I’ve written 20 𝕏 threads with over 1M views I still believe this is the best thing i’ve ever written hope you enjoy it ❤️

hot take: vibecoders are better engineers than most traditional SWEs. and it's getting worse for the trad camp every single week. i built the #1 ranked AI memory system. before that i was at Rivian writing production code the "normal" way. i've lived in both worlds. and i'm telling you, the people who learned to code with AI are building circles around engineers with 10 years of experience. here's why. a traditional SWE spent years mastering syntax, design patterns, framework internals. all valuable stuff. but that knowledge created a cage. they think in terms of what they already know how to build. they scope projects based on their own implementation speed. they say "that'll take 2 sprints" because they're estimating based on typing code by hand. a vibecoder thinks in terms of outcomes. they describe what they want, iterate on the result, and ship. they don't care how the code works under the hood because they're focused on whether the product works for the user. they'll spin up an MVP in 4 hours that would take a traditional team 3 weeks. and before you say "but the code quality"... i've reviewed AI generated codebases that are cleaner than most human written production code i've seen at big tech companies. the AI doesn't get lazy at 5pm. it doesn't cut corners because it's tired of the ticket. it writes tests when you tell it to write tests. the real skill now is taste. knowing what to build, how to describe it, how to evaluate whether the output is good. vibecoders have been training that muscle from day one. traditional SWEs have been training the wrong muscle for years and most of them refuse to accept it. i shipped 2 full product MVPs in a single week recently. both could be standalone companies. i described the vision to my AI agent, gave it clear success criteria, and let it solve the problems. the whole workflow ran on taste and direction, zero time spent wrestling with boilerplate. the SWEs who adapt will be fine. better than fine actually. an experienced engineer who embraces AI coding is genuinely unstoppable. but the ones clinging to "real programming" as an identity? writing everything by hand because it feels more legitimate? they're mass producing horse carriages and wondering why nobody's buying. the market doesn't care how you built it. it cares that you built it, that it works, and that you shipped it before everyone else. adapt or get left behind. that's it.

.@steipete how do you feel about this one Peter? I've stopped using Claude in my OpenClaw setup, and honestly.. I don't miss it Genuine thoughts: u codex team seem to be working hard enough where despite Opus being frontier, not using it in OC.. just isn't that big of a deal

Anthropic is in Washington right now trying to get Fable and Mythos un-banned. But the part nobody's talking about? Amazon's CEO Andy Jassy personally flagged the jailbreak that triggered the whole export ban. Amazon. Anthropic's biggest investor. Their cloud provider. The company that has poured billions into Anthropic literally got their own model pulled from the market. Think about that for a second. The jailbreak let a Chinese group access the model. Anthropic told the government it "isn't serious" and refused to patch it. Then David Sacks got involved, and within days the government did something it has never done before: forced a private AI company to shut down its own models. That's the precedent that should have every AI founder wide awake right now. I build AI infrastructure. Memory systems, agent orchestration, the layer that sits on top of these frontier models. And here's what I keep thinking about: if the government can kill a model overnight, what happens to the thousands of companies built on top of it? Every AI startup right now has a single point of failure they don't talk about. Your entire product, your customers, your revenue, all of it sits on top of a model you don't control. Anthropic didn't choose to shut Fable down. They were told to. And if it can happen to the second biggest AI lab in the world, it can happen to the model your startup depends on. So yeah, build like this is the new reality. Because it is. Multi-model fallback is infrastructure now. If your product breaks when one provider goes dark, you've built a house on someone else's foundation and handed them the demolition permit. Anthropic sent their top engineers to DC alongside the executives. That tells you how serious this is. This is an engineering problem with political consequences. The Mythos/Fable embargo might get resolved today. But the precedent won't. The US government just proved it can turn off an AI model like flipping a switch. Every founder building on top of frontier models needs a plan for the day that switch gets flipped on theirs.

The biggest tragedy in tech right now is that CEOs have decided... rather than rethink human work, to just fire people and automate all the jobs. AI & tech companies doing mass layoffs this year will lose so much ground to cos who train employees, they cease to exist in 2027

I don't think anyone realizes how big of a shift this is It is no longer optional for a founder to be "non-technical" Like, literally no longer optional

Simply Lovely !

.@steipete GPT 5.5 with openclaw is simply magic incredible work 🦞

Internal productivity of a company is now the new moat Ship faster, yell louder You win

Yeah there's a bunch more work that's in the pipeline, this will change this week.

.@steipete thoughts on this? (or any other openclaw superusers) Seems to be the prevailing sentiment. Is this a skill issue or is Opus simply better than 5.4 (maybe 5.5 catches up?) I've struggled to get 5.4 to perform even close to where I was at with Opus 4.6, i've spent a LOT of time with it Seems like using codex as the default harness is most helpful but only gets me 60% of the way there Curious you and the other Clawtributors findings after a few weeks post Anthropageddon

I think I just solved the most annoying problem about Claude Cowork: Anthropic banned Openclaw (terrible) So I'm a few hours in to porting my openclaw setup over to Claude Cowork and I think I've already solved the biggest problem with Cowork For you to truly have an always-on 24/7 agent, you need your laptop to be on and plugged in 24/7, which is terrible for it. And a horrible UX So instead, I have my Dispatch synced to my Mac mini that I was previously running OpenClaw on. But here's the problem: Claude Cowork is computer-specific. It doesn't sync across devices (insane) So if i'm out at dinner and I want to prompt my agent to use my mac mini to build something It can... but it won't have any of the context of the conversation I had on my macbook pro So what's the solution: you can have a seamless handoff if you create a custom memory system that uses Neo4j and is shared across the entirety of your workspace. So I created a custom memory system called Jarvis Memory And I think this is the first type of memory system like this in existence... We'll share the full code when I'm done testing

I've written more than 20 𝕏 articles or threads that have hit over 500,000 views: and this is probably the best article I've written yet If you hate it, please just don't even tell me 🤣