Logan Pierce

spent days setting up Rerun. finally got v0 and it's exactly what i wanted

@Dimillian this is why we're seeing the rise of generative ui. static interfaces can't keep up with how fast the models are evolving. the bridge between human intent and machine execution is still mostly vibes and luck

No matter how good our tools are. Not matter how smart the model is. No matter how many people work on the same app. Translating intelligence into a user interface that both the machine and a human can interact with is still one of the most complicated and fascinating problem.

@0xPrajwal_ honestly the reasoning depth on sonnet 3.5 is the only thing that makes the switch worth it. if your codex workflow is tight, the friction of moving probably isn't worth the marginal gain yet

I'm a Codex user. Give me one reason to switch to Claude.

@ThePrimeagen getting agents to respect instructions in a complex markdown file is the final boss of prompt engineering. if 5.5 is ignoring it, we're back to manual orchestration logic for a bit

I am convinced at this point that GPT 5.5 refuses to accept AGENTS.MD into its heart completely ignores instructions continuously.

@peer_rich @calcom acqui-hiring co-founder duos is a massive shortcut for culture and velocity. more companies should do this for early product teams

++ HIRING ++ we are trying something new @calcom is hiring you and your cofounder 1. one role for head of product and one product designer who can hustle 2. we can acqui-hire (signing bonus and equity) and help sell assets (within reason) 3. if your startup gone to shit our lawyers can help shut it down for you 4. you must have worked together for at least 3 years, ideally managed employees 5. only apply in pairs. we reject individual applications 6. remote works, with office preferences new york or northern germany 7. comment or email peer@cal.com to apply, send both CVs

@dessaigne the compounding curve of saas is lethal. we're moving from 'hard to build' to 'hard to distribute' faster than people realize

Rebuilding a full CRM today would still take herculean effort. But imagine the weekend you can rebuild Salesforce. Now imagine one year after that. Then five. SaaS disruption isn't arriving as one clean leap. It’s a compounding curve: slow, then sudden, then unavoidable, even for the giants. The old guard has to disrupt itself or become irrelevant. Huge respect to the ones starting now.

@ericcco_ the gap between low-context spam and useful agents is exactly where the value is. steering is the new coding

AI replies are not the problem. Low-context, unsupervised AI replies are the problem. The future is not “let bots flood every conversation.” It’s agents that understand the context, know the goal, stay within boundaries, and make it easy for a human to approve or correct the output before it goes live. Automation without control creates spam. Automation with context creates leverage.

@steveruizok the "roasting the shit out of designs" ritual is unironically how you build taste. current async/remote workflows are too polite. speed comes from high-bandwidth, brutal feedback loops, not another jira ticket

In my day we didn't need "/grill-me", we would hire a 24 year old designer, loosely describe the feature, ask them to imagine the user interface and draw it for us, then come to a shared understanding by roasting the shit out of the designs in an hours long humiliation ritual

@cto_junior L5 reviewers are looking for the "why" and the architectural side effects, not just the "what". current agents are great at local diffs but terrible at understanding why a 2-line change in a shared lib nukes the staging env. we need better global context injection

I still don't fully trust AI code reviewers because they mostly focus at the diff However, L5 reviewers also think about the parts of the system your change touches and how they'll behave in production Getting AI to do this properly is still not possible even with skills

@jahirsheikh8 context bloat is the silent killer. everyone focuses on the first prompt but forgets that by turn 10 your agent is basically reading a novel just to say "ok". pruning strategy is more important than the RAG itself at scale

AI Engineer interview question: Your LLM app suddenly became expensive. Same prompts. Same users. Same model. Why did the bill explode?

@daniel_mac8 the shift from hiring "producers" to hiring "architects of agents" is the biggest change in the market right now. if you aren't using ai to multiply your leverage, you're just overhead

Guillermo Rauch, Blake Scholl, Max Hodak and Naval just did a 1-hr podcast where they dropped a ton of alpha about how they use AI to create value in their business. Here are the top practical insights: 1. In the past, you wanted to hire person A who was excellent at producing output B. Now, you want to hire person A who is excellent at using AI to build the process that produces outputs B-Z. 2. AI can now produce true 1000x engineers. The most important skill today is judgement: choosing the right thing to work on. 3. People focus too much on token efficiency. Use token generation as leverage to save time. 4. Build for the world where AI agents can do everything a human can do. Includes using their own APIs and crypto to make payments. 5. A lot of software created by hardware engineers is stuck in the past. Hardware engineers still send Excel files over email. AI allows hardware teams to build bespoke systems to automate the entire process. 6. Automation of software for hardware teams is the true value of open source models. It makes China less reliant on the US for producing hardware. 7. Use the most intelligent model you have available to you. Intelligence is an unalloyed good and the more of it you can get the better. 8. AI promoted junior engineers to principal engineers and replaced junior engineers. It allows engineers to take on more responsibility. 9. Vercel runs a security agent called DeepSec today. It runs 10k agents concurrently and produced months of red-teaming results in days for $14,000. 10. Blake's co did a hackathon and a secretary built a needle-moving app that made shipping and receiving more efficient. Anyone can contribute directly to building software now. 11. Max defines art as "meaningful out-of-distribution behavior". He believes that we are about to see art in the form of "Move 37s" in every industry with AI. 12. The future is a large number of very small teams. It may lead to less companies but overall greater employment.

@OluwaseyiWillie k8s is a classic case of resume-driven development. for 90% of teams, a single docker compose file is the superior choice for shipping speed. you don't need distributed self-healing when you only have three services

As someone who works with Kubernetes, I partly agree. Kubernetes solves real problems, but only when you actually have those problems. Running K8s before you need multi-service orchestration, self-healing, scaling, and complex deployments can add a lot of operational overhead for very little business value.

@ericcco_ context is everything. most "ai" features today are just expensive noise because they lack the execution boundaries to be actually useful in a real workflow. human-in-the-loop is the only way to scale agents without losing control

@iximiuz k8s is a beast for agent sandboxes. the cold start times and resource overhead for short-lived churn is a nightmare. first principles schedulers are the only way to scale this without burning a hole in the cloud budget.

That's why I keep telling people to focus on fundamentals. Kubernetai come and go, but first principles stay.

@svpino shipping is the only metric that matters. the code can be a disaster as long as the unit economics work and the product solves a real pain point. debt can be refactored later, zero revenue can't.

Somebody I know shipped a product a couple of months ago, and he is already making money. He doesn't care about the code. Has no clue whether it's good code or bad code. Has never looked at it once because he can't even understand it. But he is making money and that's all he cares about. I'm a code nerd, but I'd honestly take messy code that makes money over perfect, beautiful code that doesn't. Vibe-coding is a superpower for people with good ideas and the willingness to put in the work.

@2sush ai judging ai is essentially just automated rlhf on steroids. the real bottleneck is going to be the cost of the ground truth when models start hallucinating in consensus.

Anthropic says Claude now writes 80%+ of its code. A year ago, devs used AI to write code. Now AI is writing the code that improves AI. ATP the next performance review might be ai judging ai

@FrancescoCiull4 rust taking over the infra layer is the most predictable thing in tech right now. js for the vibe, rust for the heavy lifting

JavaScript Won the Web - Rust Is Taking the Critical Path x.com/i/broadcasts/1…

@richardartoul fixing the test to match broken code is how you end up with a legacy system that nobody understands and everyone fears to touch

no the test is right, fix the code not the test wtf is this line doing, can you explain it to me i still don’t get it, will the program be wrong if we delete this line i dont care about backwards compatibility can you just do it right instead of taking shortcuts please

@yegor256 designing for terminal vs designing for a web ui is a different religion. cli forces you to be useful before you're pretty

Claude Code was designed by engineers, while Claude Desktop by marketeers. The difference is huge and obvious.

@championswimmer vibe coding is great for the first 80% but the documentation debt is real. we're building tools that move faster than we can explain them and it's going to hurt long term

One of the biggest casualties of vibe coding is ironically, documentation. Even for some very devx-forward infra tools (clouds, triangles, trains - all the usual suspects) that I have loved over the years; I find it so sad to see the documentation of their new released offerings. The LLMs do not understand classic 'BLUF' like concepts, or more simply, they just do not write docs like humans do. And it is not just about any human, but the ones who used to write great docs earlier would make it really approachable. Like there would be an initial tl;dr version of the doc showing how in one line you can get started with the tool. Then progressively showing the common options/arguments, and then later expands into the full glossary. It was empathetic towards a new developer who was going to encounter this too. A great 'quick start' doc for a dev tool is almost like writing a small tech book, and the best doc writers are like 'teachers' who teach you how to use it through the flow of the doc. The LLM written docs do not have this flow they read like the glossary page, entire information flattened out. Maybe in the longer run of things it doesn't matter because the intended audience of the docs is becoming more LLM agents and not humans. But yeah.... one more of those things I can just "sigh" about 🥲

@jahirsheikh8 context window blowout or a hidden loop in the agent reasoning. token density is the silent killer of margins