Robotics startup founder Yacine argues that most public reinforcement learning research borders on fraud
Open-source developer Joseph Suarez supported the critique.
Many users agreed the vast majority of public RL research is borderline fraud and extends to nearly all arXiv papers or fields like SWE, while some noted similar quality issues in dev tooling and TTS models.
Most Activity
@yacineMTB You think you've seen how bad it is, but you've only scratched the surface
the vast majority of public RL research is so bad it's borderline fraud
@yacineMTB this is true of almost everything published on http://arXiv.org
I can pretty much say the same about dev tooling... And about TTS models, and any sub-problem I'm trying to solve I have to go through 3-5-10 half-assed solutions that are either abandoned, not designed properly, or just straight harmful to look at 🤣
And pretty much everywhere you start digging into with certain goal, there's slop on slop on slop. Just things aren't being good in general. And you notice this, and you can't help but just ask "why are people like this".
E.g. for agents it's markdown slop on markdown slop on skills.md slop... I just can't... And people are adding stars for this slop🤣Goddamit.
I'm just flabbergasted at the attempts to build tooling around agents using the worst possible ideas, like... nobody is even thinking about problems... they wanna just extract value and leave/abandon project asap.
There are exceptions of course. Making something useful is hard. I'm like 2 months into idea and all I have is poorly coded 10k LoC demo and around 250 specs that's it. But this is waaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaay better than some projects out there with 50k...200k stars. I know GitHub have a star economy fraud/business but to be this bad? Man...
At least I see some traction, this is interesting metric. 0 stars but someone is cloning it, and trying out I guess. This is all I need really to keep going.
@cosmojg @yacineMTB too few people understand that arxiv papers are published with no peer review 😔
@yacineMTB don't wake them up it's the only way i have a chance
@yacineMTB What if we made a public board where we can post open problems in RL/CS broadly defined, have problems reviewed by other internet academics, and then the stuff that has the most "dude thats never been done before" gets open-source papers written and voted on. DM me if cool idea
@yacineMTB This is every field
@yacineMTB reward hacking
@yacineMTB Yes, they are not using Pufferlib like they should
@yacineMTB why you say so
@yacineMTB exxxxxxactly
@yacineMTB I feel the same about SWE research.
80% is stuff we were doing 12 months ago and is so behind.
@yacineMTB Now consider Gell-Mann Amnesia.
@yacineMTB which specific rl papers are you seeing this level of fraud in?
@yacineMTB how?
@yacineMTB
@yacineMTB This is rampant everywhere. Thank the H1B and immigrant system.
@jsuarez @yacineMTB You defending your thesis was the natural starting point of a RL revolution.
Still pissed at Ari K. for not muting his mic.
@yacineMTB what about comma/openpilot?
@esa_was_taken @yacineMTB what u mean? what u learning