matansok

• matansok 16 May 2026 14:27 UTC

  1 point

  0

  in reply to: Florian_Dietz’s comment on: Flo­rian_Dietz’s Shortform

  I think this has a very high chance of success, but trades off the reliability of the probe—if the NLA somehow learns how to hide misalignment semantically (and not steganographically), this type of probe becomes useless.

• matansok 31 Mar 2026 16:49 UTC

  2 points

  1

  in reply to: No77e’s comment on: No77e’s Shortform

  This seems anecdotal.

  So far, we have documented cases of Generative AI being used to subvert elections in Romania (actually causing an annulment), and to some extent in NYC. We also have this report by OpenAI from 2024, which details such influence operations facilitated using OAI’s API. Given the proliferation of significantly more capable open-source models in the 18 months since, we can be fairly confident that broader, more complex operations are taking place today.
  
  I also tend to associate AI-slop with low quality content, but we know AI is more capable than that, which leads me to believe that significant amounts of content online are parts of influence operations by malicious actors.

• matansok 3 Mar 2026 6:49 UTC

  2 points

  −1

  in reply to: ChristianKl’s comment on: Chris­ti­anKl’s Shortform

  I mean… The Iranians did try to assassinate Trump.

  I’m saying that propaganda efforts are significantly more effective against democracies. Authoritarians receive no penalty for lying, and are less susceptible to lies by their enemy. Their local media is 100% controlled. This is why they lie blatantly all the time—this is their main lever for “winning”, since they have basically zero means of military resistance.

  I think martyr death might invigorate people in the short term, but it’s much harder to argue that the war was “won” when your leader died, and the loss of key personalities can be devastating to centralized organizations. Look at Qassem Soleimani—where is Quds force now? Same for Nassralah and Hezbollah’s power in Lebanon.

• matansok 2 Mar 2026 23:56 UTC

  −4 points

  −9

  in reply to: ChristianKl’s comment on: Chris­ti­anKl’s Shortform

  Thinking about this as a war between two “countries” is, in my opinion, missing a large part of the picture. They would take out Trump if they could.

  In a conflict between a democracy and an autocracy, the democracy is usually limited by domestic support, while the autocracy can just continuously reframe the narrative and keep on going pretty much indefinitely. This is true even at an enormous military disadvantage—the authoritarian side can force the democracy to surrender purely by psychology, and this is massively exploitable in the information age.

  For example—the (obviously fake) report of a strike on an all-girls school in Iran that keeps circulating for the past 3 days with ever growing casualty counts. These videos are clearly targeting Americans in order to diminish support for the campaign.

  Taking out the leader serves both as a very tangible, undeniable “achievement” for the democracy (although even that can be framed as a desired outcome in certain belief systems). It also takes out a significant psychological driving force, as these are usually charismatic, cult personalities.

• matansok 24 Feb 2026 19:03 UTC

  7 points

  1

  on: Large-Scale On­line Deanonymiza­tion with LLMs

  Would be interesting to see if these results hold up in a more adversarial setting. In both cross-platform matching and matching split accounts, the users are not actively trying to remain anonymous, and would be less averse to sharing identifiable information.

  I would argue that if de-anonymization is possible, it’s better if a white-hat researcher does it first, a-la responsible disclosure.

• matansok 18 Oct 2025 11:30 UTC

  3 points

  0

  in reply to: faul_sname’s comment on: faul_sname’s Shortform

  I don’t necessarily disagree, but one big thing is freedom of speech. If the party line is to go big on AI, which is likely, given current investments, I’m not betting on Chinese Yudkowsky. Same with frontier lab whistleblowers, etc

• matansok 30 Sep 2025 21:31 UTC

  1 point

  0

  in reply to: anaguma’s comment on: Burny’s Shortform

  I feel like I’m missing something—the model doesn’t “realize” that its CoT is monitored and then “decides” whether to put something in it or not, the CoT is generated and then used to produce the final answer. Since the standard practice is not to use CoT output for evaluation during training, there’s no direct optimization pressure to present an aligned CoT.

  The only way I can see this working is through indirect optimization pressure—e.g. if misaligned CoT leads to misaligned output, which causes the CoT to also present as aligned