MK

@sasajuric Tests pass, linters, builds and deployments green. Same as any proof provided by human hand built built systems

"I generated a lot of code and didn't need to read any of it. The codebase is in a great state." Uhm, pardon my confusion, but how do you know the codebase is in a great state if you didn't read any of it?

@Pearticle @kyfriedfella @matt503ea5sf9z5 Not really. Distillation requires WORK - time in code and dev and GPU / compute time. Plus costs of using the API to distill from.

@psi686 @kyfriedfella @matt503ea5sf9z5 It does matter to OP's point though. They had to do far less than the frontier labs.

How is Kimi, running on a bunch of 14nm huawei toasters beating SpaceXAI with multiple datacenters of blackwell chips.

@benjamin_horne Calculate for yourself how much it costs to run deepseek or Kiki full-size and explain to me how it’s a “fraction of the price” compared to API costs ?

Can someone please explain to me how OpenAI and Anthropic capture even a single dollar in revenue in a world where Chinese models are just as good (if not better) and cost a fraction of the price? Outside of a few niche situations where orgs will pay to use American models for security reasons (e.g. a handful of US military contracts) why pay to use their models at all? In this world, how much are OpenAI and Anthropic worth? Has to be < $100 bn each, right?

@kyfriedfella @matt503ea5sf9z5 It doesn’t matter if they distilled, they were able to further reduce the loss function and teach the model something. That’s what machine learning is, so either way they created novel development.

@matt503ea5sf9z5 Stealing. It’s called distillation. They don’t have to do as much work or spend because they just distill the frontier American models

@trikcode Literally just add “be good a system design” as a bullet point and it’s solved

AI coding assistants are getting scary good at fixing isolated bugs, but they are absolutely terrible at software architecture. you accept 10 consecutive "perfect" PRs, only to realize the model quietly introduced three different state management patterns. I think we are trading short-term speed for massive, invisible architectural drift

@DennisTheProf @flavioAd It should finish the task. This isn’t the clients last session and he’s fucking off.

@flavioAd It should off board. Leave you a doc of exactly where it left off and what it was working on.

Getting rate limited is fine Killing the task halfway through is not claude code should just let the task finish like codex does

@mardehaym Hey so Mark, basically AI could add a “second payment provider” in a probably an hour.

A senior architect told me she'd stopped approving AI-generated PRs from her juniors. Her VP noticed. Approval rates dropped. Cycle times spiked. Suddenly, she's the bottleneck. So she invited the VP to a code review. Just to observe. The junior explained the PR. Confident. Articulate. Code executed. Tests passed. Then she asked, "What happens if we add a second payment provider next quarter?" Silence. The junior stared at the code. The model that generated it doesn't know the roadmap. The code didn't contain an answer, because the system design didn't exist beyond the current ticket. In 8 months, the juniors shipped 40,000 lines of code. None of them could explain how the system worked outside of their immediate task. They thought they understood the codebase. They never actually read it. "They're not bad engineers," she said. "But they've never been wrong long enough to learn anything." The VP had offered courses, certifications, book clubs. But the architect needed something else, and it wasn't more training. She needed permission to slow things down, so the juniors could build a real mental model. One that lives outside the chat window. Juniors don't develop judgment from getting things right. They develop it by sitting with wrong answers, long enough to understand why they failed. You want to test their understanding? Ask them to diagram the system. No IDE/docs/ AI. If the diagram has gaps, those gaps exist in their heads too. You won't notice, until something actually breaks.

@AdityaShips Amplify is so cheap it doesn’t even make sense

Vercel is getting way too expensive for me atp. Any good alternatives for hosting my Nextjs app without compatibility issues?

@apnmrev Product was NEVER the hard part. All college students / unemployed would have been printing money by your logic. It’s always and forever marketing and only marketing.

founders are cooked. just talked to a 26 year old who shipped 7 apps with claude in the last 30 days. all clean. all working. real problems solved. still sitting at $0 revenue. another guy same age spending 90% of his time on posting short form? already printing 6 figs with a productivity app. product used to be the hard part. not anymore. this is the new game. attention is the only thing that matters now. the world is yours if you move fast and chew glass on distribution early.

@sarahwooders You are using the word “skills” all over but we are using MCP servers for TOOLS. Not the same thing

Why are people still using MCP instead of just APIs, now that we have skills?

@GeorgeMayer Just call him bro

Unpopular opinions I hold about work/coding things: - Remote work is net bad - tracking lines of code is reasonable - a manager’s job is harder than IC - an ICs job is way more fun than a manager’s - context switching is a skill, not a burden - if you can’t study and pass the leetcode tests you probably shouldn’t work at faang (exceptions apply) - unwilling to study is not the same, but it sort of is disqualifying in a different way - leetcode tests are (were) dumb, but they served a real purpose

@Yourpop8 You’re just a cunt

Is it just me or does the World Cup just feel off this year?

@shaneparrish Couldn’t have been very challenging if an LLM could’ve written it

People have no idea what's coming with the next generation of kids who are AI native. My youngest teen just started an internship. On the first day he was given a "challenging" two weeks worth of work with very specific objectives, timelines, etc. By 10am the next morning he was done the entire list and asking for more work. They didn't think he could possibly be done. How? He used AI to help (with their permission). And he KNOW how to use AI (he's not using it like a google search bar). These AI native kids are gonna run laps around 25-40 year olds that are not using AI.

@jahirsheikh8 There is something to Claude code / codex harness context contamination. Opus has always performed better in web UI than in harness

met a dude today, he still copy pasting the code from chatgpt

@AishwaryaDevv That no longer matters

"Claude is better at coding" Nah, you are just bad at coding.

@devXritesh I’ll go with synchronous because everyone in the responses say asynchronous but they are all brokies

Backend Interview Question: A user clicks: "Forgot Password" Which approach would you choose? A) Synchronous B) Asynchronous What's your choice?

@samueljmcd Yep. The way I look at it, API pricing is likely in my eyes to be wildly profitable for them. I would not be surprised to hear about 80% margins. Subscription pricing is probably still profitable, factoring that many users underperform, and caching

Fair yeah, we don’t know their actual margins. But the $400/$8k/$14k numbers aren’t estimating provider cost, they’re “API list price for what you used.” That part’s calculable. The bigger question is whether API list price itself reflects cost, deepseek is a good example because they have similar-class models for ~1/35-100th the price So even the $14k figure likely overstates real value, not just understates uncertainty.

You’re probably underpaying for AI and don’t realise it. SemiAnalysis bought every Anthropic and OpenAI plan, ran long-horizon coding tasks to the weekly limit, and measured actual compute delivered: Claude Pro ($20) → ~$400/mo compute Claude Max 5x ($100) → ~$2,000/mo Claude Max 20x ($200) → ~$8,000/mo (40x) ChatGPT Plus ($20) → ~$340/mo ChatGPT Pro ($200) → ~$6,000/mo (30x) The bigger point: almost nobody treats model/compute allocation as a real discipline yet. Which model for which task, API vs subscription, when to switch providers. That’s about to become a core engineering competency, not a nice-to-have.

@samueljmcd Right but we can’t see the subscription plan costing - I.e actual compute cost to the provider vs actual subscription usage. So we just really don’t know

@psi686 They’re relative estimates to the cost you’re paying I believe

@ZachWLambert How many students lunches have you paid for ?

Elon Musk is now a trillionaire. More specifically, his net worth is $1.2 trillion or $1200 billion. That number doesn’t even compute for most of us, so here is some helpful context: The USDA needs $18.8 billion to feed every child in the US school system breakfast and lunch each day. Elon could pay for that 64x and still have money leftover. The World Food Program needs $13 billion to feed the 110 million hungriest people on earth. Elon could pay for that 92x and still have money leftover. The National Alliance to End Homelessness needs $9.6 billion to provide housing for every unhoused person in the United States. Elon could pay for that 125x and still have money leftover. The average American teacher makes $74,495 per year. Elon could pay the annual salary of over 16 million teachers and still have money leftover. The problem is not that we don’t have enough money. The problem is that we have built a world where one person can accumulate more wealth than the GDP of 180 countries while children go hungry, families drown in medical debt, teachers are forced to buy school supplies for their students, and people sleep on the streets. This is a complete moral failure.