Will Depue claims China has complete, unauthorized access to OpenAI and Anthropic internal GitHub repositories and Slack communications

VIEWS49.2KBOOKMARKS59LIKES745RETWEETS30

elie@eliebakouch

there is no question, none at all, that all labs has full access to all of deepseek, kimi, glm, minimax tech report/model today

will depue@willdepue

there is no question, none at all, that china has full access to all of openai & anthropic’s github/slack/docs today

no disrespect to their independent research progress, but i wouldn’t be surprised if we see plausibly-deniable stolen arch methods in chinese oss models

1d49.2K74559

REPLIES26

roon@tszzl

@willdepue Maybe they can figure out wth is going on here

will depue@willdepue

there is no question, none at all, that china has full access to all of openai & anthropic’s github/slack/docs today

no disrespect to their independent research progress, but i wouldn’t be surprised if we see plausibly-deniable stolen arch methods in chinese oss models

1d20.7K54316

Nathan Lambert@natolambert

Frontier labs are definitely SOTA at self serving nonsense.

Yes, AI is a crucial technology, but also Silicon Valley systematically spreads knowledge via talent exchanges and bars. This isn’t a national security deep decision making room.

will depue@willdepue

there is no question, none at all, that china has full access to all of openai & anthropic’s github/slack/docs today

no disrespect to their independent research progress, but i wouldn’t be surprised if we see plausibly-deniable stolen arch methods in chinese oss models

13h38.3K34346

kache@yacineMTB

A drama where a Chinese spy and an American CIA spook meet at openai while they're larping as OpenAI researchers and fall in love

will depue@willdepue

@CFGeek after talking at length with security researcher friends, and first-hand experience of openai security posture, i really can’t see this not being true

and even so, i don’t think anyone would dispute there being a handful of human intelligence sources per lab

1d18K29027

Miles Brundage@Miles_Brundage

I wouldn't put things in quiiiite so absolute of terms (there are at least some practical issues in exfiltrating everything given access IIUC) but broadly speaking, yeah it's a super serious/bad situation + I appreciate @willdepue calling attention to it

will depue@willdepue

there is no question, none at all, that china has full access to all of openai & anthropic’s github/slack/docs today

no disrespect to their independent research progress, but i wouldn’t be surprised if we see plausibly-deniable stolen arch methods in chinese oss models

1d32.4K12331

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

imagine Chinese folks at OpenAI. "Hey Tao u sure you didn't leak our architecture to the CCP? Dem reds are getting real good." - "Nah Mark, couldn't be me." - "Shengli, you? Are u letting the chicoms access our Slack via your acc?" - "No comra- sir! Sama's Interests Come First!"

Charles Foster@CFGeek

@willdepue clearance interviewer: ok last question, just a formality… have you worked with any foreign spies?

will depue: most definitely

22h12.2K10823

Susan Zhang@suchenzang

keeping a nice bamboo ceiling cap on talent here from hypothetical coercion risk is a great foot gun carpet to shove sinophobia under and places that indulge in these theories any differently than any other form of corporate espionage by any other nationality deserve to have talent driven away and back to where they supposedly came from

Miles Brundage@Miles_Brundage

Ah I see (honestly just getting back from an East coast trip so v tired, not trying to be obtuse)

I think that is one of multiple factors one might worry about in this context, forgot it was emphasized here, but I do agree it is one of the risk vectors, not bc of people being badly motivated but due to coercion risks

20h9.9K9517

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

The really absurd thing is that Americans treat their AGI industry as an honest-to-goodness Manhattan Project to cybernuke the Chinese, but don't actually expect the latter to resist in any meaningful way except try to compete in products. The CCP are hippies, thus deserve death

flaw@flawedaxioms

It is really absurd that like 60% of the staff of US companies building potentially the most vital dual use tech ever are foreigners. This has to be making the admins head explode on the daily.

1d8.2K12413

Nathan Lambert@natolambert

@willdepue I don't expect major breakthroughs to stay secret for very long, once you release the model it'll diffuse with it.

RSI is similar, it'll pull the whole industry forward slowly, is largely a function of compute and singularity isnt real. My case is here: https://www.interconnects.ai/p/lossy-self-improvement

will depue@willdepue

@natolambert i don’t think we’re disagree on that point, but i don’t think that takes any chance of major breakthroughs or RSI in the future seriously. slow lab leakage doesn’t work then

12h1.7K3812

Susan Zhang@suchenzang

@Miles_Brundage @willdepue with all due respect, it's pretty disappointing to see you encouraging this, miles

Miles Brundage@Miles_Brundage

I wouldn't put things in quiiiite so absolute of terms (there are at least some practical issues in exfiltrating everything given access IIUC) but broadly speaking, yeah it's a super serious/bad situation + I appreciate @willdepue calling attention to it

21h8.4K844

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

This is an interesting clarification so the catch-up to date is endogenous + distillation + open publications and other mundane factors. But when we get to Agent-2… some friends will turn out to not be friends

will depue@willdepue

@teortaxesTex @Miles_Brundage @suchenzang i don’t think they have yet! but us labs are doing a lot of long-term agi focused research that i think china isn’t. if any of it takes off, we’ll see exfiltration i think

20h5.1K266

Susan Zhang@suchenzang

@Miles_Brundage @willdepue no, it's about individual employees here and has always been

saying it's "ccp" is just political cover

Miles Brundage@Miles_Brundage

@suchenzang @willdepue Is the concern “China” vs. “the Chinese government”? Agree that could have been framed more precisely, though I took Will to mean the latter

20h2.5K374

Nathan Lambert@natolambert

In fact, the biggest advantage OpenAI and Anthropic has is abundant resources and talent to keep pushing the frontier and be there first. With the scale required you can only “steal” or “leak” so much.

Nathan Lambert@natolambert

Frontier labs are definitely SOTA at self serving nonsense.

Yes, AI is a crucial technology, but also Silicon Valley systematically spreads knowledge via talent exchanges and bars. This isn’t a national security deep decision making room.

13h3.8K551

Susan Zhang@suchenzang

just saving some new receipts here coincidentally on this particular thread

Susan Zhang@suchenzang

reminder to one day tell the tale from many moons ago of how a certain group within a certain lab pushed for extra surveillance/auditing of certain ethnicities (regardless of nationality)

傻逼透了

1d4.8K255

will depue@willdepue

@CFGeek after talking at length with security researcher friends, and first-hand experience of openai security posture, i really can’t see this not being true

and even so, i don’t think anyone would dispute there being a handful of human intelligence sources per lab

1d4.2K294

Nathan Lambert@natolambert

@eliebakouch banger

elie@eliebakouch

there is no question, none at all, that all labs has full access to all of deepseek, kimi, glm, minimax tech report/model today

1d2.5K490

Miles Brundage@Miles_Brundage

https://www.rand.org/pubs/research_reports/RRA4159-1.html

Miles Brundage@Miles_Brundage

I wouldn't put things in quiiiite so absolute of terms (there are at least some practical issues in exfiltrating everything given access IIUC) but broadly speaking, yeah it's a super serious/bad situation + I appreciate @willdepue calling attention to it

1d2.4K105

Susan Zhang@suchenzang

@Miles_Brundage @willdepue > china has full access

"good security" hardly covers that first key phrase

Miles Brundage@Miles_Brundage

@suchenzang @willdepue Encourage good security?

20h2.7K331

will depue@willdepue

i don’t know if this has happened yet, likely not, and they don’t seem to need the help much so far anyways. but it probably will

even just pointing at the papers to look hard at is often enough. having conviction, knowing what works is most of the work anyways

1d1.6K152

Stanislav Fort@stanislavfort

@willdepue If you see the same method in OSS, there is also a very strong base hypothesis that LLM research is just not that difficult and that you can converge to the same finding independently. I've seen this happy so many times in academic research.

will depue@willdepue

there is no question, none at all, that china has full access to all of openai & anthropic’s github/slack/docs today

no disrespect to their independent research progress, but i wouldn’t be surprised if we see plausibly-deniable stolen arch methods in chinese oss models

22h2.5K360