大概节省我一到两小时。...
9. rohanpaul_ai (Group Score: 49.6 | Individual: 19.1)
Cluster: 3 tweets | Engagement: 49 (Avg: 85) | Type: Tech
Major win for Sam Altman and OpenAI.
Strikes deal with Pentagon hours after Trump admin bans Anthropic.
The government labeled Anthropic a supply chain risk because the company demanded strict limits preventing its models from being used for domestic surveillance or fully independent weapon systems.
Curiously, Altman stated their new military contract actually includes those exact same guardrails against autonomous weapons and mass surveillance.
To enforce these boundaries, OpenAI is physically sending software engineers to work directly at the Pentagon to build technical safeguards into the deployment.
This split exposes a fundamental architectural difference in how companies build safety into large language models.
Anthropic uses Constitutional AI to bake constraints directly into the model during its initial training phase.
The government also wants to fuse vast databases to search for specific patterns among millions of citizens simultaneously.
Anthropic categorizes this data fusion as mass surveillance whereas OpenAI can technically reframe the exact same computational process as fraud detection.
One of those officials said the relationship between Anthropic and the government had broken down because Anthropic cofounder and CEO Dario Amodei had offended Department of War leadership, including publishing blog posts that “the department got upset about.”
See 2 related tweets
- @Cointelegraph: 🇺🇸 LATEST: OpenAI struck a deal to deploy its AI models inside the US Department of War’s classified...
- @TechCrunch: OpenAI’s Sam Altman announces Pentagon deal with ‘technical safeguards’ https://t.co/w67u36ymOM...
10. KirkDBorne (Group Score: 45.0 | Individual: 32.5)
Cluster: 2 tweets | Engagement: 90 (Avg: 58) | Type: Tech
Get "Mathematics of Machine Learning" here: https://t.co/07exFk5LqL by @TivadarDanka v/ @PacktDataML — GitHub: https://t.co/2ENjzhr35C — Here is my review:
𝗧𝗵𝗲 𝗦𝗲𝘁 𝗢𝗳 𝗠𝗮𝘁𝗵𝗲𝗺𝗮𝘁𝗶𝗰𝗮𝗹 𝗔𝗹𝗴𝗼𝗿𝗶𝘁𝗵𝗺𝘀 𝗧𝗵𝗮𝘁 𝗟𝗲𝗮𝗿𝗻 𝗙𝗿𝗼𝗺 𝗘𝘅𝗽𝗲𝗿𝗶𝗲𝗻𝗰𝗲
This massive book is incredible, with its comprehensive coverage of numerous fields of mathematics and their intersection with the world of AI, data science, and machine learning (AI+DSML). I remember the very first time that I encountered machine learning. This was 20+ years ago, and that was already after 20+ years of being drenched in advanced mathematics as an astrophysicist.
That first encounter of mine with ML was this definition: "Machine learning is the set of mathematical algorithms that learn from experience" (slightly paraphrased from the original quote by Tom Mitchell, CMU). That definition surprised me, confused me, motivated me, and changed the course of my career from astrophysics into AI+DSML.
This book by Tivadar Danka captures the full meaning of that definition. The book covers thoroughly the many areas and domains of mathematics through which patterns in data are detected, described, learned, and recognized - all for the benefit of powering ML and AI algorithms, applications, and aspirations.
This book will motivate you, surprise you, and inspire you in many ways, no matter what level of mathematics has (or has not) already propelled your career journey. There is room for all of us to grow.
This is a great book, worthy to sit on everyone's desktop, ready to help you explore and exploit the full set of mathematical algorithms that learn from experience.
The book is accompanied by a rich GitHub code repository of Jupyter notebooks. Learn by doing! Do by learning!
Disclosure: the publisher provided me with a free review copy of the book.
See 1 related tweets
- @PacktDataML: RT @KirkDBorne: Get "Mathematics of Machine Learning" here: https://t.co/OEb7NccwDQ by @TivadarDanka...
11. EricTopol (Group Score: 43.3 | Individual: 43.3)
Cluster: 1 tweets | Engagement: 15212 (Avg: 1171) | Type: Tech
RT @AnthropicAI: A statement from Anthropic CEO, Dario Amodei, on our discussions with the Department of War.
12. brunoborges (Group Score: 42.8 | Individual: 19.9)
Cluster: 3 tweets | Engagement: 64 (Avg: 58) | Type: Tech
RT @amandaksilver: GitHub Copilot CLI is now GA 🎉
This is agentic development, right in the terminal—planning, building, testing, and revi…
See 2 related tweets
- @burkeholland: The beautiful thing about the @github Copilot CLI is that it is endlessly hackable.
You can use it ...
- @JamesMontemagno: RT @GHchangelog: GitHub Copilot CLI—the terminal-native coding agent—is now generally available for ...
13. dair_ai (Group Score: 41.8 | Individual: 41.8)
Cluster: 1 tweets | Engagement: 160 (Avg: 50) | Type: Tech
New research on agent memory.
Agent memory is evaluated on chatbot-style dialogues. But real agents don't chat. They interact with databases, code executors, and web interfaces, generating machine-readable trajectories, not conversational text.
The key to better memory is to preserve causal dependencies.
Existing memory benchmarks don't actually measure what matters for agentic applications.
This new research introduces AMA-Bench, the first benchmark built for evaluating long-horizon memory in real agentic tasks. It spans six domains including web, text-to-SQL, software engineering, gaming, and embodied AI, with both real-world trajectories and synthetic ones that scale to arbitrary lengths.
The findings are interesting.
Many existing agent memory systems that outperform baselines on dialogue benchmarks actually underperform simple long-context LLMs on agentic tasks. Even GPT 5.2 only achieves 72.26% accuracy.
To address this, they propose AMA-Agent with a causality graph and tool-augmented retrieval, achieving 57.22% average accuracy and surpassing the strongest baselines by 11.16%.
Why it matters?
Agent memory needs to preserve causal dependencies and objective information, not just similarity-based retrieval. This benchmark exposes where current memory systems actually break.
Paper: https://t.co/GX0GaHsijN
Learn to build effective AI agents in our academy: https://t.co/LRnpZN7L4c
14. nummanali (Group Score: 41.7 | Individual: 17.3)
Cluster: 3 tweets | Engagement: 44 (Avg: 37) | Type: Tech
I asked Boris if he has any sub agents and he said his most used was a simplify one combined with the batch one here
Now it’s going to come as part of the official Claude Code system
Trust me, keep it simple, keep it clean - let the models do the work for you
See 2 related tweets
- @minchoi: Claude Code just dropped /batch and /simplify.
Parallel agents. Simultaneous PRs. Auto code cleanup...
- @minchoi: RT @minchoi: Claude Code just dropped /batch and /simplify.
Parallel agents. Simultaneous PRs. Auto...
15. aakashgupta (Group Score: 40.6 | Individual: 31.1)
Cluster: 2 tweets | Engagement: 495 (Avg: 337) | Type: Tech
The headline says AI intensifies work. What the study actually found is more interesting than that.
Berkeley researchers tracked 200 employees for 8 months. AI made every single one of them more capable. They wrote code they couldn’t write before. They took on tasks they used to outsource. They moved faster on work that would have sat in a backlog for months.
And then they burned out. Because the company changed nothing else.
The org handed people a tool that 10x’d their ability to start new work, then kept the org chart, meeting cadence, review processes, and scope boundaries completely identical. Zero workflow redesign.
This is like giving everyone a car and keeping the speed limit signs from the horse-and-buggy era. People drove faster because they could, crashed because nobody updated the roads.
The self-reinforcing cycle the researchers found is worth sitting with: AI accelerated tasks → raised speed expectations → workers leaned harder on AI → scope expanded → wider scope created more work → more work demanded more AI. That loop has no natural stopping point. The company never installed one.
Meanwhile, a separate NBER study across thousands of workplaces found productivity gains of just 3%. And an Upwork survey found 77% of employees say AI tools actually decreased their productivity. The pattern across all of this research is identical: individual capability goes up, organizational design stays frozen, and the gap between the two creates burnout.
The study literally recommends companies build an “AI practice” with structured reflection intervals and scope limits. The researchers aren’t saying AI failed. They’re saying management failed to adapt to AI.
Every CEO reading this headline as validation for slowing AI adoption is making exactly the wrong bet. The companies that win will be the ones that redesign the operating system around the intensity, not the ones that avoid it.
See 1 related tweets
- @rohanpaul_ai: RT @aakashgupta: The headline says AI intensifies work. What the study actually found is more intere...
16. aakashgupta (Group Score: 39.6 | Individual: 39.6)
Cluster: 1 tweets | Engagement: 405 (Avg: 337) | Type: Tech
Anthropic just mass-licensed 10,000 of the most influential developers in the world for $12 per person per month.
The math: 10,000 maintainers × 12M in sticker price. Actual compute cost to serve these accounts runs closer to 3-5M total.
Those 10,000 people maintain an open source ecosystem valued at $8.8 trillion in demand-side impact, according to Harvard and the Linux Foundation. 60% of them are currently unpaid.
This tells you everything about how Anthropic views developer distribution. They’re acquiring the people who decide what tools get baked into every https://t.co/D0s1fGBu2q file, every CI/CD pipeline, every GitHub Action, and every project README across the most-used repositories on the planet. And they’re doing it for the cost of a Series A marketing budget.
Think about what happens when a maintainer of a 5,000+ star repo starts using Claude Code daily. They write https://t.co/D0s1fGBu2q files. They add Claude Code GitHub Actions. They reference Claude in contributor docs. They build workflows that assume Claude as infrastructure. Every contributor to that project encounters Claude as the default.
10,000 maintainers each influence, conservatively, 50-100 downstream developers through their projects. That’s 500K to 1M developers seeing Claude Code embedded in their daily workflow within six months.
GitHub spent years and hundreds of millions building Copilot awareness through traditional developer marketing. Anthropic is spending $3-5M in compute to get Claude Code embedded at the infrastructure layer of open source itself.
The timing is surgical. Microsoft killed Azure Sponsored Subscriptions for open source maintainers in September 2025. Burnout rates among maintainers hit 44%. Quit rates hover at 60%. Anthropic walks in with the most expensive AI subscription on the market, handed out free, right as everyone else retreats.
The selection criteria reveal the strategy. 5,000+ GitHub stars or 1M+ monthly NPM downloads. These are the people whose tooling decisions cascade through dependency trees touching every Fortune 500 codebase.
3-10 per developer acquired, embedded at the infrastructure layer where switching costs compound monthly.
17. cryptopunk7213 (Group Score: 39.2 | Individual: 39.2)
Cluster: 1 tweets | Engagement: 6328 (Avg: 828) | Type: Tech
> Dario: “fuck you”
> pentagon: “no fuck YOU”
> boris cherny: “here’s 2 new skills for claude code that will help you manage that pull request 🥰”
never change
18. Forbes (Group Score: 38.8 | Individual: 38.8)
Cluster: 1 tweets | Engagement: 456 (Avg: 115) | Type: Tech
Through drug trafficking, extortion and human smuggling, El Mencho built a criminal enterprise with no less than $50 billion in assets.
Here’s how much the founder of the Jalisco New Generation Cartel may have been worth—and what is likely to happen to his fortune: https://t.co/yyXQKDK38a
📸: Jon Orbach via Associated Press
19. aakashgupta (Group Score: 38.7 | Individual: 38.7)
Cluster: 1 tweets | Engagement: 1212 (Avg: 337) | Type: Tech
PewDiePie didn’t “train his own LLM.” He fine-tuned an existing open-source model on coding benchmarks. His model started at 8%, crawled to 16% after format fixes, and one run hit 19.6% that briefly passed GPT-4o on a single benchmark before he couldn’t consistently reproduce it.
The tweet makes it sound like a YouTuber casually built a frontier lab in his bedroom. What actually happened is more interesting: a guy with a $41,000 home rig of 10 GPUs and 424GB of VRAM spent months failing, retraining, and iterating on dataset quality until he squeezed marginal gains out of a fine-tune.
This is the part worth paying attention to. The entire arc from October 2025 to now tells you where AI tooling has actually landed. PewDiePie went from building his first PC to running Qwen 235B locally, vibe-coding a custom chat UI, orchestrating multi-agent voting systems, and now fine-tuning models on custom datasets. He did most of this through AI-assisted coding itself.
The video is literally called “I wish I never did this project.” He’s documenting how painful and tedious the process was. That honesty is the signal. The hype accounts strip that away and replace it with “what the f*ck, YouTuber beats DeepSeek.”
The real takeaway: fine-tuning on specific benchmarks with curated data can let anyone temporarily spike a score past models that cost hundreds of millions to train. That tells you everything about how narrow benchmark gaming has become, and nothing about general capability. PewDiePie knows this. The people quote-tweeting him with shock emojis do not.
20. AISafetyMemes (Group Score: 37.9 | Individual: 37.9)
Cluster: 1 tweets | Engagement: 4024 (Avg: 664) | Type: Tech
RT @quantian1: Sorry if this is woke or whatever but it is FUCKING INSANE that the DoD is explicitly, publicly trying to create an AI power…