Win Wang

Founder @ ParaQuery (YC P25) | Data processing for the GPU era.
of FP, PL, physics, quantum computing, and trolling AI with math.

[Win Wang @Winium]

Indeed, AI code is 90% dogshit, even now. Correctness is not "pristine". AI routinely elides newlines, as if it's playing code golf. Massive files which can clearly be refactored, etc.
Studies even show LLMs do better with smaller files, so they're even handicapping themselves.

[Dmitrii Kovanikov @ChShersh]

> AI-generated code looks pristine

lol, lmao

[Salar Rahmanian @SalarRahmanian]

The Scala Summit 2026 recordings are out! 🎥

All 11 talks from our October 19 conference in San Francisco: AI and agents, capture checking, Chisel, ZIO, GPUs for type inference, and more.

https://t.co/FDRvqRdrAj

#Scala #ScalaSummit #BayAreaScala #SFScala #FunctionalProgramming

cc @BayAreaScala @sfscala @Winium @higherkindedtpe @adamwarski @lukasz_bialy @RomanowskiKr @liosedhel @VirtusLab @scala_lang @tstokrk

[Win Wang @Winium]

"On the (allegedly) sequential nature of unification"...
Type inference is not sequential in practice. Measured STLC on a GPU:

[Win Wang @Winium]

@ChShersh Python is a C++ service. (/s)

[Win Wang @Winium]

@mandylu @gabriel1 normative vs descriptive, strikes again

[Win Wang @Winium]

@LongFormMath @grok which transfinite field should we start with tomorrow?

[Matej Cerny @matej_cerny]

My favorite talk from Scala Days 2026 was about baking refinement types right into the compiler.

Unlike libraries like Iron or Refined, it adds a built-in solver that verifies logic and math across multiple variables at compile time with zero runtime cost.

Moving from simple checks to full verification is how Scala's type system should evolve to constrain LLMs - AI code won't compile unless the math holds up.

[Win Wang @Winium]

@jmilinovich tmux. I pin sessions to tmux session names. Typically a few per worktree.

[Salar Rahmanian @SalarRahmanian]

🎤 Speaker spotlight: Win Wang @Winium , founder of ParaQuery, ex-Twitter distributed builds & parallel Scala compilation.

"GPUs for Type Inference and Typeclasses": we love strong types, we hate compile times. Can we throw a GPU at implicit resolution?

🗓️ Mon Oct 19 · SF · 6 days to go
🎟️ Get your ticket: https://t.co/K54RCwwkn2

#Scala #Scala3 #BayAreaScala #SFScala #SanFrancisco #ScalaSummit2026

[Win Wang @Winium]

@stevebauman I thought you were going to say that AI still needs to understand the complexity as well. Higher-level languages would presumably still give us more AI per dollar.

[Win Wang @Winium]

I love it when LLMs matter-of-factly roast their own code, as if they weren't the perpetrator. Just something funny about it.
"The data shows these lists are massively oversized relative to actual usage, wasting 98%+ of their allocated capacity..."

[Win Wang @Winium]

@trq212 @Hesamation I find that Claude is really good at making easy-to-read PDFs with graphics. Better than asking it to make pedagogical code.
"Unfortunately", it's still up to us to direct the pedagogy in the output.

[Win Wang @Winium]

@philkellr I still do deep-ish technical work by having design discussions with agents. While there's wait time, do I really want to do the menial work of supervising/noting benchmarks/profiling, or do I just discuss their ramifications? I prefer the latter.

[Win Wang @Winium]

@threepointone Different flavors of problems. Not everyone likes the same flavors.

[Win Wang @Winium]

@PunkSC @bernoussama @hnasr Wait. Is this worse than multiple clients on one IPv4? I'm not sure how multiple IPs necessarily leads to bad security?

[Win Wang @Winium]

@wangzhr4 New, not so much, but maybe exacerbation. The prevalence and exaltation of purely vibecoding thoughtlessly? Potentially problematic.

[Win Wang @Winium]

@wangzhr4 Agents sometimes also aim for "insane" (complexity, requirements). Perhaps that also maps to "low standards of how to get it done", but it requires a qualitatively different approach(es).

[Win Wang @Winium]

@fwbrasil @reggieperry Honestly, nothing wrong with that, especially when expanding to broader areas. I love learning from LLMs.
But I'd rather people be upfront about issues/catches.
But blaming your agents for a performance matrix you're endorsing outright?

[Win Wang @Winium]

@wangzhr4 I mean, we could *at least* ask the agents (directly):
1. Why is the comparison bullshit?
2. What dumb things are we doing?

I find that agents will just tell you how they screwed up.

[Win Wang @Winium]

@fwbrasil @reggieperry A CUDA agent literally made a serial queue that a single thread processed while 48k other threads just sat and waited. Amazing design.