iiinova — where your digital self keeps evolving

Vish Abrams @vishvananda.bsky.social · 10/09/2026

2/ With post-run optimizations I'm within 50% of the performance of gcc/clang. Further runs will hopefully bring that down. Lot's to learn here about long-horizon agentic work. See all the details in medium.com/@vishvananda...

medium.com

Agentic C++ Compiler Redux

From Self-Hosting to State-of-the-Art

001

Vish Abrams @vishvananda.bsky.social · 10/09/2026

1/ Update from my agent-built c++ compiler project: I've had multiple end-to-end runs that reach self-hosting with no-human intervention. One was good enough for me to completely replace my reference implementation.

100

Reposted by Vish Abrams

Dan Luu @danluu.com · 21/08/2026

There's no reason for software to be slow anymore: danluu.com/perf-opt/

The other day, I saw a viral tweet saying that people talking about how LLMs are causing slow, bloated, code are going to eat crow once they re-write everything in super-optimized assembly. We're not quite at the point where we want to write everything in assembly, but some variant of what Nolan Lawson said about testing, you can choose how many bugs you want now, which I less eloquently noted here, is becoming more true for performance.

In response to a comment in my last post that the cost of formerly specialized performance work has dropped by many orders of magnitude and performance work that used to require a person or team that had a rare set of skills can be done by anyone who can type a few sentences1, which means that you can do all sorts of optimizations that used to be too expensive to be worthwhile for all but the largest scale or most lucrative projects, Marc Brooker responded with

Completely agree with your closing point. Dynamic custom software, fitted to a particular workload rather than a class of workloads, seems like a very likely outcome. (Which comes with all kinds of fun risks and opportunities of its own). Kind of reminds me of FFTW. And a ton of weird old demoscene techniques which were all about being super fast and small on a very particular problem (and often very particular hardware). For example, I remember a demo that re-used its code as textures to get great cache locality.

And Michael Malis has noted

There’s been a meme circulating about how AI doesn’t help because “code was never the hard part.” I think that’s true in some domains, but in others, writing the code absolutely was the hard part. JIT compilers are a great example of that. For many pieces of software, a JIT compiler would help a lot with speeding up the code. The rarity of JIT compilers makes me believe that implementing a JIT compiler historically was too difficult for it to be worthwhile. LLMs have lowered the barrier to entry and made it much easier to writ…

3844

Vish Abrams @vishvananda.bsky.social · 22/08/2026

The only way I was able to actually get there was to have the agent rewrite from scratch focusing on perf from the beginning. With a good foundation, 20 hours or so of micro-optimization in a loop got compile times down to state of the art.

010

Vish Abrams @vishvananda.bsky.social · 22/08/2026

I spent (far too much) time optimizing my vibe-coded c++ compiler. I learned that agents in optimization loops are amazing at micro-optimization and absolutely terrible at macro-optimization. Weeks of optimizing with a lot of guidance was barely able to get within an order of magnitude of gcc/clang.

110

Vish Abrams @vishvananda.bsky.social · 14/05/2026

9/ I wrote the whole thing up in four parts, including the repos if you want to run your own agent through the suite: medium.com/@vishvananda...

medium.com

I Spent 2 Billion Tokens Writing a C++ Compiler So You Don’t Have To

The Saga Begins (Part 1 of 4)

010

Vish Abrams @vishvananda.bsky.social · 14/05/2026

8/ My current read: agents handle tedium well. They also take the easy path. Humans need contracts and gates that make that path land in the right place.

100

Vish Abrams @vishvananda.bsky.social · 14/05/2026

7/ I ended up building the harness around the compiler: trace summaries that fit in context, tight test output, a tracker, a ratchet for old patterns, a Boost bug pipeline, and a performance gate based on instructions retired and memory use.

100

Vish Abrams @vishvananda.bsky.social · 14/05/2026

6/ Micro-optimization did not fix that. The useful fixes were structural: split the template layer, remove text fallbacks, thread typed data through, and force each hidden case into the tests.

100

Vish Abrams @vishvananda.bsky.social · 14/05/2026

5/ The hard lesson came after the compiler worked. I had optimized for correctness and left structure underspecified. The agent filled the gaps with strings, fallbacks, special cases, and one 30,000-line semantic file.

100

Vish Abrams @vishvananda.bsky.social · 14/05/2026

4/ I kept going and built the missing half of the course: 28 more assignments that take the compiler from PA9 to self-hosting C++. Each assignment has instructions, a reference implementation, and acceptance tests.

100

Vish Abrams @vishvananda.bsky.social · 14/05/2026

3/ The result surprised me less than the amount of steering it needed. A model that can write C++ on the first pass has been around for a while. A model that can run the compiler, read the error, and try again for hours feels different.

100

Vish Abrams @vishvananda.bsky.social · 14/05/2026

2/ I got curious. I asked it to finish the remaining original assignments. It did. Then I packaged the course into a repo, pointed agents at it, and watched frontier models work through the whole nine-assignment gauntlet with little steering.

100

Vish Abrams @vishvananda.bsky.social · 14/05/2026

1/ I got nerd-sniped by a C++ compiler course 13 years ago and quit when the next step turned into a tedious parser refactor. A few months ago I handed the old code to Codex. It finished the refactor in about an hour.

221

Vish Abrams @vishvananda.bsky.social · 24/04/2025

I really enjoyed digging into this topic. The concept from @kball.llc of essentially unlimited demand for software really changes the way you think about the future of software engineering.

021

Vish Abrams @vishvananda.bsky.social · 10/04/2025

Heroku has an mcp server! youtu.be/7udDus8zv-g?... anyone up for vibe operating?

youtu.be

Introducing the Heroku MCP Server

YouTube video by Heroku

110

Vish Abrams @vishvananda.bsky.social · 07/12/2024

This is great. This aligns with what I’m thinking for 12 factor.

000

Reposted by Vish Abrams

Chris Peterson @ca-peterson.com · 06/12/2024

Totally suggest reading the blog (blog.heroku.com/next-generat...) rather than slides, but sharing slides from my re:invent talk about the next-generation of Heroku for anybody inclined. speakerdeck.com/ca_peterson/...

speakerdeck.com

Heroku Accelerates the Path to Cloud Native

A talk from re:Invent 2024. See https://blog.heroku.com/next-generation-heroku-platform for more info.

031

Reposted by Vish Abrams

Betty Junod @bettyjunod.bsky.social · 03/12/2024

When in Vegas 💜 #Heroku

2336

Vish Abrams @vishvananda.bsky.social · 03/12/2024

I’m excited about the proposal to add an identity factor. Workload identity is generally a huge PITA for developers

110