Ahnaf Shahriar (@enough_ahnaf) header
Ahnaf Shahriar (@enough_ahnaf) profile picture

co-founder of @simantic YC F26 | washed waterloo intern 🇨🇦

Waterloo, ON
Joined July 2022
Photo by Ahnaf Shahriar (@enough_ahnaf)Photo by Ahnaf Shahriar (@enough_ahnaf)Photo by Ahnaf Shahriar (@enough_ahnaf)Photo by Ahnaf Shahriar (@enough_ahnaf)Photo by Ahnaf Shahriar (@enough_ahnaf)Photo by Ahnaf Shahriar (@enough_ahnaf)Photo by Ahnaf Shahriar (@enough_ahnaf)Photo by Ahnaf Shahriar (@enough_ahnaf)Photo by Ahnaf Shahriar (@enough_ahnaf)Photo by Ahnaf Shahriar (@enough_ahnaf)
Ahnaf Shahriar retweeted
Ramadan Global Buildathon space partners! If you’re in Detroit - or SF, KL, London, etc - consider joining in person Here’s a quick walk through with Medina Labs in Michigan to see how they can help me bring my Ramadan idea to life (it’s a real one! 🐦 my smart Hoopoe)
4
5
2
38
2,040
Ahnaf Shahriar retweeted
Rockstar Games 周三解雇了我。第二天 GTA 6 就发售了 我是 Rockstar North 的高级引擎工程师,从事 RAGE 性能优化。三个月前,我在遥测模块里发现了一条本不该存在的数据管道,而且我能证明它流向了哪里 此刻每一份正在运行的 GTA 6 都在收集玩家的行为数据。这部分不算新闻。每款游戏都会收集遥测数据。隐私政策里写的是"用于改进未来产品的匿名游戏数据分析"。我发现的不是这个 遥测模块的数据会汇入 Take-Two 内部的一个系统,名为 Lens。Lens 并不会把匿名使用统计数据发送到 Rockstar 开发人员的数据看板。它为每一位玩家建立一份个人行为画像。不是匿名的。与你的主机账号、你的平台 ID、你的 IP 绑定 我在8月14日向我的技术主管报告了这件事。他说这是"用于玩家体验研究的扩展分析",让我别管这事。我要求提供数据处理协议。他说这是在 Take-Two 集团层面处理的,不在 Rockstar 的职责范围之内 我调取了 Lens 数据管道的完整文档。我发现的内容如下: 遥测的内容远远超出游戏时长和崩溃报告。Lens 以毫秒级精度记录每一次输入。按键时间点、摇杆动作、反应速度、PS5 自适应扳机上的扳机压力模式。它实时运行行为分析,并生成他们称之为"玩家行为签名"的东西 - 一份对每位玩家都独一无二的画像,就像指纹对每个人都独一无二一样 Lens 会从你的游戏过程中推导出:冲动控制评分、风险承受度指数、攻击性反应模式、耐心阈值、压力下的决策速度、消费冲动评级。这些不是调试指标。这些是心理行为类别 直到我找到了对外传输的管道,我才明白它们是用来做什么的 这些数据不会留在 Rockstar。也不会留在 Take-Two。Lens 通过内部 API 将行为画像传送给 Take-Two 旗下一家名为 Vantage Analytics 的子公司。Vantage 没有公开网站,没有新闻页面,没有招聘信息。它于2023年在特拉华州注册成立 Vantage 是一家行为 AI 公司。它用玩家画像训练预测模型,并把这些模型授权给商业客户。管道文档中包含一份不完整的客户清单。两家保险集团,一家信用风险建模公司,一个人力资源筛选平台,以及一家广告数据经纪商 你的 GTA 6 游戏过程此刻正在生成一份行为评分。没人追你的时候,你开得有多快。你会不会在红灯前停下。NPC 无端攻击你时,你怎么反应。你行动之前要想多久。你花游戏内的钱有多冲动。每一个选择都在喂养一个模型,企业正在付费查询这个模型,用来对真实的人做决定
3,452
28,246
3,458
160,597
7,651,312
Ahnaf Shahriar retweeted
GTA 6 looks fun and everything, but the Li Auto L9 BUILT-IN DRIVING SIM is also insane
329
711
92
12,511
692,099
Can an ESP32-S3 run AI models locally without sending data to the cloud? That's the idea behind esp32-s3-edge-ai from Simantic Dev. The project explores running edge AI on the ESP32-S3, a dual-core microcontroller clocked at up to 240 MHz, with vector instructions for accelerating neural-network operations. The interesting challenge is fitting inference into limited RAM and flash while keeping latency and power consumption under control. For anyone working with embedded AI, this is a useful direction to explore. Think about local sensor classification, smart devices and always-on detection without depending on a cloud connection. I would be particularly interested in seeing how far INT8 quantisation, PSRAM and the ESP32-S3 vector instructions can push on-device inference. Would you try running your own trained model on an ESP32-S3?
ESP32-S3 Facial recognition running in Simulation! Edge AI and IOT is on the rise and we are here to help it get there. This is a Virtual ESP32-s3 running real firmware connected to the video feed of my webcam via serial. Of course it's insanely laggy but thats mostly because demoing using an inefficient path, you would be using the esp32s3's camera interface directly. But even then you would probably get 3 FPS, bargain deal when you know that an s3's cost like 6 bucks. Long time fan of the ESP32 series and so excited to keep sharing with everyone all the things you can do in simulation! It's gonna be a week of launches! But if you want to beat me to making cool things run in simulation, try it -> simantic.com
6
7
79
3,401
Ahnaf Shahriar retweeted
Today we released Pyrite 0.10.0. Pyrite is the most accurate and complete firmware simulation model for Arm Cortex-M/R and RISC-V based platforms, made by @simantic. Pyrite is out of stealth, and the usage documentation is available at simantic.com/docs
6
4
1
27
1,487
Ahnaf Shahriar retweeted
In the last two months, @simantic has processed over 1.5 million simulation runs equating to over 3000 hours of compute time, with an average growth rate of 70% per week. Ever since we joined @ycombinator 8 weeks ago, we've been working on creating the most accurate simulation models for a wide range of MCUs and SoCs, to close the gap in fully autonomous hardware and firmware development. Now we run in the CI pipelines, benchmarks, and agentic verification loops for engineering teams building satellites, robotics, and consumer electronics. If your firmware team needs a better CI and verification loops, DM me or email us at [email protected], or go to simantic.com
5
2
1
27
1,183
Some companies use your personal data, like browsing history, to serve different prices to different costumers. It’s called surveillance pricing, and we just banned it in New York City. This tweet will say the same thing in incognito mode, don’t worry.
1,023
20,718
982
223,906
2,956,776
Ahnaf Shahriar retweeted
Trump dumping his Nvidia stock a few weeks before this Chinese spy story dropped is the least shocking thing I’ve ever heard in my entire life. Straight to jail.
2,011
11,878
294
88,557
1,832,158
Ahnaf Shahriar retweeted
35
6,509
92
52,187
663,693
Ahnaf Shahriar retweeted
Claude is now Khalid.
57
1,858
202
22,309
417,124
Ahnaf Shahriar retweeted
Last October, I started @simantic because I saw firsthand how embedded and firmware engineers were struggling to take advantage of the recent developments in frontier AI models. In software, AI models are able to generate the code, run testing in simulation, and deploy to production, end-to-end without human input other than the first prompt. In firmware and embedded software, AI can generate great code, but any validation work is left to the engineer to run manually, on physical devboards or HIL setups. We're building the simulation layer to solve that. We're giving AI the visibility into hardware that it needs to fully close the gap in fully autonomous hardware development, without the physical bottleneck. If you're part of a firmware or embedded team looking for a better and faster way to test your code in simulation, email us at [email protected]
Simantic is the simulation layer that lets engineers and AI agents test and validate firmware at the speed of software. Frontier AI models have made generating firmware code cheap and effortless, but testing still depends on manual hardware debugging, expensive equipment, and software designed more than a decade ago. Simantic is modernizing firmware testing with simulation built for the era of coding agents. w/ @ycombinator @shong0504 @enough_ahnaf
7
3
1
28
1,596
Ahnaf Shahriar retweeted
Today we’re launching @vorelios (YC F26)! We’re a research lab building AI that can engineer anything. Models that predict a design’s behavior in seconds and use those predictions to invent and improve designs autonomously. 🧵 w/ @krishkapadia , @ycombinator vorelios.com
31
16
6
154
12,735
Ahnaf Shahriar retweeted
We just launched! So hyped about this!
Simantic is the simulation layer that lets engineers and AI agents test and validate firmware at the speed of software. Frontier AI models have made generating firmware code cheap and effortless, but testing still depends on manual hardware debugging, expensive equipment, and software designed more than a decade ago. Simantic is modernizing firmware testing with simulation built for the era of coding agents. w/ @ycombinator @shong0504 @enough_ahnaf
9
3
1
47
1,847
Ahnaf Shahriar retweeted
Simantic is the simulation layer that lets engineers and AI agents test and validate firmware at the speed of software. Frontier AI models have made generating firmware code cheap and effortless, but testing still depends on manual hardware debugging, expensive equipment, and software designed more than a decade ago. Simantic is modernizing firmware testing with simulation built for the era of coding agents. w/ @ycombinator @shong0504 @enough_ahnaf
31
17
5
73
10,147
Ahnaf Shahriar retweeted
Why do Muslims treat interest-free finance like a limitation, when history proves it's a superpower? Around 1700, Mughal India produced roughly a quarter of the world's economy, about the share the US holds today. The Islamic gold dinar was so trusted that European kings minted copies of it, Arabic script and all. Muslim merchants financed trade from Spain to China through partnerships, profit-sharing, and trust. Riba-free finance wasn't a handicap those economies carried. It was the engine that built them. The real opportunity today isn't fixing a broken system. It's reigniting belief in one that already worked brilliantly. Nobody opens a shawarma shop thinking, "It's halal, so Muslims will show up." The food wins on its own merits: Better taste. Better price. Better experience. Halal is simply assumed. It's not the pitch, it's the foundation. So imagine applying that same confidence to halal finance. History already proved it can compete, not just comply. It can lead. Once we truly believe 0% interest finance is better, we'll build something extraordinary again.
35
104
8
679
90,317
Ahnaf Shahriar retweeted
We're launching Greptile for firmware. AI code reviewers are great for catching bugs in software, but when it comes to firmware, they miss critical bugs in ISR and peripheral timings, and runtime memory. We're launching Simantic Bot to fix that. We gave AI the ability to run firmware in simulation, and test scenarios like providing external outputs from sensors and wireless communication, all inside your Github PRs. Check it out at simantic.com/
15
25
8
313
20,982
Ahnaf Shahriar retweeted
sneak peek into how your packages will move in the not so distant future (yes, the airplane does not have a cockpit)
11
28
18
272
67,569
Ahnaf Shahriar retweeted
alexandr wang is the successful version of the cluely guy
258
225
47
11,002
418,498
Your ESP32 can run AI. The trick is making the math fast enough. Espressif's ESP-NN library provides optimized neural-network functions for chips like the ESP32-S3, ESP32-P4, ESP32-S31, ESP32 and ESP32-C3. On the ESP32-S3, optimized convolution kernels can be several times faster than plain C implementations. For one benchmark, person detection drops from 2.3 seconds to 54 ms with ESP-NN. It also supports common operations used by models such as MobileNet and YOLO11n, making on-device vision much more practical. How far do you think we can push AI inference on an ESP32?
11
52
1
513
18,121
Ahnaf Shahriar retweeted
Employee: 3k, 3k, 3k, 3k, 3k Freelancer: 900, 5k, 0, 9k, 1k, 4k Founder: 0, 0, 0, 0, 0, 0, 600k Open Source Maintainer: 0, 0, 0, 0, 0 "thanks bro" Vibe coder: -$20, -$100, -$200, -$200
293
1,427
185
28,066
1,169,765