Ian Yu

121 posts

Ian Yu banner
Ian Yu

Ian Yu

@ianyu93

AI Engineering is saving my mental health

Toronto Katılım Nisan 2020
101 Takip Edilen79 Takipçiler
Ian Yu
Ian Yu@ianyu93·
@getpy I’m confused by some of the comments on what this is a strong signal to…
English
1
0
1
607
Ankur Gupta
Ankur Gupta@getpy·
Apple hiring with DSPy prominently listed "Deep experience in foundation model-based AI programming (i.e.: using DSPy for optimizing foundation model prompts, for example) and a drive to innovate in this space." jobs.apple.com/en-us/details/…
English
7
17
384
41.1K
Ian Yu
Ian Yu@ianyu93·
@getpy "product manager adds new requirement" LOL Yeah, yeah!
English
0
0
2
45
Ankur Gupta
Ankur Gupta@getpy·
@ianyu93 You can compare with raw prompt text and then show what happens when product manager adds new requirement. Plus show output of the prompt against different models from SLM to Frontier and compare that of a signature. That way you don't finger point at a framework. 2 cent.
English
1
0
2
41
Ian Yu
Ian Yu@ianyu93·
I think after this talk, I wanna take a step back and explain better on "why DSPy" beyond optimization on a blog post.
English
1
1
11
2.8K
Ian Yu
Ian Yu@ianyu93·
@getpy Absolutely. I do want to see how to illustrate that better for people from other communities, although now I think about it, it might become pure comparison rabbit hole as well. Hmmmm!
English
2
0
2
101
Ian Yu
Ian Yu@ianyu93·
Tomorrow I'll focus on my actual and personal experiences as I build many applications with @DSPyOSS, also going to be my first presentation on a TUI :) maven.com/p/f6686b/build…
English
1
5
34
3K
Ian Yu
Ian Yu@ianyu93·
@TelepathicPug @DSPyOSS Thank you! I would definitely prepare more 101 introduction in the future, but it's really a great framework :D
English
0
0
1
38
Ian Yu
Ian Yu@ianyu93·
@MaximeRivest Yup. Human level context engineering is missing out. It’s why I stopped having MCPs by default, and barely have rules/skills if any. I also have never ran out of usage for CC. And I love it that it’s a loose cannon because I already control where it can go wild.
English
1
0
1
200
Maxime Rivest 🧙‍♂️🦙🐧
Look, I have no allegiance to LLMs and harnesses, a better one comes in, I change. I changed often. I saw you all saying Codex is more careful, Claude is a loose cannon. I went to try Codex, again. No it's not even close for me. I don't know why, Claude Opus 4.5 (6) in Claude Code fits me like a glove. My only explanation thus far is that you people are letting it compact itself. Never! I am civilized, I control my compaction.
English
7
1
21
4.3K
Ian Yu
Ian Yu@ianyu93·
I'm vibecoding (true to its definition) because of client needs. And I hate it. Absolutely hate the feeling.
English
0
0
0
44
Ian Yu
Ian Yu@ianyu93·
@theo I code with neovim, so it’s perfect for me :) but I also realize it’s a smaller crowd…
English
0
0
0
40
Theo - t3.gg
Theo - t3.gg@theo·
Calling it now: all these agent coding TUIs are a phase and it will be short lived. Most devs will be back in GUIs and IDEs in a few months.
English
596
84
3.5K
606.7K
Ian Yu
Ian Yu@ianyu93·
I just realized, how much better have models gotten. Certain tricks that I used to do, probably are not as important anymore.
English
1
0
1
111
Prashanth Rao
Prashanth Rao@tech_optimist·
dspy.RLM has dropped!!! I clearly know what my next deep rabbit hole (and suite of experiments) will involve, haha. Thanks so much to the DSPy team for pushing this through, this is a super exciting development for the entire ecosystem 🚀
isaac 🧩@isaacbmiller1

The dspy.RLM module is now released 👀 Install DSPy 3.1.2 to try it. Usage is plug-and-play with your existing Signatures. A little example of it helping @lateinteraction and I figure out some scattered backlogs:

English
6
5
88
6.8K
Ian Yu
Ian Yu@ianyu93·
@isaacbmiller1 My writing is so unpolished that it’s hard to get mistaken 🥲
English
0
0
0
19
isaac 🧩
isaac 🧩@isaacbmiller1·
i get really embarrassed whenever i write something by hand and then look at it and think "yeah someone might guess this was written by AI"
English
3
0
5
688
Ian Yu
Ian Yu@ianyu93·
@jamesacowling I’m always confused by this. If there r too many lines changed, ask for split. Just better standards, better checks, already filter a lot. Comm on issues, PRs. Good software takes time to review and build. Why keep on trying to make it “fast”? Just communicate on trade-offs.
English
0
0
0
206
James Cowling
James Cowling@jamesacowling·
Every senior engineer I know is currently bottlenecked by code review, correctness, and trying to wrangle the output of their team in sound architectural directions
James Cowling@jamesacowling

If you haven't read about The Software Crisis of the 60s/70s you should: en.wikipedia.org/wiki/Software_… Productivity ground to a halt before they developed good abstractions for managing software complexity in software. Without good platforms it'll happen again.

English
58
83
1.9K
239.9K
Prashanth Rao
Prashanth Rao@tech_optimist·
Plenty of awesome tweets by Omar in the last few days on this topic. The model has symbolic access to its own prompt (the REPL alone isn't what makes it "magical"). The RLM has much more structure internally, and it can handle unlimited long context both in input and output space. Quote a paradigm shift when you think about it.
Omar Khattab@lateinteraction

Giving the model symbolic access to its own prompt is key, just like allowing models to “think” before they begin producing output was key. It might be worth thinking about all the ways one can give the model such “pointers”. A standard REPL is not strictly necessary.

English
1
0
2
122
Prashanth Rao
Prashanth Rao@tech_optimist·
Curious what folks think about the need for optimization (e.g., GEPA) if/when RLMs become more widespread - do RLMs compliment optimization? Or does the fact that LLMs manipulate their own context mean that optimization is somehow less important/useful? Can the two co-exist?
English
3
0
34
5.9K
Ian Yu
Ian Yu@ianyu93·
@lateinteraction I find that, because conversation/chat is many people’s first major interaction with NLP, it’s hard to imagine beyond its application when that’s just one of the many possible ways to frame tasks.
English
0
0
0
69
Omar Khattab
Omar Khattab@lateinteraction·
Many folks are yet to grasp the specification problem. > Wouldn't AI write better code than us? "Better" for what? Or with respect to what exactly? I'm building software because I have a specific vision/intent/need. How should that intent be specified in the first place, so that a specific piece of code is better or worse with respect to it? I'm saying: for many needs, conversation is not an efficient medium of expressing *what* I want. AI having excellent taste or horrible taste has no bearing on that.
Omar Khattab@lateinteraction

I will be so disappointed if the way we build software with AI remains vibe coding rather than a genuinely higher level of abstraction. I want to express the “code” of the system—but shorter and more pleasant—not manage an agent to write 100,000 lines of low-level slop for me.

English
25
14
140
14.4K
Ian Yu
Ian Yu@ianyu93·
@tech_optimist is my inspiration to write more, and after reading this piece on writing, it's making me think more about how to formulate my ideas into more concrete outlines better. I know my content will be scattered, but hopefully I can have another piece after NYE :D
Prashanth Rao@tech_optimist

This year, I've gotten a ton positive feedback on my writing from people in all walks of life: engineers, academics, CEOs, VCs and even non-technical folks. So I put some thought into why that is, and externalized my thinking into this blog post 👇🏽 1/3 thedataquarry.com/blog/a-framewo…

English
0
0
1
58
Ian Yu
Ian Yu@ianyu93·
@CodeByNZ Grpc endpoints. Autocorrect 😓
English
0
0
0
12
Ian Yu
Ian Yu@ianyu93·
@CodeByNZ If you use nvim, actually you probably don’t even need to be a regular user, just a bit setup, I reco github.com/mistweaverco/k… It’s light weight (just a file), I can run some rest and grow endpoints before Postman even starts up :D
English
1
1
3
256
NZ ☄️
NZ ☄️@CodeByNZ·
any alternative to postman ?
NZ ☄️ tweet media
English
1.5K
61
2.3K
463.4K