Published on

科技推文精选 - 2026-02-17

Authors

今日科技要闻：阿里巴巴发布大规模 Qwen3.5 MoE 模型及 Ling-2.5-1T，标志着其在开源人工智能领域的重大发力。OpenAI 通过收购 OpenClaw 予以回击，旨在加速个人智能体的开发；Cloudflare 则推出了用于持久性自主系统的 SDK。安全方面，谷歌披露了针对 Gemini 的模型提取攻击，引发业界忧虑。Anthropic 与美国国防部就安全问题产生纠纷。与此同时，凭借国家级峰会的助力及黑石集团对初创公司 Neysa 的 6 亿美元投资，印度 AI 产业正加速崛起。

1. FuSheng_0306 (Group Score: 101.2 | Individual: 43.4)

Cluster: 3 tweets | Engagement: 6574 (Avg: 507) | Type: Tech

RT @sama: Peter Steinberger is joining OpenAI to drive the next generation of personal agents. He is a genius with a lot of amazing ideas a…

See 2 related tweets

  • @gdgtify: RT @steipete: I'm joining @OpenAI to bring agents to everyone. @OpenClaw is becoming a foundation: o...
  • @FT: OpenAI hires OpenClaw founder Peter Steinberger https://t.co/kJ8HjKh7Lh...

2. cryptopunk7213 (Group Score: 76.7 | Individual: 46.9)

Cluster: 3 tweets | Engagement: 3325 (Avg: 740) | Type: Tech

lmao so just to recap the week:

  • openai acquired peter steinberger and openclaw committing to keep the project 100% open source and make it a core part of openai’s ecosystem. zuck is fckin fuming.

  • china’s bytedance got sued by disney and hollywood for their seedance video model being so damn good BUT turns out it doesn’t even use their IP, it’s just that good lol

  • a clawdbot (openclaw) gave birth to a replicate of itself and autonomously paid for its api access

  • some random dude discovered $1.5 trillion worth of lithium buried under a super volcano in nevada miraculously solving the USA’s battery crisis (tesla praising the lord rn)

  • anthropic confirmed spotify engineers no longer code they just prompt claude to write 100% of the platforms code. we’re talking about a $100B+ company here

  • openai launched codex spark, a faster version of codex 5.3 that slings 1000 tokens / sec insane

  • anthropic and openai both lost key AI safety staff because “the world is in peril” and they reject AI ‘adult mode’

  • xAI laid off 20-40% of staff in a vital bid to align grok, X and spaceX teams before building 100Tw of compute in space

  • matt schumer’s “something big is coming” article hit 100M+ views warning against AI except… it was co-written by an AI

  • runway raised 300Mata300M at a 5.3B val

  • europe committed $13B+ to fund next-gen AI startups to compete with the US

  • norway decided to fck their entire economy with a 36% tax on UNrealised gains

goodnight.

See 2 related tweets

  • @BrianRoemmele: I must be frank. One reason we abandoned OpenClaw was I suspected the project would be privatized.

...

  • @cryptopunk7213: we're massive fucking fans of @steipete

here's everything you need to know about

  • openai acquirin...

3. mark_k (Group Score: 63.6 | Individual: 56.1)

Cluster: 2 tweets | Engagement: 2611 (Avg: 318) | Type: Tech

Google has revealed that "commercially motivated" actors attempted to clone @GeminiApp by bombarding it with over 100,000 prompts. This "model extraction" attack aimed to steal the AI’s proprietary logic and reasoning capabilities, particularly in non-English languages, to train a cheaper, unauthorized copycat model.

The attackers systematically mapped Gemini’s response patterns to create a synthetic dataset for fine-tuning smaller, open-source models. Google’s Threat Intelligence Group detected the coordinated activity and blocked it, labeling the incident a direct attempt at intellectual property theft.

Beyond commercial cloning, Google’s report noted a rise in state-backed threats. Groups from Russia, China, Iran, and North Korea are increasingly using AI to refine phishing campaigns, perform reconnaissance, and assist in writing code for malware.

Source: Ars Technica

See 1 related tweets

  • @BlindGoose1337: RT @mark_k: Google has revealed that "commercially motivated" actors attempted to clone @GeminiApp b...

4. CloudflareDev (Group Score: 61.2 | Individual: 31.8)

Cluster: 2 tweets | Engagement: 90 (Avg: 98) | Type: Tech

Building an AI agent is easy; making it stay "smart" across sessions is the hard part.

The Cloudflare Agents SDK gives you a framework to build autonomous agents that persist state, run background jobs, and scale at the edge.

Come join @lauragift_ on Feb 19, 4 PM SGT, for an online hands-on workshop on building scalable, stateful agents. Register here 👇 https://t.co/J7vUCRhae6

See 1 related tweets

  • @Cloudflare: Build, persist, and scale autonomous agents with the Cloudflare Agents SDK!

The Agents SDK enables...

5. vllm_project (Group Score: 57.9 | Individual: 22.7)

Cluster: 3 tweets | Engagement: 272 (Avg: 147) | Type: Tech

🎉 Congrats to @Alibaba_Qwen on releasing Qwen3.5 on Chinese New Year's Eve — day-0 support is ready in vLLM!

Qwen3.5 is a multimodal MoE with Gated Delta Networks architecture — 397B total params, only 17B active.

What makes it interesting for inference:

🧠 Gated Delta Networks + sparse MoE — high throughput, low latency, lower cost 🌍 201 languages and dialects supported out of the box 👁️ One model for both text and vision — no separate VL pipeline needed

Verified on NVIDIA GPUs. Recipes for Docker, pip, and multi-node deployment 👇 https://t.co/VNbDLoDdFe

#vLLM #Qwen #OpenSource #Inference

See 2 related tweets

  • @ModelScope2022: Qwen3.5-397B-A17B is here 🚀 Open-source flagship: 397B total, 17B active. Apache 2.0.

⚡ 19x faster ...

  • @vllm_project: 🎉 Congrats to @Alibaba_Qwen on releasing Qwen3.5 on Chinese New Year's Eve — day-0 support is ready...

6. aakashgupta (Group Score: 50.6 | Individual: 33.9)

Cluster: 2 tweets | Engagement: 415 (Avg: 493) | Type: Tech

Tobi built the original Shopify codebase himself in 2004. Then he spent 20 years doing what every founder-CEO does: slowly drifting into meetings, board decks, and strategy docs while the thing that made him dangerous collected dust.

94 commits in 2024. That’s a CEO who codes occasionally on weekends. 833 in 2025. That’s someone who found a way back in. 957 in 45 days of 2026. That’s a builder again.

The pattern repeating across tech right now is something nobody predicted about AI. Everyone focused on which jobs AI would replace. The bigger story is which skills AI is resurrecting. Founders who stopped coding a decade ago are shipping features. Executives who used to need a team of three to prototype an idea are doing it over lunch. The distance between “I have an idea” and “I built the thing” collapsed from weeks to hours.

Tobi’s been porting his own tools across languages using Claude as a compiler, publishing CLI utilities, and writing open source on his personal GitHub. The CEO of a $150B company is personally shipping to package managers on weekends.

That’s the part worth paying attention to. AI gave Tobi back the part of the job he loved and had to abandon to scale. The best founders were always builders first. The ones who drifted away didn’t want to. They just ran out of hours.

Now they have the hours back.

See 1 related tweets

  • @garrytan: RT @aakashgupta: Tobi built the original Shopify codebase himself in 2004. Then he spent 20 years do...

7. MarioNawfal (Group Score: 50.2 | Individual: 24.7)

Cluster: 3 tweets | Engagement: 360 (Avg: 1358) | Type: Tech

🇺🇸 The Pentagon is reportedly threatening to cut off Anthropic over a dispute about safety safeguards tied to its Claude AI model.

It’s ultimately how the system is being used in defense contexts and whether Anthropic’s built-in restrictions line up with what the military wants.

Turns out even artificial intelligence comes with real-world politics.

Source: Axios, @zerohedge

See 2 related tweets

  • @business: The Pentagon is close to cutting ties with Anthropic and may label the AI company a supply chain ris...
  • @rohanpaul_ai: RT @rohanpaul_ai: The U.S. Pentagon is pressuring top AI labs to let the military use their models f...

8. TheEconomist (Group Score: 48.7 | Individual: 25.6)

Cluster: 2 tweets | Engagement: 49 (Avg: 102) | Type: Tech

Making traditional Chinese medicine increasingly high-tech could make sense for some clinics. But building up trust in machines and robots will be a long process https://t.co/kY1hZhziin

Photo: Getty Images https://t.co/JzwAJPgZyT

See 1 related tweets

  • @TheEconomist: Adherents hope that technology will make traditional Chinese medicine increasingly objective and sta...

9. business (Group Score: 47.7 | Individual: 47.7)

Cluster: 1 tweets | Engagement: 916 (Avg: 87) | Type: Tech

India kicks off one of the world’s largest artificial intelligence summits Monday, with Prime Minister Narendra Modi seeking to clear a path for India in a heated race to develop frontier models. https://t.co/JtISA7RZYE

10. rohanpaul_ai (Group Score: 47.2 | Individual: 31.7)

Cluster: 2 tweets | Engagement: 2 (Avg: 118) | Type: Tech

Alibaba unveils new Qwen3.5 model for 'agentic AI era', Qwen3.5-397B-A17B. Apache 2.0 license

  • It is a 397B-parameter sparse mixture-of-experts model that keeps only 17B parameters active per token.

  • 8.6x higher decode throughput than Qwen3-Max at 32K context and 19.0x at 256K, while also being 3.5x and 7.2x faster than Qwen3-235B-A22B at those same lengths.

  • On quality, it reports IFBench 76.5 for instruction following, BFCL v4 72.9 for tool-call correctness, AIME26 91.3 for contest math accuracy, and SWE-bench Verified 76.4 for fixing real coding tasks.

  • Most of the claimed jump over Qwen3 comes from scaling reinforcement learning across many harder, tool-using, multi-turn environments rather than tuning for a narrow benchmark.

Older long-context models bog down because standard attention has to compare every token to many other tokens, so memory and compute rise fast as context gets longer.

Agent post-training also gets brittle when reinforcement learning runs in a small set of environments, since the model can memorize patterns instead of learning robust tool habits.

Qwen3.5 reduces the long-context bottleneck by mixing linear attention, via Gated DeltaNet, with regular attention layers so memory growth is better controlled at very long lengths.

  • For multimodality it uses early fusion, meaning text and visual tokens are processed together in 1 native stack instead of stitching a vision encoder onto a language model.

They also announced Qwen3.5-Plus, which is the managed, hosted API version of Qwen3.5-397B-A17B on Alibaba Cloud Model Studio, positioned for people who want the same core model without running multi-GPU inference themselves.

It ships with a 1M token context window by default, which is larger than the open-weight checkpoint’s typical long-context setups in most self-hosted stacks.

It also comes with “official built-in tools” and “adaptive tool use,” meaning the service can expose tool capabilities and handle when to call them in a more production-oriented way than a raw weights deployment.

On the API side, Alibaba documents OpenAI-compatible Chat Completions, plus an OpenAI Responses-style interface that includes built-in tools like web search, a code interpreter, and a web extractor, with the service managing conversation state so clients do less bookkeeping.

The release comes as Alibaba looks to attract more users to its Qwen chatbot app in China, a landscape currently dominated by rival tech giant ByteDance's Doubao and DeepSeek

See 1 related tweets

11. chrisalbon (Group Score: 47.1 | Individual: 47.1)

Cluster: 1 tweets | Engagement: 1109 (Avg: 91) | Type: Tech

I’ve had this dream of breaking down a complex task into like 100 tests, the visualizing the test coverage like this as the app is coded https://t.co/lbnE6F8tMQ

12. burkov (Group Score: 45.9 | Individual: 23.1)

Cluster: 2 tweets | Engagement: 12 (Avg: 103) | Type: Tech

A scientific table generator app: https://t.co/rZQcMs5m12

Describe a table or paste data in any text format. An LLM will generate LaTeX code and render it as a PDF and PNG.

Build your mini-apps on my new https://t.co/zasXgEO1Tp platform without coding, and make $$. https://t.co/uALbZsJP0e

See 1 related tweets

Describe a table or paste data in any text for...

13. ModelScope2022 (Group Score: 45.1 | Individual: 30.0)

Cluster: 2 tweets | Engagement: 132 (Avg: 133) | Type: Tech

Say hello to Ling-2.5-1T: 1T params, 63B active, MIT licensed

⚡ Hybrid Linear Attention: 1:7 MLA + Lightning Linear. Beats Kimi K2 on long-context throughput. 🧠 4x token efficiency: Composite rewards match frontier thinking models using 4x fewer tokens. 📚 1M context: YaRN to 1M. Beats Kimi K2.5 & DeepSeek V3.2 on RULER/MRCR. Perfect NIAH scores. 🛠️ SOTA tool use: Agentic RL trained. Leads BFCL-V4. Native support for Claude Code, OpenCode, OpenClaw. 🎯 Follows instructions: Bidirectional RL + agent verification. High-density, zero fluff. Honest: Gap vs GPT-5.2/Gemini 3 Pro on long-horizon tasks. Next version targets real-world completion.

🔧 Ship now: https://t.co/FbCgTuhNB5 https://t.co/ja4NiMdISa

See 1 related tweets

  • @AdinaYakup: Ling-2.5-1T is now live on @huggingface 🔥

https://t.co/IGzZc9pYQx

✨ 1T / 63B active - MIT license...

14. business (Group Score: 42.0 | Individual: 42.0)

Cluster: 1 tweets | Engagement: 370 (Avg: 87) | Type: Tech

A group of investors led by Blackstone plans to make an equity investment of up to $600 million in India’s artificial intelligence cloud startup Neysa. https://t.co/oA7G27rkx5

15. sofish (Group Score: 39.5 | Individual: 39.5)

Cluster: 1 tweets | Engagement: 626 (Avg: 94) | Type: Tech

RT @unwind_ai_: China's Alibaba just opensourced the SQLite of vector databases.

zvec runs as a library inside your app and is built for o…

16. BrianRoemmele (Group Score: 39.5 | Individual: 23.5)

Cluster: 2 tweets | Engagement: 66 (Avg: 344) | Type: Tech

I show you how in the 1940s a system was developed to help guide you through the next 5000 Days of AI and Robotics rise and human deskilling.

I show you how to become MORE human as AI does the mechanical work.

It is the ultimate symbiosis feedback loop… https://t.co/Cw8OBk9p3Q

See 1 related tweets

  • @BrianRoemmele: RT @BrianRoemmele: I show you how in the 1940s a system was developed to help guide you through the ...

17. steipete (Group Score: 39.0 | Individual: 39.0)

Cluster: 1 tweets | Engagement: 3615 (Avg: 730) | Type: Tech

I signed the contract today with software my company built years ago. (now it's @nutrientdocs, still best for anything PDF)

18. rohanpaul_ai (Group Score: 38.0 | Individual: 22.9)

Cluster: 2 tweets | Engagement: 76 (Avg: 118) | Type: Tech

AI data centers are swallowing so much memory that the rest of electronics is running short.

Some DRAM contract prices rose 75% from December to January, and sellers are repricing daily.

Big buyers are pulling DRAM and high-bandwidth memory (HBM) supply toward AI servers, and prices are jumping fast.

Past chip squeezes came from surprise demand, but this one comes from memory makers shifting capacity and investment toward HBM stacks used beside Nvidia and AMD accelerators.

HBM is basically many DRAM dies stacked in 8 or 12 layers and wired for very high throughput, so one accelerator can need hundreds of gigabytes, like 192GB on Nvidia Blackwell, and an NVL72 system totals 13.4TB.

TrendForce estimates HBM demand grows 70% year over year in 2026, and HBM takes 23% of DRAM wafer output, up from 19% in 2025, which leaves less plain DRAM for phones, PCs, cars, and consoles.

Howerver, this looks great for Samsung, SK Hynix, and Micron margins, but it punishes product teams that cannot prebuy supply.

bloomberg .com/news/articles/2026-02-15/rampant-ai-demand-for-memory-is-fueling-a-growing-chip-crisis

See 1 related tweets

  • @business: The AI boom — and the buildout of data centers — is fueling a memory chip crisis that's having a wid...

19. burkov (Group Score: 37.7 | Individual: 37.7)

Cluster: 1 tweets | Engagement: 350 (Avg: 103) | Type: Tech

When you train a neural network on a new task using examples (through supervised finetuning), it tends to forget what it already knew — a well-known problem called catastrophic forgetting.

The standard fix in reinforcement learning is to have the model learn from its own outputs rather than from a fixed dataset (this is called "on-policy" learning), but that requires a reward function telling the model how good its outputs are, which is often hard to define.

This recent MIT paper finds a way around that constraint: it uses the same model twice — once with a demonstration stuffed into its input as context (the 'teacher'), and once without (the 'student') — then has the student generate text, and updates the student's weights so that its token probability distributions get closer to the teacher's token probability distributions at each position in that generated text.

The trick works because LLMs can already adapt their behavior when shown an example in context, so the teacher is essentially a better version of the model that stays close to the original, making the learning signal gentle enough to avoid wrecking existing capabilities.

Across multiple experiments, this approach lets a single model sequentially learn three different skills while keeping all of them, where standard supervised training destroys earlier skills as soon as it moves to the next one.

The math also turns out to be equivalent to a form of reinforcement learning with an implicit reward, which gives the method a clean theoretical grounding beyond just "it works empirically."

Read this paper with an AI tutor: https://t.co/FQ5canO0qE

Read it alone: https://t.co/pPM7na1vWo

20. business (Group Score: 37.5 | Individual: 20.6)

Cluster: 2 tweets | Engagement: 44 (Avg: 87) | Type: Tech

Fractal Analytics shares fell in their Mumbai trading debut on Monday, signaling continued weakness in investor appetite for technology startups amid valuation concerns in an otherwise subdued stock market https://t.co/HF0j6KYYWz

See 1 related tweets

  • @Reuters: India's Fractal Analytics drops 5% in trading debut as AI jitters keep investors cautious https://t....
Discuss on Twitter • View on GitHub