mjboy.eth @themjboy

Good interview

x.com/i/article/2065…

x.com/i/article/2057…

LLMs explained without all that yucky math stuff 0xkato.xyz/how-llms-actua…

new grads often ask me what they should be doing so they don't fall behind in the ai space. there's a lot, but its honestly super manageable. become intimate with model internals. proof based linear algebra. non-convex optimization. this is stuff you could've done in undergrad. it definitely takes some time and work, but its doable. have taste, have opinions. train a small model, then train a big one. vLLM internals, tensor parallelism. hand roll kernels. cluster orchestration. do you have opinions on synthetic data? why don't you? SFT, PPO, you should know this. learn Triton. everyone is reproducing papers now so you need to be doing more. do you know the semi supply chain? where are the bottlenecks? hardware, man, hardware. your little gpu rig erector set in your basement isnt gonna cut it. build a cluster, a big one. pretrain a 800B model. now postrain it. serve it to millions of people. you should be able to beat deepseek on some benchmarks now. its a lot to take in but it all snowballs. this what job security looks like from now on. do you want to work in tech or not

"You can outsource thinking, but not understanding." I still find writing toy code one of the best ways to build real understanding. It catches the nuances that skimming code and explanations lets you skip. So I wrote nanoRL (nanoGPT, but for post-training). SFT, DPO, GRPO, PPO: four single files, ~150 lines each, converging on a toy task in ~30 steps on a MacBook. Readable end-to-end. Then I continue RL Qwen2.5-0.5B-Instruct on GSM8K with this toy code + autoresearch. Interestingly, the accuracy improves tho it's a trained model.

if you're doing AI research at all; I recommend doing the "ETH zurich" route Train models that use a single GPU. Make sure that it takes less than a minute to train models. Pufferlib is a great example. The more models you train the more you learn

Super PINTO @PINTO03091

3 weeks ago

BiternionNet、１分で学習が終わってしまったんだが。

4 14 361 244K 200

Anthropic onboarding day: Michael Scott introducing Karpathy like he just signed Wemby in free agency.

We were able to sit down with the @GoogleDeepmind team behind the new Gemini Omni Flash model to hear all of their behind-the-scenes stories, memorable moments, and many, many (occasionally embarrassing) video generations. Watch the full Release Notes episode here: goo.gle/49adTFp

Best structured reference I've found for GPU optimization - 450 papers, 14 years of research. Some techniques will have evolved, but the mental models hold up. dl.acm.org/doi/10.1145/35…

Vivek Galatage @vivekgalatage

a month ago

Yesterday's lecture on GPU Architectures by @_onurmutlu_ youtu.be/aE0onltJlOo

1 78 551 82K 522

Spent Friday night testing @Apple ‘s new open-source ml-lito image-to-3D model running locally on Apple Silicon via Metal. No cloud, API, or uploads. Getting it working was harder than expected but got there. 👇👇👇@AppleEDU github.com/apple/ml-lito

x.com/i/article/2057…

Gemini Omni is a major leap in world understanding & multimodal editing! It can take photos, video & audio and build entirely new scenes. Over time it’ll be able to handle any input & any output - starting w/ video You can even give it your own videos & iterate on your ideas:

Check out Ariticraft 🦾 - a highly efficient agentic system that generates articulated 3D assets fully automatically at scale! 🚀 articraft3d.github.io

RuiningLi @RayLi234

a month ago

🚀 Introducing Articraft, a coding agent for articulated 3D asset creation. Articraft writes code, executes it, receives validation feedback, and refines the result into simulation-ready 3D assets with parts, joints, and motion. We’re also releasing Articraft-10K: 10,000+

22 108 747 186K 788

Unity AI is now in Open Beta 🎉💫 We believe AI has the most impact when it helps creators move faster while staying in control of the creative process. Use our built-in agent tuned for Unity workflows or connect the AI tools you prefer via AI Gateway and MCP Server.

Create custom logo animations with ChatGPT Images 2.0 and Seedance 2.0. Images 2.0 can create an animation sheet for how a logo should appear. This can be quite detailed, to help the AI video tool guide how long each section takes. More in the comments:

Jensen Huang when asked whether he would rather relive his 20s or be 20 years old today.

卧槽！华尔街传奇巨鳄德鲁肯米勒彻底破防，首次曝光自己竟然深陷自我怀疑长达15年！ 昨晚，这位宏观大师在受访时，无情地揭开了自己血淋淋的交易伤疤。他坦言在1999年纳斯达克科技股狂飙时，自己明明在一月份完美逃顶，随后却因为FOMO情绪彻底失控，竟然在绝对的最高点疯狂接盘！ 当被问及从那次史诗级错判中学到了什么时，他的回答异常扎心：“什么都没学到！”因为这些交易大忌他二十年前就懂，但在真金白银的诱惑面前，情绪依然会瞬间接管大脑。 更让人震撼的是，他承认自己每天都在跟贪婪和恐惧贴身肉搏。曾经在面临账户大幅回撤时，那股摧毁神经的焦虑感，甚至让他每周都要控制不住地狂吐一两次！ 但他最终迎来了终极顿悟：不管多牛的操盘手，犯错和情绪崩溃绝对是永远无法摆脱的宿命。既然已经用长期的傲人战绩证明了自己绝非凭空靠运气，就根本没必要在亏损后无休止地精神内耗。 他向所有交易员发出灵魂呐喊：几百次的失误不过是漫长生涯里的短暂瞬间。顶级高手唯一要做的，就是咽下血泪迅速翻篇，然后毫不犹豫地拔刀杀向下一个战场！

比特币橙子Trader @oragnes

2 months ago

卧槽！传统银行业底牌彻底曝光，宏观巨头Raoul Pal暴论整个金融系统必将全面接入以太坊！ 昨晚，这位华尔街老兵无情嘲讽了前两年全网高呼以太坊已死的言论。他凭借多年混迹金融核心圈的经验，一针见血地指出，银行这种庞然大物在做底层技术选择时，保住饭碗永远是第一法则。

2 7 34 187K 22

工程师，尤其是Agent工程师从Claude Managed Agents能学到的五条最深刻经验： 1. 永远别把“当前模型的弱点”写死在基础设施里 这是最容易翻车的点。很多人还在给Agent加一堆retry、guardrail、context engineering，结果新模型一出来，这些全成了负优化。正确的做法是把假设放在“可替换的harness”层，而非核心接口。 2. 状态要外置，Session才是真正的Context Object 不要把所有历史全塞prompt，而应该把事件日志做成可查询、可切片的外部对象，harness只负责“在当前需要的时候”做transform（压缩、总结、rewind）。这既解决长时序问题，又为未来更聪明的context engineering留空间。 3. 用极简、稳定的接口做“元抽象” Brain-Harness-Session-Sandbox之间只定义极少的接口（execute、getEvents、emitEvent、provision等）。接口越薄，越能扛住模型/工具的快速迭代，这就是“为尚未想到的程序做设计”。 4. 可靠性 = Decoupling + 按需Provisioning 别搞大而全的单体容器。失败了就让Brain通过tool call重新provision hands。无状态harness启动极快，扩展性爆炸。 5. 把Agent基础设施当成“操作系统”来设计，而不是一次性脚本 真正的长期胜负手不是今天能跑多复杂的task，而是半年后、一年后模型再强10倍时，你的系统还能不能自然受益，而不用重写一半代码。

Lance Martin @RLanceMartin

2 months ago

i co-wrote the Anthropic engineering blog on Claude Managed Agents, and wanted to share some thoughts on agent harnesses + infrastructure for long-horizon tasks ... 🧵 anthropic.com/engineering/ma…

33 116 999 95K 1K

Claude for Word is now in beta. Draft, edit, and revise documents directly from the sidebar. Claude preserves your formatting, and edits appear as tracked changes. Available on Team and Enterprise plans.