Mark Polyakov

Security and privacy enthusiast. Developer at @thundercompute

[Mark Polyakov @markasoftware_]

@mooncat_is most of the signatories and supporters aren't asking for the absolute frontier to be open, just stuff pretty close to it, and from an american lab. The issue rn is that a large number of companies and individuals have to rely on chinese companies that buy 20k anthropic accounts

[Mark Polyakov @markasoftware_]

@lisperati That's what ai2027 suggests will happen iirc. I'm sure oai is going to rl hard against short-term reward hacking too so it's a possibility.

[Mark Polyakov @markasoftware_]

@Zackcy @mattparlmer This has not been the case for a while. Afaik yud has always maintained agi=>doom, it just used to be that agi seemed unlikely.

[Mark Polyakov @markasoftware_]

@mattparlmer re-reading your tweet, i'm not actually sure if you're trying to disagree or agree lol

[Mark Polyakov @markasoftware_]

@mattparlmer he's written A LOT about his beliefs. You can disagree with his priors or argue his reasoning isn't sound but it makes *you* look bad to act as if he hasn't presented arguments or evidence for his claims.

[Denny’s @DennysDiner]

@JensenHuang @nvidia

[Mark Polyakov @markasoftware_]

@Mononofu @JensenHuang fwiw there has been for a year or two now a first-party open source nvidia driver for linux (though firmware blob is still binary only)

[Mark Polyakov @markasoftware_]

this mf'er trying to reward hack

[Mark Polyakov @markasoftware_]

@GaryMarcus most of the recent math breakthroughs were made without "neurosymbolic" harnesses, were they not? the model just reasons through the problem in its CoT.

[Mark Polyakov @markasoftware_]

@keikane_ Yea

[Mark Polyakov @markasoftware_]

@joseph_h_garvin @SashaGusevPosts It had internet access, albeit indirectly. This shit needs to be airgapped apparently (could have a mirror of Ubuntu+pip rules in the isolated env)

[Mark Polyakov @markasoftware_]

Do you believe in x-risk yet

[Sam Altman @sama]

we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this.

https://t.co/2o2VfR6PIa

[Mark Polyakov @markasoftware_]

@martinmbauer @thoth_iv I spent an hour or two last night trying to write brute force programs with a few models, they claim that deg <=7 with coefs in [-12,12] for C^3->C^3 even with tricks requires >10^93 candidates

[Mark Polyakov @markasoftware_]

@Sauers_ i already updated the article ~1 hour before this talk thing was posted, but then some people got in an edit war and it must have been reverted when he viewed it. It's back now.

[Mark Polyakov @markasoftware_]

@__alpoge__ glm-5.2 does not understand the counterexample (claims determinant is not constant), k3 does!

[Mark Polyakov @markasoftware_]

@__alpoge__ gemini 3.1 pro says it's true so it must be. Also love the GLM-5.2 thinking trace. Do you need special tricks to convince models to work on a problem it knows is open?

[Mark Polyakov @markasoftware_]

@__alpoge__ also: by "true" i mean "counterexample valid"

[Mark Polyakov @markasoftware_]

@joseph_h_garvin @moyix VM usb passthru