andré ⧉

ai ⊕ science | prev @nvidia, @esa

[andré ⧉ @andregraubner]

@ArtirKel @RetroBio_ @OpenAI This looks really cool!

[andré ⧉ @andregraubner]

@aryaman2020 @StatisticUrban This 100%. I feel like most people boasting about how difficult Mandarin is have never actually tried learning it… (this is not to say it’s *easy*, it’s just rather straight-forward to learn as long as you are persistent)

[andré ⧉ @andregraubner]

@miru_why @elyxlz banger

[andré ⧉ @andregraubner]

@juliapeng__ @KernaLabs looks interesting!

[andré ⧉ @andregraubner]

@AnimaAnandkumar @Caltech Glad to hear you’re safe!

[andré ⧉ @andregraubner]

🤯

[andré ⧉ @andregraubner]

@Nst_Egorova @bryan_johnson very good questions

[andré ⧉ @andregraubner]

@KleistLeo @mmjukic What did you get out of it?

[andré ⧉ @andregraubner]

@wordgrammer @teknium Please someone do this with lean!

It’s perfect bc it’s super powerful but what makes it powerful is often too tedious for human engineers to unleash its full potential.

I think @morph_labs is working on something similar, but haven’t heard from them in a while…

[andré ⧉ @andregraubner]

@CarnapSpeaks Also giving very tractatus-y vibes

[andré ⧉ @andregraubner]

@CarnapSpeaks Damn, any thoughts about the one on the right? I assume you haven’t read it yet, but the title is super evocative

[andré ⧉ @andregraubner]

@VictorTaelin @ritteradam Intuitively it feels like it might be possible to do something like ‘finding good subtypes’ that could replace the original domain and/or codomain while having additional structure. Is this what you’re thinking about?

[andré ⧉ @andregraubner]

@VictorTaelin @ritteradam “searching on the space of types before enumerating terms” can you elaborate?

I understand that type-driven synthesis is good because it massively restricts the search space. But if you’re trying to synthesize a function, then you presumably already know its signature, no?

[andré ⧉ @andregraubner]

@orrymr @fchollet @aidan_mclau https://t.co/p5adkBcKcf

[andré ⧉ @andregraubner]

@rlacombe what do you think should be the next big thing to fix?

[andré ⧉ @andregraubner]

@haeggee congrats Alex!

[andré ⧉ @andregraubner]

To actually answer the question though: If anything, I thought of “neural networks”, which to me just was an “artificial brain” learning by “seeing stuff”, without concern for what that means or how that could or could not be achieved.

[andré ⧉ @andregraubner]

This just made me realize that before I started working in ML, I basically didn’t think about how AI might work, basically at all. Watching AI movies at a kid, this question never even crossed my mind.
Man, I must have been one hell of an uncurious child 👀

[(((ل()(ل() 'yoav))))👾 @yoavgo]

what did y'all imagine machine learning is, before learning it and being indoctrinated that it all about optimizing parameterized functions involving many matrices?

[andré ⧉ @andregraubner]

@elyxlz whyy

[andré ⧉ @andregraubner]

@ted_engineer @rohanpaul_ai @fchollet Haven’t looked at the paper yet (planning to), but:
While you are right in principle, in the case of large transformers this bias is really small. You can almost specify the architecture and hyperparameters in a tweet!