@dubbydev Is it bad that I usually do that everytime a new model comes out??
>> refactor this entire page, the previous agent had no idea what they were doing
@jkudish Major updates to models always feel like you’re starting from scratch again, and need to relearn the model’s quirks. Sonnet 4 had a ton of weird issues versus 3 even though it was a lot smarter.
Claude Opus 5 made this beautiful painterly world in a single HTML file, everything made procedurally
It implemented a realistic grass and wind field simulation so that the millions of grass blades on-screen react to the wind
@unclebobmartin My workflow has been getting a proof of concept going, then iterate aggressively with the working demo, then go back and fully rearchitect into a clean architecture. So far it’s been very effective even for projects with 50k+ lines of code.
hot take: coding CLI tools will be mostly dead by the end of the year replaced with higher level abstractions focused on the tasks, not explicit code changes
How many pieces of software do you use daily that were designed for more than one million people? I’m fairly confident the answer is every single one. Does this matter? Yes.
Everyone thinks differently. Some prefer going for a long walk, others prefer talking it over with a friend, and some like talking to a rubber duck (I don’t judge) to just scratch the surface. If people are so different, why are we all stuck using the exact same software?
In the ages before AI we didn’t have any other option. If you didn’t know the CEO personally you were out of luck. Now that AI can compress weeks of human work into minutes the math flips. The time to develop software is almost zero, and the barrier to entry is practically nonexistent. We’re living in a world where practically anyone with an idea can build an app within hours or days.
Now pair a developer with a 20x Claude Code plan, and almost anything is possible. I have built so many random things purely for myself. The reminders screenshot below? Custom built, I was not going to miss trash day again, that was an all time person low. A custom agent task executor? Perfect for auto-resuming my sessions when the limit resets.
Neither of these products would be very useful to other people, because they weren't designed for them. They were designed with only one person in mind, myself. That level of customization offers a level of optimization unmet with conventional software products. Imagine a Jira/Asana perfectly designed around your company’s workflow. That’s where AI unlocks even greater velocity.
I think in the next 2-3 years more and more software will get written for audiences counted in the hundreds, not millions. What is blocking you from building some apps exclusively for yourself?
Shoutout to baby Norman for being way too adorable.
@sickdotdev I have no actual evidence either way, but anecdotal evidence shows that context rot can seriously degrade the response quality. Keeping chats short and sweet seems to help.
@IamAroke I used to think that until one of my cracked engineers spun up a cluster and reduced costs by 10x, while at the same time dramatically visibility into workflows and usage. It obviously can be misused, but it’s crazy how much AI lowers the barrier to entry.
@jimjamscodes@evisdrenova I wish it was possible to fork languages like we can for blockchains. It’s probably possibly with enough money thrown at Fable/5.6 Sol. Custom languages for specific domains would be interesting.
@theo Big fan, definitely feels stronger than Fable. Will be putting it through its paces this weekend. Been trying more graphical tests recently and it did great with my 3D Earth benchmark.
https://t.co/nLfWmRr5CA