Kit Ledru

Making one full stack, AI-native but centered on you @1i__is

[Kit Ledru @kitledru]

these people are all amazing at what they do
some of them have been invaluable in my leveling up.
can't be grateful enough for all the amazing work that they share

joined! 🤝 starts in 5 days
and maybe we can run nice demos directly in solveit! : )

[Shreya Shankar @sh_reya]

Building AI products is hard. But it's getting increasingly popular!

I'm really excited to share that my friends and I are putting together (the best) lecture series on AI Product Engineering this summer!! We've got an awesome lineup of talks spanning data, evals, and UX. With more to come.

The lecture series is completely free! And ~2k people have signed up already even though we haven't posted on social media yet! I can't wait. Join us and sign up: https://t.co/5DWcm4va5m

[Kit Ledru @kitledru]

totally! new expectations emerging.
https://t.co/NgKEOOC9tP

> for many things that would have been libs/mods yesterday, now I may do APIs for easy and standard interop/consumption by AI and other pieces, while decoupling codebases more granularly; so interfaces are less often in-app, I shift modularity to the network/HTTP level.
Apps become smaller, more KISS (helps a lot with AI context). I often say I'm moving "from apps to caps", where the control plane (as in MVC) is more about composing capabilities than building some complete UX for each app. Feels more like a giant terminal and piping, very flexible/ad hoc. Agents love to leverage such envs I think.

[1i @1i__is]

> Those standards are different.

I'm also exploring new codebase structures and conventions for AI. I find it's a lot of little things.

For instance,
I often run a glossary-style "sig.md" file that shows all public signatures in simple structured md reproducing the fs (with docstrings, additional notes, maybe LOC#, etc). Helps with modularity, orientation, arch, structure, expectations/contracts ("this input gives you that output" is often all you need to know, any more is noise).

Yes, you can use tools like ast, tree, some doc lib, etc., but that's indirection and as many failure points.
A simple text file seems often better; ideally containing nothing else than some pre-computed + processed output of said tools, the point being to ensure *some determinism* by removing the 'agentic' step of choosing what to run. So whatever the agent "often does", I try to make more robust in some way.

This basic sig.md may save like 90-95% of additional context for most orientation. More importantly, it lets the AI (and me) reason about the code from the arch/design standpoint much more cleanly, less noise.

It's really just a tiny example to drive my point.
The more I use AI, the more I discover many such tiny "tricks", which have disproportionate effects for the very little effort they require.

And in the end it seems to paint not alien but certainly different repo structs, with new expectations. For instance, for many things that would have been libs/mods yesterday, now I may do APIs for easy and standard interop/consumption by AI and other pieces, while decoupling codebases more granularly; so interfaces are less often in-app, I shift modularity to the network/HTTP level.
Apps become smaller, more KISS (helps a lot with AI context). I often say I'm moving "from apps to caps", where the control plane (as in MVC) is more about composing capabilities than building some complete UX for each app. Feels more like a giant terminal and piping, very flexible/ad hoc. Agents love to leverage such envs I think.

New "shapes" emerging, as Codex loves to say. A much flatter and wider buffet of simpler modules, much less vertical/monolithic complexity.

I must admit this fits my taste, so perhaps I'm biased and others are doing great with completely opposite principles!

But the point stands: I'm building differently, with AI a first citizen from day 1, and humans able to leverage it to great UX. It's a long journey with many stops, but I think that space is where the core paradigm shift of making software will happen. Structure, expectations, how we think about interop and modularity, how not just an API but a whole environment/world usage & options can be self-documenting, plainly obvious just by observing it.

[Kit Ledru @kitledru]

The good news is those GPUs will only ever run better models, so their utility currently increases with time.
That's the rarest sweet spot in any computer part cycle, I feel we get those like once per decade in a tiny few categories. And GPU is a big cat!

If I could afford it I'd make the same investment rn. I'm buying GPUs too at my level, 3090s (b/c nothing yet between that and 6000 Pro imo).
I can't help but think the "AI ensemble" is coming, because it's just sane infra; I don't see how/why you'd avoid granular routing to minimize cost. And now with models like GLM-5.2… omg. We shall experience sci-fi. There's a 1980s/1990s feel I love about this!

(It's not like resale value isn't great anyway if we make a mistake :D rn GPUs might even beat Apple or Thinkpads at their prime.)

[Jeremy Howard @jeremyphoward]

Wow.

@Zai_org GLM 5.2 is a marvel! It is *at least* as good as Opus 4.8 and GPT 5.5. It's super fast, inexpensive, and not too verbose.

It responds with nuance and judgement, & handles long context VERY well.

I've never experienced an open weights model like this before.

[David Gwyer - AI & ML Researcher/Engineer @dgwyer]

ElevenLabs have just granted API access to their Music v2 model. 🎶🧑‍🎤🥁🎶🎸🎷

Here's how you can set it up in SolveIt! All you need is an ElevenLabs account/subscription, and an API key:

https://t.co/QOJuQpUV21

[Kit Ledru @kitledru]

@Vishal_anton16 @antirez @mitsuhiko paying a team of 30-70 people world class salaries is negligible

not to mention the appeal of the mission, talk about an underdog story

but the will simply isn't there by anyone who could fund that. which is strange to put it mildly.

[Kit Ledru @kitledru]

living dangerously

[htmx.org / CEO of #Schizophreniamaxxing (same) @htmx_org]

about to make myself unimaginably furious at the state of web interfaces biab ttfn

[Jeremy Howard @jeremyphoward]

Also, those who focus on using AI to help improve the skills of themselves and their teams will be diamonds in great demand, since they will be the rare A++ players in a sea of mediocrity.

[Kit Ledru @kitledru]

> just wave a magic wand and get to feature and performance parity overnight (please don’t say AI, this is a serious discussion)

😂 real
very good piece, well worth a read from a swe/systems pov

[Greg Wedow @gregwedow]

@dreamsofcode_io @yinebebt_ It's because embedding the existing tsnet in anything but a Go project requires pulling in the whole Go runtime, which is annoying to interop with from a lot of languages.

They wrote about it here

https://t.co/i803Y5EEAK

[Kit Ledru @kitledru]

if it's anything like solving a math-like problem, i can definitely relate to the uncomfortable mental gymnastics haha (i'm more of a down-to-earth iterative mindset usually, abstractions come on a need-to basis for me).

I can't wait for those videos! You have a knack for expressing SWE ideas & concepts that just speaks to me, so I can't imagine a better way to dive into a foreign language :D

[Kit Ledru @kitledru]

Completely agree on correctness vs speed. And on that "divide" indeed, haha.

Haskell huh, nice. As in a more formal-ish type of programming I suppose? Man, I really have to dig into Rust next chance I get! (it's long overdue)

If you happen to have a resource that shines at explaining this idea "that once you lock down your abstractions and types, correctness falls into place", or if you have videos on your awesome YT channel that talk about that, please don't hesitate to wall-of-link and @-me, anytime! : )

[Kit Ledru @kitledru]

the next book you should read after The Lean Startup
https://t.co/qLHzNlQUZA

[Kit Ledru @kitledru]

But is that era truly over?

Because writing is one thing, but reading is another, and that was one of Go's main points too (easy for teams, as code looks about the same for a given spec regardless of who writes it, etc).

I wouldn't trust my debugging life to LLMs being perfect at it. And i don't think LLMs make rookie programmers suddenly senior level. Like, "You don't know <language>" right? AI is probably a formidable accelerator, granted (I'm living it as a mid-level programmer); but not a silver bullet that makes you suddenly get why this is Elegant JS while that is spaghetti.

And Rust, if IIUC, isn't particularly easy not to mess up badly (probably easier than C/C++, and I get there are safeguards and foolproof features etc; but LLMs will gladly ignore things too, human must remain in charge of the critical parts).

Idk. Anyhow, super interesting to watch the evolution of SWE colliding with AI.

[Kit Ledru @kitledru]

@eliebakouch 😂😂😂 it really can't help it

[Kit Ledru @kitledru]

if it teaches users even without being asked, could Nemotron 3 Ultra be a good fit for Solveit? 🤔

[1i @1i__is]

@_imdawon &gt; taught me

now *that* is good post-training

@NVIDIAAI is doing something right!

[Kit Ledru @kitledru]

It gets better when you feed them your raw empirical data, the unaltered results of having skin in the game. If I may say so. It "grounds" the LLM best, ad hoc, with the good and the bad.

But those datasets take human time to assemble. That's OOD for hyperhypers, but it's how we'll get there eventually.

[Kit Ledru @kitledru]

@TheAhmadOsman @vcsoares yeah totally, but for one-off questions without context, it should be doable… idk.

not for serious work certainly.

[Kit Ledru @kitledru]

@TheAhmadOsman @vcsoares oh 8? no 4?

[Kit Ledru @kitledru]

@vcsoares @TheAhmadOsman 350 GB for NVFP4 (read that earlier, haven't checked)
so 4x should do it

just $40k worth of GPUs :)

[Kit Ledru @kitledru]

Very soon everyone will be able to experience solveit first-hand
I can't wait for it! 🙌

[Pol Avec @pol_avec]

There's absolutely nothing like close-reading in SolveIT.

Being able to stop and ask questions at any point lets you engage more deeply and branch of the read at any point.

Today, reading Naur's paper "Programming as Theory Building" I got into topics like:

- @ylecun world models and @RichardSSutton's learn-from-experience are basic requirement for AI participating in Naur's Theory of software dev
- why current LLMs have a hard time being innovative
- Naur's framing of software building is a perfect parallel on Daoism's "the Dao that can be described is not the real Dao"
- why streamlined CI/CD might be killing your team understanding
- There can't be no absolute sw "complexity" metric bc it depends on the unique Theory of each program