muzani

Bubbles happen when a) someone buys something where they don't understand the value to, b) sell it off to another sucker who also doesn't understand the value.

If a is not true, the prices will fall to their value. If b is not true, overpriced goods don't move, and the price doesn't blow up as fast as it does.

Yes, they increase productivity, but how much? Is OpenAI really contributing a trillion dollars of value to the world? A thousand billion, a million million. 50 billion months of a $20 subscription. 80 million lifetime subscriptions of $20/month. That much value?

That's if you're matching about 40% of the population.

For some, it might be blue -> blue -> blue -> blue -> green -> blue -> green -> blue.

This is the first time I realized that turquoise is the gray area.

Better apps? iOS

Better prices? Android, especially with purchase parity

Most stable? iOS

Most flexibility and freedom? Android

Better camera or speed? This always enters the debate, but it's specific to the device, not OS

i) I'm sure everyone agrees with this, except maybe a few fans. The best thing he did was slowly wean the whole company off FACEBOOK, even if it's with the Meta smokescreen. Instagram is their core now.

ii) I would disagree on this. Meta is king of low value ads. Google is great at targeted ads, but sometimes you want something to reach as many eyes as possible, as cheaply as possible.

They've barely avoided many of the dark patterns like chumboxes and forced repetition of screens, just because they have enough people scrolling through it. Of course, there's other dark patterns here, but I give them a B+ for low value ads.

The marketplaces and events still provide massive value to the world, for free, as long as they lasted. In the end, it's nearly a trillion dollar company, which is a remarkable thing to build on low value ads.

They're tuned to target a certain customer demographic solving for certain problems. I've seen standard AI models to absolutely brilliant things sometimes. But the prompts to get it to perform like it did with GPT-3 seem to get lengthier and lengthier in time. At some point we'll probably just snip out smaller, specialized models to do certain things.

Most tech giants are one hit wonders. Microsoft is the exception, not the norm.

The playbook now is get product-market fit -> own 99% market share -> build, buy, exterminate everything that could threaten this market share -> when the disruption comes, pivot to it.

Zuck inspired the playbook. He's done it to perfection. He's been in the ring for 20 years. Yahoo dominated for about 5-7 years. By no means is he a bad CEO.

Facebook's decline was always a part of his strategy, and whatever the employees say about there being a bazillion active daily users doesn't reflect that their time is near. But Instagram still runs strong and Threads has half a billion active users.

Does Android Studio count as Jetbrains? It's core to mobile development still.

AI tools are not designed for code reviews for some reason. Maybe everyone is clicking "accept all". I end up reviewing in either Android Studio or even Fork.

It also does code reviews much better than GitHub. Maybe in time, Jetbrains will do a specialized tool for code review that doesn't need to use any AI.

I feel like this should be a service in itself, similar to Heroku or Supabase. Just tick which laws you want to adhere to, upload files to their buckets. Tick another box for audit logs and such and it'll ask you where you need your human in the loop and which buttons those humans need to press. So a bit like Carta or Deel in that sense.

I've had some big enterprise deals fall through because of something like this - military, insurance, fintech, etc.

For startups, giving a shit has always been a requirement. Once it get removed from requirements, it becomes a "corporation".

It's awfully hard to find people who do give a shit and even harder to interview for it. It doesn't scale with money, which is exactly why but companies don't look for it.

The results don't lie; it's why companies like Google who hire the most expensive people in the world regularly fail to release something like Antigravity in a stable form when people are building these things solo.

Corporations end up buying these startups for massive amounts of money. It works for everyone because the startups are compensated well for giving a shit and corporations just sit on the gold mine making sure the money keeps flowing.

But it's a good point you raise here. I disagree to an extent - AI will give more of a shit than many humans. It's the main reason many people lose their jobs over it.

So it stands to reason that startups will continue to be AI-resistant. There's a lot of hype that you can build "anything". But where are the alternatives to Facebook, Jira, MS Teams, Stack Overflow, Stripe, and all these other proven business models that everyone hates but is forced to use anyway?

"Your question is stupid" can be rephrased as, "This is how you can improve your question..." SO used to do this a long time ago, but they decided that the site was being "flooded" with low quality questions and removed that. This seems like something cheap AI can do just fine.

I believe everyone understands the root cause - people are afraid to ask questions, which stems from crappy moderation, which stems from the karmic system. The founders went on to found Discourse which has the exact same problems in community, which hints heavily that it's a problem with the system.

What I would fix is the karma system.

1) Don't give mod powers and responsibilities to people with high karma; moderation skill is intangible and not related to technical competency.

2) Get rid of "bullying" habits. Punish downvoting by increasing the cost to -1. Or do what other social media has done and hide votes, especially to high karma users.

3) Reset karma to 0, but recognize their peak value. Send an email out on why. This should be enough to get rid of all the people who are in it for the karma, and pull back all the people who were in it to contribute to a knowledge base.

4) Fire all the current moderators. They've collectively done a terrible job. Add their salaries together and use it to poach dang from HN.

We did it before. OKRs and such.

If there's an incentive to hit 99% non-cash rates, the app just does nothing when it would normally crash. A manager decided not to release a refactor that would fix a bug because it would have jeapordize the crash rate right near bonuses.

Efforts like R&D, clean code, and other optimizations are often punished if they don't affect the bonuses.

The QA team suddenly started making tons of tickets. Misaligned padding on top left, misaligned padding on top right, misaligned padding on the icons, colors are a little off. But they failed to catch the bug where it was not possible to enter real addresses because the address box was too short.

In the long run, the project gets 70% month on month growth and then gets deprioritized.

It's possible they don't know the actual costs. The one-time costs like hardware and training foundation models is huge. There's the ongoing costs like the PhD dudes they hire. There's the data they buy and decide not to use. There's the various offers under one payment plan - limited artifact hosting, cowork, image upload, Claude Code, Code tab on Claude. What about feature parity between the site and app? Who's working on all these?

tldr it seems really complex and by the time they've counted it they probably hired 40 new people for an unannounced feature.

That and Claude/Codex are setting up their own equivalents. The hosting companies and AI wrapper companies are rolling out their own built up clawdbots.

+1. Companies like AWS are not known for this kind of treatment. Even Google does this. The norm seems to be to randomly ban and ignore emails. It's a whitelist to build customer trust, not a blacklist of bad companies

It takes 3 weeks to get past the AI customer service bot, assuming you ever do. Good luck.

It has lost all trust from me by now. A google search for ABC will return all of ABC's competitors, sometimes ABC itself, and sometimes malware.

Yeah, I did notice a drop in quality. I dismissed it as saving the hard tickets for last, but I find myself having to explain how to do it better.

Often I ask it for approaches on how to to do something. It would give me say three options. For 4.6, two of those options were usually good. For 4.7, the options were all bad, and I'd have to explain that maybe it hasn't considered so and so option.

I never really understood these benchmarks. Unless there's an increase of 2x or so in a benchmark, the benchmark never actually reflects the real world performance.

They used to refer to it as "Do things that don't scale" rather than cheating :)

There were a lot of pitchforks back when Sonnet 4.5 was accidentally down for a few hours: https://www.reddit.com/r/ClaudeAI/comments/1rlz7kw/sonnet_45...

It goes to show that there's a very large and vocal user base using it for writing, and yet it's not part of the benchmark for Anthropic.

Anyway, try Sonnet 4.5 while it's still available?

Yup, they recommend a minimum of high for coding now, and cranked the default up to extra high.

For me, making it high effort just fixed all the quality problems, and even cut down on token use somehow

Claude doesn't even make the prompts secret or even yell at you for jailbreaking them.

They're always peak when released, then degrade over the next few months. Which doesn't matter because in 5 months, there's a new record-breaking model.

Gemini Pro, once voted the best model in 2025 by Polymarket, just spouts nonsense half the time now. I say this as an AI Pro subscriber. It built this amazing system for me in a weekend and suddenly it acts like an old horse.

OpenAI often has the opposite effect - they're disappointing when released and everyone goes back to them whenever Gemini/Grok/Claude breaks down and discover that hey, it's pretty good.

We shouldn't separate them.

Capitalism and democracy is often sold to us under the assumption that humanity is intelligent and can calculate depth.

Let's say Company A makes $100 products but creates 5000 units of pollution. Company B sells their product at $95 but creates 15000 units of pollution.

In this case, the penalty for pollution is built into the free market. Pollution is still necessary, but at what point is it just reckless? But this can only be the case as long as the market is aware of this cost.

"Political opinion" is an umbrella term for something that should be abstracted out. Global warming is a political opinion. People can just have the "opinion" that the pollution doesn't exist. Same for sweatshops, child labor, bribery, corruption, women's rights, and all these other pollution-like effects that damage the world.

Yes, I like my cheap, high quality t-shirts from Bangladesh, and I like the low cost labor that bring me my $0.16 tea bags. I like the water-guzzling, plagiarism machines that write my code for me.

But we have to decide at what point we tolerate these costs and for what output.

The basics always matter. It's communication in the end, and you have to speak the same language. Know what O(log N) means. Know when you pick a BFS vs DFS. Why do we use hash maps and why they work. Why Markov chains and state machines simplify design. How you can use Monte Carlo to make an efficient search in a game.

Many people think AI is going to be a singularity. But Amdahl's Law and Gustafson's Law suggest otherwise. You get diminishing gains with more power, and yet more power unlocks new things. Better CPUs didn't lead to instant simulations and word processing, they led to social networks and TVs that spy on you.

But you have to internalize these algorithms well enough to understand how the world works.

A lot of early games were more like file systems and they worked. I'm surprised nobody had mentioned ACID yet.

The API revolution might be another thing - you were able to swap out a database with any other. Risky decisions are fine when they're reversible. Databases were a more reversible way to deal with scaling and architecture.

It was always a strategy. More so with B2B. Would you use a payment gateway or database that has only received seed funding?

I worked at a B2B unicorn once that raised funds, not so much because they needed the money, but because it symbolized to the market that they were here to stay. That they're not going to run out of money and enshittifies when the market downturns.

Style 1: Question is so tough that you can only use AI to solve it. But for part to, you have to edit the code without AI.

Style 2: You are free to use any tool you like as long as you build this full app in an hour. Evaluation is that it works and that it follows appropriate architecture.

Style 3: You're given a dumber model, like Gemini Flash or Claude Haiku, and need to solve this problem.