
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.

Distinguished Engineer at Google. Previously NASA Nebula Technical Lead during the creation of OpenStack. Opinions are my own.

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.

For those who are following along with my ultra-fast C++ compiler, I've created an interactive demo showing the ~4x speedup vs. Clang: incpption.com The demo does C++ compilation in the browser using wasm. The inception² demo compiles llvm completely in the browser!

4/ I'm going to tackle rust compile times next. All of the details are in medium.com/@vishvananda/how-…

3/ My most recent gpt-6.1-astra run achieved self-hosting for less than $500 in API costs. If you want give your agent/harness a shot, the scaffold is at github.com/vishvananda/cppgm…

2/ The scaffold is instructions and tests for agents that teaches them to build the compiler step-by-step. I've done more than a dozen of autonomous agent runs through the course. The reference implementation takes the best ideas from each

1/ I've been working on a scaffold to teach an agent how to build a self-hosting C++ Compiler from scratch for the past 6 months. My reference implementation can now compile C++ code 4x faster than GCC and Clang

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/agen…

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.

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/i-sp…

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.

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.

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.

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.

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.

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.

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.