Beyang
6.8K posts
Beyang
@beyang
Building @ampcode @sourcegraph
San Francisco, CA Katılım Eylül 2008
1.9K Takip Edilen11.8K Takipçiler
@beyang I bet a caveman would loooove a dump truck can you imagine
English
Something like this study has been published for every generation of LLMs (from autocomplete to RAG-chat to agents). A few things continue to hold true:
- New tools require new training. You don't plop a caveman into a dump truck and expect good results.
- That's not to absolve the tool builders from all responsibility—good tools guide the user to good outcomes.
- That being said, there's going to be large differences between the top 0.1% most effective (not necessarily the highest token consumer) users and the median user.
- New tech creates new problems. The question is whether it is better to embrace new tech and figure out how to mitigate the problems or to avoid altogether.
Charles 🎉 Frye @ GTC@charles_irl
We find that the adoption of Cursor leads to a statistically significant, large, but transient increase in project-level development velocity, along with a substantial and persistent increase in static analysis warnings and code complexity. arxiv.org/abs/2511.04427
English
Yes that is all true, but what I'm saying is that where we're going is gonna make how folks currently do things (agile, xp, whatever you wanna call it) look like waterfall in hindsight... Maybe someone will come up with a fancy zeitgeist-y name for the new thing, but perhaps it won't even have a name, since it feels quite possible that we will do the full circle as one person with one tool.
English
@beyang Ser, that was called “waterfall” model, it’s like 30+ year old and even it’s introductoy paper concluded that it’s not functional. For the past 20+ years everybody was doing itwrative agile, which is exactly what you just proposed as a novelty.
English
This behavior seems to map roughly to the intuition I have for the recent GPT vs Claude model families in coding. The latest GPT models seem quite good at reasoning and complex planning where latest Claude now feels (relatively to GPT) like it’s more narrowly pattern matching within a lower dimensional sub space of code (most notably frontend web dev for me). Within that subspace, Claude is better, but everything else, GPT (deep mode in Amp, recommend with ^3 reasoning).
Lisan al Gaib@scaling01
GPT-5.4 is special LisanBench: GPT-5.4 vs Opus 4.6 vs Gemini 3.1 Pro
English
The delight of novel experiences will be shadowed by nostalgia for what we leave behind. It’s normal. Pour one out for the Old Ways, then move on. Figure out the new world. Do things. The New Way is learned through doing.
There will be highs and there will be lows. Great movements are built on emotion, not always positive ones. There will be those who lean into the negative, and the world has no shortage of demagogues who will gladly trade you more cortisol for influence and power over you.
You decide which path you want for yourself. A big part of that is choosing your community. For me, I want to be among those who are leaning into the delight and joy and wonder of it all. Someday, you'll tell your grandkids: I was there when the world changed. And I set my hand to the stone to build the one you’re living in now.
Mo@atmoio
I was a 10x engineer. Now I'm useless.
English
Endorse. The white pill here is that in the before times, every human dev was neck deep in the code mines, so only a small amount of attention could focus on the architecture and systems design. Hence the state of legacy codebases. Now, clean code feels within reach!
Mario Zechner@badlogicgames
your biggest enemy is still complexity. it's also your agent's biggest enemy. but it has no holistic view of your code base, so it keeps adding complexity. and you think that's how it's supposed to be, because the clanker shat it out, and you don't know the stack. glhf!
English
@Awiliii We touch upon a lot in these conversations: ampcode.com/podcast
English
It's interesting how the evolutionary path of the GPT and Claude model families influenced their relative strengths in coding today. Claude was first to robust tool use, whereas GPT opted to focus on reasoning. So the failure mode of Claude is overeagerness to use tools without thinking things through, while the primary complaint with GPT is waiting for it to think.
Initially, robust tool use mattered more, but now it seems OAI's bet on reasoning has paid off, at least in my experience. Remains to be seen how easily Claude can catch up here.
English
@beyang how do you review diffs you didnt write tho? reviewing your own ai output is one thing. reviewing someone elses ai output where you dont know the intent is way harder
English
Think deep, think different.
Amp's oracle is now powered by GPT-5.4, and the oracle is now available to deep mode (still GPT-5.3-Codex).
Think of it like this: thinking can be a model attribute, but it's also useful as a subroutine (the oracle). Think about it!
Amp@AmpCode
There's a new oracle in Amp: GPT-5.4. ampcode.com/news/gpt-5.4-t…
English
Remember that in times like this, doing >> overthinking
Thorsten Ball@thorstenball
You scroll through here and it's • more layoffs will follow • no one's hiring right now • there will be more software than ever • there will be more software engineers than ever • software is dead • this is depressing • i'm having more fun than ever No one knows.
English