Ray Shih

A duck

[NVIDIA @nvidia]

Taipei runs on innovation and great food. 🥟🍧

If you're at GTC Taipei at Computex, visiting the night markets is a non-negotiable.

#NVIDIAGTC

[luthira @luthiraabeykoon]

We implemented @karpathy 's MicroGPT fully on FPGA fabric.

No GPU.
No PyTorch.
No CPU inference loop.

Just a transformer burned into hardware, generating 50,000+ tokens/sec.

The model is small, but the idea is not: inference does not have to live only in software 👇

[Zain Shah @zan2434]

Imagine every pixel on your screen, streamed live directly from a model. No HTML, no layout engine, no code. Just exactly what you want to see.

@eddiejiao_obj, @drewocarr and I built a prototype to see how this could actually work, and set out to make it real. We're calling it Flipbook. (1/5)

[Waymo @Waymo]

London, we’re taking the next step! 🚙 We’re officially beginning autonomous driving with a trained specialist behind the wheel. We can’t wait to offer Londoners a quiet, convenient, and magical way to connect to the Tube, bus, or their final destination later this year.

[Ray Shih @existential_ray]

@lovecankill Why?

[Andrej Karpathy @karpathy]

Wow, this tweet went very viral!

I wanted share a possibly slightly improved version of the tweet in an "idea file". The idea of the idea file is that in this era of LLM agents, there is less of a point/need of sharing the specific code/app, you just share the idea, then the other person's agent customizes & builds it for your specific needs.

So here's the idea in a gist format: https://t.co/NlAfEJjtJV

You can give this to your agent and it can build you your own LLM wiki and guide you on how to use it etc. It's intentionally kept a little bit abstract/vague because there are so many directions to take this in. And ofc, people can adjust the idea or contribute their own in the Discussion which is cool.

[fox hsiao @pirrer]

這是我看到第二個人把 AI 時代比成《咒術迴戰》的，茲翻譯之XD

--
前幾天，動畫第三季剛播完，我在重看《咒術迴戰》的時候，突然想到一件事。

只要用 Claude Code，不會寫程式也能做出 App。對 Codex 下指令，外行人也能讓一個像樣的產品跑起來。這種「任何人都能做出東西」的狀況——

這不就是死滅迴遊嗎。

我不是在開玩笑，從結構上來看非常吻合，今天想聊聊這件事。讀者裡在這個產業的人可能早就發現了，但看咒術迴戰時覺得死滅迴遊莫名其妙而棄坑的人應該也不少，所以我會盡量寫得讓大家都能理解。

先說「死滅迴遊」是什麼

《咒術迴戰》漫畫裡有一個叫做死滅迴遊（Death Game）的章節。

粗略地說，故事是這樣的：一個活了超過一千年的危險角色（羂索），強制改寫全日本一般人的大腦，讓他們覺醒出「術式」——也就是各自獨有的超能力——然後把他們丟進生存遊戲裡。

重點是「術式」而不是「咒力」。咒力每個人都有一點點，但術式不一樣，那是能發動特定能力的「固有技能」。羂索用真人的「無為轉變」把非術師的大腦結構直接重塑成術師的樣子，強制讓術式覺醒。

昨天還在超商打工的人，某天早上醒來就能使用固有的超能力。可是規則跟戰鬥方式都不懂，卻被丟到跟身經百戰的術師同一個戰場上。

聽起來很不講理，但在 2026 年的現在，不是正在發生幾乎一模一樣的事嗎？

相似點①：對一般人強制賦予術式

死滅迴遊的起點，是羂索改寫一般人的大腦，讓他們覺醒術式。當事人願不願意不重要，某天突然就變成能使用特定能力的狀態。而且這個術式是什麼、要怎麼發動，本人完全沒被告知。

AI 程式代理完全是同一個結構。Claude Code 和 Codex 出現之後，就算沒寫過程式，「做出產品的能力」也突然落到手上。這種能力以前只屬於工程師。

而且跟羂索做的一樣，被賦予這個能力的人沒有選擇權。就算你不用，隔壁的人會用，競爭對手會用，整個市場會在「所有人都持有術式」的前提下開始運作。實質上就是強制覺醒。

相似點②：光是覺醒還贏不了

死滅迴遊真正殘酷的地方，不是「賦予了術式」，而是光靠覺醒完全無法獲勝。

就算得到術式，發動條件不知道、射程也不知道，甚至連自己的術式是什麼都搞不清楚。另一方面，跟羂索簽約、從一千年前復活的身經百戰術師們也在同一場遊戲裡。剛覺醒的外行人跟擁有幾百年戰鬥經驗的術師之間，就算同樣「持有術式」，差距也是天壤之別。

AI 時代完全就是這樣。

Claude Code 確實可以做出「能動的東西」。但是使用者體驗的設計、架構的判斷、甚至「到底該做什麼」這個問題——這些都不是工具能代勞的。「會動的原型」跟「別人想用的產品」之間，隔著一道像死滅迴遊結界那麼厚的牆。

換句話說，覺醒術式，跟能用術式戰鬥，完全是兩回事。這是我覺得最像的一點。

相似點③：遊戲主辦者對個人勝負沒興趣

羂索發動死滅迴遊的目的，不是享受玩家的勝負。他真正的目的是讓天元和人類同化，讓整個人類「進化」。死滅迴遊只是為了這個目的而進行的大規模熱身。讓泳者們互相殘殺累積咒力，活化結界，為同化創造條件。個別參加者的生死，只是過程中的一環。

這個結構，跟 AI 平台商的樣子不是一模一樣嗎？

不管是 Anthropic 還是 OpenAI，對我們個別做出什麼東西，老實說應該都無所謂。他們真正在看的，是人類怎麼使用 AI、怎麼適應、會在哪裡卡關——這整體的行為模式。我們在 Claude Code 上苦苦掙扎的每一個片段，都成為加速 AI 進化的資料點。

你可能會想「可是他們有提供工具啊，這樣不是很好嗎？」當然很好，羂索也確實把術式好好地發下去了。只是，發放的一方跟使用的一方，看到的風景完全不同——這件事最好心裡有數。

相似點④：規則中途會改

在死滅迴遊裡，玩家可以消耗 100 點來新增遊戲規則。日車實際上就追加了「泳者之間可以互相交換點數」的規則，直接改變了遊戲結構。遊戲進行到一半，遊戲規則自己就被改寫了。

AI 時代比這更極端。

新模型每個月都在出，API 規格一直變，昨天還能用的方法今天就不能用，上個月的最佳實踐到這個月就過時。羂索每週都在改規則的狀態，已經變成日常。

那，差別在哪裡

講了這麼多「相似」的部分，其實「不同」的地方才有希望。

最大的差別是，死滅迴遊是零和遊戲，AI 時代不是。

死滅迴遊裡，想得分只能幹掉其他玩家，一個人的勝利等於另一個人的死亡。但 AI 時代因為全員都拿到了術式，市場本身正在膨脹。

以前因為「開發成本不划算」而被擱置的小問題、利基太窄沒人願意做成產品的不便，這些突然都變成個人也能解決的東西。戰場變多了。所以不是「被既有的強者獵殺」，而是誰先找到還沒人碰過的戰場，誰就贏。

這是個巨大的差異。死滅迴遊是「強度」的遊戲，但 AI 時代是「眼光」的遊戲。

另一個差別：不會死（大概）

死滅迴遊裡，19 天內沒得分就會被剝奪術式——也就是會死。當然啦，用 Claude Code 做出蹩腳的 App 是不會死的。

我不是在開玩笑，失敗成本趨近於零這件事，從結構上來看是非常重要的差異。在死滅迴遊的世界裡謹慎是美德，但在 AI 時代，上場的次數才是一切。

做 100 個只要中 1 個就夠，而且做的成本幾乎是零。這個環境要是被死滅迴遊的玩家聽到，會哭著羨慕。

怎麼在這場「死滅迴遊」活下來

好，接下來才是重點。在這個全員都拿到術式的世界裡，要怎麼活下來？從咒術迴戰角色們的行動，可以很認真地提煉出線索。

了解自己的術式

死滅迴遊一開始要做的事，就是掌握自己術式的發動條件跟性質。日車寬見迅速理解了自己「審判」術式的機制，並發揮到極致，打出了超過 100 點的成績。高羽史彥則是到中途都還沒發現自己的術式是「把自己深信有趣的事情變成現實」這種破格能力。

AI 時代也完全一樣。

「用 AI 什麼都能做」是謊言。更精確地說，用 AI 什麼都能做出「能動的東西」，但要做出有價值的東西，需要你自己獨特的脈絡。你的產業知識、日常感受到的不滿、特殊的領域經驗——那些就是你的「術式」。AI 只是把這些轉換成實作的工具。

先搞清楚「自己的術式是什麼」，這是第一步。

選擇戰鬥的場所

就像前面寫的，這不是零和遊戲，沒必要在有 GAFA 的紅海正面對決。套用死滅迴遊的講法，避開強力泳者聚集的結界，到人少的殖民地穩穩得分，這個策略比較聰明。

你的產業、你的現場、你身邊五公尺內的問題。那裡就是巨人們不會出手的「空的殖民地」。

鍛鍊「適應力」

在規則每週都在變的世界裡，對特定工具很熟練這件事本身沒有價值。你把今天的 Claude Code 用法練到頂，三個月後可能一切都變了。

死滅迴遊裡能活下來的角色，共通點不是術式特別強，而是狀況改變時還能應對的底子。AI 時代的「底子」，是發現問題的能力、把問題語言化的能力、對變化保持樂趣的心態，跟你對工具的熟練度無關。

不要成為實驗的棋子

最後，最重要的一點。

咒術迴戰裡，乙骨跟日車了不起的地方，是他們雖然參加死滅迴遊，卻沒有變成羂索實驗的棋子。他們在理解遊戲結構的基礎上，為了自己的目的利用這個遊戲。日車透過追加規則替同伴開闢了脫逃路線，乙骨把得分轉讓給隊友組織整體戰略。對於羂索預期的「透過互相殘殺累積咒力」這個劇本，他們沒有上鉤。

AI 時代也一樣。平台商發下來的工具就用，在他們打造的生態系裡運作，這都沒問題。但是，要做什麼、要往哪裡去這件事，只能由自己決定。

在全員都拿到術式的世界裡，能不能繼續玩「自己的遊戲」。我覺得，這是在這個時代的死滅迴遊活下去的唯一方法。

---

《咒術迴戰》完結了，但我們這邊的死滅迴遊，連總則都還沒集齊。遊戲還在繼續。你的術式是什麼？大家有沒有察覺到呢？

在這個 AI 時代的死滅迴遊裡，我很想像五條老師那樣，至少講一次——

「放心吧，我最強。」

[Elon Musk @elonmusk]

Next flight of Starship and first flight of V3 ship & booster is 4 to 6 weeks away

[Andrej Karpathy @karpathy]

LLM Knowledge Bases

Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So:

Data ingest:
I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them.

IDE:
I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides).

Q&A:
Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale.

Output:
Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base.

Linting:
I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into.

Extra tools:
I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries.

Further explorations:
As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows.

TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.

[Claude @claudeai]

Computer use is now in Claude Code.

Claude can open your apps, click through your UI, and test what it built, right from the CLI.

Now in research preview on Pro and Max plans.

[Dave Morin 🦞 @davemorin]

ClawCon rises in Tokyo. March 30th.
@steipete and I will be there.

🦞🗼

https://t.co/EZo0vZP2YR

We already need a bigger venue.

[Ray Shih @existential_ray]

I tried the same thing in past two weeks :p

[Peter Steinberger 🦞 @steipete]

Testing a new feature for Microsoft Foundry support in @openclaw. Their website is a jungle, I used to make screenshots so codex can guide me through it, but now Chrome has an MCP so codex can simply connect and drive my browser session and do all of that for me.

The human is no longer the bottleneck.

[Andrej Karpathy @karpathy]

Software horror: litellm PyPI supply chain attack.

Simple `pip install litellm` was enough to exfiltrate SSH keys, AWS/GCP/Azure creds, Kubernetes configs, git credentials, env vars (all your API keys), shell history, crypto wallets, SSL private keys, CI/CD secrets, database passwords.

LiteLLM itself has 97 million downloads per month which is already terrible, but much worse, the contagion spreads to any project that depends on litellm. For example, if you did `pip install dspy` (which depended on litellm>=1.64.0), you'd also be pwnd. Same for any other large project that depended on litellm.

Afaict the poisoned version was up for only less than ~1 hour. The attack had a bug which led to its discovery - Callum McMahon was using an MCP plugin inside Cursor that pulled in litellm as a transitive dependency. When litellm 1.82.8 installed, their machine ran out of RAM and crashed. So if the attacker didn't vibe code this attack it could have been undetected for many days or weeks.

Supply chain attacks like this are basically the scariest thing imaginable in modern software. Every time you install any depedency you could be pulling in a poisoned package anywhere deep inside its entire depedency tree. This is especially risky with large projects that might have lots and lots of dependencies. The credentials that do get stolen in each attack can then be used to take over more accounts and compromise more packages.

Classical software engineering would have you believe that dependencies are good (we're building pyramids from bricks), but imo this has to be re-evaluated, and it's why I've been so growingly averse to them, preferring to use LLMs to "yoink" functionality when it's simple enough and possible.

[sam @SamuelBeek]

Nobody:
Absolutely nobody:

Hardware people: lets get doom running on a pregnancy test

[Andrej Karpathy @karpathy]

Three days ago I left autoresearch tuning nanochat for ~2 days on depth=12 model. It found ~20 changes that improved the validation loss. I tested these changes yesterday and all of them were additive and transferred to larger (depth=24) models. Stacking up all of these changes, today I measured that the leaderboard's "Time to GPT-2" drops from 2.02 hours to 1.80 hours (~11% improvement), this will be the new leaderboard entry. So yes, these are real improvements and they make an actual difference. I am mildly surprised that my very first naive attempt already worked this well on top of what I thought was already a fairly manually well-tuned project.

This is a first for me because I am very used to doing the iterative optimization of neural network training manually. You come up with ideas, you implement them, you check if they work (better validation loss), you come up with new ideas based on that, you read some papers for inspiration, etc etc. This is the bread and butter of what I do daily for 2 decades. Seeing the agent do this entire workflow end-to-end and all by itself as it worked through approx. 700 changes autonomously is wild. It really looked at the sequence of results of experiments and used that to plan the next ones. It's not novel, ground-breaking "research" (yet), but all the adjustments are "real", I didn't find them manually previously, and they stack up and actually improved nanochat. Among the bigger things e.g.:

- It noticed an oversight that my parameterless QKnorm didn't have a scaler multiplier attached, so my attention was too diffuse. The agent found multipliers to sharpen it, pointing to future work.
- It found that the Value Embeddings really like regularization and I wasn't applying any (oops).
- It found that my banded attention was too conservative (i forgot to tune it).
- It found that AdamW betas were all messed up.
- It tuned the weight decay schedule.
- It tuned the network initialization.

This is on top of all the tuning I've already done over a good amount of time. The exact commit is here, from this "round 1" of autoresearch. I am going to kick off "round 2", and in parallel I am looking at how multiple agents can collaborate to unlock parallelism.
https://t.co/WAz8aIztKT

All LLM frontier labs will do this. It's the final boss battle. It's a lot more complex at scale of course - you don't just have a single train. py file to tune. But doing it is "just engineering" and it's going to work. You spin up a swarm of agents, you have them collaborate to tune smaller models, you promote the most promising ideas to increasingly larger scales, and humans (optionally) contribute on the edges.

And more generally, *any* metric you care about that is reasonably efficient to evaluate (or that has more efficient proxy metrics such as training a smaller network) can be autoresearched by an agent swarm. It's worth thinking about whether your problem falls into this bucket too.

[Ray Shih @existential_ray]

So in theory, pantheon is achievable

[Dr. Alex Wissner-Gross @alexwg]

https://t.co/N2AwpI9f8m

[Tesla @Tesla]

First Cybercab off the production line at Giga Texas

[Andrej Karpathy @karpathy]

I think it must be a very interesting time to be in programming languages and formal methods because LLMs change the whole constraints landscape of software completely. Hints of this can already be seen, e.g. in the rising momentum behind porting C to Rust or the growing interest in upgrading legacy code bases in COBOL or etc. In particular, LLMs are *especially* good at translation compared to de-novo generation because 1) the original code base acts as a kind of highly detailed prompt, and 2) as a reference to write concrete tests with respect to. That said, even Rust is nowhere near optimal for LLMs as a target language. What kind of language is optimal? What concessions (if any) are still carved out for humans? Incredibly interesting new questions and opportunities. It feels likely that we'll end up re-writing large fractions of all software ever written many times over.

[vittorio @IterIntellectus]

it’s happening

[Anthropic @AnthropicAI]

New Engineering blog: We tasked Opus 4.6 using agent teams to build a C compiler. Then we (mostly) walked away. Two weeks later, it worked on the Linux kernel.

Here's what it taught us about the future of autonomous software development.

Read more: https://t.co/htX0wl4wIf