wayne

[Noah Smith 🐇🇺🇸🇺🇦🇹🇼 @Noahpinion]

Maroon cars are the best. We need maroon cars back.

[Daniel Jeffries @Dan_Jeffries1]

I think I finally figured out why OpenClaw is amazing and took off like wild fire and why Peter is a genius, as Altman called him.

And it's actually a different way of looking at it.

It's not a DeepSeek moment for agents.

It's a Napster moment.

And just like Napster it will eventually force the industry to change. In essence when Napster came out the entire world told the music industry we don't want to buy CDs anymore and if you don't provide us a digital download experience we are just going to take it until you do.

It forced the industry to create Apple Music and eventually Spotify. Both essentially killed most music piracy by making it ubiquitous and cheap and good.

But it forced change.

The same will now happen to software. Here's why:

In essence OpenClaw lets you take what vendors don't want to give you: Unified access to countless applications.

We all want a personal assistant that can talk to freaking everything and do anything for us in the digital world.

But vendors don't want this. They want you locked into their bullshit.

For example, none of the messaging platforms want bots on there. None. They all have explicit policies against them and make it hard to do this. WhatsApp doesn't want you on there. Signal. Telegram's bot father is garbage. It's all designed to keep bots out.

They were designed for a pre-agentic era when bot = spam.

Many other things are like this. The API layers are gated, hoop-jumping bullshit. Go get an enterprise account and wait for approval and yada yada. Want access to WhatsApp? Get a business account and attach a number (what small business has a real number anymore 😂) and messages can't come from a person, etc. Google ads? It's not just an auth, it's go get a special manager account and create an enterprise key and blah blah blah.

It's a horrible experience because it was all designed for corporations to control access.

Now people are saying, make your app easy to access and accessible to me and my machine avatars and do it in a headless way or you will be dead.

Peter hacked around all this by making everything command line in the classic Linux style and using
things like an open source library that reverse engineered the web version of WhatsApp. It's all a bit house-of-cards-y because he had no choice.

At my company we had a similar idea early (and failed). Basically we wanted to make the best multimodal/computer using model because then it doesn't need an API or access hoops. You just go through the human interface layer and ain't nobody going to stop you. We failed because we weren't big enough and it's really a job for the mega-labs to solve because it is a hard problem and costs a shit ton of money.

Peter was much smarter. Make it all command line because that is ready now. Use any reverse engineered library or project or proxy available come Hell or high water and make it work by any means necessary even if it is hacky.

In short, he signaled to the software world that they better change and change fast or we are going to do this anyway and you can't stop us.

Of course some are foolishly trying. Meta is banning Claws on WhatsApp, etc.

They will all try to build their own gated, controlled, enshittified version of this thing.

They will fail.

And eventually everyone will offer a clear, easy way to get access via API for agents or they will be gone.

In essence OpenClaw gave people what they wanted, which was an app connected to everything, even when most of the vendors don't want you to have this.

[wayne @wceder3]

Diving into Karpathy’s Autoresearch, by @MrPrompts https://t.co/gIZxQDV4R5

[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.

[Anish Moonka @anishmoonka]

Marc Andreessen just dropped ~105 mins on Lenny's Podcast covering AI, jobs, careers, and why everyone is panicking about the wrong thing.

Just the clearest macro framework I've heard on where AI actually lands.

My notes:

𝟭. 𝗔𝗜 𝗶𝘀 𝗮𝗿𝗿𝗶𝘃𝗶𝗻𝗴 𝗮𝘁 𝘁𝗵𝗲 𝗲𝘅𝗮𝗰𝘁 𝗺𝗼𝗺𝗲𝗻𝘁 𝗵𝘂𝗺𝗮𝗻𝗶𝘁𝘆 𝗻𝗲𝗲𝗱𝘀 𝗶𝘁.

US productivity growth has been running at half the rate of the 1940-1970 era and a third the rate of 1870-1940. The global population is declining below replacement in dozens of countries, including China. Without AI, we would be panicking about economies shrinking from depopulation, not job loss.

The timing is almost miraculous. This is what Andreessen means when he says the real boom has not started yet. We have been in a 50-year productivity drought, and most people do not even realize it.

𝟮. 𝗔𝗜 𝗶𝘀 𝘁𝗵𝗲 𝗽𝗵𝗶𝗹𝗼𝘀𝗼𝗽𝗵𝗲𝗿'𝘀 𝘀𝘁𝗼𝗻𝗲.

Isaac Newton spent decades trying to transmute lead into gold and never succeeded. AI does something more powerful: it converts sand (silicon) into thought. The most common material in the world is the rarest output.

This one metaphor reframes the entire AI conversation. You do not have a job loss problem. You have a philosopher's stone sitting on your desk that you are not using enough.

𝟯. 𝗔𝗜 𝗺𝗮𝗸𝗲𝘀 𝗴𝗼𝗼𝗱 𝗽𝗲𝗼𝗽𝗹𝗲 𝘃𝗲𝗿𝘆 𝗴𝗼𝗼𝗱, 𝗮𝗻𝗱 𝘃𝗲𝗿𝘆 𝗴𝗼𝗼𝗱 𝗽𝗲𝗼𝗽𝗹𝗲 𝘀𝗽𝗲𝗰𝘁𝗮𝗰𝘂𝗹𝗮𝗿𝗹𝘆 𝗴𝗿𝗲𝗮𝘁.

The best coders right now are not reporting 2x productivity. They are reporting 10x. The gap between "pretty good with AI" and "elite with AI" is widening, not narrowing.

This is the most important signal for career planning right now. If you are just using AI to do the same job slightly faster, you are leaving the real leverage on the table.

𝟰. 𝗧𝗵𝗲𝗿𝗲'𝘀 𝗮 𝗠𝗲𝘅𝗶𝗰𝗮𝗻 𝘀𝘁𝗮𝗻𝗱𝗼𝗳𝗳 𝗯𝗲𝘁𝘄𝗲𝗲𝗻 𝗣𝗠𝘀, 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝘀, 𝗮𝗻𝗱 𝗱𝗲𝘀𝗶𝗴𝗻𝗲𝗿𝘀.

Every engineer now thinks they can be a PM and designer. Every PM thinks they can code and design. Every designer knows they can do both. And they are all correct, because AI enables each role to absorb the tasks of the other two.

I have seen this firsthand in the investing world. The analyst who can build models and write narratives is 5x more valuable than someone who can do only one. The same convergence is happening in the product.

𝟱. 𝗙𝗼𝗿𝗴𝗲𝘁 𝗧-𝘀𝗵𝗮𝗽𝗲𝗱. 𝗕𝘂𝗶𝗹𝗱 𝗮𝗻 𝗘-𝘀𝗵𝗮𝗽𝗲𝗱 𝗰𝗮𝗿𝗲𝗲𝗿.

Scott Adams could not have created Dilbert by being the world's best cartoonist or the world's best business mind. He needed both. The additive effect of two skills is more than double. Three skills are more than triple. Larry Summers puts it differently: don't be fungible.

The person who can code, design, and ship a product is no longer a unicorn. They are the new baseline for "extremely valuable." If you are only one of those three things, you are increasingly replaceable.

𝟲. 𝗝𝗼𝗯𝘀 𝗮𝗿𝗲 𝗯𝘂𝗻𝗱𝗹𝗲𝘀 𝗼𝗳 𝘁𝗮𝘀𝗸𝘀. 𝗧𝗮𝘀𝗸𝘀 𝗰𝗵𝗮𝗻𝗴𝗲. 𝗝𝗼𝗯𝘀 𝗽𝗲𝗿𝘀𝗶𝘀𝘁.

Executives never typed their own emails in the 1970s. Secretaries printed incoming emails and hand-delivered them. Both roles survived the transition, just with different task sets. The same will happen with AI and coding, PM work, and design.

Everyone obsessing over "will my job disappear" is asking the wrong question. The right question is: which tasks in my job are about to rotate, and am I ready to pick up the new ones?

𝟳. 𝗔𝗜 𝗰𝗼𝗱𝗶𝗻𝗴 𝗶𝘀 𝗷𝘂𝘀𝘁 𝘁𝗵𝗲 𝗻𝗲𝘅𝘁 𝗮𝗯𝘀𝘁𝗿𝗮𝗰𝘁𝗶𝗼𝗻 𝗹𝗮𝘆𝗲𝗿.

We went from human calculators to machine code to assembly to C to scripting languages. Each layer was dismissed by the previous generation. Each time, the new layer won, and total coding employment grew. AI coding is the same pattern, not a rupture.

The Perl programmers of 2005, laughing at JavaScript, are the C programmers of 1995, laughing at scripting. History rhymes, and it always rewards the people who adopt the next abstraction first.

𝟴. 𝗔𝗜 𝘁𝘂𝘁𝗼𝗿𝗶𝗻𝗴 𝗱𝗲𝗺𝗼𝗰𝗿𝗮𝘁𝗶𝘇𝗲𝘀 𝗲𝗹𝗶𝘁𝗲 𝗲𝗱𝘂𝗰𝗮𝘁𝗶𝗼𝗻.

One-on-one tutoring is the only method proven to move a student from the 50th to the 99th percentile (Bloom's two sigma effect). It used to require being born into royalty. Alexander the Great was tutored by Aristotle. Now, any kid with a phone can access the same quality of personalized instruction.

This is the most under-discussed consequence of AI. Every parent reading this should be supplementing their kid's education with structured AI tutoring right now. Not next year. Now.

𝟵. 𝗣𝗲𝘁𝗲𝗿 𝗧𝗵𝗶𝗲𝗹 𝘄𝗮𝘀 𝗺𝗼𝗿𝗲 𝗿𝗶𝗴𝗵𝘁 𝘁𝗵𝗮𝗻 𝗔𝗻𝗱𝗿𝗲𝗲𝘀𝘀𝗲𝗻 𝗼𝗿𝗶𝗴𝗶𝗻𝗮𝗹𝗹𝘆 𝗮𝗱𝗺𝗶𝘁𝘁𝗲𝗱.

Progress in bits masked stagnation in atoms. The built world is barely different from 50 years ago. Same bridges from the 1930s, same dams from the 1910s. Cartels, monopolies, unions, and regulations prevent the rate of change that people had 100 years ago.

This is also why AI will not transform everything overnight. Institutional sclerosis is real. Healthcare alone could take a generation. If you are building in atoms, budget for a war of attrition, not a blitzkrieg.

𝟭𝟬. 𝗠𝗼𝗮𝘁𝘀 𝗶𝗻 𝗔𝗜 𝗮𝗿𝗲 𝗴𝗲𝗻𝘂𝗶𝗻𝗲𝗹𝘆 𝘂𝗻𝗸𝗻𝗼𝘄𝗻.

Within a year of ChatGPT's launch, five American companies, five Chinese companies, and open-source all had roughly equivalent models. DeepSeek emerged from a hedge fund in China and basically replicated the American labs' work. The smartest AI insiders privately admit there aren't many real secrets among the big labs.

This is the most honest take I have heard from a top-tier VC. No one knows if the value accrues to models, apps, or infrastructure. Anyone who tells you otherwise is selling you certainty they do not have.

𝟭𝟭. 𝗔𝗜 𝗜𝗤 𝘄𝗶𝗹𝗹 𝗯𝗹𝗼𝘄 𝗽𝗮𝘀𝘁 𝗵𝘂𝗺𝗮𝗻 𝗹𝗶𝗺𝗶𝘁𝘀.

Human IQ caps around 160 because of biology. Current AI models test around 130-140. There is no theoretical ceiling stopping AI from reaching 200, 250, or 300. The concept of AGI as a "human equivalent" will be a footnote because AI will race past that threshold.

This is the frame that makes the "will AI take my job" debate feel small. We are not building a replacement for human thought. We are building something that will be better than the best human thought has ever been.

𝟭𝟮. 𝗧𝗵𝗲 𝗯𝗲𝘀𝘁 𝗳𝗼𝘂𝗻𝗱𝗲𝗿𝘀 𝗮𝗿𝗲 𝗿𝗲𝘁𝗵𝗶𝗻𝗸𝗶𝗻𝗴 𝘄𝗵𝗮𝘁 𝗮 𝗰𝗼𝗺𝗽𝗮𝗻𝘆 𝗲𝘃𝗲𝗻 𝗶𝘀.

Layer one: AI redefines products. Layer two: AI redefines jobs within companies. Layer three, which has not dropped yet: AI redefines the very concept of having a company. The holy grail is the one-person, billion-dollar outcome, and the best founders are chasing it.

Satoshi did it with Bitcoin. Instagram and WhatsApp came close with tiny teams. The question is no longer if this is possible with software. The question is how many of these we will see in the next five years.

AI is the philosopher's stone. The question is whether you pick it up.

The full podcast is worth your time. Link in replies.

[wayne @wceder3]

@robin_j_brooks China’s Yuan devaluation is a bold chess move, signaling worry about U.S. tariffs while risking capital flight and regional tensions. What do you think Beijing’s next play will be if Washington doubles down?

[wayne @wceder3]

@Sofia_Rani34 Happy birthday!

[wayne @wceder3]

@Whalestox Done

[FRONTROW @frontrowjp]

ハシ（Chopsticks）とフォーク（Fork）を合体させた“チョーク（Chork）”が欧米で話題沸騰中。西洋と東洋の文化のコラボ？http://t.co/kPQnHted http://t.co/wmLkZIU8