Max Niederman
354 posts
Max Niederman
@MaxNiederman
Head of Quality @ Mechanize | 19M
San Francisco, CA Katılım Ekim 2019
192 Takip Edilen117 Takipçiler
Economically it’s considered a subsidy because it doesn’t capture the social cost.
Nicholas Decker@captgouda24
Note that when you look into these sorts of claims, the “subsidy” is simply that it is not taxed as much as they think it should be. You can argue the status quo is wrong, but you can’t really call it a subsidy, now can you?
English
@GuiveAssadi @dfrsrchtwts You’re definitely not supposed to spoof this, but it’s a very frequently violated norm. Most web scrapers are at least this dishonest.
English
@dfrsrchtwts This sounds fine to me? I guess I'm not a SWE so IDK if it is actually bad. @MaxNiederman weigh in.
English
My biggest gripe with persona selection stuff (or maybe just the pop version) is that it feels like it doesn't grapple with the way LLMs are actually pretty unlike any human persona.
English
@tomkXY @tomieinlove Yes, but they are debts of the bank not of the payer.
English
@MaxNiederman @tomieinlove Banknotes were originally much the same. They weren’t standardised by region so there was no way of knowing it if was legitimate. They basically just promised you get gold of equivalent value if you went to their bank.
English
Astonishing that we could ever have a system like checks. We gave everyone the ability to write themselves arbitrary-denomination banknotes on the hope—with no way of checking—they’ll have enough funds to cover it later on. How did this not immediately collapse under fraud?
English
Speaking as someone who studied joint mathematics and computer science, this right here is how you tell apart the mathematicians from the computer scientists.
English
More people should know just how bad all public SWE evals are. METR Time Horizons was better but now it’s saturated too. I’m not aware of any truly good public SWE evals atm.
Joel Becker@joel_bkr
new @METR_Evals research note from @whitfill_parker, @cherylwoooo, nate rush, and me. (chiefly parker!) we find that *half* of SWE-bench Verified solutions from Sonnet 3.5-to-4.5 generation AIs *which are graded as passing* are rejected by project maintainers.
English
@dfrsrchtwts @1a3orn I think the problem is that “emergent” is vague enough that you can’t replace “misalignment” with other properties that generalize across training domains in this way. “misalignment from cross-domain generalization” would be better, but idk how to make it catchy.
English
@1a3orn But for example, I don't know how you could write this paper while not using the term "emergent misalignment", because that's the instance of (what may be) persona selection that they're actually interested in. x.com/safe_paper/sta…
AI Safety Papers@safe_paper
Emergent Misalignment is Easy, Narrow Misalignment is Hard Anna Soligo (@anna_soligo), Edward Turner, Senthooran Rajamanoharan (@sen_r), Neel Nanda (@NeelNanda5)
English
has anyone tried to justify "emergent misalignment" as a specific term
rather than just "personal selection model" or some other less morally loaded frame
it seems to me either entirely historically accidental, or part of AI safety community's preference for frightening terms
English
@MaxNiederman I think this is an idea that makes sense, but it is also not a good descriptive definition of how this term is actually used.
English
The idea of "luxury beliefs" makes no sense. My opinions about policing don't affect anything. Neither do yours. Poor people can absorb the costs of irrational political beliefs just as easily as rich people, because those costs are zero.
English
Max Niederman retweetledi
what the military had access to was a claude with the us constitution prepended to any prompt they asked - you can see why this pissed them off so much
English
Feel the pain of working in eval development 😂😂😂
adi@adonis_singh
decided to run 5.3 codex on xhigh as well, its 90%... rip IBench, survived 3 months.
English
@sichuan_mala I don’t think so, but that bet seems better to me than a simple acquisition play.
English
@MaxNiederman Right. But is their data worth their current valuation?
English
@yacineMTB large companies are often dysfunctional, but it’s not generally because no smart people work there. it’s just really really hard to coordinate large groups of people
English
Max Niederman retweetledi
@EigenGender @mitduckmaster It does if you have access to this machine!
English
you get a black box that can find a proof to any lean theorem you formalize (or state that no such proof exists). can you do anything useful with it?
English
@EigenGender You can use this to compute the result of any computation in time proportional to the number of bits in the output.
English