Michael Vanier 🇺🇦🇺🇸🦅

DSLs aren't optional anymore they're the supervision gate you're missing

your AI safety strategy shouldn't rely on prompt engineering if you're using JSON as your AI-to-execution interface, you're one hallucination away from disaster

@dougboneparth I did that with three hash cakes once, eaten a half hour apart. I learned my lesson.

At some point in your life you will eat an edible and think it’s not working. It’s very important that you do not eat a second one.

@yminsky @eestokesOSS We’ve been told over and over again that verification techniques are too cumbersome for people to actually use. But when an AI is generating the code, it’s not too cumbersome anymore!

@eestokesOSS Proof oriented techniques that seemed bad from a cost-benefit ratio perspective in the past may very soon become much more practical.

Not news, exactly, but an interesting observation about our rapidly changing world

@MIT_CSAIL @racketlang His book (SICP) had a bigger influence on my career than anything else I’ve read. I literally owe my job to that book. I got to meet Gerry and his wife Julie, and they are both lovely people. Happy birthday indeed!

Happy 79th birthday to Gerry Sussman, the MIT prof. who co-wrote "the Wizard Book” (Structure and Interpretation of Computer Programs) w/Hal Abelson & Julie Sussman in 1984. Read it for free here: bit.ly/4jtzw7g

@yminsky It seems like a very natural extension to first glass modules.

So good

@Caltech Congratulations Adam!

Congratulations to Adam Wierman, Carl F Braun Professor of Computing and Mathematical Sciences, on being named a 2025 ACM Fellow. awards.acm.org/fellows

@ClueHeywood He sure picked a hell of a day to give up amphetamines.

Sitting next to a couple obvs on their first date, late 20s, and he won’t stop talking about “Airplane” (1980), she’s clearly not interested, so he’s doing bits from Airplane to see if THAT breaks the ice. Just did “surely you can’t be serious…” I have to admit his game is fire.

@LawrPaulson Looks awesome!

Nice to see it in real life

@code_report @Max_ErinMills “Hosing a meetup?” That’s a meetup that will definitely “take off”!

If you are in Toronto on Jan 26 and interested in array lanuages, @Max_ErinMills and I are hosing a meetup 🥳 meetup.com/programming-la…

@alpha_convert Who?

@ThatEricAlper Nothing worth having comes without some kind of fight / Got to kick at the darkness till it bleeds daylight

@lady_valor_07 Pasta the kitty!

You have to name him the last thing you ate

@stevekrouse @JustDeezGuy

"What is a 'console'? Why would I want to 'log' to it? What is a 'log'?" Very good questions from my newest student

@mathladyhazel Visual Complex Analysis?

This sentence.

@oxcrowx Essentials of Compilation by @jeremysiek is a great book with an unusual (and good) approach.

> Dragon book How you know someone does not develop compilers.

tell me one

@opdroid1234 @JustDeezGuy @SpeakezTech @jeremysiek’s compiler book is an easier on-ramp to the same material.

@JustDeezGuy @SpeakezTech I have an easier path for folks - read Dybwigs three implementation strategies for scheme, read compiling continuations, read Ghuolums nano pass paper, study chez codebase and tada ...

Hot take: if you have to use LLVM, you don’t actually know how to write a compiler. You know how to snap LLVM parts together. Today, that isn’t anywhere near enough. @SpeakezTech and I have spent months of afternoons and weekends talking about the rise of MLIR, just to name the most obvious example. Why is LLVM just a “dialect” in MLIR? Why is MLIR a good idea, now 7 years old? What is a good one-sentence description of MLIR, and why is @SpeakezTech a perfect Southern gentleman for not rolling his eyes at me every time I say it? And finally, if MLIR really does reduce to that one sentence, why does it exist? I have answers for all of these. I submit that, if you don’t, you don’t know how to write a compiler.

@JustDeezGuy @tetsuo_cpp @SpeakezTech I’ll bite: what are some of the “simpler, better implementations in languages no one uses”?

1. LLVM is an MLIR dialect because some targets you want to hit with MLIR are still best achieved by delegating to LLVM (as of this writing, WASM tends to be such a target). 2. MLIR is a good idea because LLVM is sufficiently old and monolithic that it's brutally hard to do e.g. polyhedral optimization with it, and that's increasingly important when your target isn't a CPU (GPU, FPGA, dataflow processor...) 3. MLIR is "a nanopass compiler framework and dataflow manipulation framework written in C++," which is a good idea for dramatically increasing the reach of both kinds of frameworks (which already have much simpler, better implementations in languages no one uses). 4. @SpeakezTech has the patience of a saint, and is far too polite to point out the myriad ways point 3) is an oversimplification.

Newton didn’t discover gravity at Cambridge. He did it after leaving due to the plague. Einstein didn’t discover relativity in academia. He did it while working a low-effort white collar job. Maybe relentlessly grinding within the system is incompatible with genius.

@alonsosilva @miniapeur

Honestly mom what do you think?