It would be very useful to understand more about the government safety concerns associated with frontier AI releases so we could (a) know what risks everyone will face if/when open source reaches Mythos class & (b) whether they are doing enough or too much to prevent those risks.
Many users doubted prospects for real transparency on government safety concerns for frontier AI models, suspecting motives like ensuring lab loyalty or bureaucratic delays instead.
No Digg Deeper questions have been answered for this story yet.
If it was just military implications, that would be one thing, but if this is about info security at scale (as Five Eyes suggests), publicly understanding risks and mitigations is incredibly important to tens of thousands of firms that need to prepare for what is next.
It would be very useful to understand more about the government safety concerns associated with frontier AI releases so we could (a) know what risks everyone will face if/when open source reaches Mythos class & (b) whether they are doing enough or too much to prevent those risks.
@emollick I don't think they have safety concerns.
I think they have loyalty concerns.
They want to make sure the AI labs know the State is their master. Not sam, not dario, but USG/Trump. Making that clear is the goal.
@emollick Not sure open source reaches frontier levels if top models aren’t released. At least that should be the case if the distillations rumors are true.
@emollick We've seen the headlines saying it will take >3 months to "upgrade/protect" our core systems
Now that the gvmt is aware, they'll try to drag their feet to somehow brute force a complete infra overhaul in an unprecedented amount of time.
ezpz
@emollick Apparently only China can save us now according to some of the discourse. The irony of such a statement is profound . #FreeGPT5.6
@emollick How much impact do you see that this measure will have on the Chinese opensource models? Some experts say, that Chinas success was heavily because of their distillation of prime US frontier models. If true, that would slow them down now. Others say, China is just innovative. wdym
@emollick Fair ask, right now its all vibes and leaks, zero actual transparency on what the concerns even are
@emollick Hard to expect much transparency/clarity from this admin unfortunately
@emollick It needs to be transparent - so that lawmakers make the correct decisions. Otherwise there is the opportunity for a political party to take advantage of gate-keeping access potentially, as well as having other countries (China etc) getting a leap.
@emollick The government wants to fix their paperclip shortage before releasing the models to the public.
@emollick I wouldn't expect to get this kind of transparency. We'll be told the approved narrative and nothing more.
@emollick Do you want to know what the government thinks, for predicting/triangulating/Kremlinology/etc, or do you want to know what's actually happening/going-to-happen in cybersecurity? Because for the latter, don't listen to anything the government says. I'm serious.
@emollick Something will need to be figured out soon. The uncertainty is already complicating the frontier labs future plans. Reuters is reporting that OpenAI is looking to postpone its IPO.
@emollick Part of the issue is nobody wants to share their risk framework openly
ur just flying blind until someone leaks it
The opacity isn't accidental. Disclosing specific risks often means disclosing the capability that creates them — a roadmap dressed as a warning. Labs face the same bind: the honest red-team report is also the attack manual.
So we're stuck grading the safety response without seeing the test. Hard to know if it's overcaution or theater when the whole exam is sealed.
The tell will be open source. When a Mythos-class model ships with weights public, we'll finally see which concerns survive contact with reality and which were just leverage.
@mreiffy @emollick There is an RSI feedback loop i'm gathering, once the models hit a certain level of capability gpt 5.2 (ability to do loops) they can aid in the development of the next model. This is why Anthropic seemed to have broke from the pack.
@emollick This really sucks
@emollick The asymmetry is the hard part. A closed lab can stagger a rollout, but once that capability lands in open weights there's no controlled-preview button left to press. At minimum, naming the actual concerns out loud would let the rest of us plan instead of guess.
@Art_If_Ficial @emollick lol, lamo, a decade even with ai. no one knows shit about fuck or where things are, then you have to deal with every locality, entrenched leaders who don't want to change things, ancient hardware. shit a decade is an optimistic outlook.
@emollick Yep. Hard to judge the risk if they won't say what the risk is. Otherwise it's policy by vibes.