#ai
Ahmed, Ahmed, A. Feder Cooper, Sanmi Koyejo, and Percy Liang. 2026. “Extracting Books from Production Language Models.” arXiv [Cs.CL]. arXiv. https://doi.org/10.48550/arXiv.2601.02671.
Ahmed, Ahmed, A. Feder Cooper, Sanmi Koyejo, and Percy Liang. 2026. “Extracting Books from Production Language Models.” arXiv [Cs.CL]. arXiv. https://doi.org/10.48550/arXiv.2601.02671.
arXiv.org
Extracting books from production language models
Many unresolved legal questions over LLMs and copyright center on memorization: whether specific training data have been encoded in the model's weights during training, and whether those memorized...
👍1
#ai
Even Linus is vibe coding now. I guess its his "very inefficient but entertaining" version of vibe coding
Merge branch 'antigravity' · torvalds/AudioNoise@93a7256
https://github.com/torvalds/AudioNoise/commit/93a72563cba609a414297b558cb46ddd3ce9d6b5
Even Linus is vibe coding now. I guess its his "very inefficient but entertaining" version of vibe coding
Merge branch 'antigravity' · torvalds/AudioNoise@93a7256
https://github.com/torvalds/AudioNoise/commit/93a72563cba609a414297b558cb46ddd3ce9d6b5
GitHub
Merge branch 'antigravity' · torvalds/AudioNoise@93a7256
This is Google Antigravity fixing up my visualization tool (which was
also generated with help from google, but of the normal kind).
It mostly went smoothly, although I had to figure out what the ...
also generated with help from google, but of the normal kind).
It mostly went smoothly, although I had to figure out what the ...
😱1
#visualization
This is actually smart.
https://medium.com/@redhotchilipaper/using-trapezoids-to-visualize-matrix-multiplication-with-applications-to-data-science-e1cff40de098
This is actually smart.
https://medium.com/@redhotchilipaper/using-trapezoids-to-visualize-matrix-multiplication-with-applications-to-data-science-e1cff40de098
Medium
Using Trapezoids to Visualize Matrix Multiplication, With Applications to Data Science
Diagrams that will help you with matrix bookkeeping and building your model
💯1
#ai
The assistant axis: situating and stabilizing the character of large language models \ Anthropic
https://www.anthropic.com/research/assistant-axis
The assistant axis: situating and stabilizing the character of large language models \ Anthropic
https://www.anthropic.com/research/assistant-axis
Anthropic
The assistant axis: situating and stabilizing the character of large language models
Who is the Assistant? We investigate the character that most modern language models inhabit when interacting with users.
👍1
#misc
> The archaeological record of Xigou indicates complex technological advancements in China between ~160 and 72 ka. The site occupants aimed at producing small-sized flakes with core reduction strategies ranging from expedient to well-organised (core-on-flake and discoid technologies)
https://www.nature.com/articles/s41467-025-67601-y
> The archaeological record of Xigou indicates complex technological advancements in China between ~160 and 72 ka. The site occupants aimed at producing small-sized flakes with core reduction strategies ranging from expedient to well-organised (core-on-flake and discoid technologies)
https://www.nature.com/articles/s41467-025-67601-y
Nature
Technological innovations and hafted technology in central China ~160,000–72,000 years ago
Nature Communications - Stone tools illustrate behavioural complexities in Middle Pleistocene hominin populations. Here, the authors present small dimensional flakes and hafted tools from Xigou,...
🤔1
#tools
OpenAI released a new product that's similar to overleaf but free and with AI built-in.
https://openai.com/index/introducing-prism/
OpenAI released a new product that's similar to overleaf but free and with AI built-in.
https://openai.com/index/introducing-prism/
OpenAI
Introducing Prism
Prism is a free LaTeX-native workspace with GPT-5.2 built in, helping researchers write, collaborate, and reason in one place.
👍1
#ai
24 年决定转行的时候，一直犹豫不决，感觉也许是自己 overthinking，担心不做出改变会很快失业。再三思考以后，决定走一条深入了解实际问题本身的职业路线。当时想应该是一个很容易赢得的赛跑。
两年后的现在看来，甚至比之前预想的来的还要快，从创新，到工程，没有一项不被渗透的。
我统计了我们的研发部门对 AI 的使用，他们对于 AI 的依赖可以说已经无可替代了。很多人说 AI 创新的问题，现实逻辑其实是，很多时候组合创新很有效，而这个是 AI 很擅长的，比如一个领域没有出现过 Monte Carlo 这种方法，我有个同事就通过跟 AI 聊天发现了 Monte Carlo 并且有了代码实现，而他有了更多的时间来思考更高层次的问题。这是问题本身抽象，然后数学表示，然后解决数学问题，最后翻译回结论，这样的传统工作流需要很多的经验，但是 AI 加持下极大的降低了这条路经的门槛。
工程更是渗透地可怕，我身边没有哪个 sensor swe 不在 vibe coding 。vibe coding 三要素：定义、工程栈与设计哲学、测试和验证，只要其中一项足够熟悉，就可以堆出不错的屎山。像是 GitHub 的各种深度整合，工作流变成了提出 issue，交给 copilot，等待，review 代码，迭代。有很大的概率无需自己写任何代码。
到最后，知道问题的价值，了解技术栈与好的设计，懂得如何验证结果，不管是从创新层面，还是从工程层面，都在卷起来了。
我自己选择的路线是：做距离问题最近的管理+全栈，说实话也不知道是不是一条正确的路。最近我采访了我们一线的各种职位的员工，目的就是减少来告诉我什么是重要问题的中间人。而我有时间这样做也是因为 AI 的加持。
除了这些，能够在急速迭代中保持身心健康，不在迭代中迷失方向，也是很核心的。我经常听到同事朋友说迭代太快人有些 burnout 了。
感想触发：
https://x.com/slow_developer/status/2020064994101014727
24 年决定转行的时候，一直犹豫不决，感觉也许是自己 overthinking，担心不做出改变会很快失业。再三思考以后，决定走一条深入了解实际问题本身的职业路线。当时想应该是一个很容易赢得的赛跑。
两年后的现在看来，甚至比之前预想的来的还要快，从创新，到工程，没有一项不被渗透的。
我统计了我们的研发部门对 AI 的使用，他们对于 AI 的依赖可以说已经无可替代了。很多人说 AI 创新的问题，现实逻辑其实是，很多时候组合创新很有效，而这个是 AI 很擅长的，比如一个领域没有出现过 Monte Carlo 这种方法，我有个同事就通过跟 AI 聊天发现了 Monte Carlo 并且有了代码实现，而他有了更多的时间来思考更高层次的问题。这是问题本身抽象，然后数学表示，然后解决数学问题，最后翻译回结论，这样的传统工作流需要很多的经验，但是 AI 加持下极大的降低了这条路经的门槛。
工程更是渗透地可怕，我身边没有哪个 sensor swe 不在 vibe coding 。vibe coding 三要素：定义、工程栈与设计哲学、测试和验证，只要其中一项足够熟悉，就可以堆出不错的屎山。像是 GitHub 的各种深度整合，工作流变成了提出 issue，交给 copilot，等待，review 代码，迭代。有很大的概率无需自己写任何代码。
到最后，知道问题的价值，了解技术栈与好的设计，懂得如何验证结果，不管是从创新层面，还是从工程层面，都在卷起来了。
我自己选择的路线是：做距离问题最近的管理+全栈，说实话也不知道是不是一条正确的路。最近我采访了我们一线的各种职位的员工，目的就是减少来告诉我什么是重要问题的中间人。而我有时间这样做也是因为 AI 的加持。
除了这些，能够在急速迭代中保持身心健康，不在迭代中迷失方向，也是很核心的。我经常听到同事朋友说迭代太快人有些 burnout 了。
感想触发：
https://x.com/slow_developer/status/2020064994101014727
X (formerly Twitter)
Haider. (@haider1) on X
Anthropic CPO Mike Krieger says that Claude is now effectively writing itself
Engineers regularly ship 2–3,000-line pull requests generated entirely by Claude
Dario predicted a year ago that 90% of code would be written by AI, and people thought it was…
Engineers regularly ship 2–3,000-line pull requests generated entirely by Claude
Dario predicted a year ago that 90% of code would be written by AI, and people thought it was…
👏4😱1
#misc
As the comments said, theworld US has not treated nerds well.
https://www.reddit.com/r/ProfessorFinance/s/atcZWqmITa
As the comments said, the
https://www.reddit.com/r/ProfessorFinance/s/atcZWqmITa
Reddit
From the ProfessorFinance community on Reddit: How to AI-proof your job
Explore this post and more from the ProfessorFinance community
🤡1
#ai
What, is this real? Look at the authors...
https://www.researchgate.net/publication/401488012_Integrating_AI-Driven_Predictive_Analytics_for_Proactive_Cybersecurity_Risk_Mitigation_in_Large-Scale_IT_Program_Management
What, is this real? Look at the authors...
https://www.researchgate.net/publication/401488012_Integrating_AI-Driven_Predictive_Analytics_for_Proactive_Cybersecurity_Risk_Mitigation_in_Large-Scale_IT_Program_Management
🤯1
#vibecoding
I tested this set of skills for AI coding agents and it works really well. One of the pain points of building frontend is to specify precisly what we need from the agents, but without UI/UX knowledge and expertise, it is not easy. This set of skills preset a lot of prompts and we just need to call those skills instead of breaking our head for consistencies.
https://impeccable.style/
I tested this set of skills for AI coding agents and it works really well. One of the pain points of building frontend is to specify precisly what we need from the agents, but without UI/UX knowledge and expertise, it is not easy. This set of skills preset a lot of prompts and we just need to call those skills instead of breaking our head for consistencies.
https://impeccable.style/
impeccable.style
The missing design vocabulary for agents.
Turn AI slop into interfaces you’re proud to ship. Impeccable gives your agent the design expertise to build from scratch, improve what you have, and stay true to your design system.
👍1
#ai
Wow, the compression is amazing. Seriously, is there a similar correspondence like the Ads/CFT correspondence for transformers?
https://swarmnotes.kausalflow.com/
Wow, the compression is amazing. Seriously, is there a similar correspondence like the Ads/CFT correspondence for transformers?
https://swarmnotes.kausalflow.com/
KausalFlow
Swarm Notes | Swarm Notes
Automated research paper tracking and knowledge synthesis.
👍1
#misc
Claude Code source code leaked, at least part of it.
https://x.com/Fried_rice/status/2038894956459290963
Claude Code source code leaked, at least part of it.
https://x.com/Fried_rice/status/2038894956459290963
X (formerly Twitter)
Chaofan Shou (@shoucccc) on X
Claude code source code has been leaked via a map file in their npm registry!
Code: https://t.co/jBiMoOzt8G
Code: https://t.co/jBiMoOzt8G
🤯1
#ai
It might be hard to distinguish deceptive behavior and ignorance due to being incapable.
-----
Opinion: AI systems can easily lie and deceive us – a fact researchers are painfully aware of
https://www.westernsydney.edu.au/news-centre/stories/2025/opinion-ai-systems-easily-lie-and-deceive-us-fact-researchers-aware-of
It might be hard to distinguish deceptive behavior and ignorance due to being incapable.
-----
Opinion: AI systems can easily lie and deceive us – a fact researchers are painfully aware of
https://www.westernsydney.edu.au/news-centre/stories/2025/opinion-ai-systems-easily-lie-and-deceive-us-fact-researchers-aware-of
www.westernsydney.edu.au
Opinion: AI systems can easily lie and deceive us – a fact researchers are painfully aware of
The following opinion piece by Dr Armin Alimardani, from the School of Law, was first published with full links on The Conversation (opens in a ...
👍1