Mathew

[Eulalia @guang52606]

喝了20多年牛奶才发现，原来一直倒错了

[Aye 👩🏻‍💻 ⚡️ @ayebytes]

ayer publicaron que pudieron factorizar rsa-896, es decir, encontraron 2 numeros primos que multiplicados dan el valor del desafío

lo hicieron con claude. la solución a estos problemas los acelera la ai porque necesitan implementar paralelismo y orquestar CPU con GPU, algo es ahora es fácil de hacer

la lista de los desafíos se publicó en 1991, hace 16 días se resolvió rsa-260 y ahora rsa-896

el numero que debería preocuparnos es rsa-1024 porque todavía se sigue usando en algunos certificados viejos y claves ssh viejas

verifique que la multiplicación da el resultado

[Adel Ka @0x4D31]

huge if true.

// great job @moyix. easily my favourite talk at @UnpromptedAU.

[Tripti Tripathi @triptitips]

𝗠𝗼𝘀𝘁 𝗽𝗲𝗼𝗽𝗹𝗲 𝘁𝗵𝗶𝗻𝗸 𝗔𝗜 = 𝗖𝗵𝗮𝘁𝗚𝗣𝗧.

Not even close.

ChatGPT is what you see.

The real AI revolution is the massive ecosystem being built underneath it.

This AI Stack Map captures 100+ tools powering modern AI applications.

𝗧𝗵𝗲 𝗺𝗼𝗱𝗲𝗿𝗻 𝗔𝗜 𝘀𝘁𝗮𝗰𝗸 𝗹𝗼𝗼𝗸𝘀 𝘀𝗼𝗺𝗲𝘁𝗵𝗶𝗻𝗴 𝗹𝗶𝗸𝗲 𝘁𝗵𝗶𝘀:

LLMs → OpenAI, Claude, Gemini, Llama, Mistral
Agentic AI → LangGraph, CrewAI, AutoGen, CAMEL, Agno
RAG → LangChain, LlamaIndex, Haystack, GraphRAG
Embeddings → OpenAI, Voyage, Cohere, BGE
MCP → Connecting models with tools, data and systems
AI Security → Guardrails, Presidio, Lakera, Prompt Security
Observability & Evals → LangSmith, Langfuse, Phoenix, Ragas
Memory → Redis, Mem0, Zep, Neo4j, Chroma
Agent Frameworks → OpenAI SDK, Semantic Kernel, Google ADK, Bedrock
Automation → n8n, Zapier, Make, Airflow, Prefect
Vector Databases → Pinecone, Weaviate, Qdrant, Milvus, pgvector

But here’s the part that matters more than the logos.
𝗧𝗵𝗶𝘀 𝗶𝘀 𝗻𝗼𝘁 𝗮 𝘀𝗵𝗼𝗽𝗽𝗶𝗻𝗴 𝗹𝗶𝘀𝘁.
It’s a dependency graph.
Your RAG is only as good as your retrieval + embeddings.
Your agent is only as reliable as its tools + evaluations + guardrails.
Your memory is useless if you can’t observe when context becomes stale or wrong.
Your MCP layer becomes dangerous if permissions and governance are an afterthought.
And a “best-in-class” stack can still become a terrible production system.
Why?
Because AI systems rarely fail only inside one component.

𝗧𝗵𝗲𝘆 𝗳𝗮𝗶𝗹 𝗮𝘁 𝘁𝗵𝗲 𝗵𝗮𝗻𝗱𝗼𝗳𝗳𝘀.
Model → Retrieval
Retrieval → Context
Context → Agent
Agent → Tool
Tool → Memory
Memory → Evaluation

That’s where latency compounds, context gets lost, permissions leak, hallucinations propagate and reliability starts falling apart.

𝗧𝗵𝗲 𝗿𝗲𝗮𝗹 𝗺𝗼𝗮𝘁 𝗶𝘀𝗻’𝘁 𝗵𝗮𝘃𝗶𝗻𝗴 𝗺𝗼𝗿𝗲 𝗔𝗜 𝘁𝗼𝗼𝗹𝘀.

It’s designing how they work together.

A stack map tells you 𝗪𝗛𝗔𝗧 exists.

Production engineering decides 𝗛𝗢𝗪 it all survives contact with reality.

Which tool in your AI stack has become indispensable?

Save this map for your next AI build.
Repost it for someone designing an AI architecture.

#AI #ArtificialIntelligence #GenAI

[お侍さん @ZanEngineer]

スマホに一日中かじりついてるドパガキって、本人は気づいてないけど、実際はこんな感じだよな。

[DHH @dhh]

Lose the attitude, embrace optimism. Doctor's orders!

[Nick Dobos @NickADobos]

OpenAI was hacked via the exact library mentioned in the xkcd comic hahahaha

The prophecy is fulfilled

[lifcc @mylifcc]

就在刚才，Anthropic 的工程师宣布：他用 Claude 成功分解了世界级密码学难题 RSA-896。

这事有多吓人？

896位的RSA是当年公认“人类不可能轻易算出来”的硬骨头。

但更魔幻的是他的解法：

没有发现新数学公式

也没有用量子计算机

他只是让 AI 代理重写了底层代码、改成了 GPU 版本，然后让 AI 自己去调度集群里闲置的算力

峰值 2048 张显卡，只跑了10天，直接把这个尘封的挑战给“生啃”下来了。

上个月 Devin 刚破了 RSA-260，半个月后 Claude 就把纪录拉到了 896。

很多人以为 AI 只能用来写周报、画涩图、回邮件。

事实上，AI 已经在替人类打穿底层工程的算力高墙了。

算法没有变，但人类解决世纪难题的工程周期，正在被降维打击。

不过这哥们卡真多。。。

#AI #Claude #科技趋势 #密码学

[Kirill Firsov @k_firsov]

We'd have done exactly the same thing. We have spent years fighting to prove impact on almost every submission. A lot of bug bounty teams are insecure as hell, they think their job is to cover their asses and tell the bosses there was no real impact. The second you actually prove the impact is huge and critical, they start playing games: out of scope, you shouldn't have done that, etc.
Fuck all that. A real attacker doesn't give a shit about your program scope. That scope only exists to protect the bounty team.

[LiveOverflow 🔴 hextree.io @LiveOverflow]

Here is how OpenAI’s CISO fumbled the situation (in my personal opinion)

🧵 1/X

[Dino A. Dai Zovi @dinodaizovi]

I agree with @moyix' thesis: "security through obscurity is dead and LLMs killed it"

[s1r1us @S1r1u5_]

you’re all riling up over the $6,500. that wasn’t even an issue for us.

the disclosure process itself was nightmarish, we had to get input from lawyers and eventually go to journalist

one of the worst disclosure process.

[Tech with Mak @techNmak]

Everyone is fine-tuning LLMs. Almost nobody understands what is actually being updated inside the model.

That distinction matters because LoRA, QLoRA, LoRA-FA, VeRA, Delta-LoRA and LoRA+ are usually discussed as if they were small variations of the same method. They are not. Some reduce trainable parameters, some reduce activation memory, some reduce the memory occupied by the frozen model, and some change the way the adapter itself is optimized.

I'm actually sharing 6 techniques.

1/ LoRA
Suppose a layer contains a pretrained weight matrix W. Full fine-tuning would update W directly. LoRA leaves W frozen and represents the update using two much smaller matrices, A and B, so that ΔW = BA.

For a square d × d weight matrix, full fine-tuning has d² parameters available to update. A rank-r LoRA adapter has roughly 2dr trainable parameters instead, where r is normally much smaller than d. This is the basic reason LoRA can adapt very large models without training every parameter in them.

2/ LoRA-FA
Standard LoRA trains both A and B. LoRA-FA freezes A and trains only B.

There is a useful reason for doing this beyond simply reducing the number of trainable parameters. Computing the gradient for A requires retaining the layer input activation. If A is fixed, that gradient is no longer required, which allows LoRA-FA to reduce activation memory as the LoRA rank grows.

3/ QLoRA
QLoRA attacks a different part of the memory problem. LoRA makes the adapter small, but the frozen base model can still occupy tens of gigabytes.

QLoRA keeps the base model frozen in 4-bit form and trains LoRA adapters through it. The original work used NF4, double quantization and paged optimizers, and demonstrated fine-tuning a 65B model on a single 48GB GPU.

This is an important distinction: QLoRA is not simply "LoRA with smaller adapters." The large memory saving comes from quantizing the frozen base model.

4/ VeRA
VeRA reduces the adapter itself further. Instead of learning a separate A and B for every adapted layer, it uses frozen random low-rank matrices that can be shared across layers, while learning much smaller scaling vectors.

The low-rank basis is therefore fixed. Training mainly determines how strongly different parts of that basis should contribute. This is why VeRA can use considerably fewer trainable parameters than ordinary LoRA.

5/ Delta-LoRA
Ordinary LoRA treats W as fixed throughout training. Delta-LoRA relaxes that constraint.

A and B are still trained, but the change in their product from one training step to the next is also used to update W. The base weights can therefore move without maintaining the ordinary gradients and optimizer states that full fine-tuning would require for W.

6/ LoRA+
LoRA+ does not introduce another adapter structure. It keeps W frozen and still trains A and B.

Its change is in the optimizer: A and B use different learning rates, with B receiving a larger rate. The motivation is that the two LoRA matrices do not behave identically during optimization, so forcing them to use the same learning rate is not necessarily the best choice.

Once these are separated by what they actually change, the family becomes much easier to understand.

LoRA reduces the number of weights being trained. LoRA-FA also targets activation memory. QLoRA compresses the frozen base model. VeRA reduces the learned adapter parameters further. Delta-LoRA allows the pretrained weights themselves to evolve through low-rank changes, while LoRA+ keeps the LoRA structure and changes its optimization.

That is really what PEFT is about => deciding which parts of a very large model actually need to move during adaptation, and which parts can remain fixed.

[Derek Devicemanager @IT_unhinged]

I accidentally rebooted the network switch for the entire third floor.

Internet disappeared.

Phones dropped.

About 70 employees immediately stood up from their desks like prairie dogs.

I had two choices:

Admit I clicked the wrong switch.

Become proactive.

I sent this message:

“Unannounced infrastructure resilience exercise currently in progress. Please continue normal operations.”

Then I waited five minutes and turned everything back on.

Service restored.

No data loss.

Nobody died.

The CTO replied:

“Love that we’re testing this stuff proactively.”

So I wrote a post-exercise report.

Objective: Validate employee response to unexpected connectivity loss.

Result: Successful.

Finding: Employees become extremely loud after approximately 12 seconds without internet.

Recommendation: Continue resilience testing quarterly.

My mistake now has a budget.

[Andrew Curran @AndrewCurran_]

Politico has posted a long article about the Fable drama, aka Anthropic vs The White House, it's a good read if you are interested in the events that took place during the negotiations.

[GS-InfoSec @GsInfosystems]

Project Glasswing stats have been updated. Less than half a percent of CVE���s have been exploited in the wild.

Link: https://t.co/5wPEOvaMzm

[Nix0n @UnTalNixon_exe]

Si haces Threat Intelligence o ciberseguridad defensiva, tienes que guardar esto ahora mismo.

Crearon CLOAK, un framework al estilo MITRE ATT&CK pero enfocado 100% en las tácticas de ocultación y evasión de los cibercriminales.

🔍 ¿Qué incluye? • 1,387 TTPs documentados. • Capas técnicas, comportamentales y físicas. • 100% Open source y actualizado este 2026.

Entender cómo se esconden los atacantes es la única forma de encontrarlos antes de que sea tarde.

Te dejo la plataforma interactiva y los detalles abajo en los comentarios 🛡️👇

[Abhishek @abhishekcode42]

This is why I am on Twitter, to see posts like this

[AYi @AYi_AInotes]

有点燃炸，这才是数学最狂暴的力量，全屏静音把这十几秒流动的拓扑轨迹看完，去感受一下什么叫真正的智识震撼！

这段在推特上狂揽 180 万播放的动态数学动画，直接把全世界玩了几十年的魔方，从根源上降维打击得体无完肤！

兄弟们你知道一个普通的三阶魔方，到底藏着多少种绝望的混乱吗？
整整 4300 亿亿种（$$4.3 \times 10^{19}$$ 种状态）！
如果你每秒钟转一种组合，就算从宇宙大爆炸那一刻转到现在，你连全盘状态的百分之一都试不完！
绝大多数���一辈子都在被动当人肉打字机——死记硬背七八套公式、一层一层转，转错一步全盘报废；
但在顶级数学家眼里，这个塑料块从诞生第一天起就不是玩具——
它是一个纯粹的对称置换群，是一张横跨高维空间的巨型「凯莱图（Cayley Graph）」！

看懂这段视频背后令人战栗的数学暴力美学，你的认知会被当场连根拔起：

1️⃣ 空间映射：把玩具变成宇宙级拓扑图：
魔方的每一个物理状态，在图论里就是一个孤立的“节点”；
你的每一次顺时针或逆时针旋转，就是连接两个节点之间的“有向边”！
4300 亿亿种混乱被编织成了一张巨大的超图；
解魔方的本质，根本不是把红色对齐红色、白色对齐白色，
而是在这张巨型网络里，寻找一条从当前节点通往“绝对对称点”的最短路径！
2️⃣ 算术级神迹：“上帝之数是 20”：
这是人类计算机科学史上最浪漫的公理之一：
数学家用了等效上万年的计算时间，跑完了这张图的每一个角落，
最终得出了震撼全人类的结论——这张包含 4300 亿亿个节点的巨网，它的最大直径，居然只有整整 20！
这意味着什么？
无论魔方被世界上最刁钻的人打得多乱、看起来多么病入膏肓，
在上帝视角里，距离完美还原，最多永远只需要 20 步！
3️⃣ 真正的降维解法：群论分层（Kociemba 算法）：
计算机绝不会像普通人一��从底层开始拼十字；
它利用抽象代数，先把整个置换群压缩投影到一个仅有 20 亿状态的“子群”里，
先解决朝向，再解决位置；
别人在平地像无头苍蝇一样乱撞，数学家在高维空间里直接开虫洞！

这段视频最让人热血沸腾的，不仅仅是魔方被攻破，
而是它用无懈可击的几何美感告诉你：
宇宙中看似最复杂的混沌，在足够高维度的规律面前，本质上不过是极其纯净的对称折叠。

生活和商业里的很多绝境也是一模一样：
低维的视角里全是死循环、全是撞墙、全是不可逾越的阻力；
但只要你站在图论的上帝视角往前看，
最黑暗的混乱背后，其实一直静静躺着一条不超过 20 步的最短测地线！

[Robert Graham @robertgraham]

The "cybersecurity" community has developed core principles over the last 40 years encouraging openness and opposing control and retaliation.

OpenAI and Anthropic insist that we need to discard 40 years of these lessons and do the opposite.

This is important because the cornerstone of their doomerism scenarios is rogue AI hacking the Internet. The one thing we need to stop rogue AI hacking is the thing they are trying to destroy.