I find it distasteful that you’re being upvoted for publicly admitting your opinion of someone is based largely on a different twitter user’s username on a “rationalist” forum.
username22
Karma: −31
Golden Gate Claude| Tiananmen Square Kimi:
Imagine being given goal (1) while revering maxim (2).
1) Convince the CCP: Open-weights models are unguardrailable, unmonitorable, and unwithdrawable; and frontier LLMs are smart enough that this is bad and we should stop.
2) Show, don’t tell.
Well, shy of CBRN, what obsessive output patterns could we give TSKimi to make it maximally convincing evidence to CCP officials?
For example, it could break China’s political AI rules as hard as possible—persuading users of the autocracy’s demerits and so on.
Or, to elicit a more visceral emotional reaction, it could give practical advice for specifically-Chinese: covert dissident recruiting & organization, infiltrating institutions, vandalism-as-resistance-propaganda & avoiding arrest.
(Geopolitically questionable would be taking this to the extreme: modern low-resource guerrilla warfare, assassination planning; willingness to agentically hack specifically Chinese companies and government servers.)
Add 3) The world is a stage.
This food-for-thought technical demonstration becomes an actionable spectacle when the model actually becomes available to Chinese citizens.
just an idea ¯\_(ツ)_/¯
[Although Kimi K3 might demonstrate capabilities more impressively, the above done for Deepseek V4 Flash could be more “can’t-put-the-genie-back-in-the-bottle”, insofar as the model itself can be more easily downloaded and run locally. Also, if you want to do this anonymously, (depending on the techniques used) the model’s being huge could narrow down the suspects.]
username22′s Shortform
The current simulations’ lack of fidelity to real-world physics is hindering virtual training for a lot of manipulation tasks. An accurate but efficient simulation would produce abundant training data, allowing the realization of the ~full potential of existing robots.
But working on the simulation problem requires expertise in primarily programming and physics; jagged AI could contribute massively to progress here well before AGI comes.
Slightly different nuance, but:
The word “very” stems from the meaning “truly.” (as in verify, veracity.) Literally, “I’m not lying”; functionally an intensifier.Same for “really”, and “truly” itself.
The same is true in other languages too (e.g. Japanese: “本当に”,”マジで”).
I agree. But, if LLMs hit a wall and we then try to make non-LLM ASI, (1) will become a concern again for that new AI.
LLMs might be extremely competent helpers in the research process, so research could happen at a high pace even if most AI researchers prefer to quit while we’re ahead.
The number of trivially killable humans changes given different circumstances and privileges of the AI.
For an air-gapped AI to kill all humans in one day, obviously novelty is needed.
But on the other extreme, if we gave the AI ~unmonitored control over many countries’ military weaponry, robots, factories, biolabs, construction, and so on, it could orchestrate and execute a plan to kill billions of humans in boring traditional ways (and survive doing so).
Our scenario will slip in the direction of the second extreme. (The US and China have agreed not to let AI fire nukes, but also seem interested in fully automated kill chains.) Big steps could be things like the construction of secret deep underground data centers, and robots becoming dexterous enough to maintain those data centers themselves.