Daniil Tsevkov

Working on long-term human progress.
AI safety · Technology · Institutions · Global coordination

[Daniil Tsevkov @Tsevkov]

@Tony_Diver glad it's back, but safety testing access that depends on lobbying is a favour, not a system

[Daniil Tsevkov @Tsevkov]

@peterwildeford @PunchbowlNews interesting that both the labs (per axios) and congress now point at cyber as the first big one. might end up being AI's ozone hole

[Daniil Tsevkov @Tsevkov]

sam altman says society will have to accept "some fairly severe cyber incidents" from open models as the price of liberty.
we don't say that about planes, medicines or nuclear plants. we certify them first.
why is AI the exception?

[Jaynit @jaynitx]

Sam Altman says society will have to accept severe cyberattacks from open-source AI as the price of preserving freedom

"I support open source models. There is a coming tidal wave of cybersecurity problems that are going to come from these open source models"

"There are people in the industry who say we should ban open source, because otherwise a bunch of systems are going to be hacked"

"We're going to, as a society, I think, have to just accept some fairly severe cyber incidents from open models, in exchange for the liberty that comes with that"

[Daniil Tsevkov @Tsevkov]

@peterwildeford who decides how many severe incidents society accepts though? that’s usually what elections and treaties are for

[Daniil Tsevkov @Tsevkov]

@Miles_Brundage same thing happened to "AI safety". in 2023 it was literally the name of the summit, by 2025 vance opened his speech by saying he wasn’t there to talk about it

[Daniil Tsevkov @Tsevkov]

In july, 29 countries set up a new global AI cooperation body in Shanghai. The diplomat called it “china’s new AI club”.

AI governance is splitting into rival clubs. Will countries come to their senses in time and start working together instead of competing?

[Daniil Tsevkov @Tsevkov]

@_NathanCalvin the difference between a commitment and a policy is whether you can change it on your ow

[Daniil Tsevkov @Tsevkov]

@ohlennart finally a compute governance proposal that doesn't need a treaty

[Daniil Tsevkov @Tsevkov]

wild that the labs are already planning for the day after an AI catastrophe.
would be nice if someone was planning for the day before

[Andrew Curran @AndrewCurran_]

Axios is reporting that the top labs are wargaming scenarios for the aftermath of a catastrophic event in the next 6 - 12 months, that some executives now see as being inevitable.

[Daniil Tsevkov @Tsevkov]

@Yoshua_Bengio if the safety-minded people all leave, who's left inside to catch the next incident?

[Daniil Tsevkov @Tsevkov]

@j_asminewang whatever actually happened, this is why labs need an independent process for safety disputes. right now it's just "trust us" vs "trust us"

[Daniil Tsevkov @Tsevkov]

@hamandcheese @deanwball @tylercowen ask him which existing institution he'd actually trust to handle advanced AI, given how worried he is about institutional decay

[Daniil Tsevkov @Tsevkov]

@ohlennart nuclear test bans got lucky here. the thing you can verify (a seismic signal) is almost the same thing that matters (a test). AI doesn't get that luxury

[Daniil Tsevkov @Tsevkov]

@GarrisonLovely if sharing with an outside safety org is a firing offense, outside oversight of AI labs is basically optional

[Daniil Tsevkov @Tsevkov]

The treaty banning nuclear tests was signed in 1996 and still hasn't entered into force.
But its global monitoring network was built anyway, and it caught North Korea's nuclear tests.

With AI we keep waiting for the treaty. maybe the monitoring should come first.

[Daniil Tsevkov @Tsevkov]

@russianforces interesting that the push would have to come from the non-P5. feels like middle powers are the ones with the most reason to defend these norms now

[Daniil Tsevkov @Tsevkov]

@nicholaswu12 AI incident in July, liability bill in October. the incident -> rule pattern, playing out in real time

[Daniil Tsevkov @Tsevkov]

The UN Security Council passed 44 resolutions in 2025. lowest since 1991.
at the same time there were 130+ active armed conflicts, the most since WWII.

The body built to stop wars is deciding less while there are more of them.

[Daniil Tsevkov @Tsevkov]

@HumanHarlan assurance contracts are such an underrated fix for coordination problems. nobody wants to be the only one who shows up

[Daniil Tsevkov @Tsevkov]

@deanwball the annoying part is we need those exact institutions to be at their strongest right when AI shows up