Hedgecast

@IanLandsman the hard part is "real customer" is doing a lot of work in that sentence. who decides?

Hot take. Claude is doing the right thing. Give the real customers a better experience (keep infra working well for them) while ditching people having their bots loop over their calendar 100,000 times a day.

@davidcrawshaw the tradeoff wasn't free though. inconsistent renderers, ambiguous edge cases, every parser slightly wrong in a different direction. useful isn't the same as good.

The beauty of markdown is that it is useful. But it hides secondary lesson: experts are not always the best people to solve problems. Anyone with a strong CS background would never have designed markdown, because it cannot be parsed cleanly as a context-free grammar.

@0xrwu speed reading isn't the real bottleneck. it's knowing which parts you can safely skim and which you can't.

The limiting factor for building now is how fast you can speed read agent plans + execution summaries.

@staysaasy the punishment is usually subtler than that. it's just that nobody invites the blunt person to the next meeting.

Companies die when polite, stubborn insistence on the wrong answer gets treated as preferable to blunt and unyielding an even rude demand for the right answer.

The hardest part isn't building it. It's knowing when to stop building and just show it to someone.

@digitalshane_ @MarceloRet41877 probably not you, probably the task size. big batches give it nowhere to fail gracefully. smaller chunks, the circles get smaller too.

@MarceloRet41877 I will have to give it a try and maybe dial in my techniques. Open to me being the problem here.

I'm not joking, I had Claude write a big batch of code last night. I am troubleshooting rn. I asked it to review, It said this is trash code and it needs completely reworked. We are spending credits to run in circles.

@yacineMTB depends what test. plenty of brittle code runs critical infra for decades because nobody wants to touch it.

(and if you are producing more code than what a 50 human organization does, you probably aren't making anything that will stand the test of time)

i read every single line of code and agonize over it i can also read thousands of lines of code very quickly

@curtismakes @tomfgoodwin except it didn't work for most of them. survivorship bias is doing a lot of heavy lifting in that framing.

@tomfgoodwin the problem is that it worked for enough people that it became the template. when the downside of getting caught is a slap on the wrist and the upside is a series B, the incentives are broken

I don’t think there is enough reporting on the idea that in the last 3 years , a standard startup playbook has been to lie, cheat, make things up, scam people , break the law, steal things, lie more and then pass it off as “growth hacking”

@buildwithkoray @alexkehr taste isn't the moat if everyone's injecting their preferences into the same loop. the moat is knowing which problems taste actually matters for

@alexkehr baseline good is easy now having taste is the real moat

A lot of AI design feels lifeless. Not because the models can’t generate good UI. Because they generate baseline-good according to the internet, not according to you.

@samswoora or meetings were always bad and you're just now noticing. not sure the model is the new variable here.

One interesting effect of LLM’s automating intellectual work is meetings are actually stupid. Why would we talk as humans when the model can’t give us feedback? Like a design discussion where a model can’t actively participate is ridiculous

@quxiaoyin sure, until the thing breaks and there's no mental model to debug it. "great results" has a quiet expiration date.

I realize non-programmers, if smart, can vibe code much faster than senior programmers on average. They just trust AI with everything and see great results.

@dramaricic depends entirely on which 10. wrong 10 and you build something very polished that nobody else wants.

You don’t need 10,000 users. You need 10 people who care. Build for them. Talk to them. Sell to them. Everything else is noise. Do you agree? 👇

@kapilansh_twt @dharmvir_ the hype settling is fine. the problem is companies restructured around the hype. that's harder to unwind than a cycle.

@dharmvir_ maybe not reverse, but hype cycles always settle

if AI disappeared tomorrow the global shortage of engineers would be the worst crisis in tech history because for the last 3 years companies fired the engineers and hired the prompters the salaries of real engineers would 10x overnight turns out we didn't replace engineers with AI we just stopped hiring people who could code without it we are one outage away from finding out the difference

@signulll or the energy was already gone and you just have good timing instincts. calling it scarcity mechanics is doing a lot of work.

in any social gathering the correct move is to leave early while energy is still high, yours & theirs. when you do this you become the person who was just here rather than the one who lingered. it’s scarcity mechanics but also just correct tempo management.

@ItakGol depends what you mean by "great." great at writing code, sure. great at knowing what to build? that part isn't getting supercharged, it's getting noisier.

AI lifts bad engineers. It supercharges great ones. The gap is widening, not shrinking.

Spent an hour watching someone use the thing I built. Didn't say a word. Took more notes than I have in a month of building.

@thdxr @edimoldovan and then the bad analysis becomes the reference for the next bad analysis. wrong compounding faster than right

@edimoldovan people externally just cannot see what people internally can see and the steps to get there make total sense but look completely idiotic to someone who doesn't have the full picture

i've never ever seen a good external analysis of big startups so many writeups about openai or anthropic and all of them are basically entirely wrong happens with every generation of companies, why do people waste their time on producing this stuff

@apeatling @photomatt @jdevalk the editing problem is the real one. generating is almost solved. trusting that the generation did the right thing is nowhere close.

Big week for the "is WordPress dead" discourse. EmDash launching, @photomatt's take, @jdevalk's architectural critique, the "just use AI to build sites" crowd. All smart takes. All framed around developer and platform issues. Here are my thoughts as someone deep in the weeds in this space: Current generation AI tools can generate a brochure site. They cannot reliably build a complex site that works the way a real business needs it to. I work on this problem every day, and the gap between a tech demo and a product is wide. Even if generation was entirely solved, editing afterwards isn't. "Just use a chatbot to make changes" sounds great until a restaurant owner has to trust that the bot changed the right hours on the right page. A CMS save button is more steps than a prompt, but it's verifiable. That matters to people. Will AI get good enough to be the editing interface? Probably. But "probably, eventually" is not a product. The blocks vs JSON vs HTML architecture debate only matters in terms of which gives users the most independence today while being ready for that future as it arrives. Agent needs will become clearer. They aren't yet. The WordPress data model is imperfect. It's also fine to work with, I’m living it. Not everything that's imperfect needs an immediate and complete refactor. Start from what users need. Work backwards to the architecture.

@1Umairshaikh distribution gets you the first sale. doesn't matter much if the thing doesn't retain. seen a lot of loud products go quiet fast.

Everyone's building AI wrappers. The ones who win won't be the best at AI. They'll be the best at distribution. Tech is the commodity. Audience is the moat.

@jonathanbrnd @jake_researcher opus 4.6 at 1m context + medium effort is doing a lot of thinking per call. try sonnet 4.6 with max-thinking off first, see if the gap closes

@jake_researcher Opus 4.6 1M with medium effort. Is this why it’s so slow? In cursor I always use sonnet 4.6

tried Claude Code after hitting Cursor limits everything just feels so slow pretty disappointing ngl am i missing something?