It’s not overengineering if it works. The thing to care about is over featuring or building features that are not worth their human cognitive cost, or just engineering too fast.
It’s only gonna become more of a problem in the future.
Replying to @miguelbranco80 @k74l3d
vibes. I like claude.ai for UI work and chatgpt.com for research. For coding, I'm sol/terra/luna/glm rn. I can't really clock fable/opus and claude code drives me nutty.
1
1
85
Replying to @k74l3d @miguelbranco80
I haven’t, I’m a very minimalistic dev. I use pi with OpenAI and glm, and claude.ai/chatgpt and copy paste 😂
1
1
54
Replying to @H1manshuSharmaa
Yeah but it’s not my code, it’s the ai. I don’t read it anymore, I read the parc usually
1
1
13
Replying to @ProgramWithAi @H1manshuSharmaa
HTTPS://parc.yolo.scapegoat.… is our research institute and a more interesting version of the GitHub work.
1
1
41
Now that models are superhuman in basically every dimension related to programming, trying to keep a “traditional” grasp of our code (which was flawed anyway) is just capping the model’s capabilities. And if you disagree, just wait for the next generation then.
Meaning: we have to come to terms with the fact that we will reach (imo have reached) a point where we struggle to understand good code, and spend a lot of effort to turn it into puny human code. The same way I don’t nitpick a compilers assembly in 99.99999% of the cases, i don’t nitpick architectural decisions of the model.
However, I steer it into architectural dimensions, and I think my deep interest in the structure of code (as in mathematical structures) and abstraction allows me to find the right keywords.
For example: this is a window manager like UI to do ggplot style manipulations and graphs, using a clever “anything is a semantic object” technique. Look at the code and you’ll see the core engine of all of this is about 150 lines of code. _superhuman_ . But all it took me to summon this, which could have taken ages of “human brained engineering”, was “make me a grammar of graphics tiled UI using presentation based UI principles (ciccarelli 1984), using a brutalist retro look”
claude.ai/public/artifacts/5…
Think about it, if adding a lease mechanism to catch straggling transactions and the 200 lines of code that come with it feel like too much, how are you going to feel about the 200 lines it’ll take to remove it, and the countless iterations to fix straggling transactions once you run at scale?
76
Replying to @ProgramWithAi @H1manshuSharmaa
2
1
45
github.com/go-go-golems/go-g… goddamn
138
Replying to @travisennis
I posted an example in a side comment, but here’s the introduction of a owner ref and lease (and maybe more, but GitHub is impossible to browse on mobile) pattern which might feel like overengineering but pay off almost immediately. github.com/go-go-golems/go-g…
1
1
93
Replying to @H1manshuSharmaa
I’m on mobile but here is a common owner ref + lease pattern that it added retroactively (this replapi thing has grown over months) because I had a stuck repl session and couldn’t figure o it where it came from. Had I had a lease and an owner ref, it would been a single log line and it would have been handled : github.com/go-go-golems/go-g…
So now when I see Lease and Owner and Capability, even for a menial feature, I’m like “sure bro, I get it”. But… I had to sit down and learn these patterns and how SOL implements them. Thankfully it’s a great teacher.
1
1
21
Replying to @H1manshuSharmaa
I posted an example in the side comment. I have many many more in my current work on GitHub.com/wesen but at this point I think my architecture is out of reach for most newcomers.
2
1
52
Replying to @miguelbranco80 @k74l3d
I never use ultra. It’s too much for my brain. I use sol low or medium, usually. That keeps me plenty busy.
2
2
42
Replying to @ProgramWithAi @travisennis
Here’s a concrete example: chatgpt.com/share/6a5e1665-d…
I wanted to instrument my function calls in my js sandbox. I was thinking of a simple hook that would increment some counter somewhere, but it started designing a probe structure and a set of actions structs that this probe could take and what not. I decided to a) yolo build it and b) follow the rabbit hole further (see conversation and the final textbook) and c) dude everything works and it’s so nice to already be able to assert “this probe won’t modify anything” or see a debug list of what probe has what action type.
1
17
Replying to @travisennis
Many. For example, modelling relations as explicit objects. Instead of putting a pointer to an object, say the owner of a file, sol/terra will often introduce a OwnerRef object that has a set of attributes attached to it (say, a fingerprint of the transaction contents that created said ownerref). That will allow it to log the identify of this references when debugging, and see for example when it changes, something that can easily be missed otherwise. If your system is distributed, this can be the entire difference between “cannot figure a tricky race condition out based on logs” vs “ownerref recreated with transaction hash x” and instafix the bug.
The model doesn’t care that it now has to write foo.owner = ownerref{…} and add a transaction id hash function (which probably already exists).
We only think of it as overengineering because we’re lazy and would think “nah later”. But having sound transactions that turn messy structures into declarative atoms is OP.
Now, when you start in a feature, new models will do all of that at once. Instead of thinking “do I need all this machinery”, think “will I need this machinery later on” and usually the answer is “yes. Damn son, how could you already think of all of this”.
That takes time, if you are first confronted with a domain, and there might of course be stuff you don’t need “in the big scheme of things”. But if you build production software, I usually found 5.6 to be wayyyy ahead of me. It takes time to digest, and I’ll easily spend 3-4 days learning and dissecting an architecture I don’t understand. But it always paid off.
1
2
63
Replying to @miguelbranco80 @k74l3d
Yeah but we are not LLMs and don’t have the capability to transplant our colleagues thinking into our brain. Removing information from the context window will only help in as much as it might remove “load” from the attention mechanism. But our models are good enough that it usually doesn’t matter.
Even then, thinking of an agent (ie making you draw an analogy towards how we think as individual humans, ie real agents with closed context) will make you miss better opportunities to for example reduce the context window size / attention load. Thus antipattern.
1
2
44
Replying to @k74l3d
I’ll usually have Terra/luna on low or medium thinking do the work, then switch to high or xhigh thinking, do the review and the fixing, and _leave_ the whole thing in the context window, might as well continue with that info. Context windows don’t degrade as quickly as they used to.
15
Replying to @k74l3d
They’re basically just context window management techniques, and we lose a lot of flexibility and caching opportunities by pretending they are autonomous.
One concrete example: Assume you have a context window filled with writing file a, including all the thinking and tool calls and skills used to get there.
Why strip all that information away by giving it to a fresh “agent” which basically amounts to changing the system prompt, and potentially another model. It’s akin to doing a code review without the original developer explaining their thought process and tools they used.
Within the original context window, just switch to say “you are now big kahuna reviewer, review the work” which is usually more insightful, _and_ is able to answer “why did this mistake happen? Which skill is wrong? Which tool can be improved”. Models are good at “cosplay”.
1
2
55

