这绝对是今年最让人拍案叫绝的硬件逆向神作:安全研究员 Kate 发现英伟达旗舰计算核心 GA100 芯片内部存在一个电压故障注入漏洞(Voltage Fault Injection),并亲手把一张当年矿潮遗留下来的阉割版 CMP 170HX 8GB 矿卡,原地解锁成了满血的 A100-80GB!
当年英伟达为了防止矿工抢占数据中心显卡,拿 GA100 大核心切出了 CMP 170HX 矿卡,不仅在硬件熔丝(Fuses)层面砍掉了显存容量、禁用了 NVLink 高速互联、把 PCIe 降到 Gen1,还锁死了张量计算吞吐。
Kate 顺着 GA100 上电启动读取板载熔丝配置寄存器的微妙时间窗口,用精密控制的电压毛刺(Voltage Glitching)去干扰芯片的熔丝传感放大器——让芯片在自检瞬间把原本烧断的限制熔丝全部误读为 0(未熔断状态),再配合重新植球焊满的高带宽显存,一举解开了英伟达在硅片层面的全部封印:不仅显存认满 81,920 MiB(80GB),连原本被阉割的 PCIe Gen4、NVLink、MIG 切分 也全数复活。
更离谱的是实测性能:因为英伟达在卖零售版 A100 时为了保证晶圆良率,统一锁在了 108 个流多处理器(SM);而这张被毛刺攻击彻底放飞自我的 GA100 核心,居然把晶圆上全部完好的 120 到 126 个 SM 单元 全点亮了,跑出来的真实算力比官方正版 A100 还要快出整整 1.2 倍。
3BHU1 retweeted
Sri Lanka has abolished pensions for its parliamentarians. Public office is a responsibility, not a lifelong entitlement. Taxpayer money must serve the people, especially with pressing needs in healthcare and education.
Time to rethink political privileges in India?
3BHU1 retweeted
IAS Officer: Do you have an appointment? Come back later.
Taxpayer: You are a public servant not a private employee. Serving the people who pay your salary is your duty not a favour.
3BHU1 retweeted
One bad policy can ruin an entire industry:
In 2022, India was the world’s 2nd-largest sugar exporter.
In 2026, India is set to import sugar for the first time in a decade.
Sugar prices have risen by 40% in just two months.
Rice prices are also hitting a one-year high.
The government cannot refuse to roll back E20 simply because some businessmen have invested in ethanol factories. That is a joke.
E10 will also require ethanol.
3BHU1 retweeted
震惊瘫坐了,怎么把 Linux on Android 做得这么流畅的,还有 GPU accelerated kwin_wayland,还做了音量、亮度、Wi-Fi设置、剪贴板、输入集成、触屏,还预装了 flatpak (
评价为这帮人用 AI 做出来的东西比以往十几年的 Linux 容器易用得多
(就是DroidDeck,但是没想到它的 Linux 部分也做得如此精细)
3BHU1 retweeted
We present PhysAg3D! 🚀 From a single image, it generates physically interactive 3D assets ready for Embodied AI simulation.
Bridging 3D generation & Agentic models, it delivers image-aligned fidelity with rich physical interactions (articulated joints, soft bodies, etc.).
⌨️ 太卷了,OpenIDEA 照着 IntelliJ IDEA 的手感重做了一个 IDE,打开项目 0.9 秒,内存只占 169MB。
OpenIDEA 的仓库 10 月 21 日才建,10 月 28、29 两天就连发了四个版本,issue 区已经有六十多个开着。
用过 IDEA 的都懂:项目一大,开机等索引、风扇狂转,内存两个 G 起步。
OpenIDEA 用 Rust 做内核,键位、工具窗口、全局搜索都沿用 IDEA 的习惯,肌肉记忆不用改;编辑、调试、Git、终端、数据库和 HTTP 请求也都在一个窗口里。
项目拿 freqtrade 这个 Python 项目做了对比:从打开到 CPU 平稳,OpenIDEA 1.1 秒,IDEA 13.2 秒;内存 169MB 对 2011MB。文档也老实写了,IDEA 对 Python 的分析更深,而且项目和 JetBrains 没有任何关系。
熟悉的手感,轻得多的身子。
GitHub:github.com/l0ng-ai/openidea
3BHU1 retweeted
Thank you Constitution of India, its politicians and its people.
A general category candidate scoring 550 marks with an All India Rank of 13,000 in NEETPG cannot get MD Anaesthesiology at Banaras Hindu University.
But an SC candidate with just 70 marks out of 800 and rank 2,23,863 walks straight into the same MD Anaesthesiology seat at BHU.
This is not counselling.
This is the open murder of merit.
You are stuffing operation theatres with people who scored lower than random guessing, then expecting them to anaesthetise patients.
Patients will pay with their lives. Genuine doctors who actually studied will keep rotting outside. And you will still call this “social justice”.
Keep destroying the system.
History will remember who sold merit for votes.
3BHU1 retweeted
"How Google humiliates Indian Developers with Google Play BillDesk Verification"
Scary to read stories of people having to pay bribes of 50K to 2 lakhs to BillDesk. @RBI what is this?
reddit.com/r/india/comments/…
3BHU1 retweeted
Alright folks github.com/Acelogic/Snapinto… here it is, have fun. Expect for this to be very rough, you may have to point some agents at it. Working on de-slopifying the codebase and making it more adaptable. But hey Hackintosh Lives 🥳
3BHU1 retweeted
Hacked and reverse engineered @atomberg_tech fan to run using Siri / Apple home
With matter over wifi
Going to write full blog and tear down tomorrow
But wtf! Took me like 4 hours
3BHU1 retweeted
最近我的推特时间线:
移植switch游戏到iOS/MacOS
破解PS5主机
移植iOS到安卓(真假未知)
在安卓上安装纯血Linux
把现代软件搬到win98/旧Mac系统
自制kindle固件
给电视刷机
玩ESP32的
拥抱AI,迎来了一个肆意折腾的好时代
3BHU1 retweeted
🦆🔫
He convertido un AirPod en un mando para jugar al Duck Hunt.
Lo giras para apuntar y das un toque en el micrófono para disparar.
Sin tocar el ordenador, full vibecodeado por voz al mientras dormía a la bebé 😅
3BHU1 retweeted
I don't think people realize the movement happening with "Reverse Engineer Anything," which just crossed 300k GitHub stars. It's not about "oh, let's make an open source Photoshop" or "the death of software." It's about unlocking human creativity. It's about what this UNBLOCKS in our brains, what new building blocks we have, and the big question: WHAT WILL YOU BUILD?! cc @reatools @morluto
3BHU1 retweeted
Quick Question:-
People say India is more FREE than China is
Having lived for 68 years in India I know
I. Courts don't offer equitable justice to all citizens
II. Police, Officials and Politicians abuse their power massively
III. Any Indian can be jailed for 7 days without being presented before a magistrate
IV. Any Indian can be interrogated outside the presence of his lawyer by the police
V. Properties can be encroached by thugs without the slightest recourse barring an expensive settlement or a long drawn court battle
VI. Citizens can be forced to wait for 20 minutes, 30 minutes even 45 minutes for convoys of ministers to pass through
VII. The Average time to prove ones innocence is 6 1/4 years
VIII. The Average time to establish ones guilt is a single FIR registered in 20 minutes
IX. Bail is decided based on wealth of the accused rather than the weight of the case. A Theft accused could spend 15 months without bail and a rape accused could walk out in 3 months with bail
X. Corrupt Politicians & Officials are so frequent that people now regard Honest Politicians & Officials as fools
Where is this FREEDOM everyone talks about?
(If you say "Freedom to tweet this" - I believe the same freedom is available in every nation except North Korea maybe)
3BHU1 retweeted
People are now using AI agents to reverse engineer and patch games. The legal side is a grey area, but the engineering is a great lesson in building verifiable AI systems :) Let us focus on that and see how it is done...
Warning: Long and detailed post ahead!
A shipped game is a binary. Decompilation turns it back into readable code, which is what you need to patch the game. However, decompilation produces assembly, so...
- Types and structs are gone. The binary only tells you to "read 4 bytes at offset 16", not "player->health".
- Names and comments are gone. Everything becomes jargon like "sub_4012A0" and "v7".
- The structure is gone. The optimizer inlines functions and reorders instructions, so the assembly looks nothing like the original source.
Tools like Ghidra generate rough pseudo-C, but someone still has to figure out what the developer probably wrote. That is a guess, and it is the part agents are asked to handle.
Now that the context is set, here is what the process and flow look like...
A game has tens of thousands of functions, so the work is split up. Each function becomes a ticket (e.g., a GitHub issue), and multiple agents work in parallel. Each agent queries the disassembler through an MCP server, guesses types and struct layouts, and writes reasonably clean C++.
The trouble is that wrong code can look right. For example, the original "health" field sits 16 bytes into a struct, but the agent guesses 20.
The code compiles and the game starts, but the function now reads the wrong memory. Agents can also invent logic when the assembly is confusing or delete odd-looking checks that were actually bug fixes.
A reviewer agent cannot reliably catch this. It has no ground truth. The fix is a check that returns only PASS or FAIL. This is called matching decompilation:
- Compile with the same compiler and flags the original developers used.
- Compile the agent's function into an object file.
- Extract its bytes and compare them with those of the corresponding function in the original executable.
Identical bytes mean identical machine code and, therefore, identical behavior, bugs included :) Agents are still guessing, but we now have a way to verify the output and retry until the guess is correct.
By the way, this applies well beyond games. You need to define correctness in a way a machine can check, then let the agents iterate against it. Tests, golden files, whatever suits your use case.
Hope this helps.























