michael

PhD student at @BoulderNLP @lecslab, intern at @AmazonScience

Studying NLP for rare+endangered languages, multilinguality, automata, synthetic data

[michael @WasianBayesian]

Beyond honored that PolyGloss has won an Outstanding Paper award at #ACL2026

We’re really excited about NLP for endangered language documentation, and it’s inspiring to see that the community is too!!

[michael @WasianBayesian]

Excited to announce that the PolyGloss paper has been accepted to @aclmeeting!

Previously, we trained models to help in endangered language documentation workflows by automatically predicting interlinear glosses. But real-world user studies revealed crucial issues...

[michael @WasianBayesian]

This can actually be a negative, especially in educational settings!

[George Millo @georgemillo]

One underrated aspect of coding with AI is that I never get stuck anymore.

In the olden days I'd sometimes hit a bug or other difficulty that could leave me spinning my wheels for days. I'd google fifty variations of the same query, read poorly-written docs I barely understood, post on StackOverflow and get no answer, pore through obscure technical blogs and bang my head against the screen but end up no closer to victory than when I started. Sometimes I'd give up completely.

This basically never happens anymore. AI doesn't magically solve everything but it cuts through all the crap so I quickly that I never feel like I'm making no progress.

Kids these days don't know how good they've got it!

[michael @WasianBayesian]

@georgemillo This is a good thing in the short term and a bad thing in the long run

[michael @WasianBayesian]

@januarycomputer That’s just “human preference tuning”, and it’s by definition not verifiable

[michael @WasianBayesian]

“Coding is solved” is a weird phrase to me

“Coding” is not a math conjecture!

It’s the broad process of translating informal, underspecified requirements into a formal representation. That’s not really something that has a definable “solved” state

[Boris Cherny @bcherny]

@dedene Coding is solved, bugs are not yet solved. Fix incoming

[michael @WasianBayesian]

@alexolegimas @edzitron And what data did you fit that curve on?

[michael @WasianBayesian]

@pourfairelevide I think the Ayn Rand interpretation is pretty surface level and reductive

[michael @WasianBayesian]

@CasinoJoe92 @coldhealing I don’t think music is anywhere close actually!

[michael @WasianBayesian]

@qwts_chris_dev @ericclemmons You don’t think 80% of engineers are incentivized to write code that, eg, doesn’t leak customer data?

[michael @WasianBayesian]

@spaceforceguy @zebassembly Why do you have such a strong emotional response to this

[michael @WasianBayesian]

@JoshSaintJacque @aarondotdev They like to add in fallbacks that hide cases where there’s actually some deeper underlying issues

[michael @WasianBayesian]

@scaling01 Guy who thinks you should make new benchmarks that are already saturated

[michael @WasianBayesian]

@IMakeStupidGame @spirodonfl Yes, new tools tend to face resistance from people who don’t like change.

No, that doesn’t mean every new tool is exactly the same.

[michael @WasianBayesian]

@ConsumerRick @hellosvapnil Boring, try a new response

[michael @WasianBayesian]

@yong_zhengxin I’d push back on the second point; that’s absolutely true for ML conferences but I’m not sure it’s true for math

[michael @WasianBayesian]

@paradite_ None of that will help you understand bespoke idioms where you have to infer the intended meaning, which is one of the big problems.

[michael @WasianBayesian]

@industriaalist Lowkey rl might be all you need

[michael @WasianBayesian]

@repligate That’s a perfectly consistent position

[michael @WasianBayesian]

@ohfault_ It’s probably just like AI mastering (ozone or whatever) which has been around long before generative ai

[michael @WasianBayesian]

@AdrianThonig @Piyush_Goenka_ @karpathy @ChrisJMcCormick For sure! But now you’re relying on your own test coverage, instead of the robust coverage that already exists for torch.

Not to mention, this only works because PyTorch exists—if everyone starts spinning up their own impl with AI, you’d have no definitive source of truth.