Sabitlenmiş Tweet
dev
2K posts
dev
@Devinda_me
variable name consultant. software engineer + dota + anime stuff can go here. he/him.
Katılım Ekim 2020
386 Takip Edilen126 Takipçiler
@m7l5_dz I’ve a release skill that I run before any upgrade to check if my extensions will get messed up with. Especially useful when @badlogicgames goes on his weekend refactors
English
Ayo the new pi.dev version broke my vibe coded extension (sessions sending messages between them, previously a message would wake up the agent, now its not). Since I have no idea how the extension is built I now have to vibe code the fix as well. Thanks @badlogicgames :3
(but for real thank you. I wouldnt have extracted that much benefits from LLMs without your tool)
English
@infinterenders Wow a ruby simp now ? Though on a serious note this is prolly why dhh speaks highly of using llms with rails
English
rust/ts have high syntactic entropy. like for every unit of logic (the "what") , you have to provide several units of type definitions (the "how")
so this is high entropy noise for agent
ruby is semantically dense( e.g., has_many :orders) expands into hundreds of lines of latent database logic and method definitions so agent can use compression shortcut here sends a single high-value token to the prompt, and the framework handles the low-entropy boilerplate
coming to ts/rust are infinite ways to structure project . so it creates a flat manifold where the agent has no clear direction, this may leads to hallucinate architecture . rails look like English convention over configuration
i recommend reading this ( aclanthology.org/2024.findings-…) Ruby aligns heavily perfectly with the Vector Space of an LLM
Rails migration from 2012 still looks the same rather then comparing fast-changing frameworks we had today for web they shift paradigms so fast
DHH@dhh
Agents don't need types. They're perfectly capable of pulling off incredible refactorings without. Give them a linter and a test suite, and you have all you need. Token efficiency is where it's at.
English
@mattpocockuk Should come with an eval. This is if you’re using skills as auto fed into agents and not commands
English
Skills should be:
- Concise
- Responsible for one thing, not multi-step
- Composable
- Progressively disclosed
- Harness-agnostic
What else? Or - what did I get wrong?
English
It’s pretty good at touching up a regular photo as well
before and after
sunil pai@threepointone
ok maybe ai art is good
English
@nateberkopec @RhysSullivan I’m curious to see how this goes, traditionally this should be an edge case and thus would require attention. However at scale this can happen more frequently so, I wonder how this will go
English
@RhysSullivan Strongly disagree. Those are highest value, highest risk commits, so you can afford expensive human tokens. Medium value low risk changes will be automated far earlier than fixing your SEV.
English
Agents solving on call is probably going to be where the first auto merges happen
Solving the pain of never having to be paged again is worth so much
Will start gradually, being paged and seeing a fix prepared, then eventually letting it auto merge small changes & so on
English
dev retweetledi
~I wrote it in Markdown first. Easy marks easy work.~
plannotator.ai
plannotator@plannotator
html spec, activated
English
@Devinda_me What? I heard YouTubers talk about it. What’s the deal.
English
We need start looking at MacOS hardening. By default the firewall is open and anyone who is your WiFi can access your services. Also windows can write things into your disk.
English
@Devinda_me Shooting this out to you tonight. Would love your feedback and or if you want to work on the multi-chat capability From this.
English
Could use something like this in @plannotator where I can multiple concurrent chats about a plan. It has all the right primitives, I need to compose a ui from this. I usually have conversations with multiple points at a time, managing it via terminal is painstaking
English
dev retweetledi
people mislead user experience with performance lot of react features are never meant to be the most performant at all they meant for bring top dx and ux
these tech Dev YouTubers making it slop
With ugly thumbnail and calling this features makes react faster nothing in react makes it faster you never gonna escape the model of react reconciliation is the blessing and curse for react
I’m not going to Talk about single feature of react that is hyped everything here is envolve like spectrum and see browsers never meant to be fast and instant doesn’t meant it’s fast
all react ytbers are in illusion expect few you guys are in illusion of react your hallucinate and slopper then a llm/ agent
react features never reduce work it reorder the work
browsers are never meant for fast they are just distributed hyperlinks
don’t ever think if the code is easier to write it’s fast 10x just imagine how the runtime uplifting things that makes people easier to use
English
@threepointone Feeling slightly burnout and your post resonates hard. It’s not a joke post fr
English
I can’t take your opinion on taste in software seriously if you don’t watch movies, listen to music, read books, get brunch with friends, enjoy baths and long drives, kiss someone under the moonlight, eat a flaky croissant with a bitter coffee…
this is not a joke post.
English
So I got a proof of concept working. Extended plannotator, uses a tool so it could work with other harnesses too ? It works on my pi setup. Might try this on an infinite canvas.
plannotator@plannotator
@Devinda_me Almost there, thanks for being patient and great idea about multiple chats, hadn't considered that.
English
Preview - Plannotator will provide a long running review surface. This will enable you to control and change between multiple reviews simultaneously. It also enables a nicer integration with agents (Pi & OpenCode auto-connect if in the same project)... and other cool things to come on the plan/annotate side.
Here's how I would launch a review on the entire stack of this work. UX after this would enable launching AI reviews across the entire stack and other things.
English
@plannotator One of these days, you’re gonna say why not just put the agent in here too. And phew you’ve built yourself your own harness.
English
@jamonholmgren If I was to get an intern right now, I’d have them own a specific complex domain of my codebase. Have them understand it across the board, user to implementation and take responsibility for its development.
English
My current gut feel is that juniors will start at small companies as the first or second developer hire, for developing in-house software.
They’ll build low stakes apps for those companies, like estimating, automations, websites, task management.
It’ll teach them the strategic layer (since they own the whole stack) in a lower pressure environment. They’ll need to think about all of it, but the projects are small and the user bases tiny.
This is as opposed to being a junior on a larger tech team, where you’re given small tactical slices to work on under close supervision. That route is drying up, I’m sad to say.
These juniors will be affordable to a small business and AI will give them instant productivity that the business can leverage for operational efficiencies.
My son @cedricholmgren is currently doing exactly this route; working at a local fence and deck company as their in-house programmer. He owns probably a dozen apps in 5 different languages and frameworks. It’s amazing. He’s been doing it for a couple years and at some point someone’s going to discover that he’s a lot smarter and more disciplined than his old man and snap him up.
Matt Pocock@mattpocockuk
Tactical vs Strategic Programming, and why I'm nervous for juniors: Good programming involves a mix of tactical and strategic decision-making: - Tactical: on the ground, short-term. The soldier doing the fighting. - Strategic: high-view, long-term. The general planning the war. You need to be a tactician to write good code. To choose the right syntax. To figure out the file structure. To figure out how best to test your changes. But you need to be a strategist to build code that lasts. To design the architecture. To automate away problems. To think beyond today. Agents have eaten the tactical part of programming. When you can pay below minimum wage for code, there's no point going into the trenches yourself. But AI cannot code strategically. Agents need someone at the top of the pyramid to tell them what to do. They need oversight. So, a developer's day-to-day job has become 100% strategy. Long-term thinking, all the time. (maybe this is why I'm so tired all the time now) If you identify as a tactical programmer - a code monkey - then you are out of luck. The job has changed. Personally, I like it. I always preferred thinking strategically about code. If you asked me what my job was about, I'd say 'building apps', not 'writing code'. But what makes me nervous is that we've pulled down the only bridge that brought juniors into the industry. We used to train juniors like this: 1. Give them only tactical tasks 2. Let them build up their strategic experience slowly Eventually, they are a good enough strategist that they are no longer a junior. But what happens when all tactical code is written by AI? What is the point of a junior? We obviously need juniors. We need new lifeblood coming into the industry. We need to leave paths open for extraordinary hires to enrich our companies. But how do we train them? How do you train strategic thinking? These are the questions I'm thinking about. I'd love to know your thoughts.
English
@isaac_ts_way Is there a school of study dedicated to the classification of intelligence ? It seems like no one has the vocabulary to describe intelligence. Like how do I describe why I prefer a certain model without using examples ?
English
None of these coding benches actually capture intelligence and IDK why.
Yes composer 2.5 is great, and scores really high, but when you compare it to actually using Sonnet/Opus day to day there are certain things that Claude does 100x better.
Like yesterday I had opus build a new e2e testing command that reuses containers. It reduced the run time for local test runs from 40s to like 7s.
I initially wrote it with composer 2.5, and it had a bunch of massive issues that made it unusable and glitchy. But Opus reviewed that code and immediately fixed all of the issues with it, now it works perfectly.
Idk how to categorize it exactly, but for anything that’s a lot more complicated, Composer 2.5 does not cut it. It’s not 90% as good in these situations, as suggested by the benchmarks, it’s completely unusable in these situations.
I don’t think any benchmarks are effectively capturing the differences that really matter between coding tasks. at least not in a way that’s easily discernible by the final score
Theo - t3.gg@theo
Oh my god it scored worse than Composer 2! Not even 2.5! And it cost 4x more to run!!! This might be the worst major lab model drop of all time. Llama 4 tier. Insane.
English