Kevin Z

@[email protected]

[sleepy.md @sleepy0x13]

这是我写过最重的一篇文章之一。

它从一本被切掉书脊、送进 Claude 的旧书开始，最后写到硅谷试图接管人类文明的解释权。

为了写清楚这件事，我重新梳理了硅谷科技右翼几十年来的思想演变。

技术为什么从一种工具变成信仰，增长为什么被赋予道德，��速为什么被写成宇宙规律，人类又为什么逐渐从未来的主角，降级成智能进化过程中的临时载体。

Nick Land、Andreessen、e/acc、Peter Thiel、Curtis Yarvin，这些名字看起来离普通人很远。但他们的思想已经进入模型的训练数据、系统规则和「宪法」，决定 AI 能说什么、怎样理解世界，又会把怎样的历史和价值观讲给下一代。

古代焚书，是让知识消失。

AI 时代更高效。它保留全部答案，却让来源、语境和争论过程逐渐消失，最后由少数公司的模型替人类决定如何记住文明。

[Team Canada @TeamCanada]

HISTORY! The Canadians’ 1-0 win at Los Angeles Stadium in the Round of 32 match means they have advanced to a Round of 16 clash with either Morocco or the Netherlands.❤️🇨🇦 https://t.co/ukhYYbolw2

[Kevin Z @SailKevin]

@Motor_Sport

[Eric Xu (e/Mettā) @xleaps]

https://t.co/inXqWdQSNy

[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.

[Turkic Archives @TurkishArc]

Meanwhile in Istanbul

[Javi @rameerez]

I made this product launch video over the weekend with just prompts

It's all vibe coded

https://t.co/V9i8Z6N8ow

There's something you should know, though:

Like everyone else, a few days ago my timeline started getting full of videos like this when Remotion launched their Claude skill, so I decided to give it a go

I was captivated by all the examples, so I started like everyone was saying: "just write a prompt"

I typed the prompt, and it created an extremely bland, untasteful, stock-looking video

10 prompts in and it was not getting better. It was very, very bland. But at least it was something, so I kept going at it

I ended up spending my entire weekend on this, 2-3 days of work. Only to realize my original reference videos that inspired me to get started were all fake

Everyone was outright lying about their results. They all claimed "I made this with just one prompt", but it was just bait, they didn't really use Remotion or code at all, it was just a normal, human-made motion video

Then you expand the X post and read the replies and they're all like "haha joke" in the comments, but their main post already got 1.5 million views and bamboozled everyone who didn't read further

And this is a problem: when a viral trend happens, these posts flood your timeline, and you only realize that they're all noise and bait (and that they haven't even used the tools they claim) when you click through the post and read its comments. But 90% of people (like me, initially) just see the post on their timeline while scrolling, and assume it's all real. You don't go in to check every single post you see: you just like it, or save it for later, and carry on with your day, thinking what you saw was the real thing, and that it's all outstanding results, and that motion designers are really done

And it's so anxiety inducing, because everyone is hyping their results, but most of it is just not true. I have stopped reading X lately because going in makes me so anxious, everyone is claiming extraordinary outlier results just for the views and clicks, and you feel like you're lagging behind and you're not good enough because you don't get those results

So for this video I decided to actually take the tech out for a spin, and see what results I could really get out of it

I used Remotion and Claude Code 4.5, but contrary to what everyone was claiming, this video was not "just a prompt". It was fully vibe coded, but it required much more than a prompt. It was multiple days worth of work

Here's what I learned:

- Making vibe coded videos with Remotion is ~10-20x slower than building app code. I've been wasting my Claude limits on this video

- Everything takes a lot of manual work and reprompting. You often need to go frame by frame correcting tiny things

- It makes very silly mistakes

- Even Opus 4.5 has very very limited knowledge of spatial / visual things. It doesn't understand well z-indexes, layers, compositions, proportions, temporal coherence, etc. Claude Code feels extremely dumb when creating code for Remotion videos, which surprised me a lot, beacuse I had been mind blown by how incredibly well it worked with my Ruby on Rails SaaS codebases

- You need to have some design knowledge to adjust things manually, you need to ask for exactly what you want, in the technical jargon it expects. You can't just say "make this more beautiful" or "animate this better" because it just creates slop

- Right now vibe coded videos are promising, but I think I could have done this video faster just by doing it manually in After Effects. It really took that much work

- If you have a creative idea for something you want to animate, it takes multiple hours of back and forth prompting to create just one or two seconds worth of **good** animation

- Tip: PARAMETERIZE everything! It tends to hardcode magic numbers everywhere in the code, so if you change something earlier in the video timeline, everything else breaks. You want to essentially be creating "key frames" with code by telling it to parameterize every frame where something important happens, and calculate the rest of the keyframes based off that. This comes in handy when you need, for example, to adjust keyframes to match the music

So in summary: vibe coded videos are promising, but right now it only works for very stock-looking videos unless you put in a ton of effort

Maybe actually useful for 1-2 second web animations though, I'll try that next

It will obviously get better, this feels like the quality of code generation in 2023-2024, you need to hold its hand and correct it at every step along the way. But even if video code generation was better, you would still need someone with motion design knowledge to at least set the creative direction, lay out the overall script and composition, etc. It's not completely hands-off unless you want slop

And a word on caution: especially here on X, there's 90% hype and 10% reality, nothing is what it seems. Do not believe what you see online, people are constantly baiting and then just laughing it off in the comments

[Kevin Z @SailKevin]

@JamesGunn

[汗青 HQ @hq4ai]

AI原生MV系列第二弹：SURREAL
相比3个月前，AI的进步让我们能比「白色皮卡丘」完成更多的效果。也使用了几乎市面上所有的AI模型来制作，它们每个人都在承担最擅长的部分。
真的感谢这些工具，让个人创作者生活在一个黄金年代。

[Hashem Al-Ghaili @HashemGhaili]

I did more tests with Google's #Veo3. Imagine if AI characters became aware they were living in a simulation!

[Mark Teasdale ★ @MarksGonePublic]

Today's December Fog in Vancouver: 5 Hours of This Fascinating Weather Compressed into 2 Minutes. Sound on / High res #Vancouver #Fog #StanleyPark #Downtown

[hope hopes hoping @hopes_revenge]

bye

[False Creek Ferries @FalseCreekFerry]

We had some special visitors in the creek today.

[Formula 1 @F1]

MAX VERSTAPPEN IS THE 2024 F1 WORLD CHAMPION! 🏆

A brilliant fourth title in-a-row! Incredible!!!

#F1 #LasVegasGP

[Kevin Z @SailKevin]

@2021CANyo @lovesickdoe 🥰

[Kevin Z @SailKevin]

@MaoDie8964 恭喜！🥳

[Alex Wiltschko @awiltschko]

Well, we actually did it. We digitized scent. A fresh summer plum was the first fruit and scent to be fully digitized and reprinted with no human intervention. It smells great.

Holy moly, I’m still processing the magnitude of what we’ve done. And yet, it feels like as we cross this finish line we are instantly at a new starting line. I’ll have more to share about what’s in store that we’re building on top of this.

A huge HUGE congrats to the entire team across scientific, engineering, operational, and creative disciplines. It takes a village named Osmo to do this.

I don’t know if this is embarrassing, but I carry the plum scent with me a lot of places and smell it constantly. It makes me smile.

I’m curious, if y’all want to smell it? If we made a limited release fragrance of the first teleported scent and dedicated the proceeds to science, would you want it?

[Kevin Z @SailKevin]

@2021CANyo 又让你用坏一台电脑吗？😂

[Jamon @jamonholmgren]

Me reviewing your pull request

[Kevin Z @SailKevin]

@katelyntiann 似乎真的可以应付。上��去感觉他们的出餐效率变高了，非常厉害。😂